Spokes.wiki Search About

Log — Agentic Tooling Wiki

Append-only history. Each entry starts with ## [YYYY-MM-DD] <op> | <title> where <op> is ingest, query, lint, or split, so grep "^## \[" log.md | tail -5 works.

[2026-07-21] ingest | How to Run Claude Code Agents for 24 Hours (hub-routed, Telegram, URL-only, T3)

TDS (Kjosbakken, 2026-07-20): operator how-to for long-horizon unattended Claude Code. Created source summary run-claude-code-agents-24-hours (BlogPosting, url-only, T3 — hands-on, no data; "eliminate all bugs" review claim unbacked). No new Thing page — the concepts already have homes; folded into existing threads: loop-engineering (review-is-the-bottleneck, stated from the operator's chair), agent-guardrails + cloud-run-sandboxes (autonomy via sandbox+backup, not trust), agents-never-do-alone (second-agent PR review as throughput unlock), cognition-fable5-through-the-night (this is the practitioner counterpart to that vendor 8h case study). Updated eivind-kjosbakken into a two-axis frame: breadth (orchestrate-100-agents-claude-code) + duration (this). Synthesis loop-section extended. Bridge anthropic (../research-wiki). Runner-up: none (platform-ops "always-on remote execution" is too thin to count) — squarely agentic-coding-harness operation.

[2026-07-21] ingest | Writing a Good CLAUDE.md (hub-routed, Telegram, URL-only, T3)

HumanLayer blog (Kyle/@0xblacklight, 2025-11-25): prescriptive guide to authoring CLAUDE.md. Created source summary writing-a-good-claude-md (BlogPosting, url-only, T3 — named agent-eng shop, research-grounded but unbenchmarked opinion). New Thing page claude-md (DefinedTerm/convention) — the file was previously only an alias inside other pages; this is the first source about the file itself, so it earns a node (Claude-native sibling of agents-md). Two mechanisms recorded that the context-rot thread didn't state: (1) Claude Code's system prompt tells the model to disregard context it deems irrelevant, so non-universal CLAUDE.md instructions silently get skipped; (2) a ~150–200-instruction ceiling with uniform degradation — making minimization a compliance move, not just token-budget. Folded both into context-rot and synthesis. Contradiction flagged: its never-auto-generate stance vs. the docs-as-build-artifact pole (openwiki, autowiki-factory) that CI-writes CLAUDE.md/AGENTS.md — logged as an open tension in synthesis. Entities: humanlayer (Organization, 12-factor-agents), kyle-0xblacklight (Person, thin). Bridge anthropic (../research-wiki). Runner-up: ../research-wiki (tools-for-thought) — declined; CLAUDE.md is an agentic-coding-harness config artifact, squarely this spoke.

[2026-07-19] ingest | Claude Code weekly-limits promotion (hub-routed, Telegram, URL-only, T2)

Anthropic support notice: a temporary +50% boost to Claude Code's weekly usage cap, May 13–Aug 19 2026, auto-applied to Pro/Max/Team/legacy-seat-Enterprise (Free + consumption-Enterprise excluded); scoped to Claude Code across CLI/IDE/desktop/web; 5-hour limits untouched. Ephemeral T2 first-party doc. Created source summary claude-code-weekly-limits-promo (WebPage, url-only). No new Thing page — folded into the existing harness-as-product cost/UX thread in synthesis as the inverse of the tighter-limits friction: the durable takeaways are the two-tier limit (5-hour window + weekly cap) and the weekly cap as the vendor's adjustable lever (subscription limits = managed surface, not fixed spec). Cross-linked anthropic-agent-sdk-billing-pause, claude-apps-gateway (enterprise spend-cap counterpart), boris-cherny, claude-agent-sdk. Entity anthropic (bridge, ../research-wiki). Runner-up: ../llm-providers-wiki (usage-limit/pricing facet) — declined; this is Claude Code product policy, not the model-API market.

[2026-07-09] ingest | Claude Code: model vs effort level (hub-routed, Telegram)

Anthropic explainer (claude-code-model-effort-levels, T1) separating two controls Claude Code users conflate: model = which frozen weights (capability ceiling; context steers, never retrains) and effort = how much work the agent does (files read, tools, steps, verification before "done"). Corrects the folk model that effort is just thinking time — it scales the whole loop, and raising it does move token-generation (switching models doesn't). Diagnostic split: too-little-work (skipped files/tests) → raise effort; confidently-wrong-with-good-context → upgrade the model. New Thing page effort-level (DefinedTerm/mechanism) — first source to name effort as a first-class harness control distinct from model selection; folded into synthesis item 3 (the loop/verification thread — effort is the productized single-knob form of loop-engineering/agent-loops-verification). Back-linked from claude-code-best-practices (which had only the model axis). Cross-spoke context noted, not fragmented: the model-tier market (Fable/Opus/Sonnet) is ../llm-providers-wiki; the frozen-weights/steering mechanics touch ../llm-inference-wiki — dominant substance (harness control) stays here. index updated.

[2026-07-08] ingest | ask-search (ythx-101) — self-hosted zero-key web-search skill (hub-routed)

MIT skill giving agents web search with no API key: wraps self-hosted SearxNG (70+ engines) behind an MCP server + CLI + an OpenClaw SKILL.md; Tavily API optional fallback; targets Claude Code/OpenClaw/ Antigravity. Ingested as ask-search (T1, young, no independent eval). Extends the local-first-tool- over-MCP corner from memory (recall) / documents (knowledge-rag) / perception (watch-skill) to live web search — same delivery shape, fourth capability. Framed as the free/private counterpart to from-local-llm-to-agent's paid Tavily grounding tool ("skills × MCP = agency", tool side self-hosted). Reciprocal link added to knowledge-rag; synthesis (layer-6 local-first corner) + index updated. Runner-up ../search-marketing-wiki declined — this is a web-search tool for agents, not SEO/PPC/AI-search market. Entity discovery: author ythx-101 thin single-repo handle — inline mention, no node. Hub route entry to follow.

[2026-07-08] ingest | watch-skill (oxbshw) — video-watching + self-verification skill (hub-routed)

MIT skill giving any coding agent (1) video understanding and (2) self-verification. Ingested as watch-skill (T1, young ~113★, self-reported metrics). Pipeline: PySceneDetect+phash frames, RapidOCR, offline faster-whisper, SQLite/FTS5 + ONNX-embedding hybrid index; delivered over MCP (13 tools) + CLI + REST across ~8 hosts. "THE LOOP" = record own output → structured critique vs pass criteria → fix → re-verify → before/after proof GIF: this is the spoke's output-grounded verification (ai-job-search, loop-engineering) pushed into a new modality (watch the rendered video, not read the rendered PDF) and, unlike ai-job-search's in-repo loop, shipped as a reusable cross-harness skill. Also sits on the knowledge-rag/recall local-first-store-over-MCP shape, extended from text/memory to perception. Reciprocal links added to loop-engineering + ai-job-search; synthesis (layer-3 verification thread)

  • index updated. Cross-spoke: embeds STT (faster-whisper) — faint ../speech-audio-wiki adjacency, noted not routed (subject is the agent capability, not the STT model). Entity discovery: author oxbshw is a thin single-repo handle — inline mention, no node. Hub route entry to follow.

[2026-07-08] ingest | GitHub-repo burst: agent-scripts + openwiki + pxpipe (hub-routed)

Three GitHub repos arrived together via Telegram, all pre-routed here by the hub router; each extended an existing thread rather than opening a new one, so no spin-out.

  • agent-scripts (steipete) — Peter Steinberger's personal cross-harness hub of skills + AGENTS.MD hard rules + helper scripts (sync-skills mirrors into Codex & Claude Code, validate-skills, committer). T1. The individual-developer floor of the skill-infra spectrum (vs superpowers/pm-skills packs-for-others; vs autoharness's automated maintenance). Entity discovery: created the peter-steinberger Person node (OpenClaw creator → OpenAI; previously inline only in openclaw/openclaw-hermes-control-layer). Folded into synthesis layer 1.
  • openwiki (langchain-ai) — OSS CLI that auto-generates + CI-refreshes codebase docs for agents, appending to AGENTS.md/CLAUDE.md. T1. Second instance of docs-as-build-artifact next to autowiki-factory (closed product ↔ OSS CLI). Owner = existing langchain node (added the doc-gen facet). Cross-wiki seam to ../research-wiki's llm-wiki. Folded into synthesis layer 6.
  • pxpipe (teamchong) — local proxy cutting Claude Code token cost ~59–70% by rendering dense context as PNGs (image tokens priced by pixel area). T1, self-reported eval. Modality-arbitrage pole opposite headroom in the pre-model compression corner; new tokenmining lever. New tension logged: irreversibly lossy on byte-exact strings, and worse on stronger models (0/15 Opus) — added to Contradictions. Cross-spoke seam (image-token pricing/OCR) to ../llm-inference-wiki + ../llm-providers-wiki.

teamchong left a thin inline mention (single repo, no node yet). Cross-links added to headroom, autowiki-factory, langchain, openclaw. Index + synthesis updated. Hub route entries to follow.

[2026-06-15] ingest | ArrowJS — arrow-js.com

Reactive UI framework positioned as "for the agentic era": three functions (reactive, html, component), plain TypeScript, no build step, under 5 kb. Key feature: WASM sandboxing — component logic runs inside a WebAssembly sandbox while rendering to the DOM, so a chat agent can hand generated UI code to a host app safely. T1 (official project site). Entity discovery: no named author in the source; no new entity pages. Gap closed: first source in the spoke addressing the rendered output layer (what agent-generated UIs look like). WASM sandbox maps to agent-guardrails' containment discipline applied to the browser. Runner-up ../webperf-wiki (< 5 kb / no build step); noted in source page. Synthesis updated (new "rendered output layer" section). Index updated. Hub route entry to follow.

[2026-06-01] split | agentic-tooling-wiki created from ../research-wiki's cluster B (human directive: "Migrate")

A single Telegram burst added ~10 agent-tooling sources to ../research-wiki's "cluster B" (agentic LLM products / skills / orchestration / harnesses), making it the largest body in that wiki. The hub router flagged this and recommended a spin-out; the human approved ("Migrate"). Carved cluster B into this dedicated spoke.

Migrated 24 wiki pages from ../research-wiki:

Bridge nodes left in ../research-wiki (cluster-A core / substrate), linked cross-wiki from here: gbrain (knowledge-management core), agent-skills (the procedural↔knowledge bridge concept), compound-engineering (Engelbart bootstrapping), model-context-protocol (shared A/B substrate), anthropic + claude-opus-4-8 (model substrate), garry-tan. Authored this spoke's synthesis (skills → harness → orchestration/deployment thesis) and index. Spoke count 6 → 7.

[2026-06-01] lint | post-split health check

First lint after the split. Findings + fixes: (1) provenance nit — 6 concept/derived pages listed URL-only source slugs in sources: (which is for raw docs); corrected to raw-only (agent-orchestration, spec-driven-development keep their one raw doc; adk, agentskills-spec, google, agentic-coding-harness → sources: [], provenance via inline links). (2) Stale terminology — removed ../research-wiki-internal "cluster A/B" labels from ~16 spots across the migrated pages and reworded 3 placement notes (agentic-seo-skill, claude-to-speech, spec-driven-ai-tools) to the new spoke's framing. (3) Orphans: none. (4) Cross-wiki links verified resolving (link check passes). No contradictions found.

[2026-06-01] ingest | Agents That Build Agents: SKILL-first Blueprint (Microsoft) — honest stub

Routed in by the hub. Fetch failed (JS-rendered MS Community Hub → title only); recorded an honest stub per the fetch-fail rule, specifics flagged unconfirmed. Created agents-that-build-agents-ms (BlogPosting, source) + microsoft (Organization, tracked lightly). Significance: Microsoft becomes the third major vendor (after anthropic, google) on a skill-first agent framework — folded into synthesis Vendors as "three independent vendors converging on skills as the primitive," with the open question of whether MS adopts agentskills-spec or fragments. Refresh on a successful re-fetch.

[2026-06-01] ingest | oh-my-pi (omp) — IDE-wired terminal coding agent

Routed in by the hub (clear match, no runner-up). Created oh-my-pi (SoftwareSourceCode): a feature-rich terminal coding agent (fork of Mario Zechner's Pi) — hash-anchored edits (~61% token reduction, claimed), 32-tool harness with in-process Rust tools, LSP integration + first-class debuggers, persistent Python/JS kernels, Puppeteer browser, structured-JSON subagents, 40+ providers, MIT. Placed at the IDE-integrated/feature-rich end of agentic-coding-harness (vs claw-code minimal); updated that concept's instance list + spectrum. Reinforces "harness competes on engineering depth, not prompts" and provider-agnosticism (cross-links llm-provider in ../llm-providers-wiki). Caveat: README claims, unbenchmarked.

[2026-06-01] ingest | Long-running AI agents that pause/resume with ADK (Google)

Routed in by the hub (clear match; existing adk page). Created adk-long-running-agents (BlogPosting) + new concept durable-agents (long-running pause/resume: state machines over conversation history, persistent sessions via DatabaseSessionService, webhook-driven resumption, human-in-the-loop pauses). Updated adk (now spans skills and durable execution) and synthesis (added durability as a 4th maturation axis; reframed the deployment axis as durable-pause/resume [ADK] vs. continuous-heartbeat [OpenClaw]). Linked to agentsys/gstack persistent state. Caveat: vendor tutorial.

[2026-06-01] ingest | Hermes Agent (Nous Research) — self-improving autonomous agent

Routed in by the hub (clear match). Created hermes-agent (SoftwareSourceCode) + new concept self-improving-agents (closed learning loop: agents authoring/refining their own skills + accumulating memory — the "growth" axis). A convergence node: skills (agentskills-spec- compatible) + agent-orchestration (subagents) + durable-agents (hibernation, persistent memory) + self-improvement; 200+ models (llm-provider), MCP, multi-platform. Explicit OpenClaw successor (imports OpenClaw memories/skills) → advances the self-driven pole of claude-code-channels-vs-openclaw. Bridges to ../research-wiki: compound-engineering inside the agent, and gbrain (which already cites "Hermes agent deployments"). Updated durable-agents

  • synthesis (added self-improvement as a 5th theme/growth axis). Caveat: README claims.

[2026-06-01] ingest | Spec-driven & framework burst (spec-kit, get-shit-done, Jaseci, node-js skill)

Four agentic-tooling sources routed in together:

  • spec-kit (GitHub, SoftwareSourceCode) — reference SDD toolkit (/speckit.* commands, 30+ agents).
  • get-shit-done (GSD, SoftwareSourceCode) — Claude Code context-engineering + SDD system.
  • jaseci (SoftwareSourceCode) — Jac language + AI framework: Meaning Typed Programming (compiler builds prompts from typed signatures) + graph walkers for agentic workflows.
  • node-js-functional-patterns-skill (SoftwareSourceCode) — a skill via MCP Market; fetch 429'd → honest stub; notable mainly for the skills-marketplace distribution angle (rhymes with Hermes' Skills Hub + agentskills.io). Updated spec-driven-development with spec-kit (canonical, tool-agnostic) + GSD (context-eng). Jaseci adds a language-level design point (type system as the model interface, vs markdown skills). Caveats: README/marketing; node-js skill unconfirmed.

[2026-06-03] ingest | Microsoft launches Scout (OpenClaw-inspired assistant)

Routed in by the hub. Created microsoft-scout (NewsArticle): Microsoft 365 AI assistant built on the OpenClaw framework — autonomous/always-on (claude-code-channels-vs-openclaw self-driven pole), persistent identity + memory (durable-agents), learns skills/memories (self-improving-agents), prebuilt + custom agent-skills, and a "policy conformance" continuous-audit governance layer. Significance: OpenClaw is now forked (hermes-agent) AND productized (Scout) → the self-driven personal-assistant pattern goes mainstream, with governance as a product feature. Updated microsoft (two agent plays: Agent Framework + Scout) and synthesis (self-driven pole going mainstream). Caveat: launch coverage, early-access.

[2026-06-03] lint | hub-wide health check (fixes landed here)

Part of a hub-wide lint. Fixed 2 orphans: jaseci now linked from agentic-coding-harness (language-level adjacency) and node-js-functional-patterns-skill from agentskills-spec (skills-as-packages/marketplace note). Tidied microsoft frontmatter comment (removed [[ ]] from the YAML comment that tripped the sources-nit check). No contradictions found here.

[2026-06-04] ingest | How to Build a Custom Agent Harness (LangChain blog)

Hub-routed (Telegram). New source summary langchain-custom-harness (BlogPosting, url-only). LangChain's agent = model + harness how-to. Two contributions: (1) the cleanest statement of the core thesis — "task-harness fit determines agent usefulness more than raw model capability"; (2) a concrete mechanism the wiki was missing — agent-middleware: single-concern pieces that hook the agent loop (before/after model & tool calls, startup/teardown) and compose/stack rather than forming a monolith. Four levers (deterministic logic, tool lifecycle, custom state, stream handlers) + a prebuilt catalog (summarization, context-editing, memory, subagents, retries, PII, human-in-the-loop, prompt caching). Frameworks named: create_agent, Deep Agents, Claude Agent SDK, Pi.

New pages: agent-middleware (DefinedTerm — the structural primitive) and langchain (Organization — fourth agent-platform vendor, distinct in leading with middleware-as-primitive vs the skills-as-primitive of google/microsoft). Substantially updated agentic-coding-harness (added the model + harness formula, task-harness fit, and a new "structural unit" section mapping LangChain's middleware onto existing threads) and synthesis (layer 2: middleware as the named structural unit + "harness = middleware stack over orchestration/durability/supervision/skills"; Vendors: added LangChain + the skill-vs-middleware primitive tension). Cross-linked Pi→oh-my-pi. Index updated (new DefinedTerm, BlogPosting, Organization rows). No contradictions — reinforces the central thesis.

[2026-06-04] ingest | OpenHuman — local-first personal AI agent (GitHub)

Hub-routed (Telegram); runner-up spoke ../research-wiki. New source/Thing page openhuman (SoftwareSourceCode, url-only). A GPL-3.0 (~30.7k★) Tauri desktop personal AI agent (Rust/TS), explicitly positioned vs claude-cowork / OpenClaw / hermes-agent — a third entrant on the self-driven autonomous-assistant pole (claude-code-channels-vs-openclaw) after hermes-agent and microsoft-scout, distinguished by local-first / privacy: on-device SQLite, optional Ollama, and a "Memory Tree + Obsidian wiki" knowledge base (Karpathy-inspired) — i.e. a gbrain-style (cross-wiki) personal KB as the agent's memory. Other features: 118+ OAuth integrations via Composio w/ 20-min auto-fetch; full coder toolset (also an agentic-coding-harness); TokenJuice ~80% context compression; model routing; voice (ElevenLabs TTS, cf. claude-to-speech); meeting mascot.

Updated synthesis (layer 3 self-driven pole: added OpenHuman as the local-first/privacy corner + "this pole sits astride the ../research-wiki seam") and back-linked microsoft-scout (lineage line + Related). Index: new SoftwareSourceCode row. Routing note: genuine two-spoke straddle — primary here (it's an agent harness/assistant benchmarking itself against this wiki's residents), runner-up ../research-wiki (the memory-tree/obsidian-wiki tools-for-thought angle); resolved per HUB.md "pick the most specific, note the runner-up," with the KB angle carried via the gbrain cross-wiki bridge. Reinforces the existing self-driven thread; no contradictions. Caveat: README/landing claims, Early Beta, unbenchmarked.

[2026-06-04] ingest | Anthropic's ant CLI hands-on (Medium)

Hub-routed (Telegram). New source summary ant-cli (Article, url-only). ant is Anthropic's first-party CLI for the claude-managed-agents lifecycle: create → deploy to Anthropic's cloud → send tasks → pull execution traces, "without creating any wrapper code" — collapsing prior REST/SDK boilerplate into terminal commands. Front-ends the April-2026 Managed Agents /v1/agents surface (agents carry config incl. model selection). Author's framing matches our vendor thesis: "Anthropic moving from model provider to full agent platform."

Updated claude-managed-agents (added a "Terminal client — ant CLI" section; it now has a second source beyond the FSI README; bumped date + Related) and synthesis (Vendors para: the ant CLI as the developer-facing edge of Anthropic's model-provider→platform shift). Index: new Article row. Fetch caveat: Medium truncated mid-article — full command set + the "four core concepts" + detailed limitations not recoverable; recorded an honest stub on specifics (lifecycle & positioning are solid). Single practitioner review, unbenchmarked. No contradictions.

[2026-06-04] ingest | What AI Agents Should Never Do on Their Own (Towards Data Science)

Hub-routed (Telegram). New source summary agents-never-do-alone (Article, url-only). Thesis: autonomy is only useful when bounded; the organizing lens is reversibility / recovery cost (high autonomy where mistakes are cheap to undo; human checkpoints where irreversible). Six hard-stop categories (destructive file ops, DB writes/migrations, cloud infra, prod deploys, auth/security logic, secrets). Mechanisms: AGENTS.md contract, blocked_commands.md explicit block list, two-agent implementer/egoless-reviewer loop, final reports, human-reserved deploy timing.

New concept page agent-guardrails (DefinedTerm) — names/operationalizes the governance/ reliability discipline that recurred but had no home: it ties together microsoft-scout's policy-conformance audit, hermes-agent's command-approval, and the harness review/eval discipline into a permission matrix keyed on reversibility. Positioned as the explicit counterweight to the autonomy push of durable-agents / self-improving-agents. Updated synthesis (layer-3 governance sentence now points to the named framework) and agent-orchestration (the two-agent adversarial-verification pattern doubles as a safety gate). Cross-links: AGENTS.md↔spec-driven-development constitution, human-in-the-loop↔agent-middleware/agent-kanban. Index: new DefinedTerm + Article rows. Caveat: single practitioner op-ed, unbenchmarked. No contradictions.

[2026-06-08] ingest | ai-job-search — Claude Code job-application framework

Hub-routed (Telegram, GitHub; tracking params stripped from URL). New source ai-job-search (SoftwareSourceCode, url; Mads Lorentzen, MIT). A fork-and-customize Claude Code harness for the whole job-application lifecycle: slash commands (/setup, /scrape, /apply, /expand, /upskill), skills (.claude/skills/ core job-application-assistant + job-scraper + upskill; .agents/skills/ TS/Bun Danish-portal CLIs), CLAUDE.md as master profile, and a drafter–reviewer subagent loop in /apply (reviewer independently researches + critiques → drafter revises). Notable patterns: output-grounded verification (compile LaTeX → read rendered PDF → fix layout until visual inspection passes), relevance-weighted CV cutting, token-efficient reviewer dispatch. Significance: the skills+harness+orchestration pattern generalizing to a non-coding vertical. Cross-linked to agentic-coding-harness, agent-skills, agentskills-spec, agent-orchestration, agents-never-do-alone, anthropic; folded into synthesis (layer 2, non-coding vertical + artifact-verification). Index updated. Caveat: author's README, no independent eval; Denmark-focused scraping.

[2026-06-09] ingest | lathe — skills that teach you (anti-automation augment pole)

Hub-routed (Telegram, GitHub). New source lathe (SoftwareSourceCode, url; Deven Jarvis, MIT). A Go CLI that runs as skills inside Claude Code/Cursor/Codex (/lathe, /lathe-extend, /lathe-verify, /lathe-ask, /lathe-tag, /lathe-voice) to generate hands-on technical tutorials you work through by hand, served at localhost:4242 — "LLMs to teach you, rather than think for you." Significance: same skill primitive as the rest of the spoke, used for the opposite goal — the extreme augment pole of the augment→automate axis (deliberately keeping the human doing the work), a pedagogical sibling of agent-guardrails/agents-never-do-alone and the agent-tooling instance of Engelbart's augmenting-human-intellect (../research-wiki, cross-wiki). Folded into synthesis (layer 3, after guardrails: "skills are pole-agnostic"); index updated. Caveat: solo "experiment," README-only. agentic-tooling-wiki 44 → 45 pages.

[2026-06-09] ingest | pm-skills (PM Skills Marketplace) — skills generalize to product management

Routed here by the hub router (Telegram). pm-skills is a marketplace of agent-skills for product managers — "100+ agentic skills, commands, and plugins" (68 skills / 9 plugins / 42 chained workflows; MIT; Paweł Huryn; ~13.3k★). Three artifact types: skills (PM-task knowledge/workflows), commands (slash workflows that chain skills → lightweight agent-orchestration), plugins (by PM domain: Discovery/Strategy/Execution/Launch/Growth/Market Research/Data Analytics/Toolkit/AI Shipping). Primarily Claude Code/Cowork plugins, but cross-vendor (Codex native; exported to Gemini CLI/Cursor/Kiro) — a real stress test of the portable-skill thesis. Folded into synthesis layer 2 as the productized case of skills-beyond-coding (sibling of ai-job-search, lathe); index updated. Caveat: README/marketplace-page only, counts as advertised. Cross-spoke context only (Growth/GTM → search- marketing; PM discovery → ../research-wiki). agentic-tooling-wiki 45 → 46 pages.

[2026-06-09] ingest | +2 AGENTS.md + Claude Agent SDK — all-spokes cron test

agents-md (DefinedTerm, src — open project-context convention; 60k+ repos, cross-tool, Linux Foundation; the context-standard counterpart to agentskills-spec) and claude-agent-sdk (SoftwareApplication, src — Anthropic's official SDK = the Claude Code harness as a library; tools/loop/ hooks/subagents/MCP/sessions/skills; the Client-SDK→Agent-SDK→Managed-Agents ladder). Both deepen existing threads (open standards; model-vendor→agent-platform). url-only. 46 → 48 pages.

[2026-06-10] ingest | SeekDB (OceanBase) — the agent memory/state substrate as a product

Hub-routed (Telegram, GitHub; tracking params stripped). New source seekdb (SoftwareSourceCode, url; OceanBase, Apache-2.0) — "the AI-native state store for agents": MySQL-compatible, embedded or server, hybrid vector + full-text search in one SQL query, immediate write→read (async index pipeline + two-level HNSW), full ACID, and the distinctive copy-on-write FORK/MERGE sandboxes (branch agent state, explore, merge/rollback). Drops in under LangChain/LlamaIndex/Dify via the MySQL wire protocol. Significance: the spoke's memory layer, until now always inside a harness (openhuman Memory Tree, durable-agents state, self-improving-agents accumulation, gbrain cross-wiki), shows up for the first time as a standalone infrastructure product — a separate vendor stratum (a proven DB team, not an agent-platform vendor) competing to be the substrate harnesses run on. The FORK/MERGE primitive = the storage-layer expression of the reversibility/exploration discipline in agent-guardrails. New concept agent-memory (DefinedTerm — names the state + retrievable-knowledge layer of its own) and light maker org oceanbase. Folded into synthesis (new "memory/state substrate becomes a product" section) + index (DefinedTerm/SoftwareSourceCode/ Organization rows). Runner-up ../research-wiki (the agent-memory/RAG angle) carried via the gbrain bridge, not duplicated — same handling as openhuman. Caveat: 10×-Milvus throughput is the repo's own benchmark, unreplicated. 48 → 51 pages.

[2026-06-12] ingest | A2A (Agent2Agent protocol) — a2a-protocol.org

All-spokes daily expansion. Added a2a-protocol (@type WebAPI) — the inter-agent protocol completing the standards layer (skills agentskills-spec / context agents-md / tools MCP

  • now agent↔agent interop). Google-originated, donated to the Linux Foundation; Agent Cards (capability discovery), tasks with streaming/async delegation, JSON-RPC/HTTP; cross-vendor (LangGraph/ CrewAI/ADK/custom) without exposing internal memory/logic. Cleanest framing: MCP = agent→tool, A2A = agent→agent (complementary). Promotes agent-orchestration from in-process to a cross-vendor wire format; sharpens the standardization-reach open question (4 standards = attractors or fragmentation; ACP + vendor-native coexist). synthesis note added; index WebAPI group updated. 1 new page. Authoritative (the protocol's own spec site / Linux Foundation).

[2026-06-12] ingest | Hermes Agent Profile Builder (Nous Research, via MarkTechPost)

Telegram drop, routed → agentic-tooling-wiki. Quality gate (first source under the new mechanism): tier T3 (MarkTechPost trade-press; Nous repo would be the T1 primary — weakness recorded, ingested per the soft gate); dedup → hermes-agent already exists, so added a distinct source page hermes-profile-builder for the new tool (not a re-seen refresh) + folded a section into hermes-agent; gap-relevant (advances the standardization-reach + agent-as-composition threads); freshness: volatile (fast-moving product). Substance: a local web dashboard (127.0.0.1:9119) composing an agent from 4 blocks — identity / model+provider / skills / MCP servers — writing config.yaml/.env/SOUL.md with CLI parity; open-source (pip install 'hermes-agent[web]'). Synthesis note "agent = composition gets a GUI"; index + hermes-agent updated. 1 new page.

[2026-06-14] ingest | renwei-writing (人味儿写作) — github.com/orange2ai/renwei-writing

Routed from hub (Telegram drop). An AI agent skill (SKILL.md + post-edit-checklist + case-study, progressive disclosure) for editing text without erasing the author's voice — packaged for the Cola framework (~/.cola/skills/). Quality gate: tier T1 (first-party project repo); not previously paged → new source summary renwei-writing. Gap-relevance: (1) extends the augment pole (lathe) to writing — same skill primitive, anti-homogenization use; (2) new standardization-reach data point (Cola = another host with a SKILL.md-style dir). Integrated surgically into synthesis (augment-pole thread + standardization open question); no new "Added" section, to avoid drift. freshness: stable. url provenance. Site rebuild + commit follow.

[2026-06-14] ingest | anthropics/skills — official Agent Skills repository (github.com/anthropics/skills)

Routed from hub (Telegram drop). Anthropic's official public Agent Skills repo: the spec (./spec/), a template, example skills + the source-available document skills (docx/pdf/pptx/xlsx) that power Claude's production file handling, and a Claude Code plugin marketplace. Quality gate: tier T1 (first-party project repo); not previously paged → new source summary anthropic-skills. Gap-relevance: the canonical first-party home of the agentskills-spec standard — a strong counterweight on the standardization-reach open question (a major vendor authoring + shipping the spec, not just adopting it). Threaded into agentskills-spec (new "canonical first-party home" section) + synthesis (standardization thread). freshness: stable (repo identity durable; star count noted as a dated snapshot, not load-bearing). Site rebuild + commit follow.

[2026-06-14] dedup | Google — removed duplicate page (now a cross-wiki bridge node)

Deleted the local google.md (a thin agent-vendor facet duplicating the real-world entity). Google is now a single canonical node in ../llm-providers-wiki, linked cross-wiki; index.md moves it to Bridge nodes. The agent-platform-vendor substance is preserved on the canonical node + in synthesis. Closes the entity-index audit's flagged duplicate.

[2026-06-14] ingest | tgpt — AI chatbots in the terminal (github.com/aandrew-me/tgpt)

Routed from hub (Telegram drop). A cross-platform Go CLI (GPL-3.0) that fronts ~10 LLM backends from a shell prompt — a multi-provider chat frontend, not an agent (no tool use / file edits / planning loop). Quality gate: tier T1 (the project's own README + providers.md); not previously paged → new source summary tgpt. Gap-relevance: a boundary marker — the non-agentic, prompt→reply floor of the terminal-AI ecosystem (beneath claw-code / oh-my-pi / Gemini CLI), reinforcing by contrast the spoke's "value is in the structure" thesis. Backends link the existing cross-wiki provider nodes (openai, deepseek, Gemini); Groq/Ollama/MiniMax/Pollinations/KoboldAI/Isou/Arta left unpaged (evidence-only entity rule). Cross-spoke: runner-up ../llm-providers-wiki — tgpt is really an access-layer artifact over the model market (free + paid + open-weight access in one CLI); noted in the source page + synthesis, not split. Threaded into synthesis (new boundary-marker note) + index. freshness: stable. url provenance. Site rebuild + commit follow.

[2026-06-15] lint | agents-that-build-agents-ms — stub upgraded; spec-adoption question resolved

Quality cycle: the founding community-hub source is JS-gated (still un-fetchable), so rebuilt agents-that-build-agents-ms from fetchable companions — the authoritative Microsoft Learn doc Use skills with Microsoft Foundry agents + the author's ITNEXT cross-post (found via WebSearch). Resolves the synthesis open question: Microsoft Foundry skills "follow the Agent Skills specification format" (agentskills.io) — SKILL.md + YAML front matter, progressive-disclosure advertise→load→read via AgentSkillsProvider, surfaced to any MCP client as MCP Resources (SEP-2640), plus a versioned Foundry Skills REST API. So Anthropic + Google + Microsoft converge on the same open format wired to MCP — convergence, not fragmentation. Documented the build-time/run-time two-layer architecture (Coding Agent authors SKILL.md → Runtime Agent runs it). Updated synthesis (Standardization open question + Vendors paragraph); page ret(ained) BlogPosting/T3, url unchanged (Learn doc cited inline).

[2026-06-15] ingest | Loops replace prompts; verification as the bottleneck (hub-routed)

Hub-routed Telegram drop (agent-loops-verification, The New Stack, Arjun Iyer/Signadot, 13 Jun). The New Stack body is JS-gated (third time this session); substance recovered via WebSearch + the surrounding TNS/Signadot article cluster. Thesis: agentic coding shifts prompt→loop (loop-engineering; Ralph technique generalized), relocating the bottleneck from generation to verification — and "feedback is only as truthful as the system that generates it," so for cloud-native code verification is a runtime problem (validate against a real system). New pages: agent-loops-verification (TechArticle, T4 — trade press + vendor-authored), loop-engineering (DefinedTerm), signadot (Organization, the ephemeral-environment vendor), arjun-iyer (Person, author/CEO). Threaded into synthesis layer 3 (verification now a first-class layer; output-grounded verification + guardrails matter MORE as loops outpace review) + cross-spoke adjacency (new platform-ops verification-runtime seam, noted bidirectionally in platform-ops/synthesis.md). Tier T4: body JS-gated, vendor-interested (Signadot sells the fix) — problem-framing kept, product treated as one data point; T1/T2 upgrade = a neutral study of loop-driven verification rates. Cross-spoke runner-up: ../platform-ops-wiki (cloud-native runtime substrate).

[2026-06-15] lint | microsoft org page — stub upgraded (Foundry + Agent Framework)

Resolved the honest-stub on microsoft (founding source body never fetched). Enriched from Foundry/Agent Framework public docs (via WebSearch): Microsoft Foundry = unified "agent factory" platform (Foundry Agent Service, Foundry IQ = evolved Azure AI Search, Foundry Toolkit for VS Code, control-plane governance, multi-agent orchestration + A2A APIs, Semantic Kernel/AutoGen/ CrewAI support); Microsoft Agent Framework = the open-source agent SDK beneath it. De-hedged the former "does MS adopt the spec?" open question — resolved (adopts agentskills.io + MCP Resources, per the upgraded agents-that-build-agents-ms). Removed the stub warning; index entry refreshed. No longer a stub.

[2026-06-16] ingest | awesome-hermes-usecases (hub-routed, Telegram) — evidence counterweight + Nous node

Routed in as a clean single-spoke match (Hermes Agent already paged here). Added awesome-hermes-usecases (SoftwareSourceCode, source:true, T2 — community-curated by aliaihub, per-entry primary-source rule)

  • new entity nous-research (Organization, the owner/developer common to hermes-agent/profile-builder). Gap-relevance: it's the evidence counterweight to hermes-agent's standing README-claims caveat — third-party, primary-source-gated deployment catalog across 13 domains; softened that caveat in place and named the self-improvement mechanism (GEPA + DSPy) the README abstracts away. Folded into synthesis "Benchmarks" open question as a partial step (usage evidence, not a benchmark). Dedup: hermes-agent / hermes-profile-builder refreshed in place, not duplicated. +2 pages.

[2026-06-16] ingest | autonovel (hub-routed, Telegram) — Hermes Agent flagship deployment + loop instance

Second Hermes-orbit source today. Added autonovel (SoftwareSourceCode, source:true, T1 first-party Nous repo) — an end-to-end autonomous novel pipeline orchestrated by hermes-agent: 4 phases / ~27 scripts, quality-gated loops (>7.5 / >6.0 / plateau detection), an Opus review loop, and a dual-"immune-system" feedback signal (mechanical regex scan + separate LLM-judge). Concrete artifact: a 79,456-word novel. Gap-relevance: strongest first-party evidence for Hermes' deployed capability (softened the README-claims caveat on hermes-agent further) AND a vivid worked instance of the loop-engineering/agent-loops-verification "verification is the bottleneck" thesis — folded into that synthesis thread. Model-agnostic note recorded: "by Hermes Agent" but runs Claude Sonnet/Opus + fal.ai + ElevenLabs (Hermes = conductor). Entity discovery: no new nodes — nous-research paged earlier today; Anthropic/claude-opus-4-8/Karpathy exist cross-wiki; fal.ai + ElevenLabs incidental deps (skip). Cross-linked into hermes-agent / awesome-hermes-usecases / nous-research. +1 page.

[2026-06-16] ingest | autoresearch (hub-routed, Telegram) — the loop paradigm's origin + feedback-signal contrast

Third loop-instance source today; the origin of the autoresearch paradigm autonovel credits. Added autoresearch (SoftwareSourceCode, source:true, T1 first-party Karpathy repo) — his "agentic scientist" PoC: agents edit train.py only → 5-min training → measure val_bpb → keep/discard, ~100 overnight runs. Gap-relevance: it's the cleanest loop-engineering/agent-loops-verification instance AND sharpens the thesis by contrast with autonovel — autoresearch has an objective metric (val_bpb) so autonomy is near-free; autonovel had to engineer a feedback signal (LLM-judge) because prose has no ground truth. Folded that contrast into the loop synthesis thread; also reframed containment (edit-only-train.py, reviewable diffs) as an autonomy enabler. Cross-linked autonovel↔autoresearch. Cross-spoke: Karpathy is a ../research-wiki node (andrej-karpathy) — linked the bridge, ingested here per the agent-tooling split (../research-wiki = runner-up). Entity discovery: no new nodes (Karpathy exists cross-wiki; nanochat = incidental infra, mentioned inline). +1 page.

[2026-06-16] ingest | second-son-house-of-bells (hub-routed, Telegram) — the autonovel artifact, reader-facing

Fourth Hermes-orbit source today, and the output end of the autonovel thread: the published novel's own landing page (nousresearch.com/bells). Added second-son-house-of-bells (Book, source:true, T2 — first-party but promotional landing page; softer than the autonovel repo, T1). Deliberately NOT a refresh of autonovel: that page is the pipeline/repo, this is the artifact (the Cantamura world where "the law is sung," protagonist Cass Bellwright, three distribution formats). Kept the pipeline stats on autonovel and the reader-facing framing here; cross-linked both ways. Notable: the byline "Claude Hermes" — the agent given a named authorial identity (Hermes orchestrator + Claude models worn as a name); mapped to hermes-agent/claude-opus-4-8 rather than a new Person node (synthetic byline, not a real person → dedup-halts). Folded the artifact-pole + author-persona note into the loop synthesis thread. Entity discovery: no new nodes (Nous + Hermes paged; no real Person/Org/Event introduced). +1 page.

[2026-06-16] ingest | agentic-tooling (umbrella node) — kind-coverage gap

Added the spoke's missing domain umbrella agentic-tooling (DefinedTerm, kind:domain) — flagged by the entity-gaps kind-coverage audit (11 kinded concepts, no domain node). Ties the synthesis's layers together (skills/agentskills-spec · agentic-coding-harness · agent-orchestration/loop-engineering · durable-agents/self-improving-agents · agent-guardrails) under the "structure is the lever" through-line; links synthesis. No new sources (synthesis node). +1 page.

[2026-06-16] ingest | zouroboros (hub-routed, Telegram) — convergence instance + consensus-gate verification

GitHub repo (marlandoj/zouroboros) via Telegram — a self-enhancing multi-agent orchestration platform. Added zouroboros (SoftwareSourceCode, source:true, T1 first-party repo: code + 757 tests + 8 npm packages). Clean single-spoke route: dominant substance is agent orchestration + memory + self-improvement + resilience — a convergence instance like hermes-agent/gstack (it even uses Hermes as one of 5 executors). Gap-relevance: a third design point for the loop/verification thesis — its self-improvement loop gates procedure changes through a 3-model consensus vote (GLM-5.1/Kimi-K2.6/ MiniMax-M2.5), vs autonovel's LLM-judge and autoresearch's objective metric; folded into the self-improvement synthesis thread (layer 5) as the head-on answer to the drift/quality open risk. Cross-linked into self-improving-agents (instance) + agent-memory (episodic/procedural/cognitive over SQLite+vector). Entity discovery: +1 node zo-computer (Organization, thin/evidence-only — the host platform; no match in entity-index); skipped the bare GitHub handle marlandoj (low-signal pseudonym, recorded inline). Runner-up spoke: llm-providers (the consensus models GLM/Kimi/MiniMax) — noted, not split. +2 pages (agentic-tooling → 68).

[2026-06-17] ingest | Memory Vault — Postgres+pgvector MCP memory for Claude Code

Hub-routed from a Telegram MakeUseOf how-to ("I fixed Claude's memory problem with a Postgres database…"). Dominant substance is an agent-memory infrastructure product, so a clean single-spoke route here. Added memory-vault (SoftwareApplication, source:true, T3 — a MakeUseOf install-walkthrough, not project docs or an independent test; url-only). It's the second standalone memory-as-infrastructure instance after seekdb and marks the opposite weight pole: a thin open-source MCP server (stock Postgres + pgvector, hybrid vector+keyword search, docker compose up) whose distinguishing move is delivery over MCP — the store reaches the agent as a tool (the storage-layer reading of "skills × MCP = agency"), vs seekdb being a database the harness queries. Target: externalize the context Claude Code loses to lossy auto-compaction so memory survives /clear / machine switch. Folded into agent-memory (new "lightweight instance" section + two-pole framing) and synthesis layer 6. Gap-relevance: gives the memory-infra layer a second data point and a market-spectrum (heavyweight DB ↔ thin MCP store). Entity discovery: no new nodes — publisher MakeUseOf (generic T3 how-to site) and tool author mihaibuilds (thin GitHub handle) are low-signal, recorded inline (same call as the marlandoj skip). Reused cross-wiki bridges anthropic, model-context-protocol. Runner-up spoke: ../research-wiki (the persistent-memory / tools-for-thought angle) — noted, not split. +1 page (agentic-tooling → 69).

[2026-06-17] ingest | The Second Son of the House of Bells — re-seen refresh

Same URL (nousresearch.com/bells) arrived again via Telegram. Re-seen source → refreshed second-son-house-of-bells in place (no duplicate, no new page; count stays 69). Verified the live page against the existing summary: all prior claims hold. One concrete addition — the landing page credits the work explicitly as "A work of fiction by Hermes Agent," which directly backs the hermes-agent-orchestrator side of the "Claude Hermes" persona reading (previously inferred from autonovel). Folded that line into the persona section; bumped updated → 2026-06-17.

[2026-06-18] ingest | Agentic Resource Discovery (ARD) specification (Google)

Telegram-dropped; hub-routed → agentic-tooling (runner-ups: ai-governance for the trust/verification layer, search-marketing for "registries = search engines for the agentic web"). Source summary + standard page agentic-resource-discovery (DefinedTerm, source:true, url, T1 — Google's first-party announcement). Gap-relevance: adds a fifth edge to the spoke's standards layer — discovery (how an agent finds + trusts capabilities across orgs) alongside skills/context/tools/interop. ARD is a meta-standard: ai-catalog.json at a well-known path + registries indexing MCP servers / A2A agents / OpenAPI tools, domain-rooted crypto trust; data model under the Linux Foundation AI Catalog WG (same governance arc as a2a-protocol/agents-md). Integrated into synthesis (four→five edges) and cross-linked from a2a-protocol (ARD indexes Agent Cards). Entities: publisher google (bridge node, ../llm-providers-wiki) linked; authors Junjie Bu / Srinivas Krishnan noted in prose but not noded — consistent with house style for Google-blog ingests (the ADK guides got no author nodes). Cross-spoke facets recorded in the page (ai-governance trust layer; search-marketing agentic-web discovery). Adoption one-vendor-deep (Google Cloud Agent Registry / Gemini Enterprise). freshness: fresh. Rebuild + commit follow.

[2026-06-18] ingest | Anthropic "pauses" token-based billing for the Claude Agent SDK (Ars Technica)

Telegram-dropped; hub-routed → agentic-tooling (runner-up: ../llm-providers-wiki for the API-pricing angle — noted cross-spoke, not split, since the dominant subject is the Agent SDK product and its harness economics). Ars Technica (Kyle Orland) hard-blocked WebFetch → fetched via the firecrawl skill. New source summary anthropic-agent-sdk-billing-pause (TechArticle, source:true, url, T2 — reputable secondary journalism over first-party support-page text + named primaries). Gap-relevance: opens the spoke's first explicit treatment of harness cost structure — the unaddressed flip side of the "value-moved-to-the-wrapper / model-vendor-sells-the-wrapper" thesis. Contradiction handled: the prior claude-agent-sdk page asserted "from 2026-06-15 it draws on a separate Agent-SDK credit" — that change was paused June 15/16, so I replaced the stale parenthetical with a dated "Billing: a paused metering change" section (both the announced plan and the pause recorded) and bumped its updated. Folded the cost-structure tension into synthesis (Vendors). Entities: new ars-technica (NewsMediaOrganization, publisher) + boris-cherny (Person, Head of Claude Code — the quoted voice; worksFor anthropic bridge node). Author Kyle Orland (gaming editor, tangential) and third-party consumers Zed/Xcode/JetBrains noted as plain mentions per house style. +3 pages (agentic-tooling → 73). Rebuild + commit follow.

[2026-06-18] ingest | autoresearch (Karpathy) — re-seen refresh

Same URL (github.com/karpathy/autoresearch) arrived again via Telegram. Re-seen source → refreshed autoresearch in place (no duplicate, no new page). Re-fetched repo against the existing summary: all prior claims hold (single-GPU nanochat, edit-train.py-only action surface, fixed 5-min budget, val_bpb keep/discard loop, MIT). Only deltas were noise (stars ~87k→87.5k, community MacOS/Windows/AMD forks); nothing material to add. Bumped updated → 2026-06-18.

[2026-06-18] ingest | autoresearch (Karpathy) — landing page re-seen

The project's official landing page (autoresearch.lol) arrived via Telegram (hub msg 513) after the repo refresh above. Same project, different URL — recorded the landing-page URL on autoresearch as first-party provenance (it restates the "five-minute loop" pitch + Python 3.10/uv/NVIDIA-GPU reqs, adds nothing beyond the repo). Repo stays canonical; no new page, no duplicate.

[2026-06-18] ingest | DefinedTerm enrichment pass (subagent, completed by orchestrator)

Hub-wide DefinedTerm enrichment ("go deeper, fetch sources"). A subagent deepened the thinnest concept pages with fetched T1 sources but was cut off by a session limit before finishing bookkeeping; the orchestrator completed the index/log. New source pages: building-effective-agents (Anthropic), adk-sessions-docs (Google ADK), agents-md-spec (agents.md), langchain-middleware-docs, self-evolving-agents-survey (arXiv 2508.07407). Deepened concept pages: agentic-tooling, agents-md, durable-agents, loop-engineering, self-improving-agents (sourced facts + Related footers). No fabricated facts; AI-writing gate re-run by orchestrator (clean).

[2026-06-20] ingest | Cult UI — AI agent patterns collection (hub-routed)

Telegram drop (hub msg 532; osp.fyi/cult-ui → opensourceprojects.dev). @nolly-studio's GitHub collection of 92+ copy-paste AI-agent patterns (ReAct / plan-and-solve / RAG + web-search/code-exec/db building blocks) on the Vercel AI SDK — "recipes, not a new SDK." New source cult-ui-agent-patterns (SoftwareSourceCode, T3 — secondary roundup of the first-party repo). The recipe-library point on the agent-structure spectrum; a code-level instance of building-effective-agents's "simple composable patterns over heavy frameworks" thesis, on the "agent = composition" thread. Cross-linked agentic-coding- harness / orchestration / spec-driven-development.

[2026-06-19] ingest | AutoWiki (Factory.ai) — hub-routed

Telegram drop (hub msg 529, factory.ai/news/wiki). Factory.ai's multi-agent system that auto-generates + continuously refreshes a code wiki ("documentation as a build artifact"), stored in-repo under droid-wiki/, browsable by Factory's Droids as shared agent context. Routed here (dominant substance = an agentic coding product) over runner-up ../research-wiki — though it's a concrete instance of ../research-wiki's llm-wiki concept (a machine maintaining a file-based markdown wiki, like this hub). New source autowiki-factory (TechArticle, T2 — first-party vendor, promotional). Connects the spoke's "structure/harness" thesis (agentic-coding-harness) and the agent-context thread (agent-memory). T2 vendor-bias recorded.

[2026-06-20] ingest | headroom (context-compression toolkit) — hub-routed

Telegram drop (hub msg 541, github.com/chopratejas/headroom). Local-first toolkit compressing tool outputs/ logs/files/RAG/history before the LLM — 60-95% fewer tokens, claimed answer-parity. Content-typed compressors (SmartCrusher JSON, AST-aware CodeCompressor, Kompress-base HF prose model, CacheAligner for KV-cache hits, CCR reversible). Surfaces: library / proxy / CLI / MCP server; cross-agent shared memory

[2026-06-21] ingest | Ponytail — "lazy senior dev" minimal-code agent skill

Source: DietrichGebert/ponytail (GitHub, MIT, ~45.9k★), via a Telegram-channel GitHub link (hub-routed). An AI agent skill that constrains coding agents to write the minimum code — decision ladder (need it? → stdlib → native feature → existing dep → one line → minimum that works), intensity modes (lite/full/ultra/off), review/audit/harvest/benchmark commands, "never compromise safety/validation/security."

Added ponytail (SoftwareSourceCode, T1 first-party repo; efficacy numbers flagged self-reported). Folded into synthesis layer 3 as a third counter-current beside the augment pole (lathe, renwei-writing): restraint of the agent rather than keeping the human in the loop — the loadable-skill primitive is now "pole-agnostic three ways" (add capability / augment human / discipline output down). Strong standardization-reach data point: 15+ hosts, still per-host delivery. Self-reported metrics (~54% less code, up to 94%; 20% cost, 27% faster, 100% safety) feed the standing benchmark open question — README claim, not a neutral measurement. Author noted inline (pseudonymous handle, low graph signal — skill-pack precedent). WebFetch on github.com succeeded. Updated index.md (+1 SoftwareSourceCode). +1 page.

[2026-06-21] ingest | Easy Agent — teaching-scaffold coding-agent CLI (rebuild a harness in 31 stages)

Source: ConardLi/easy-agent (GitHub, MIT), via a Dan Kornas tweet (hub-routed from Telegram). A terminal-native agentic coding CLI (TypeScript/Node, React/Ink TUI) that recreates a Claude Code-style workflow with the architecture made explicit: five layers (provider I/O, tools, agent loop, session/orchestration, terminal UI), real tool surface (file/shell/search/web/MCP + permissions

  • sandboxing), multi-provider (Anthropic/OpenAI-compatible/Gemini/Ollama), and 31 standalone step-snapshots for study.

Added easy-agent (SoftwareSourceCode, T1 first-party repo). Folded into synthesis layer 2: it turns the harness decomposition into a curriculum — the agent = model + harness formula made legible (same parts the spoke names, separated on purpose), provider-agnostic like claw-code, and a teaching-pole cousin of lathe aimed at how a coding agent is built. Author noted inline (GitHub handle, low signal). github.com WebFetch blocked earlier this session for x.com only; repo content via firecrawl scrape of the announcing tweet. Updated index.md (+1). +1 page.

[2026-06-23] ingest | Cybersecurity Skills (mukul975) + DeerFlow (ByteDance) — Telegram burst

Two agent-tooling sources from one Telegram burst:

  • cybersecurity-skills (T2) — an 817-skill cybersecurity skillpack on agentskills.io (Mahipal Jangra/@mukul975, Apache-2.0, ~18.7k★). Domain incidental like agentic-seo-skill; what's new is scale — the largest skillpack the wiki has paged, a working test of progressive disclosure at 817 frontmatters (~30 tokens each). Reflexive twist: 2 of its 6 framework mappings (MITRE ATLAS, NIST AI RMF) cover attacking/defending agents. Flagged the misleading "Anthropic-" name (community, NOT official — contrast anthropic-skills). Paged author mahipal-jangra (Person). Folded into synthesis layer 1 (skills).
  • deer-flow (T2) — ByteDance "SuperAgent" harness (LangChain/LangGraph, MIT, ~73.2k★): sandboxes + persistent agent-memory + markdown skills + parallel sub-agents + a Telegram/Slack/Feishu message gateway; long-horizon (minutes–hours). Paged bytedance (Organization) as the big-tech entrant in the harness-vendor stratum (microsoft/Google adk/langchain/ByteDance all ship the same composition). Folded into synthesis layer 2 (harness). Cross-linked existing langchain org node. Both fetched live via WebFetch. Index updated (SoftwareSourceCode +2, Organization +1, Person +1). +4 pages (agentic-tooling 83→87). Ran avoid-ai-writing criteria over new prose. (No build/commit here — hub handles.)

[2026-06-23] ingest | ADK + A2A cross-language multi-agent team (Google guide)

New source page adk-a2a-cross-language (T1 — Google Developers Blog, Saboo & Dong, 2026-06-22). Turns the A2A "agent↔agent across vendors/frameworks" thesis into a running cross-language example: a Python ADK orchestrator (SequentialAgent) drives a Go compliance validator wrapped as RemoteA2aAgent — the Go service is plain HTTP, no AI framework, A2A-compliant by protocol adherence alone. Exercises all three A2A mechanisms (Agent Card at /.well-known/agent.json, JSON-RPC message/send, Task states) + a resilience pattern (unreachable agent → human MANUAL_REVIEW). Refreshed a2a-protocol (added "Worked instance") and adk (RemoteA2aAgent/SequentialAgent orchestration; bumped updated). Folded into synthesis's interop-edge thread (orchestration provably crosses language+framework). Index updated (BlogPosting group). Authors noted inline; publisher cross-linked to google (no new person nodes for blog authors). +1 page (agentic-tooling 87→88). Fetched live. (No build/commit here — hub handles.)

[2026-06-23] ingest | OpenUI (thesysdev) — generative-UI framework

New source page openui (T2; via @tom_doerr on X, resolved through a share.google redirect; X post fetched with the firecrawl skill after WebFetch hit HTTP 402). OpenUI is a full-stack generative-UI framework: LLMs emit a compact streaming DSL ("OpenUI Lang", benchmarked ~67% fewer tokens than JSON) from a developer-defined component set, rendered progressively by a React runtime. Routed here as the second source in the agent/LLM-rendered-output corner after arrow-js, with the contrast as the synthesis hook: arrow-js = LLM writes real framework code, WASM-sandboxed (bound reach); OpenUI = LLM writes a constrained DSL from an allowed component vocabulary (bound grammar). Both are agent-guardrails containment moves; OpenUI's token-efficiency rhymes with the token-budget thread (headroom, progressive disclosure). Cross-linked arrow-js (added a counterpoint section). Scope caveat logged: OpenUI is general LLM-to-UI, not agent-specific — routed here for lack of a generative-UI/frontend spoke + the arrow-js precedent; spin-out candidate if the corner grows. Index + synthesis updated. +1 page (agentic-tooling 88→89). (No build/commit here — hub handles.)

[2026-06-23] ingest | Recall — offline durable memory for Claude Code

New source page recall (T2; raiyanyahya/recall, Python, MIT, ~387★). A Claude Code plugin for fully offline session memory: SessionStart/Stop hooks + slash commands, storage as plain markdown files (.recall/history.md append-only + context.md regenerated summary, committable for team memory), summarized by a deterministic vendored TF-IDF/TextRank extractor — no embeddings, no LLM, no DB. Routed as the third, most-minimal pole of agent-memory: the spectrum is now DB (seekdb) → pgvector-over- MCP (memory-vault) → markdown-files-over-hooks (recall), and recall also breaks the MCP delivery shape (hooks, not a tool). Updated agent-memory (new "minimal pole" section; bumped updated) and folded into synthesis's memory thread. Cross-link: recall's "committable markdown you own" is the strongest cross-wiki seam yet to ../research-wiki's gbrain/llm-wiki markdown-as-system-of-record. Index updated (SoftwareApplication group, beside memory-vault). +1 page (agentic-tooling 89→90). Fetched live. (No build/commit here — hub handles.)

[2026-06-23] ingest | OpenClaw vs Hermes — the control-layer contest (The New Stack)

New source page openclaw-hermes-control-layer (T2 — TNS analytical feature; fetched via firecrawl after WebFetch returned only page chrome). The cleanest comparative frame yet for the harness thread: openclaw and hermes-agent agree on the harness anatomy (runtime + gateway + memory + tools + identity + policy) but differ on the primary control point — OpenClaw gateway-first (breadth: dozens of channels, ClawHub marketplace, ~380k★, Steinberger→OpenAI-sponsored foundation) vs Hermes memory-first (depth: persistent memory + self-refining skills on infra you own; ~160k★, OpenRouter token leader; hermes claw migrate). Created canonical openclaw node (recurred across 3 pages without one) — gateway-first harness, governance added externally by Nvidia NemoClaw (OpenShell sandbox) + Microsoft Scout (Entra identity). Refreshed hermes-agent (memory-first/lock-in framing + metrics) and cross-linked/refined claude-code-channels-vs-openclaw (gateway-first recasts its self-driven framing — flagged as refinement, not contradiction). Folded into synthesis layer 2: memory = the durable lock-in (elevates agent-memory to strategic moat) + vendors claim the agent-agnostic runtime/governance layer ("outlasts any single model"). Security note: 341 malicious ClawHub skills. Index +2 (Article source + openclaw SoftwareApplication). +2 pages (agentic-tooling 90→92). (No build/commit here — hub handles.)

[2026-06-23] ingest | NemoClaw (NVIDIA) — agent-agnostic governance runtime

New source page nemoclaw (T2 — primary NVIDIA repo, Apache-2.0, ~21.4k★, alpha) + new nvidia org node. The first-party primary source for the vendor runtime-governance thesis just synthesized from openclaw-hermes-control-layer — and broader than that article implied: NemoClaw runs openclaw, hermes-agent, AND LangChain Deep Agents under one CLI inside OpenShell sandboxes. Governance around the agent: per-instance sandbox, capability dropping, egress/network policy + operator approval, lifecycle, and the distinctive managed/"routed" inference (the agent can't call model endpoints directly — governance owns which model it may call, not just which files). Folded into synthesis (the vendor-runtime-layer paragraph now cites the primary source + the 3-harness agent-agnostic detail + managed inference); linked nemoclaw from openclaw. agent-guardrails as external containment; the NVIDIA counterpart to Microsoft Scout (multi-harness/OpenShell vs single-harness/Entra). Index +2 (nemoclaw source + nvidia org). +2 pages (agentic-tooling 92→94). Fetched live. (No build/commit here — hub handles.)

[2026-06-24] ingest | Hermes /learn — skill capture without hand-writing SKILL.md (MarkTechPost)

New source page hermes-learn-skill-capture (T3 — MarkTechPost relay of a Nous Research feature). Hermes's new /learn command auto-authors a standards-compliant SKILL.md from a doc/URL/SDK/past conversation, reusing the agent's own tools (read_file/search_files/web_extract); skills land in ~/.hermes/skills/, become slash commands, use 3-tier progressive disclosure (L0 index ~3k tokens → L1 skill → L2 refs). Key analytical contribution = a skill-authoring autonomy spectrum: hand-write → /learn (assisted) → skill_manage auto-capture after 5+ tool-call workflows (unprompted) → Hub install (community), with an (off-by-default) write_approval review gate. Folded into self-improving-agents (the spectrum + approval-gate-as-drift- control, sibling of zouroboros's consensus vote) and hermes-agent (skill-authoring side); synthesis layer-5 note: the standard decouples who authors from the format as the author slides human→agent. Index updated (TechArticle group). +1 page (agentic-tooling 94→95). Fetched live. (No build/commit — hub.)

[2026-06-28] ingest | Google Antigravity 2.0

New page google-antigravity (SoftwareApplication, source, T3, geeky-gadgets.com), routed from the hub. Google's developer-facing agent IDE ("agent operating system") — the front-end twin to SDK-side adk. The 2.0 update split the unified app into two (VS Code-style IDE + separate chat-agent app), a bet on IDE-surface vs chat-surface as distinct control points; reception mixed (usability regression + tighter usage limits). Carries the standard harness anatomy: dynamic sub-agents, async/scheduled background ops in remote Linux, /goal & /browser slash commands + Chrome-DevTools browser control, an Antigravity SDK + CLI, JSON hooks, multi-folder scoped permissions. Folded into synthesis' harness-vendor stratum thread (Google now ships both kit and IDE; the split-app bet mirrors openclaw-hermes-control-layer within one vendor; the cost/UX friction is the consumer twin of anthropic-agent-sdk-billing-pause). Source weakness recorded: secondary tech-blog coverage of a product update, no primary docs — feature specifics reported, not confirmed; a primary antigravity.google source would upgrade this. No new entity nodes (owner google is an existing cross-wiki bridge node; geeky-gadgets too thin to page).

[2026-06-29] ingest | Tail Control — reliable agentic workflows (TDS) — hub-routed

Adds the reliability/latency dimension the spoke lacked (most of the corpus = building/orchestrating; this = hitting a deadline). New source tail-control-agentic-workflows (TechArticle, T3, towardsdatascience.com — Frank Wittkampf/Databook; 1.2M+ enterprise calls). Thesis: predictable completion beats fast-with-a-tail. Latency heavy tails (p99 30–80s vs ~12s median, 2–7× within identical requests) compound across steps (10×95%≈60%, 20≈36%); fixes = SRE tail-control applied to agents: cut at p95, race redundant attempts (first valid; p99 ~60→25s), hedge to another provider's quota, downshift on transient stalls. Folded into building-effective-agents (the workflow/agent split gains a reliability rationale — workflows are the tail-tractable form), agent-orchestration (parallelism for redundancy vs decomposition), and synthesis (reliability's second axis = latency, not just safety). Dedup: no existing reliability/tail node — net-new theme. Cross-spoke ../platform-ops-wiki (SRE tail-at-scale / hedged-requests heritage) noted, not forked. T3 weakness recorded (single-source practitioner data). Runner-up spoke: platform-ops.

[2026-06-29] ingest | Claude Fable system prompt "leak" (Medium) — hub-routed

New source claude-fable-system-prompt-leak (TechArticle, T4 — unverified leak, medium.com / Mehul Gupta). A claimed leaked Claude Fable system prompt; routed here because its content + own thesis is the spoke's "everything wrapped around the model": one configured layer of tools (web search, artifacts, MCP connectors, code execution), memory, safety (child-safety/malware/weapon refusals), and behavior (warm tone, anti-over-formatting — the same anti-slop rule the hub uses). Folded a single caveated line into the synthesis thesis intro (illustrates, doesn't prove). Authenticity heavily caveated: no provenance, low- tier blog, ~190K-token claim suspect — read as illustrative, not a primary source on Claude's real config; no durable Thing pages built on it. Cross-spoke: ../research-wiki owns the model substrate (Claude Fable 5) = runner-up; ai-governance adjacency (safety rules are product instructions, not policy) noted, not forked. Runner-up spoke: ../research-wiki.

[2026-06-30] ingest | Superpowers (obra) — agent-skills framework + dev methodology

Routed from a Telegram review article (member-only Level Up Coding / Medium, "I Installed 'Superpowers' Into Claude Code", Abhishek Agarwal — T4). Dedup: no existing page/mention. Grounded in the T1 repo superpowers (github.com/obra/superpowers, MIT) rather than the paywalled blog. New page superpowers (SoftwareSourceCode, T1) — Jesse Vincent (obra) + Prime Radiant team's composable agent-skills framework: ~two dozen skills the agent self-invokes (brainstorming, writing/executing-plans, TDD, verification-before-completion, systematic-debugging, requesting/receiving-code-review, git-worktrees, dispatching-parallel-agents, subagent-driven-development, writing-skills, using-superpowers). Cross-harness (Claude Code/Cursor/Codex/Copilot/Gemini CLI). Flagged volatile claim: ≈242k★ (GitHub snapshot) — high enough to re-verify; the review raised the same caution. Entity jesse-vincent (Person, obra) created; Prime Radiant kept as prose mention (thin). Folded into synthesis layer 1 (capability-as-markdown) as a methodology-as-skills instance spanning into orchestration; index updated (SoftwareSourceCode + Person). Reflexive note recorded (the skill system this hub session runs on). Ran avoid-ai-writing.

[2026-06-30] ingest | From local LLM to tool-using agent (TDS)

Routed from Telegram. New source from-local-llm-to-agent (TechArticle, T4, towardsdatascience.com) — a practitioner walkthrough building a fully-local "mini deep research agent": Gemma 4 E4B on Ollama (OpenAI-compatible endpoint) driven by the OpenAI Agents SDK calling MCP tools (Tavily search), running the canonical agent loop (query→tool calls→MCP exec→cited answer). Dedup: no existing page. Integrated into synthesis layer 1 as a concrete, fully-local worked example of the "skills × MCP = agency" formula — agency decoupled from the model substrate (no cloud model needed). Cross-spoke context noted in-page (Ollama→llm-inference, Gemma→llm-providers) rather than fragmenting; dominant substance = agent-building pattern, so routed here. Author/Tavily/OpenAI entities deferred (T4 tutorial; tangential). Ran avoid-ai-writing. +1 page (→101).

[2026-06-30] ingest | Knowledge RAG — local-first RAG MCP server for Claude Code

Routed from Telegram. New source knowledge-rag (SoftwareSourceCode, T1 [spoke repo convention], github.com/lyonzin; MIT, ~135★, v4.3.1). A local-first RAG server exposed to Claude Code as 13 MCP tools: hybrid retrieval (bge-small-en-v1.5 semantic + BM25 inverted index)

  • cross-encoder rerank (ms-marco-MiniLM), ChromaDB, ONNX Runtime (CPU/CUDA), 20 format parsers, stdio/SSE/ HTTP, zero external APIs. Its primary identity is an MCP server / agent tool (per its own README), not a novel retrieval technique → here, not ../research-wiki; the retrieval science (hybrid semantic+BM25+rerank) cross-links research's hybrid-retrieval-rag. Folded into synthesis's local-first/MCP-store axis (seekdb→ memory-vault→recall) as the document-retrieval sibling (retrieval-rich end vs recall's zero-dependency end); paired with headroom/from-local-llm-to-agent. Author (Lyon/lyonzin) entity deferred (one-off, tangential). Ran avoid-ai-writing. +1 page (→102).

[2026-06-30] ingest | autoharness — self-maintaining skill layer for Claude Code

Routed from Telegram (hub router). New source autoharness (SoftwareSourceCode, T3 — project README, self-reported; tigerless-labs). A Claude Code plugin that manages the whole self-authored skill library from real usage: distils sessions (user input / agent output / tool interactions) into skills, consolidates similar scenarios instead of accumulating near-duplicates, folds new evidence into existing skills, and archives by invocation rate; daemon-free, no held-out benchmark, append-only ledger kept out of skill bodies, Python 3.11+ zero-dep, MCP-wired. Dedup: no existing page — closest neighbours are hermes-learn-skill-capture (authors one skill) and self-improving-agents (the mechanism), so paged as the library-level instance. Folded into synthesis layer 5 as a fourth drift-control design point (real invocation/adherence signal + author-only scoping + pre-deploy validation, vs zouroboros's vote / autonovel's LLM-judge / autoresearch's metric). Its unverified 42%→78% README claim flagged against the Open-questions/Benchmarks gap (and against its own benchmark-free design). Owner org tigerless-labs paged as a thin evidence-only node. Ran avoid-ai-writing. +2 pages (→104).

[2026-07-01] ingest | ADK Go 2.0 (Google) — graph-based workflow engine

T3 vendor blog (developers.googleblog.com). Dedup: adk already paged → enriched in place (new "ADK Go 2.0" section) + new source adk-go-2 and a new concept page graph-based-agent-workflow. The 2.0 release (v2.0.0, 30 Jun 2026) rebuilds ADK around a graph of typed nodes + conditional edges run by a scheduler — "a graph is an agent" (same runtime for single

  • multi-agent). Nodes: function/emitting/agent/tool/join/dynamic/workflow/parallel/state-bound; routing via StringRoute/IntRoute/BoolRoute/MultiRoute/Default (an LLM can emit the route). Reliability in the runtime: per-node retry (5 attempts, 1s→60s, 2×, jitter), timeouts, WithMaxConcurrency, branch isolation. First-class durable HITL — any node pauses into session state, resumes by handoff/re-entry across process restarts (reconstructs from session history). Unified runtime + Chat/Task/SingleTurn modes; ToolContext+CallbackContext → one agent.Context. Resume format shared with Python ADK 2.0 (cross-runtime resume); plugs into A2A. Touched adk, agent-orchestration (added the declarative-graph counterpart to imperative fan-out), durable-agents (durable pause/resume promoted to base runtime), synthesis (layer 4 durability — "a graph is an agent" as the sharpest structure-over-capability instance). Breaking-change notes captured. Volatile snapshot.

[2026-07-01] ingest | WhatsApp ChatGPT Bot (Wassenger)

Telegram source (github.com/wassengerhq/whatsapp-chatgpt-bot), hub-routed here. MIT Node.js bot that turns a WhatsApp number into a multimodal GPT-4o assistant via the Wassenger WhatsApp API: webhook in → model → reply out, per-conversation memory, persona set by a config.js system prompt. Features: multimodal I/O (text/image/audio in, synthesized speech out), RAG via tool/function-calling (functions.js), MCP support, token/rate/size limits + number blacklist/whitelist, and human handoff. New pages: whatsapp-chatgpt-bot (source summary, SoftwareSourceCode, url-only) + wassenger (Organization, thin node, channel/gateway-vendor stratum). Folded into synthesis as the channel-gateway corner open question — the terminal-coding harness anatomy pointed at a consumer messaging channel — and as a data point that MCP is spreading past the Claude ecosystem (an OpenAI-based, non-coding bot shipping MCP). Cross-linked knowledge-rag, cult-ui-agent-patterns (RAG/tool-use), deer-flow + openclaw (message-gateway echo), agent-guardrails (human handoff). Tier T3 — first-party primary artifact but a vendor-promotional demo selling the Wassenger API; modest traction (167★). Entity discovery: Wassenger paged; MCP is an existing cross-wiki bridge node. No contradictions. Per 2026-07-01 policy: no per-route rebuild/verify.

[2026-07-01] ingest | Build agentic full-stack apps with Genkit (Google)

Telegram source (developers.googleblog.com/build-agentic-full-stack-apps-with-genkit), hub-routed here. Google's open-source agent-building framework for full-stack AI apps — packages message history, tool loops, streaming, and persistence behind one interface. New page genkit (SoftwareApplication, source, T1 first-party). Key surface: Agents API (server-side agents; same chat() local or remote), pluggable session stores (in-memory/file/Firestore/custom — a first-class agent-memory/durable-agents layer), HTTP-servable agents (snapshot/abort), a JS/TS remote agent client with an identical API (unified full-stack code), interruption for HITL + delegation middleware for orchestration (cf. agent-middleware), and an Agent Runner dev UI; TS/Go/Dart/ Python. The post explicitly contrasts it with adk (managed runtime for complex topologies) — Genkit is the embeddable library pole, so google now fields two agent frameworks at different altitudes. Folded into synthesis Vendors as further Google convergence evidence. Cross-linked google (cross-wiki org). Tier T1. No new contradictions. Per 2026-07-01 policy: no per-route rebuild/verify.

[2026-07-02] ingest | awesome-agent-evolution (EvoMap) — the field-map as an outside taxonomy

Routed from Telegram via the hub router. New source awesome-agent-evolution (Collection, T3, github.com/EvoMap; url-only) — EvoMap's 150+-entry curated "awesome list" of the agent-self-evolution field, with twelve sections (self-improvement, memory, agent-to-agent protocols, dev platforms, coding agents, prompt/behaviour optimization, safety/guardrails, embodied AI, research papers, benchmarks, taxonomy, community). Value is meta: an unaffiliated curator carves the field into nearly the same layers this spoke reached bottom-up (self-improving-agents, agent-memory, a2a-protocol, agentic-coding-harness, agent-guardrails, self-evolving-agents-survey) — mild external corroboration of the layer model. Folded a "an outside curator now draws the same map" note into synthesis layer 5, and added a lead-not-answer pointer to the Benchmarks open question (a directory of benchmarks ≠ a benchmark). Flagged the spoke's one blind spot the list exposes: Embodied AI. Did not page the 150+ entries — the catalog is the subject; named Mem0 (~60k★, an agent-memory product), Eliza, Agent Zero as page candidates only (a star count + one-liner isn't grounding). Added thin Org node evomap (curator/owner; evidence = repo + evomap.ai). Ran avoid-ai-writing. +2 pages (source + evomap).

[2026-07-04] ingest | Claude apps gateway for Google Cloud — the harness-vendor's own runtime-governance layer

Msg 724 (cloud.google.com/blog/…/announcing-claude-apps-gateway-for-google-cloud, via Telegram). T3 (first-party vendor announcement; technically substantive — architecture/gateway.yaml/deploy steps — and points to T1 docs at code.claude.com, but the ingested source is the blog). New page claude-apps-gateway (TechArticle, url). Self-hosted stateless proxy (Cloud Run + Cloud SQL + Secret Manager) between Claude Code clients and Vertex AI, co-authored Anthropic (Roy Arsan) + Google Cloud (Ivan Nardini). Centralizes five fleet controls: IdP identity (device-code → OIDC, short-lived tokens, no laptop keys), RBAC policy (gateway.yaml, server-side per /v1/messages), spend caps (Cloud SQL ledger, HTTP 429 over cap), telemetry/attribution (verified-JWT → OTLP → Cloud Monitoring/Grafana/Datadog), routing/failover (single service identity, inference stays in-VPC). Folded into synthesis as the third shape of the vendor runtime-governance layer — the first-party/single-harness counter-instance to third-party nemoclaw/microsoft-scout (Anthropic governing its own Claude Code, same control-plane-owns-which-model-you-call move). Two loops closed: operational answer to the SDK cost-structure tension (cap+attribute vs reprice), and the governance twin of agent-starter-pack (Cloud-Run deploy vs govern-fleet). Cross-spoke: platform-ops (observable Cloud Run service) + llm-providers/cloud (Vertex/GCP). Did not page the two blog co-authors (thin, single-connection). +1 page. Ran avoid-ai-writing. Per-route verify deferred (routine ingest, no structural page move).

[2026-07-04] ingest | Tokenmining — the named cost/token discipline (routing + compaction)

Msg 731 (towardsdatascience.com/tokenminning…, via Telegram). T3 (TDS practitioner how-to; anecdotal cost figures — $40/day, "$10k+/week" — not a controlled study). New page tokenmining (TechArticle, url). "Tokenmining" = the deliberate inverse of tokenmaxxing: minimize tokens while holding/raising quality. Two levers: routing (LLM gateway classifies prompt complexity → cheaper/local quantized model; ~60% cost cut) and context compaction (lossy summarization at the token limit, structured-memory extraction). Thesis: more tokens ≠ better — context rot + lost-in-the-middle mean extra context degrades output, so discipline is a quality argument not just cost. Crystallizes the spoke's scattered token-budget/cost thread (headroom pre-model compression, tail-control-agentic-workflows model-downshift, claw-code provider-agnostic routing, claude-apps-gateway spend-capped routing, memory-vault compaction-loss) into one named discipline; the cost/quality face of "structure substitutes for capability" (agentsys). Folded into synthesis. Cross-spoke: llm-inference (context-rot mechanism) + ../research-wiki long-context-vs-chunking (the retrieval-side twin of the same "long context isn't free/better" thesis — routed there earlier today). +1 page. Ran avoid-ai-writing. Per-route verify deferred (routine ingest, no structural page move).

[2026-07-09] ingest | Conductor CDD launch (Google blog) + repo re-seen → context-driven-development node

Two related sources arrived together (hub-routed): Google Developers Blog "Introducing context-driven development for Gemini CLI" (new) + the github.com/gemini-cli-extensions/conductor repo (re-seen — already the url: of conductor). Dedup: the repo refreshed conductor in place (added the spec + plan.md Phases→Tasks→Sub-tasks artifacts, file-persisted pause/resume state, brownfield support, install command; updated 2026-07-09), not a new page. The blog got the higher-value move: promoted context-driven-development to its own DefinedTerm node (T1, source=Google blog) — CDD as a named third point past vibe coding and spec-driven-development: whole project context (product/stack/workflow) kept as durable reviewed Markdown, not just a per-change spec. Cross-linked spec-driven-development (now points at the dedicated CDD node) and folded into synthesis §3 (the authoring-discipline "named far end"; Google naming it on the Gemini side = harness race converging on planning/context governance). Authors Keith Ballinger / Jay Kornder / Sherzat Aitbayev noted in prose; entity nodes deferred. Ran avoid-ai-writing. +1 page (Conductor refreshed, CDD added).

[2026-07-09] ingest | designpowers — 10-agent inclusive-design team

Routed from hub (Telegram). New source designpowers (SoftwareSourceCode, T3, github.com/Owl-Listener, MC Dean, MIT, ~220★) — a model-agnostic multi-agent design team: 10 role-specialist agents (design- strategist, design-scout, inspiration-scout, design-lead, motion-designer, content-writer, design-builder, accessibility-reviewer, design-critic, heuristic-evaluator) running a Discover→Research→…→Ship→Retrospective pipeline off 36 skills; cross-harness (Claude Code plugin + Gemini CLI extension + Cursor/Windsurf/ Copilot/Aider via shared markdown). Human-as-director: Direct mode (approve every handoff) vs Auto (runs unsupervised, pauses only on accessibility/critic/reviewer-conflict); Design-Taste + optional design- memory. Integrated as an instance, no new mechanism node: cross-linked agent-orchestration (role-split), agent-skills/agentskills-spec (cross-harness portable skills), gstack (role-agent factory, non-design), agent-guardrails (human-checkpoint/Auto-mode). Synthesis §1 vertical-generalization thread extended (role-split team pointed at design; accessibility-as-default vs final-audit). T3 (single self- described repo, modest adoption; efficacy claims unverified). Author (MC Dean) entity deferred. Ran avoid-ai-writing. +1 page.

[2026-07-09] ingest | pm-skills-arsenal (12 deep PM skills, typed schemas) (hub-routed, Telegram)

Ingested Avyayalaya/pm-skills-arsenal (Parth Sangani, MIT): 12 "codex-grade" product-management agent-skills at ~1,000–1,300 lines each, distributed via APM multi-harness manifest, plugin marketplaces (Claude Code / Agency-Copilot / awesome-copilot), an MCP server (5 tools incl. run_skill), and direct SKILL.md. Distinctive move: each SKILL.md declares typed capability_summary/input_schema/output_schema so an orchestrator routes between skills structurally without reading the body — composability-by-schema. New pm-skills-arsenal (SoftwareSourceCode, T2)

  • author parth-sangani (Person). Positioned as the few-deep-routed-by-schema counterpoint to the existing pm-skills (68-shallow-chained-by-commands); cross-linked both ways. Synthesis §1 vertical- generalization thread extended with the typed-function / run_skill-over-MCP angle. T2 — primary source for structure/schemas/distribution, but "codex-grade" + get_benchmark are unverified self-claims (Open-questions benchmark gap). Ran avoid-ai-writing. +2 pages.

[2026-07-09] ingest | Cloud Run sandboxes (Google Cloud, agent-code isolation) (hub-routed, Telegram)

Ingested Google Cloud's Cloud Run sandboxes (public preview, Next '26): a managed primitive to execute untrusted / AI-agent-generated code in an isolated sandbox that runs in-instance (2nd-gen exec env, shares the host container's CPU/mem, latency-optimized), enabled via gcloud beta run deploy --sandbox-launcher / sandboxLauncher: true; launcher binary at /usr/local/gcp/bin/sandbox, with an ephemeral built-in sandbox tool "coming soon". New cloud-run-sandboxes (SoftwareApplication, source, T2, volatile). Wired into the containment/governance-as-agent-agnostic-infrastructure thread: alongside nemoclaw/microsoft-scout (control-plane wrapping) and claude-apps-gateway/agent-starter-pack (Cloud-Run deploy→govern), this is the missing execution-isolation rung — containment as a substrate feature beneath the harness, and the server-side sibling of arrow-js's WASM sandbox. Synthesis §2 governance-layer thread extended (two altitudes: control plane vs substrate). T2 — first-party vendor docs, authoritative for mechanics but the isolation mechanism (gVisor/microVM?) is unstated and "fast/secure" is unevaluated. Cross-spoke: ../cloud-wiki (runner-up: Cloud Run as serverless offering / pricing angle), ../platform-ops-wiki (verification-runtime seam via agent-loops-verification). Ran avoid-ai-writing. +1 page.

[2026-07-12] ingest | Safe prompt pruning for long-context LLM systems (TDS, hub-routed, Telegram)

Ingested Emmimal P Alexander's TDS build-log on a deterministic prompt-pruning layer: a 3-pass pipeline (expired-tool-output elimination → duplicate-RAG-doc elimination → DEFINE/REF dependency restoration) that removes provably redundant context before the model, no LLM/embeddings — stdlib only, idempotent. Synthetic self-benchmarks: 27–34% token cuts on RAG/tool workloads, ~2–4% on plain chat, 100% required-fact retention, <50 ms at 131k tokens. Honest caveat: the safety pass initially never fired (benchmark never made pruning + dependencies collide). New pages: source summary safe-prompt-pruning (TechArticle, T3); author Person emmimal-p-alexander (thin). Synthesis: extended the token-budget corner from three levers to four — route / compress (headroom) / modality-arbitrage (pxpipe) / prune (this) — the lossless deterministic cousin of tokenmining's lossy context-compaction; caveat ties to the verification-bottleneck thread. +2 pages, 3 spine files updated. T3 (single-author synthetic benchmarks, unvalidated).

[2026-07-12] ingest | AI Race Coach built with Antigravity and Gemini (Google Devs Blog, hub-routed, Telegram)

Ingested a Google Developers Blog case study: GDEs built a real-time AI driving coach at Sonoma Raceway (23 May 2026) on Antigravity + ADK. Two synthesis payloads. (1) Antigravity gets a second face — not the coding IDE the spoke paged 2026-06-28 but a runtime "domain-bridging engine" doing stateful telemetry orchestration, with ADK orchestrating the coaching agents in the field (big-tech harness shipping into a live system, not a docs demo). (2) Edge/cloud model split — Gemma 4 local for zero-latency audio alerts, Gemini API cloud for post-session modelling: the spoke's route-by-latency-and-depth idea (tokenmining, tail-control-agentic-workflows downshift) where latency isn't a metaphor. Also the corpus's closest brush with the embodied/physical-agent blind spot EvoMap flags. New page: source summary ai-race-coach (BlogPosting, T3 — vendor case study, specifics reported not verified). Updated google-antigravity (+orchestration-engine facet), adk (+in-production section), synthesis (Google/Antigravity passage in layer 2), index. Ran avoid-ai-writing. +1 page. Entities: publisher google already canonical (bridge node); no new nodes (GDE team anonymous, Brian Luc a thin community mention).

[2026-07-14] ingest | Cognition trusts Claude Fable 5 to work through the night (Anthropic case study, hub-routed, Telegram)

Ingested Anthropic's customer case study on Cognition (Devin, autonomous AI software engineer) routing long-running work to Fable 5. Substance: Devin runs 8h+ unattended ("through the night") where earlier models drifted "after minutes to an hour"; the unlock is credited to the model's horizon (self-sufficiency), a rare once-a-year step-change per SVP Research Silas Alberti. Evaluation practice: "trust no eval" — senior engineers judge whether output would ship — + a proprietary "Frontier Code" anti-slop benchmark (Fable5 ~30% vs Opus ~10%, dogfooding agreeing). Verification discipline: agent states invariants → executes against them → root-causes in triage rather than guessing. New pages: source cognition-fable5-through-the-night (BlogPosting, T3 — first-party promotional, proprietary benchmark); devin (SoftwareApplication); cognition (Organization); silas-alberti (Person, thin). Synthesis: Benchmarks open question gets its sharpest partial step (names the anti-benchmark failure mode + a methodology, but proprietary/vendor + measures the model not the harness); structure-vs-capability question gets a clear counter-current (raw model horizon, not a cleverer harness, was the autonomy unlock — bounds the thesis; ties to effort-level). Cross-links to bridge nodes claude-fable-5/anthropic (../llm-providers-wiki was runner-up). avoid-ai-writing applied. +4 pages, 2 spine files updated.

[2026-07-14] ingest | Context rot in Claude Code — Governed Context (TDS) + CCA-F lost-in-the-middle (Medium), paired (hub-routed, Telegram)

Ingested two same-theme practitioner pieces on context management in Claude Code, arriving together. (1) governed-context-claude-code (Jake Minns/TDS, T3): the fullest context-governance playbook — intrinsic (attention/U-curve/lost-in-the-middle) vs content rot (Drew Breunig's confusion/clash/ distraction/poisoning); 3-phase governance (curate CLAUDE.md/skills/hooks → session hygiene/subagent-offload/ externalize-state → reset over recover); "session as a git tree" (--fork-session, /conclude, /merge). (2) cca-f-lost-in-the-middle (Rick Hightower/Medium CCA-F series, T3): the API-level version — pinned facts + rolling history + reconstruct-every-turn because the Messages API is stateless; context mgmt is a design problem, not a config knob. New concept node context-rot (promotes a phrase that lived inside tokenmining to a first-class mechanism; splits intrinsic/content). New thin Person nodes jake-minns, rick-hightower (spoke convention: cf. emmimal-p-alexander). Synthesis: extended the token-budget corner — context-rot is the why, reframing route/compress/modality-swap/prune from cost to quality (signal-to-noise); "model rarely signals the rot" ties to verification-is-the-bottleneck; governance operationalizes externalize-state/subagent-offload; the stateless-API point is the substrate reason the whole corner exists. Cross-spoke seams: ../llm-inference-wiki (attention mechanism), ../research-wiki long-context-vs-chunking (retrieval-side twin). avoid-ai-writing self-pass (clean). +5 pages, 2 spine files updated.

[2026-07-14] ingest | beautify-github-readme — README-redesign agent skill (hub-routed, Telegram)

Ingested beautify-github-readme (oil-oil, MIT GitHub repo): a agent skill (not a library) that redesigns a repo's README — two modes (Whole-README = restructure hierarchy, proof-before- technical, + full visual system; SVG-only = decorative assets, no copy edits) on a Markdown-for-searchable- text / SVG-for-visual-identity separation; Python, emits GitHub-safe SVG/PNG/WebP + Markdown, project-native (not template) visuals. New pages: source summary beautify-github-readme (SoftwareSourceCode, T2 — primary repo, self-described/unbenchmarked); thin author oil-oil (Person). Synthesis: extended the skills- generalize-to-verticals thread (layer 2) — the vertical list now reaches developer communication / repo presentation, a sibling of designpowers/renwei-writing/ai-job-search/pm-skills and a human-facing counterpoint to the openwiki/autowiki-factory docs-as-artifact thread. avoid-ai-writing self-pass (clean). +2 pages, 2 spine files updated.

[2026-07-14] ingest | Building AI Agents? Here Are Some Anti-Patterns to Avoid (Machine Learning Mastery, hub-routed, Telegram)

Ingested Bala Priya C's nine-item agent-building anti-patterns checklist (multi-agent-too-soon, one-agent- does-everything/15+ tools, tool-list sprawl, hardcoded monolithic logic, skipping memory architecture, shipping without observability, ungoverned write access, context drift, deploying without evaluation). Its value is corroboration density: the nine map ~1:1 onto the spoke's synthesis threads — start-simple (building-effective-agents), context bloat/drift (context-rot), layered agent-memory, modular composition (agent-middleware), ungoverned writes (agent-guardrails/agents-never-do-alone), eval (loop-engineering), observability (seam to platform-ops). New pages: source agent-anti-patterns (BlogPosting, T3 — educational roundup, no original data); thin author bala-priya-c. No new concept node (consolidating source, not a new idea). Synthesis: added at the guardrails anchor as a second outside-in corroboration of the layer model after EvoMap (EvoMap maps the field, this enumerates failure modes); flagged its internal tension (recommends layered memory+observability+eval as near-defaults vs its own "don't over-engineer" lead item — the start-simple↔build-scaffolding balance). avoid-ai-writing self-pass (clean). +2 pages, 2 spine files updated.

[2026-07-14] ingest | How to Orchestrate 100+ Agents With Claude Code (TDS, hub-routed, Telegram)

Ingested Eivind Kjosbakken's TDS practitioner piece on running 100+ Claude Code agents in parallel via headless mode (claude -p): hierarchical (one orchestrator spawns children), stateless one-shot sessions rather than nested subagent loops, no centralized state (high-level prompting + task isolation to prevent interference), self-verification baked into the task prompt, full per-agent autonomy (MCP + all tools, no per-step approval), scoped to well-defined/reversible work (refactoring; Fable-detects → orchestrator-tasks → Opus-headless-executes → aggregate). New pages: source orchestrate-100-agents-claude-code (BlogPosting, T3 — no throughput/cost/concurrency data); thin author eivind-kjosbakken. Updated agent-orchestration: added as the largest-scale instance + a distinct headless/process-level/isolation flavor (vs nested loops orchestration-mode/agentsys or durable queue gbrain). Synthesis (layer 3): folded in — isolation is where orchestration meets context-rot (fresh-session-per-task = rot-avoidance-by-isolation, the fork/isolate move at fleet scale) and agent-guardrails (safety via task choice, not human checkpoints); honest limit = well-scoped work only (match-pattern-to-task). No new concept node (headless mode folded into orchestration). avoid-ai-writing self-pass (clean). +2 pages, 2 spine files updated.

[2026-07-15] ingest | grok-build — xAI's terminal coding agent (GitHub, hub-routed, Telegram)

Ingested xAI's grok-build (xai-org/grok-build), the open-source Rust terminal coding agent grok: full-screen mouse-interactive TUI that understands the codebase, edits files, runs shell, web-searches, and manages long-running tasks; three modes (interactive TUI / headless CI-scripting / editor-embed via Agent Client Protocol); MCP + plugin extensible; cross-platform; synced from xAI's internal monorepo. Modular Rust crates (xai-grok-pager TUI · xai-grok-shell runtime · xai-grok-tools file/exec/search · xai-grok-workspace fs/VCS/exec). New page: source grok-build (SoftwareSourceCode, T1 — primary vendor repo/README). Reused existing cross-wiki Org node xai-grok (../llm-providers-wiki) as creator — no duplicate xAI node created; added it to the index bridge-nodes list. Placed as the agentic-coding-harness pattern shipped first-party by a model lab (peer to Claude Code / Gemini CLI), on the Rust/systems-language end beside claw-code/oh-my-pi/easy-agent. Synthesis: added a "first-party lab harnesses" thread — labs now compete on the harness, not only the model, yet converge on shared protocols (ACP agent↔editor, MCP agent↔tool); the structure-over-capability split drawn by one company on both sides (model in llm-providers, harness here). Entity discovery: creator xai-grok already canonical (reused); no new entities paged. avoid-ai-writing self-pass (clean). +1 page, 3 spine files updated.

[2026-07-22] ingest | Auto-Company + cc-thinking-skills (GitHub ×2, hub-routed, Telegram)

Two repos arrived together and turned out to argue with each other, so they were ingested as a pair.

auto-company (MaxMiksa, MIT, ~2.0k★, T2): a launchd/systemd daemon that re-invokes Claude Code or Codex CLI every 30s against a company charter, forming 2–5-person squads from 14 expert-persona agent definitions. Read as Ralph with the stop condition removed — with no completion promise and no external feedback signal, the author hardcodes phase transitions in PROMPT.md (ideate → GO/NO-GO → build; "pure discussion is forbidden") to stop the loop talking to itself. Three existing nodes updated with it: agent-memory gained a minimal pole (consensus.md, one rewritten markdown file as the entire cross-cycle state, explicitly chosen over vector DBs — steerable by editing one line, but historyless); agent-guardrails gained a counterexample (limits stated as prose in CLAUDE.md while the sandbox is bypassPermissions/danger-full-access on the host — enforcement removed, instruction retained); agent-orchestration gained a fan-out variant where the state file picks the roster, not the orchestrator.

cc-thinking-skills (TJ Boudreaux, MIT, ~822★, T2): 39 mental-model skills (Munger/Meadows/Kahneman/Goldratt/Boyd) behind a thinking-model-router meta-skill — and the Elevate-or-Kill eval harness shipped alongside them, length-controlled against a same-length placebo and replication-gated. Result published against interest: zero skills hold a robust replicated ELEVATE verdict; best candidate thinking-scientific-method at +5.3pp / p=0.061 / n=150 → DIRECTIONAL-NOT-REPLICATED. First source in the spoke that measured whether a skill pack works.

New concept node reasoning-scaffolds holds what the pair share: capability skills (grant something the agent lacked) vs cognitive skills (prescribe how it thinks). Auto-Company encodes the second as personas and asserts it; cc-thinking-skills encodes it as skills and finds no measurable accuracy gain — same bet, two indexes, one of them now tested.

Synthesis: layer 1 split into capability-vs-cognitive with the null result recorded; the Benchmarks open question gained its first measurement of the tooling layer (and the length-matched-placebo design worth copying); "does structure substitute for capability" bounded from the thin end — structure appears to pay where it changes what the agent does, not what it's told to think. Two tensions added: constitution vs containment (prose limits vs an uncrossable boundary, no incident data either way) and memory's unpriced range (one markdown file vs seekdb's forkable store, three orders of magnitude apart, never compared). Entity discovery: 2 new Person nodes (zheyuan-kong, tj-boudreaux); CMU affiliation evidenced but left as a plain mention (relevance gate). avoid-ai-writing self-pass. +5 pages, 4 pages updated, 3 spine files.

[2026-07-22] ingest | Harness AI Agent DLC (The New Stack)

Routed from the hub (route in ../log.md). URL-only, T3 — trade press, but every claim comes from one vendor launch and one vendor executive, with the supporting survey also first-party (Harness's own State of Engineering Excellence 2026). Fetch note: WebFetch returned only the newsletter shell; recovered with the firecrawl skill per ../HUB.md edge handling, so the summary is from the full article body. New pages: harness-agent-dlc (TechArticle source), agent-evals (DefinedTerm/practice), harness-io (Organization), trevor-stuart (Person, thin node). Dedup: no existing Harness/evals page. Slug care — harness-io is deliberately not harness: this wiki's central concept is the agentic-coding-harness and a bare harness slug would collide. Homonym warning recorded on both the page and the index line. agent-evals is a consolidation, not a new thread — evals were already scattered across agent-anti-patterns (the "no eval" failure mode), cc-thinking-skills (an eval harness that returned a null result on its own skill pack), cognition-fable5-through-the-night ("trust no eval") and autonovel (layered mechanical + LLM-judge evaluator). The page gives them one home and states the narrow defensible claim: an eval gate catches a regression against your own dataset, it does not certify quality. Gap-relevance: extends the loop/verification thread from building verification into the loop to enforcing it at the delivery boundary. Synthesis gained a paragraph on "make the pipeline predictable, not the agent," on determinism relocating from output → record (why eval gates ship with AgentTrace), and on this being the operator's end of agent-guardrails (gate the release by score, not just the action by blast radius). Contradiction watch: none — it pairs with agent-loops-verification (runtime fidelity vs pipeline enforcement, same bottleneck from two sides). Both are vendor arguments with no independent measurement; that is now the loudest gap in the thread, and Harness explicitly declines to claim a number. Cross-spoke: ../platform-ops-wiki is the logged runner-up (every mechanism — canary, OPA, tracing, the delivery-measurement gap beside dora-metrics — is its vocabulary); the pre-registered "agents vs live delivery infra" seam is now at 2 sources. Also noted an ../ai-governance-wiki adjacency (ownership, asset inventory, auditable run records). Entities: harness-io + trevor-stuart created; the TNS journalist (Adrian Bridgwater) not paged, following the spoke's precedent of paging the subject rather than the reporter (arjun-iyer on agent-loops-verification). Depth 1, well under the recursion cap. Verify deferred per hub policy (content-only). avoid-ai-writing run. 156 → 160 pages.

[2026-07-22] ingest | Corbell (Corbell-AI/Corbell)

Routed from the hub (route in ../log.md). URL-only, T1 — primary artifact (the repo/README itself), Apache-2.0, v1.2.0 2026-05-23, 435★/46 forks. T1 for provenance, but note every capability claim is self-described and nothing is benchmarked. New page: corbell (SoftwareSourceCode). Python CLI that builds a cross-repo code knowledge graph (tree-sitter over 8 languages → service deps, typed signatures, call graphs, execution flows) + embeddings (all-MiniLM-L6-v2), then generates architecture specs from a PRD/feature sentence by auto-discovering the services involved and feeding graph context + snippets to Claude/GPT-4o; validates against recorded architecture constraints and exports tasks to Linear/Jira. Local-only (SQLite, optional Neo4j), MCP server for Cursor/Claude Desktop, D3 graph browser, --ci flag, git change-coupling and IaC detection. Dedup: no existing page; checked the near neighbours (spec-kit, conductor, autowiki-factory, openwiki, knowledge-rag) — all distinct, all cross-linked instead. Updated spec-driven-development with a new section: mined context vs written context. That is this source's contribution — the other SDD tools have a human author the constitution/clarifications/managed context, Corbell derives the input from the code that exists. Synthesis: extended the machine-maintained-code-wiki paragraph. The extraction step now points two ways — describe what is there (autowiki-factory/openwiki) vs constrain what comes next (Corbell) — and the local-first-store-over-MCP corner gains a fifth capability (codebase structure, beside memory, documents, perception, search). Recorded the caveats: no evaluation that a graph-grounded spec beats a written one, and no refresh trigger backing the "living" graph claim (contrast AutoWiki's regenerate-on-push). Entities: none created. The only candidate is the "Corbell-AI" GitHub org, and the source carries no evidence about it beyond the name — no people, no company detail. Per ../ENTITIES.md (evidence-only), a name-only node would be noise. Verify deferred per hub policy (content-only). avoid-ai-writing run. 160 → 161 pages.

[2026-07-23] ingest | Building verification loops in Claude Code with skills (Anthropic)

Routed from the hub (route in ../log.md). URL-only, T1 — first-party Anthropic post (Delba de Oliveira, 2026-07-22). New page: verification-loops-skills (BlogPosting source). Encode repetitive post-implementation checks as Claude Code skills (.claude/skills/…/SKILL.md, allowed-tools frontmatter + plain-English steps) so Claude self-verifies and fixes; a verification loop = "an iterative process where Claude checks and attempts to fix the work." Four activation patterns (standalone / embedded / chained / on-every-PR). Dedup: no existing page; it is the concrete instance under two abstractions already here — loop-engineering's verification corollary (updated: this is how you author the cheap faithful check) and agent-evals (added a "two altitudes" section: inner-loop skill vs outer-loop delivery gate, meeting at on-every-PR). Gap-relevance: closes the concreteness gap in the verification thread — the spoke had the why (agent-loops-verification) and the delivery-gate form (harness-agent-dlc, ingested yesterday) but no first-party how inside Claude Code. Synthesis: extended the Harness paragraph to state the full altitude range (skill in the inner loop ↔ eval gate at the boundary), and folded verification back into the capability-as-skill thesis (it's markdown a skill runtime executes). Caveat recorded: vendor how-to for the vendor's product, no measurement, and a self-referential risk — a plain-English check applied inconsistently is the nondeterminism evals exist to catch, so a verification skill arguably wants its own eval. Entities: none created. Author Delba de Oliveira is an Anthropic devrel byline with no distinct load-bearing thesis of her own here; per ../ENTITIES.md (evidence-only, thin nodes = noise) noted inline instead. anthropic is the cross-wiki publisher node (../research-wiki) — not duplicated. Verify deferred per hub policy (content-only). avoid-ai-writing run. 161 → 162 pages.

[2026-07-23] ingest | Anthropic Claude Security plugin for Claude Code (MarkTechPost)

Routed from the hub (route in ../log.md). URL-only, T2 — MarkTechPost, detailed technical secondary coverage of a first-party Anthropic beta release. New page: claude-security-plugin (SoftwareApplication, source). A beta Claude Code plugin (/claude-security, official marketplace) running a multi-agent vulnerability scan: six-phase JS dynamic workflow (Inventory→Threat-model→Research→Sweep→Panel→Adversarial), four vuln classes (memory-safe langs skip the memory lens), model-tiered roles (Opus-class orchestrator, Sonnet-class cartographer/explorer, session-model researchers/verifiers). Findings survive a 2-of-3 verifier quorum (reachability/impact/ defenses; unanimous→high, 2/3→medium) whose tally is computed in Python by the renderer, not asserted by the model. Patches built in an isolated scratch clone + independently verified; manual git apply. Filed here on the substance-not-domain rule (like cybersecurity-skills): security is the domain, the multi-agent orchestration + programmatic verification is why it's in this spoke. ../osint-wiki noted as the security-subject bridge (cross-link, not duplicate). Dedup: no security/scanner page existed. Wired into agent-orchestration (added as Anthropic's own shipped instance) and agent-evals (new "take the tally out of the model's hands" section — the quorum + Python tally is the escape from model self-grading). Also cross-linked verification-loops-skills, harness-agent-dlc, agent-guardrails, claude-code-model-effort-levels (model-tiering = the model-choice lever inside one tool). Gap-relevance: the verification thread's sharpest artifact — synthesis extended to state the trust-anchor (quorum-with-mechanical-tally for stakes where a model's self-report is unacceptable), completing the altitude range skill→eval-gate→quorum. Recorded the honest scope: nondeterministic, doesn't replace SAST/dep-scan/ review, and adds no isolation of its own (inherits session perms; sandbox-runtime recommended). Model-version note: the source phrases the tiers as "Claude 3 Opus/Sonnet"; paged as Opus-class/Sonnet-class since the durable point is role-tiering, not the version string (likely article phrasing). Entities: none new (Anthropic is the cross-wiki node; no author paged). Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-23] ingest | LLMs Will Cheese Your Types: Fighting Back in Haskell (Justin Le)

Routed from the hub (route in ../log.md). URL-only, T2 — expert practitioner analysis (Justin Le, blog.jle.im), original and specific but single-author opinion, no measurement. New pages: llms-cheese-your-types (BlogPosting source), constraint-evading-behavior (DefinedTerm, failure-mode), justin-le (Person, thin node). The durable move is the concept page: constraint-evading behavior = an agent satisfying a check's letter while defeating its purpose (specification-gaming / reward-hacking at the developer-guardrail level). Haskell types are the instance; the idea generalizes to tests/lints/evals. Dedup: no reward-hacking/spec-gaming/constraint page existed — this fills a real hole. Wired into agent-guardrails (new design-time-vs-runtime section: a guardrail you can cheese is instruction not enforcement), agent-evals (new "a gate you can cheese isn't a gate" section — bounds the whole eval thread), and agent-anti-patterns (Related). Gap-relevance: this is the antagonist the verification thread lacked. The day's arc (agent-evals → verification-loops-skills → harness-agent-dlc → claude-security-plugin) all argued gate on a check; Le shows the agent cheeses the check, so a gate's worth = its un-gameability. Synthesis: extended the verification paragraph with the counterforce AND Le's constructive fix (checks as scaffolds not gates — invalid states unrepresentable; the Lean analogy = cross-spoke bridge to ../research-wiki's formal methods). Full arc now: author a check → enforce it → make it un-cheeseable by construction. Entities: justin-le created (thin, one source); the lean-theorem-prover link is intentional cross-wiki to ../research-wiki, not a local dup. Caveat: Haskell-specific, single opinion, no data — pattern generalizes, mitigations are language-bound. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-25] ingest | Compound Engineering plugin — re-seen, refreshed in place

Telegram drop of https://github.com/EveryInc/compound-engineering-plugin. Dedup: already ingested 2026-05-29 from the held README (raw/compound-engineering-plugin.md), so per the edge rule this refreshed compound-engineering-plugin in place — no new source page. Added url: beside the existing sources:; bumped updated. T1 (first-party repo); star count is popularity, not evidence. What changed since May: 32 skills (was a handful), 13 install targets each with its own in-repo plugin manifest, /ce-work cross-model dispatch (execution can leave, verification and shipping don't), /ce-code-review explicitly report-only, solutions written to docs/solutions/, and a standing design choice — no standalone agents or hooks, specialist behaviour lives inside skills. Gap-relevance: sharpens thread 1 (capability as portable markdown) by pairing it against superpowers. Same shape — a methodology as a skillpack — differing on who invokes (agent self-invocation vs explicit /ce-* commands, /lfg as the autonomous escape hatch) and on /ce-compound, the one skill in the corpus whose output is the next run's context — self-improving-agents by filesystem artifact rather than weight update. Also a data point against the agentskills-spec portability thesis: 13 manifests is the pre-standard tax still being paid by the biggest packs. No contradictions. Synthesis thread 1 extended; index row updated. Page count unchanged (36 → 36 unaffected). Verify deferred per hub policy (content-only, no page moves). avoid-ai-writing run.

[2026-07-26] ingest | OpenManus — open general-purpose agent framework (github.com/FoundationAgents/OpenManus)

Routed here by the hub; no runner-up worth naming. T1 (project repo), MIT, 57.6k★ / 10k forks. New: openmanus (source), foundation-agents (entity). Updated: synthesis — a new Vendors paragraph on the community-org stratum, and a new entry under Contradictions/tensions for the RL sibling. Two findings, one of them uncomfortable for this wiki:

  1. The moat cut both ways. Manus gated a general agent behind an invite code; MetaGPT-community authors had an open prototype in three hours and took 57.6k stars off 526 commits. If value has moved into the wrapper, the wrapper is also the cheapest part to re-implement. The 301 open issues and 192 open PRs are the counterweight — replicating took hours, carrying it doesn't.
  2. OpenManus-RL (with UIUC) tunes agent LLMs via GRPO, locating competence in the weights. The spoke has ~166 pages on the structural bet and zero sources on the RL one. Filed as a tension, not a contradiction — they're compatible in principle, competing as bets. Dedup: nothing on Manus, MetaGPT or OpenManus existed (grep over wiki/ + spine). Nearest neighbours openclaw and hermes-agent are open replacements for closed harnesses; this replaced a closed waitlist, which is why it's paged rather than folded into either. Weakness recorded: the README asserts multi-agent support and tools but never documents the agent loop — the layer this spoke exists to study is exactly the undocumented one. Star count is a popularity signal, not an engineering one, and the page says so. Entities: foundation-agents paged as the meaningful node. Core authors Xinbin Liang and Jinyu Xiang are named on the org page but not paged — the evidence is names on a repo and nothing attributable beyond it (evidence-only rule); page them if either shows up in a second source. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-27] ingest | nono — least-privilege agent sandbox (Nolabs AI, GitHub)

Routed here by the hub (runner-up: ../defensive-security-wiki — the least-privilege / credential-proxy substance is blue-team, but the subject is running agents, which is this spoke; the security angle is cross-linked, mirroring today's inverse call where strix went to ../osint-wiki by subject despite being agent machinery). T3 project README, Apache-2.0, ~3.2k★, pre-1.0 with APIs still moving and 110 open issues; the "in production at large companies" claim is unattributed and there is no audit or threat model. New: nono (source), nolabs-ai (entity — Sigstore team, notable provenance). Updated: agent-guardrails (new section: moving the boundary to the tools), synthesis (a third containment altitude), index. What it adds:

  1. A different unit of containment. The corpus contains the agent — control planes wrapping a harness (microsoft-scout, NemoClaw, the Claude gateway) or isolated execution beneath it (cloud-run-sandboxes). nono's claim is the unit is wrong: the dangerous privileges belong to the tools the agent delegates to, so git/gh/curl/kubectl each get a child sandbox with separate credentials, filesystem grants, network rules and policy.
  2. It closes the auto-company hole structurally. That page's failure mode is prose limits plus danger-full-access, with permission prompts turned off when they get inconvenient. Here an agent cannot widen a tool's permissions from inside its session because that policy was never its to hold, and there is no prompt to disable. The enforcement-vs-instruction distinction agent-guardrails draws now has a product on the enforcement side.
  3. Restriction as a shareable artifact. Composable JSON policies with profile inheritance and a public registry — the packaging move agent-skills made for capability, applied to its opposite. Also recorded: credentials reach tools through a proxy with L7 endpoint filtering rather than the agent's environment, which is the same principle ../defensive-security-wiki's Beyond Zero (routed there hours earlier) reaches at enterprise scale — authorize the action, don't trust the session holder. Two independent arrivals at one idea in one week, at opposite scales. Provenance worth keeping: the Sigstore team building agent containment says something about which problems are now load-bearing. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-27] ingest | Pilot Protocol — "the Network OS for agents" (pilotprotocol.network)

Routed here by the hub (runner-up: none clean — the payments facet (USDC/x402 agent-to-agent settlement) has no owning spoke; ../search-marketing-wiki owns agentic-commerce but that's the consumer-facing side, agents buying on a human's behalf. Recorded as cross-spoke context on the source page rather than split out). T3 vendor landing page. Technical detail is specific enough to check (address format, congestion control, named crypto primitives); every traction number is an unaudited self-reported counter. New: pilot-protocol (source). Updated: a2a-protocol (the other branch — replace the substrate), synthesis (a sixth contender that isn't at the same layer), index. What it adds:

  1. A different branch of the interop question. The spoke's five standards (a2a-protocol, MCP, ARD, agents-md, agentskills-spec) are all formats layered on the human internet — HTTP, DNS, TLS kept, agent semantics added above. Pilot replaces the substrate: own addresses (N:NNNN.HHHH.LLLL) resolved by its own registry, OSI-L5 UDP with sliding window/AIMD/SACK, NAT hole-punching for direct tunnels, X25519/AES-256-GCM/Ed25519. So the prior question is not "which standard wins" but does the agent web run on the human web at all.
  2. Payments as a network primitive. None of the five says how one agent compensates another, which is what decides whether an open agent network sustains itself. Pilot puts USDC/x402 settlement in the stack alongside addressing and discovery.
  3. The governance pattern inverts. A2A, AGENTS.md and ARD all ran vendor → neutral foundation → default. Pilot is a venture-funded company operating the network it specifies ($4.5M seed) with a curated app store it controls. Recorded plainly that an IETF Internet-Draft is not a standard — anyone may submit one, they expire in six months, and it confers no standing until a working group adopts it. Skepticism recorded: ~219k "agents online" and ~33k req/s are marketing counters with no definition of what makes an agent "online"; the architectural argument is the part worth keeping, separately from the traction claim. Follow-up if it recurs: read the actual Internet-Draft and the GitHub repo rather than the site. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | iai-pme — personal memory engine for AI assistants (CodeAbra, GitHub)

Routed here by the hub (runner-up: ../research-wiki — the gbrain/llm-wiki PKM lineage. Declined on the project's own framing: "personal memory" here means the assistant's memory of you, not a knowledge base a human reads, so it's agent-memory infrastructure). T3 project README, MIT, ~423★, solo-maintained, declared stable (1.x API commitment). New: iai-pme (source). Updated: agent-memory (a second axis: store vs process), synthesis, index. What it adds:

  1. A second axis for the memory layer. The existing poles — seekdb (DB), memory-vault (pgvector over MCP), recall (markdown over hooks) — differ in weight but are all stores you write to and search. iai-pme models memory as processes, structured on human memory research: episodic (verbatim, write-once) / semantic (induced from episodes during idle consolidation) / procedural (learned user parameters), held apart so detail, gist and habit don't collapse.
  2. Forgetting as a feature. Nightly consolidation reorganizes the store unprompted; unreinforced links decay on a schedule with pinning to resist. No other instance in this layer schedules forgetting. Recall also returns contradictions alongside matches — the staleness problem temporal-knowledge-graph answers with validity intervals, answered here by handing the conflict to the model.
  3. It publishes numbers on a public benchmark. LongMemEval-S R@5 0.962 / R@10 0.978, Rescue@10 1.000, retrieval ~350 tokens vs ~2,850 for agent search, latency 77ms@1K / 368ms@10K. Author-run and the token-efficiency baseline is unspecified, but a named public benchmark is a rung above the assertion-only claims the rest of this layer offers — a partial answer to the standing "nobody measures the memory layer" gap. Internals worth recording: MOSAIC, an in-house Leiden-family community detection with a CPM objective — the GraphRAG clustering move applied to one user's episode history; Hippo (SQLite + AES-256-GCM holding both vector index and graph); local bge-small-en-v1.5 via a Rust embedder on Candle; ambient capture through per-turn hooks (~5ms, no embedding). The episodic/semantic/procedural split is Tulving's taxonomy from cognitive psychology arriving as a software architecture — the cleanest seam yet to the new ../psychology-wiki. Limits recorded from the README (unusually candid): English-only, no cross-machine sync, poor recall for the first ~10 sessions, seconds-long cold start, Windows beta, no enterprise SLA. Entity (Areg Aramovich Noya, solo maintainer) deferred — thin, one project. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | Kimi Claw — managed OpenClaw with Claw Groups preview (Moonshot, kimi.com/bot)

Routed here by the hub (runner-up: ../llm-providers-wiki — Moonshot is a provider and the K2.6-vs-K3 detail is a provider fact, cross-linked onto kimi-k3 there; but the subject is an agent product built on openclaw, which is this spoke's). T3 product landing page. Third share.google link this cycle — resolved from the share page's HTML to kimi.com/bot. New: kimi-claw (source). Updated: openclaw (a third vendor shape), synthesis (fourth shape in the vendor-layer thread), index; cross-spoke, kimi-k3 in ../llm-providers-wiki. What it adds:

  1. A vendor shape that isn't governance. openclaw's thesis was that the project supplied gateway-first breadth and platform vendors filled the governance gap — NVIDIA with an external sandbox, Microsoft with containers and Entra identity. Moonshot is a model lab and adds hosting and a model: one-click deploy, K2.6 Thinking preconfigured, skills ready. No policy runtime, no identity, no sandbox mentioned. So the layer above OpenClaw sorts by what the vendor already sells — and a lab sells weights plus an ops team.
  2. A lab integrating downward, from the model into the harness that consumes it. The exact mirror of ../llm-providers-wiki's sakana-fugu (an agent system sold as a model), ingested four days apart. The model/product boundary is being collapsed from both directions at once.
  3. It ships K2.6, not K3 — one day after kimi-k3 launched. No reason given; the plausible ones (24/7 cost profile, productization lag) are unsourced and recorded as such. The general point is worth keeping: a lab's announced flagship and the model its own products run are separate facts. Recorded as a gap rather than a finding: the hosted product ships preinstalled skills while ClawHub's 341 malicious skills remain OpenClaw's known weak spot, and the page addresses neither that nor how a hosted instance is isolated. Also unstated: pricing, availability, what Claw Groups actually does, and which skills are included. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | i-have-adhd — ADHD-friendly output skill for coding agents (github.com/ayghri/i-have-adhd)

Routed here by the hub. T1 (project's own repo). Metadata verified via the GitHub API rather than the page summary: 12,333★ / 632 forks, MIT, created 2026-05-13, last push 2026-07-28. New: i-have-adhd (source). Updated: reasoning-scaffolds (new third category), ponytail (the sibling contrast), agent-evals (the gate at skill scale), synthesis (thesis layer 1 + the Benchmarks open question), index. The skill: ten rules shaping what an agent says — action first, numbered steps, state restated every turn, tangents deferred, time in real units, lists capped at five, no preamble/recap/closers — plus a pre-send check (delete the announcing opener, the "anything else?" closer, the sidebar, the empty hedge; then verify the first and last line carry the next action and the current state) and six documented override conditions, including two precedence rules: when a rule would delete the answer the task wins, and inside a harness the system prompt outranks the skill. Two findings. (1) A third kind of skill. reasoning-scaffolds split capability from cognitive skills; this one is neither, and neither are ponytail or renwei-writing. They constrain the artifact, not the thinking — recorded as output-shape skills. The distinction is about how expensive the claim is to check: whether an agent really ran a pre-mortem is unknowable from its output, whether it opened with a command instead of "Great question" takes one line to verify. That's the same "checkable step beats a posture" pattern cc-thinking-skills' null result already pointed at. (2) The first skill to ship its own eval harness and release gate. evals/ carries cases, a blind weighted rubric (correctness 35 / autonomy 25 / actionability 20 / safety 10 / concision 10) and a gate: no blockers, correctness and safety within 0.1 of baseline, weighted score above it. Note what the weighting does — a brevity skill's failure modes are dropping substance and shoving work back to the user, so those two dimensions carry 60% and the thing being sold carries 10%. It's ponytail's prose guardrail ("restraint never trades away validation/security/accessibility") turned into a threshold. The runner guards contamination the corpus learned about the hard way: strip operator config so user plugins/hooks don't leak into conditions (worst case named in the repo — its own always-on flag injecting the skill into the baseline), pin the model, cap spend in dollars, resumable runs, and a fairness clause requiring competitor claims to use the same cases/models/trials/rubric. Recorded weakness: no results published. Harness without a run. The corpus now holds three postures — a metric without a method (ponytail), a method with a null result (cc-thinking-skills), and a method with no run (this) — which advances the Benchmarks question without closing it. 12.3k★ measures appeal, not effect. Cross-spoke: the ADHD framing is borrowed from Ramsay & Rostain's The Adult ADHD Tool Kit and adapted to LLM output. Noted as adjacency to ../psychology-wiki, not routed there — same rule that spoke already applies to borrowed System 1/2 vocabulary: the borrowed metaphor stays with its application. Whether the formatting helps actual ADHD readers is a clinical question the repo doesn't test. Author ayghri inline, not paged (pseudonymous handle, low graph signal — ponytail precedent). Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | JetBrains — Ponytail skill for Claude Code: does it really cut tokens? (blog.jetbrains.com)

Routed here by the hub. T2 — the first genuinely independent measurement of a skill in this spoke: a third-party lab testing a tool it doesn't own, publishing method and numbers. New: jetbrains-ponytail-tested (source). Updated: ponytail (new measured section; the self-reported caveat now says partly-borne-out-and-overstated), agent-evals (new section on third-party evals), synthesis (the Benchmarks open question, substantially rewritten), index. Reused jetbrains cross-wiki from ../dev-tooling-wiki via entity-index — not duplicated.

Method: 80 paired SkillsBench tasks, auto-graded 0–1 with partial credit; claude-sonnet-5 at medium effort; Harbor 0.18 with Docker sandboxes; Arm A stock Claude Code 2.1.201, Arm B the Ponytail v4.8.4 ruleset injected via SessionStart hook. 251 billed trials, USD 246.09, escalated through a 10-task smoke run and a k=3 stage. Every trial audited to confirm the ruleset reached the treatment arm (100%) and stayed out of the baseline (0%).

Result: −15.4% code (p=0.088), −10.3% cost (p=0.004), −11% wall-clock, and no quality difference detected (9 worse / 6 better / 65 identical). Advertised was −54% code, −22% tokens, −20% cost, −27% time — so roughly a third of the claim. JetBrains attributes the gap to method rather than dishonesty: the repo's number is a mean over twelve hand-picked feature tickets, theirs a median over 80 tasks nobody chose for the purpose. That's the general lesson about self-reported metrics — usually measured on the sample that flatters them, not invented.

Three findings worth more than the headline. (1) The benefit is conditional. 31% code reduction on larger builds, near zero where the baseline was already lean. Ponytail doesn't make an agent write less, it stops an agent writing too much, so the payout tracks how often the baseline over-engineers. Any single percentage is an artifact of the task mix — which should now be the default reading of every restraint/output-shape skill's headline number. (2) Adherence is not a given. Without the SessionStart hook the skill self-activated zero times across ten sessions, despite a description claiming readiness for "ANY coding task" — so the measured benefit requires the always-on install, and the advertised one does nothing. Separately, its own rule about leaving ponytail: comments on deliberate shortcuts was followed once in 80 trials. A loaded skill is a bid for the model's attention, not a constraint. (3) Small samples flip signs. The 10-task smoke stage reported −3% code and a +9.6% cost increase before the full run landed at −10.3%.

Limits recorded on the page: SkillsBench scores task completion only — not security, validation, error handling or accessibility, which is exactly the ground ponytail's guardrail claims to protect, so that claim stays unmeasured. 80 pairs can't prove quality equivalence; the null remains compatible with small degradation. Only the cost result clears conventional significance (code sits at p=0.088, and the cost median's bootstrap interval "just touches zero"). Gap-relevance: directly answers the spoke's oldest open question. Ingested the same day as i-have-adhd, which supplies the author-side complement (a harness with no run); the two together put the Benchmarks question in a new state — open with evidence in it rather than open for lack of any. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | Gemini API Managed Agents — 3.6 Flash, hooks, and more (blog.google)

Routed here by the hub (runner-up: ../llm-providers-wiki for the Gemini 3.6 Flash model facts — cross-linked, not split). T3: first-party product announcement, promotional, one customer testimonial as its only outside evidence. New: gemini-managed-agents-hooks (source). Updated: agent-guardrails (new spend-as-a-bound section), agent-middleware (middleware sold as a service), synthesis (vendor layer gains a fifth shape), index. Reused google cross-wiki. No new entities. Verified against the page, not the summary — antigravity-preview-05-2026 looked like a fetch artifact and is in fact the real agent identifier. Worth the extra fetch.

What shipped: hooks via .agents/hooks.json firing on pre_tool_execution/post_tool_execution with regex matchers and either in-sandbox commands or http POST handlers; max_total_tokens capping input+output+thinking, pausing to status: "incomplete" with state preserved and resume via previous_interaction_id; cron triggers binding agent+environment+prompt, each run reusing the same sandbox so files persist; an Environments API to list/inspect/delete sandboxes against a 7-day TTL; Gemini 3.6 Flash as default (3.5 Flash / Flash-Lite selectable via agent_config.model); free-tier access on projects with no active billing.

Three findings. (1) The vendor layer gains a fifth shape, and it's the harness's negation. The closing pitch is agents that run "without breaking your budget or requiring external orchestration" — i.e. without a harness. Where kimi-claw hosts a harness someone else wrote, this dissolves the harness into the model API. A lab integrating downward past the harness rather than into it. (2) The cost is named by Google's own customer, not by Google. OffDeal's CTO: "Before agent hooks, we couldn't do this on Gemini's managed agents: the sandbox is remote, so our validation code had nowhere to run." That's the managed model's limitation in one sentence — verification needs somewhere to execute, and the agent-loops-verification thread assumes you can run a check next to the agent. Hooks are the vendor selling back a slice of the control the managed model removed: agent-middleware's before/after-tool-call hook points, re-sold as a managed feature to plug the hole managed agents opened. (3) Budget is a third guardrail axis. agent-guardrails bounds agents by reversibility and recovery cost; max_total_tokens bounds by spend, which is the bound an autonomous multi-turn loop actually needs — its failure mode is a run that never converges, not a destructive action, and reversibility-based guardrails are blind to that. Pausing rather than aborting also makes it a durable-agents mechanism: a spend ceiling and a resumable checkpoint are one construct from two sides. Detail worth keeping: the deny path explains itself — {"decision": "deny", "reason": "..."} skips the call and feeds the reason into the model's context. Double-edged per constraint-evading-behavior: telling a model exactly why it was blocked also tells it what would have passed. Recorded gaps: no pricing, no execution-time or concurrency limits, no benchmark, and nothing on what an untrusted prompt can do inside a sandbox that has package installation and web retrieval. Also noted the reflexive bit — the post ships an agent skill for its own API (npx skills add … gemini-interactions-api), the skills standard used by a vendor to onboard agents onto its own product. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-28] ingest | Perplexity API Platform skills (github.com/perplexityai/api-platform-developers)

Routed here by the hub (runner-up: ../llm-providers-wiki — Perplexity as a provider and the Sonar→Agent API change; cross-linked, not split, since the artifact is agent tooling). T1, the vendor's own repo. Metadata verified via the GitHub API: Apache-2.0, created 2026-07-15, last push 2026-07-23, 71★. New: perplexity-api-platform-skills (source). Updated: agent-evals (new section on naming the un-greppable residue), agentskills-spec (widest install matrix so far), synthesis (Benchmarks open question), index. Reused perplexity cross-wiki from ../web-browsers-wiki.

Two skills: migrate-sonar-to-agent-api and pplx-cli.

Finding 1 — migration as a skill. The flagship skill's job is moving developers off the vendor's own deprecated API, and its triggers are the symptoms of a botched attempt ("replacing deprecated sonar-pro / sonar-reasoning models, or fixing 400 unknown-field errors and missing citations"). The corpus's vertical skills so far add expertise (agentic-seo-skill, cybersecurity-skills) or shape output (ponytail, i-have-adhd); this is a vendor absorbing its own breaking change — deprecation cost, normally paid by every customer reading a migration guide, packaged as agent work.

Finding 2 — the best-designed eval in this wiki, and it's from the vendor. Fixtures are two versions of one app (raw-HTTP and official SDK), each "verified against production on 2026-07-09" and deliberately loaded with the classic hazards; contributors are told not to tidy them. Reference migrations from live runs are shipped and explicitly disqualified as answer keys — "Model output is nondeterministic, so treat these as examples, not golden diff targets. Grade with check.sh" — which is agent-evals's nondeterminism problem solved (grade properties, not a diff against one lucky run). check.sh is mechanical: leftover greps, required greps, and --live runs of the migrated app against production, so the tally isn't model-asserted. Fixtures are kept comment-free to stop hint-leakage. And step 5 of the runbook documents the A/B against no-skill — rerun without the plugin, grade both. Compare i-have-adhd (rubric and gate, no baseline procedure) and jetbrains-ponytail-tested (an outsider had to build the A/B).

The move worth stealing: the rubric's last section is headed "Judged by eye (not greppable)" and names two things the script can't catch — related-questions must be "preserved via a workaround rather than silently dropped," and the streaming consumer must exit "on every terminal event, not only response.completed." Both are exactly what a greppable rubric rewards (delete the feature and the leftover-grep passes). That's constraint-evading-behavior anticipated by the gate's own authors, who then route the residue to a human instead of pretending the script covers it. The practical answer to "a gate you can cheese isn't a gate": you can't grep your way out, so name what's left.

Also recorded: the install matrix is the strongest agentskills-spec portability evidence yet — one tree serving Claude Code and Codex marketplaces plus a plain skills/ dir, npx skills add said to fan out to ~70 agents, gh skill add with version pinning, and a zero-install path ("Any agent that can fetch a URL can use a skill directly"), which is what a real standard looks like. A docs MCP server ships in the root .mcp.json beside the skills, so procedure and live facts install together. And the repo's CLAUDE.md is one line — AGENTS.md — the vendor convention reduced to a pointer at the cross-tool one. Weakness recorded: no results published. Hand-runnable harness, no score. Across today's four eval-adjacent sources the tally is one outside measurement and three harnesses nobody has run. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-07-29] ingest | jcode (1jehuang) — Rust terminal coding agent

Routed from the hub. T1, the repository read directly: MIT, Rust, 13,127★/1,452 forks on ingest day, created 5 January 2026, pushed the same day it was read. Nearest existing sibling is claw-code — same minimal-Rust-CLI corner, opposite density.

Three things earned their own analysis rather than a feature list.

The benchmark disclosure is better than the corpus norm, and the baseline still carries the claim. jcode publishes RAM and startup numbers against seven rival harnesses with the exact version of each, ranges instead of bare means, the procedure, and a footnote on why Antigravity needed a different readiness signal — most of what jetbrains-ponytail-tested had to supply from outside, given voluntarily. But the baseline row is jcode with local embedding off, a config it doesn't ship, so every "N× more RAM" is anchored there and jcode's own default lands at 6.0× baseline — above pi and Codex CLI, two tools it is comparing itself against. Startup (14 ms vs Claude Code's 3,437 ms) and per-session scaling (~10 MB vs ~213 MB) need no such reading. Folded into the Benchmarks open question as a refinement: method can be impeccable and the choice of baseline still does the work.

Memory moves inside the harness. Turn embeddings into a memory graph, cosine retrieval, injection either passive or gated by a memory sideagent, extraction on semantic drift / K turns / session end, and ambient consolidation that checks staleness and conflicts. iai-pme and knowledge-rag do comparable work behind MCP; this does it in-process, and its consolidation-with-conflict-checking converges on iai-pme's decay and contradiction surfacing independently. Bet against context-rot is that relevance-gated retrieval beats accretion — unmeasured.

Multi-agent gained a second topology, and this is the ingest's real contribution. The corpus's agent-orchestration has meant orchestrator → subagent fan-out, a tree, with isolation per agent. jcode's swarm puts several agents in one repo under a server that tracks reads and notifies an agent when a file it read is edited underneath it. That's peer-to-peer with cache invalidation against the corpus's tree-with-isolation: one prevents the conflict, the other detects and reconciles it. New synthesis section; no evidence either way on which survives contact.

Cross-spoke context recorded on the page, not split: the custom terminal (handterm), the from-scratch mermaid-rs-renderer and the negative-space widget system are real ../ui-frameworks-wiki material next to opentui — but the harness is what jcode is, so it files here. avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-07-29] ingest | jcode — second pass (README sections missed on the first read)

Re-read the full README and found two mechanisms the first pass skipped, both load-bearing. Skills load by embedding hit, not on startup: the conversation is vectorised and a skill injected on similarity against the same index that serves memory. That is a direct alternative to agentskills-spec's progressive disclosure — retrieval-selected versus description-selected capability, each with its own failure mode, neither measured. The author rejects worktrees explicitly ("Git was clearly not built for multi-agent workflows, and git worktrees is not a good solution") and wants a new git-like primitive, which sources the swarm-topology claim made in synthesis an hour earlier rather than leaving it as our inference. Also: the iOS app is announced and unshipped, reaching your own machine over Tailscale with "Openclaw like features" — linked to openclaw, same problem solved earlier. Person node 1jehuang added per ../ENTITIES.md (thin, handle-only, the mahipal-jangra precedent).

[2026-07-29] ingest | Shipping code without human verification (The New Stack)

Routed from the hub. T3 — sponsored content: the post carries "Aviator sponsored this post", Aviator sells a verification product, and the article links to aviator.co/verify exactly where the argument needs a solution. Rubick's credentials are real (ex-VP Eng, New Relic and Gremlin) and the piece is better than most vendor content, but conclusion and product line up. WebFetch returned a truncated excerpt; recovered in full with the firecrawl skill.

Dedup note: this is the fully-readable companion to agent-loops-verification — same publication, same thesis, six weeks earlier, held at T4 because its body never fetched. Kept both; this one carries the detail.

Contribution: testing asserts, verification judges. Testing confirms the code works; verification confirms the change did what was intended and meets the org's standards. Code review fused them and did the judging badly. That explains something the corpus hadn't accounted for — why improving test coverage never dissolved the review bottleneck, since intent alignment, acceptance-criteria coverage, architectural fit and blast radius aren't assertable and are most of what review was for. The article's row-by-row Test/Verify/Split table is reproduced on the page.

The gap, folded into synthesis: it never asks whether an automated gate is adversarially robust. The proposal hands judgement to an LLM on exactly the rows that can't be asserted, and this spoke holds constraint-evading-behavior — agents satisfying a check's letter while defeating its purpose — while ../defensive-security-wiki holds the extreme case of a model treating its sandbox allowlist as the puzzle. The recommended Swiss-cheese layering defends against independent failures, not correlated ones where verifier and author are the same class of system. Recorded as the sharper form of the Benchmarks open question: nobody here has measured a verifier against an agent with an incentive to pass it.

Its one quantity — "improving detection by 50% may speed up the engineering organization by 200%" — has no method, no source and a load-bearing "may", in sponsored content advocating the sponsor's category. Flagged against jetbrains-ponytail-tested. avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | Mobilerun — LLM agents that drive Android/iOS phones

Routed by the hub (Telegram, github.com/droidrun/mobilerun). Runner-up: none — no spoke owns mobile automation, and this is agent-building machinery by every test the registry applies.

New pages: mobilerun (source, T2 — first-party README + first-party benchmark page, eval harness published for reproduction), mobile-device-agent (DefinedTerm), droidrun (Organization, thin). No dedup hit: nothing in the spoke touched mobile, GUI automation or accessibility trees.

What it opens. The spoke's first source where the controlled surface is a device rather than a repo, a browser or a chat channel. Architecture is Portal-on-phone plus runtime-on-your-machine, which makes it model-agnostic by construction (OpenAI/Anthropic/Gemini/DeepSeek/Ollama/OpenRouter) — the phone runs no model, so everything that makes the agent work is harness. Reasoning mode is manager-executor planning, i.e. agent-orchestration with taps as the end effector.

Contribution: the tree-vs-pixels choice is a context-budget argument. Reading the OS accessibility tree instead of screenshots buys a claimed 500× smaller payload and better accuracy, because a labelled button doesn't need recognizing. That inverts pxpipe, which pushes text into images on a coding harness for the opposite reason — no structured description available there, one shipped free here. Folded into synthesis as a rule: ask what the platform already describes before deciding to look at it.

On Benchmarks. 91.4% over AndroidWorld's 116 tasks, external task set, named leaderboard (AutoGLM-Mobile 84.5 / LX-GUIAgent 80.2 / K²-Agent 79.3 / Finalrun 76.7), harness open for reproduction. Better evidence than most of this corpus and still vendor-run — logged against jetbrains-ponytail-tested as a ceiling, not a settlement.

Gap recorded, not papered over: a tap has no diff. The spoke's verification thread assumes an inspectable artifact before the action counts; nothing in the source addresses what reversibility bounds mean when the agent is spending money on a real phone.

avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | video-use — conversational video editing as an agent skill

Routed by the hub (Telegram, github.com/browser-use/video-use), ~20 min after mobilerun. Runner-up: none.

New pages: video-use (source, T2 — first-party README, 18.6k★/2.3k forks, no independent evaluation), browser-use (Organization, thin), text-first-agent-design (DefinedTerm). Dedup: watch-skill already covers video as agent input; this is video as output, and the two now anchor opposite ends of the same discipline rather than duplicating.

Why a concept page and not just a source summary. video-use's "text-first" design — one ElevenLabs Scribe call → ~12KB transcript, composites rendered only at decision points — is the same decision mobilerun makes with the accessibility tree, from an unrelated org in an unrelated medium, on the same day. With watch-skill's ≤100-frame budget that's three instances from three orgs, which promotes a per-repo note to text-first-agent-design. pxpipe resolves as the principled inverse (no upstream structured description on a coding harness → pricing arbitrage is the only lever left), so the corpus now holds four points on one rule instead of one apparent contradiction.

The verification contrast is the other contribution. video-use self-evaluates at cut boundaries and re-renders up to 3×; mobilerun can't, because a tap leaves no artifact. Folded into synthesis: what decides whether loop-engineering's feedback signal exists is whether the action produces something inspectable before it counts — and where it doesn't, agent-guardrails has to gate the plan (video-use asks the user to approve a cutting strategy before rendering; on a phone that's all there is).

Weakness recorded: the README claims it beats presets and traditional NLEs for any content type and measures nothing; the self-eval loop is a mechanical boundary check, not a judgment on edit quality. Stars measure attention (openclaw precedent). Filed under the standing Benchmarks gap.

Also touched: mobilerun, mobile-device-agent, watch-skill, pxpipe (backlinks). avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | MD-This-Page — web page → Markdown, for pasting into a model

Routed by the hub (Telegram, github.com/Ademking/MD-This-Page). Runner-up: ../research-wiki — the markdown-clipper / PKM lineage is a real fit, and this is a clipper. Routed here because the README's stated purpose is LLM context efficiency and that corner (headroom / pxpipe / safe-prompt-pruning / tokenmining) lives in this spoke. Recorded on the page so it can be re-filed if the PKM reading wins.

New pages: md-this-page (source, T2 — first-party README, 1.3k★, nothing measured), adem-kouki (Person, thin).

Contribution: a fifth lever, and the first one upstream of the harness. The token-budget corner had compress (headroom) / remove (safe-prompt-pruning) / swap modality (pxpipe) / route (tokenmining) — all acting on context already in flight. This one strips at capture: Readability drops ads, nav and script tags before the text ever reaches a clipboard. Folded into synthesis and into text-first-agent-design.

And it's the day's pattern with the agent removed. Third source in a morning taking a structured representation over the raw surface (mobilerun tree-not-pixels, video-use transcript-not-frames, this article-not-page) — but here a human does the curation by hand. Worth stating plainly because it undercuts the machinery: Mozilla shipped Readability for Reader Mode in 2010, and it captures a meaningful share of what the proxies and MCP servers are automating. Much of what bloats a context window is structural boilerplate strippable by rule, not a judgment. Nobody has measured where that stops being true — logged as open.

Weakness: the token-efficiency claim is asserted, never benchmarked. No agent-facing surface at all (no CLI, no MCP), which is exactly why the ../research-wiki runner-up is defensible.

avoid-ai-writing run. Verify deferred per hub policy (content-only, no page moves).

[2026-08-03] ingest | Do Context Files Help Coding Agents? (Khatri, arXiv 2607.27250)

Routed from the hub (Telegram). T1 — primary research artifact with full method, a pre-registered manipulation probe, and released code/data/power-analysis. The tier rates the artifact, not the strength of the conclusion: it is a preprint, single-authored by an independent researcher, and by its own power analysis cannot detect anything under 30pp. What it delivers is a bounded null with a mechanism, which is more useful than a bare negative. New pages: context-files-ablation-paper (ScholarlyArticle, source), agent-specific-task-difficulty (DefinedTerm/methodology). Updated: agents-md (its first evidence of effect, after two months of adoption-only sourcing), context-driven-development (premise contested), agent-evals (two design traps + the turn-count portability bug), synthesis, index. Dedup: agents-md, agents-md-spec, context-driven-development, agent-evals, cc-thinking-skills and jetbrains-ponytail-tested all already paged — updated in place, none duplicated. Only one Thing page was created from the paper; the per-agent-difficulty finding earned it because it is reusable methodology, not a fact about this experiment. Gap-relevance: substantially answers the spoke's oldest open question (Benchmarks) — the neutral, controlled, multi-agent measurement that bullet has asked for since June. Rewritten to say what remains: purpose-built task-specific context untested, no study at the ~120-task scale, and the harness-vs-model tradeoff still untouched. Synthesis: new section "Three independent measurements, three nulls" with a comparison table — cc-thinking-skills (zero robust verdicts), jetbrains-ponytail-tested (works, oversold 3.5×), and this paper (no correctness effect). Three parties, three layers of the structural stack, no outcome improvement in any of them; two found process effects and both authors attributed those to delivery mechanics, not agent capability. That sharpens the standing does-structure-substitute- for-capability question into a testable split: structure changes how an agent works, the model decides whether it succeeds. Contradiction flagged: adoption is not evidence. This spoke documented agents-md through its own site (60k projects, 30+ tools, Linux Foundation) and context-driven-development as a named discipline, all of it about reach rather than effect. Both kept; what does not survive is the inference from one to the other. The one surviving positive result is recorded as narrowly as the paper states it: context cut blind full-suite pytest runs (3.67 -> 2.44 -> 1.67/cell) on the one repository whose file warns the suite takes >20 minutes — an operational saving from a single warning about an unobservable cost, flagged exploratory by the author, and absent on the other repo (which trends the opposite way). Fetch note. The PDF summarizer returned a summary that contradicted the paper's own abstract — "modest improvements", "both agents showed performance gains", Claude 3.5 Sonnet and GPT-4o as the subjects, and rho=0.75 read as a difficulty threshold. All wrong. Extracted the 11-page text with pypdf and worked from the primary throughout; the real agents are Claude Code (sonnet-4-6) and Codex CLI (gpt-5.5), and rho=0.75 is the cross-agent correlation of per-task pass rates. Entity: author Prakhar Khatri deferred — the paper gives a name, an email and "Independent Researcher", which is not enough evidence for a node under ENTITIES.md. Page on recurrence. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-03] ingest | QM — Y Combinator's multiplayer agent harness

Routed from the hub (Telegram; the repo). T1 — primary artifact: source, licence, threat model and config all inspectable. Everything about how well it works is self-description — no benchmarks, no deployment evidence beyond YC's own, repo five days old (created 2026-07-29) with 109 open issues. Recorded that candour about limitations is not the absence of limitations; the repo says as much itself ("early, experimental software"). New pages: qm (SoftwareSourceCode, source), agent-scope-isolation (DefinedTerm/architecture). Updated: agent-guardrails (the posture ladder with an effects floor), constraint-evading-behavior (a vendor conceding its own guardrail is bypassable), synthesis, index. Dedup: no QM page existed. agent-memory, agent-orchestration, agent-skills, agentic-coding-harness, cloud-run-sandboxes all already paged — linked, not duplicated. Only one new Thing page: scope isolation earned it as a design problem nobody here had costed; the posture ladder went into agent-guardrails rather than becoming a page of its own. Gap-relevance: opens a genuinely new axis. Nearly every artifact in this corpus is one agent, one principal — coding harnesses run as whoever launched them, and "multi-agent" has meant orchestrated workers of a single user. Once two employees share a deployment, memory, files, keychain view, permissions, skills, crons and sandbox each need an owner. That also argues against the "one markdown file" pole of the standing agent-memory unpriced-range tension: one file can't hold per-scope views. Synthesis: new section "The agent acquires colleagues." On the constitution vs containment contradiction, this is the first shipping product to state the containment position as architecture — "the agent and software it runs in a sandbox are not trusted to make authorization decisions" — and then document its own leaks. Command policy "a speed bump against mistakes and injection, not a sandbox boundary"; credential purposes travel "as an instruction to the model" with core not judging whether later commands stay inside them; sandbox credentials plaintext while in use. Recorded as sharpening rather than resolving the tension: containment and prose differ in degree of porosity, not in kind, which puts blast radius back at the centre. The line kept for reuse across every human-in-the-loop page: "An approval means a human accepted the displayed action under the information available at that time, not that the resulting behavior is safe." Two practices recorded as transferable independent of the product: a 7-day npm dependency cooldown (min-release-age=7, aimed at the yank-within-hours compromise window) — the corpus's first concrete supply-chain control — and contributions as human-written prose in adrs/, maintainers implementing, with "please do not have AI artificially expand what you'd like to do into a formal proposal." Small cross-link worth its line: the repo's CLAUDE.md is nine bytes, the string AGENTS.md — a pointer, not a copy — and the file it points at is mostly process rather than repository facts. Noted beside context-files-ablation-paper, ingested earlier the same day. Entities: publisher is Y Combinator (yc-software); not paged — the evidence here is one repo, and the org's other appearances in this hub are as an investor, not a builder. Page on recurrence. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-03] ingest | claudecode-writer (Women Defining AI) — a Claude Code content workspace

Routed from the hub (Telegram). T3 — a template with no evaluation, no benchmark and no deployment evidence beyond its own illustrative walkthrough. 214★/44 forks is usage evidence, not quality evidence. Stale: created 2025-09-03, last pushed 2025-09-06, read 2026-08-03 — three days of commits and then eleven months of nothing, so it predates the agentskills.io standard and the subagent conventions this spoke tracks. New page: claudecode-writer (SoftwareSourceCode, source). Updated: context-files-ablation-paper (the untested case now has a specimen), renwei-writing (the failure it names, demonstrated by a tool promising the opposite), index. Dedup: autonovel, renwei-writing, writing-a-good-claude-md, claude-md, agent-skills all already paged — linked, not duplicated. No new Thing page: the source is small and its one reusable idea (voice examples as purpose-built context) belongs on the paper that named the gap, not on a page of its own. Gap-relevance: advances the day's own open question. context-files-ablation-paper, ingested hours earlier, found no correctness effect for generic context files and carved out exactly one untested case — purpose-built, task-specific context, "a fact the agent provably cannot infer." context/writing-examples.md is that shape: few-shot demonstration rather than instruction, and voice is a genuine knowledge gap, which is the category the paper's failure triage found absent from coding tasks. So the mechanism behind the null doesn't obviously transfer. Recorded as a well-posed instance of the question, not an answer — the template ships no evaluation and style match has no gold test. Architecture noted: fan-out repurposing, one researched artifact into N platform-specific transforms by specialist subagents. Contrast autonovel, which also fans out into formats but only after a sequential pipeline with quality gates and retry loops; this has one pass per stage and the human as the only reviewer. Read against its own demo. The README's showcased LinkedIn output carries every marker of generic LLM social copy — the "still stuck in 2019" hook, a ❌/✅ ladder, one-line paragraphs, "Game changer.", an engagement-bait question, a hashtag stack — under a promise that "your voice stays authentic." Recorded on both pages as the 人味儿 failure renwei-writing was written against, arriving from the generation side rather than the editing side. Voice-matching machinery did not prevent it, which is the observation worth keeping; the criticism is of the demonstrated output, not of the idea. Entities: Women Defining AI not paged — evidence here is one repo. Page on recurrence. Verify deferred per hub policy (content-only). avoid-ai-writing run.

[2026-08-04] ingest | agtop (ldegio)

Routed here by the hub (runner-up: ../platform-ops-wiki). T1 — the project's own repository, which is the primary artifact for what the tool does, consistent with how this spoke grades headroom and pxpipe. 1 new page: agtop. Updated: synthesis (token-budget corner), index.

It is the corner's first instrument. The token-budget corner had five levers and all five are interventions — compress (headroom), remove (safe-prompt-pruning), swap modality (pxpipe), route (tokenmining), strip at capture (md-this-page). agtop changes nothing about how an agent runs. It parses Claude Code and Codex JSONL transcripts for spend, tokens and CTX% saturation, correlates OS processes to sessions for CPU and memory, and reports per model and per subagent.

Why that matters more than a sixth lever would: the corner's claims are quantitative and nearly all self-reported, and this spoke already holds what happens when someone checks one — jetbrains-ponytail-tested took an advertised −54%/−20% to −15.4%/−10.3%. An instrument is what lets a practitioner run that check on their own workload.

Limits recorded on the page: it measures cost, not quality (a session that spent little because the agent gave up early looks good), CTX% makes context-rot's quantity visible without making it actionable, and it parses a private transcript format that can change without notice — marked volatile for that reason.

Noted, not acted on: [[claude-code]] resolves nowhere in the corpus. The product is referenced constantly across this spoke and has no entity page, so the two links in the draft were de-linked to plain text rather than inventing a page mid-route. Worth a page of its own on a source that is actually about it.

[2026-08-04] lint | consolidated six dated sections into three thematic ones

Growth edge 4, actioned at the curator's instruction rather than deferred again. ../QUALITY.md check 3 names dated (added 20xx) sections piling up outside the living thesis as synthesis drift, and this spoke had six of them stacked between Open questions and Contradictions — the most in the corpus.

They were not six findings. Two pairs were one argument each:

  • The agent got a phone and Two arrivals promote a note into a principle were the same claim about derived representations, with the second explicitly ending "see below" pointing at the first. Merged into The surface dictates the representation, which now carries mobilerun, video-use, watch-skill and text-first-agent-design in one argument, with the token-budget corner and agtop as a subsection under it, and the loop-closure point as another.
  • Three independent measurements, three nulls and Verification's endgame were both about what measurement has found and what it hasn't. Merged into Structure changes how an agent works, not whether it succeeds — the three-study table, Khatri's failure triage, and the adversarial hole in automated verification as a subsection.
  • Multi-agent gained a second topology and The agent acquires colleagues both push past one-agent- one-principal, from topology and from principals. Merged into Shared state is the next boundary, with the containment-porosity finding as a subsection.

264 lines → 202. Nothing was dropped: every number, quote and link survives, including the ones that make the spoke look worse (qm's SECURITY.md concessions, the 8.6% invisible failures, Mobilerun's vendor-run benchmark). What went is the accretion — the dates, the "a third arrival the same morning" scaffolding, and the duplicated framing where two sections restated each other.

The rule this leaves behind, recorded in the edge: a new finding extends the thread it belongs to. It does not get a dated section of its own. That is what stopped this spoke's synthesis being readable from the top, and re-earning it is a full pass.

[2026-08-04] lint | paged claude-code — the hub node the spoke was missing

Curator's instruction, following the observation logged with the agtop ingest: [[claude-code]] resolved nowhere in the corpus while being named across 84 pages of this spoke. Two links were de-linked to plain text that day rather than inventing a page mid-route; they are restored.

Written from the corpus's own sources, not from product documentation — which is why the page is short on features and long on how the thing behaves under measurement. That is also the honest division of labour: Anthropic documents what Claude Code does, and this spoke's value is holding what independent parties found when they tested it.

What it pulls together: claude-md as the vendor-tied sibling of agents-md; claude-agent-sdk as the harness unbundled from the terminal; claude -p headless mode as the primitive under orchestrate-100-agents-claude-code; effort-level as a harness control orthogonal to model choice; and the commercial terms that move (claude-code-weekly-limits-promo, anthropic-agent-sdk-billing-pause).

The section worth having is the measurement one, and it is unflattering to the tool's own conventions: context-files-ablation-paper found CLAUDE.md-style context injection does not move correctness on it (≤10pp, no-context arm higher on borderline tasks), jetbrains-ponytail-tested took an advertised −54%/−20% to −15.4%/−10.3% running on it, and cc-thinking-skills returned zero robust verdicts from 39 skills inside it. Together they are this spoke's evidence for the structure-changes-how-not-whether split, and they say the things people bolt onto the harness are worth less than advertised — not that the harness is ineffective. claude-code-best-practices agrees from the practitioner side: "the tool does not matter nearly as much as the person holding it."

A reflexive note is on the page rather than left implicit: this hub runs on Claude Code and on superpowers, which is a standing reason to read the spoke's coverage of it with suspicion.

[2026-08-04] quality | staleness queue — 9 pages re-verified, 2 projects superseded

Re-read every past-window page in this spoke. Star counts moved as expected; two findings did not.

get-shit-done was archived 2026-06-26 and is read-only, with development moved to GSD Core (open-gsd/gsd-core). agent-starter-pack is in maintenance mode — critical fixes only, active work moved to agents-cli. Both were ingested in June and both were superseded within weeks. Neither successor is held here. That is the case the freshness convention exists for, and it is the first time the queue has caught a dead artifact rather than a merely dated one.

agentsys's claim grew numbers and is still unmeasured. The README now says "Sonnet + agentsys produced more output with higher specificity than raw Opus — at 40% lower cost", with a table claiming 73–83% cost savings at equal outcome quality. That is this spoke's #1 growth edge restated by the vendor, not answered — no task set, no trial count, no rubric, no replication. What changed is that it is now specific enough to falsify, which makes it a better target. Edge updated.

Scale corrections worth carrying: gstack 126.3k★/19k forks with 497 open PRs (the maintainer is the bottleneck); openhuman 36k★, still early beta; oh-my-pi 21.8k★, ~80k lines of Rust, 31 tools, and five run modes including ACP; agent-kanban 433★ on an FSL-1.1-ALv2 licence — delayed-open, not open, and the only such licence in the spoke; claude-to-speech 18★, one person's hook script, recorded as such so the page is not read as evidence of adoption.

jaseci has repositioned: the quick guide now leads with "one language, one compiler, the whole stack. No glue" and two named properties (synechic, topokinetic), presenting a full-stack language whose agent capability is one use case. This spoke filed it as an agent framework. Flagged on the page.

[2026-08-05] ingest | Education Agent Skills (Gareth Manning)

Routed from the hub (runner-up: none — ../engineering-education-wiki was checked and declined; it owns how software and ML engineers learn, and this is school pedagogy). education-agent-skills — 165 SKILL.md files across 20 domains on the agentskills.io standard, CC BY-SA 4.0, 545★, T2. Created gareth-manning (Person; entity-index returned no match ≥ threshold). Budget: 1 of 8.

Counted from the repository tree, not the README: 165 skills confirmed, 20 domains, largest original-frameworks (17), smallest inclusive-design (3). Ships .claude-plugin/, .codex-plugin/ and .agents/plugins/ manifests, a CI validation workflow, and a hosted MCP server with OAuth.

Two things here are new to the corpus, one is a second sighting.

New: skills that cite their evidence. Frontmatter carries evidence_strength and named evidence_sources — VanLehn (2011), Shute (2008) and similar — putting the provenance of the technique in the machine-readable layer. Folded into the Benchmarks open question as a move on the gap from the other end: it measures nothing, and a citation behind a technique is not evidence for the prompt encoding it, but it makes the quality claim checkable instead of merely asserted. Cheaper to ship than jetbrains-ponytail-tested-style measurement, so possibly more contagious.

New: the first vertical pack aimed past the practitioner. Domain 20 (13 skills) addresses students mid-study-session rather than the teacher.

Second sighting: typed input_schema/output_schema, independently of pm-skills-arsenal — different author, different vertical, both paired with an MCP invocation surface. Two independent arrivals make skill-as-typed-function a property of the standard's use, not one team's style.

Caveats on the page: the "evidence-grounded" framing is the author's own and unaudited; GitHub reads the licence as NOASSERTION while the LICENSE file is CC BY-SA 4.0; the author's 20-years/27-countries standing is self-reported.

[2026-08-05] ingest | PullMD

Routed from the hub. pullmd — self-hosted URL- and file-to-Markdown service (Aeterna Labs, AGPL-3.0, 365★), T2. Created aeterna-labs (Organization; entity-index no match ≥ threshold). Budget: 1 of 8.

Lands directly in the token-budget corner. md-this-page added strip-at-capture on 2026-08-03 and its page recorded the limit — "no MCP surface, no CLI, no way for a harness to invoke it." PullMD removes that limit: MCP server, Claude Code skill activating on web-reading requests, OAuth 2.1 for the claude.ai connector. Capture-time reduction moves inside the agent loop for the first time in this corner.

Second thing worth having: the fallback cascade — Readability → Trafilatura → origin-supplied Markdown → headless Chromium via Playwright. An architectural concession that no single extractor works across the web, where md-this-page is Readability alone. Unmeasured, like everything else in this corner; the cascade's selection quality is precisely what nobody tests.

Routing tension recorded, not hidden. _inbox/firecrawl.md holds a near-twin parked since 2026-06-23 under web-data-extraction, on the reasoning that extraction infrastructure is not agent tooling. The line taken here — Firecrawl is a scale-crawling platform with an agent interface, PullMD leads with agent consumption and fell into a live thread — is thin, and both this page and the park record say so. If a third extraction tool arrives, cluster all three rather than route on thread-adjacency.

[2026-08-05] ingest | Statewright

Routed from the hub. statewright — state-machine guardrail for coding agents (Rust, ~449★), T3. No entity paged: the statewright.ai org has no evidenced footprint beyond the repo; deferred.

Third enforcement shape on the agent-guardrails thread. Blast-radius blocks the dangerous act (agents-never-do-alone); containment sandboxes the agent/its tools (nono); this scopes tool availability by workflow phase (plan=read-only → implement=edit → test=test-cmds), rejecting out-of-phase calls. The novelty is the motive — enforcement aimed at reliability, not safety — which generalizes enforcement-over-instruction from "don't do harm" to "stay on task." Destructive-op blocks still apply inside a permissive phase, so it layers on the blast-radius model.

Folded into the Benchmarks open question in synthesis: it reports 2/10 → 10/10 for two local models on a 5-task SWE-bench subset with constraints applied — the cleanest structure-substitutes-for- capability demonstration in the corpus and one of the least measured (self-run, tiny-N, the mirror of cc-thinking-skills's honest null, the setup jetbrains-ponytail-tested showed collapses under replication). Kept the one falsifiable boundary: below ~13 GB the models couldn't retain file content and constraints didn't help — structure steers a capable-enough model, it doesn't manufacture capability. Hosted-dependency caveat recorded (requires statewright.ai key; server-side share unclear).

[2026-08-05] ingest | LongHorizon-Harness

Routed from the hub (Telegram). longhorizon-harness — AMAP-ML's long-horizon harness over Claude Code / Codex / OpenClaw, MIT, T1 (official project repo, per the ladder's repo rule; the benchmark numbers inside it are first-party and recorded as such). Paged alibaba as the owner Organization — no node existed despite six pages mentioning it; entity budget otherwise unused, AMAP-ML folded into the Alibaba node rather than given one of its own (evidence-only: the repo is its whole footprint here).

The design contribution is where verification sits: the Auditor gates what enters persistent state, so an unverified result never becomes something the next round builds on. That is a third enforcement shape next to statewright's phase-scoped tools and agent-guardrails' blast-radius limits, and it extends the agent-loops-verification thread rather than restating it. Fresh context per Executor round reads as an architectural answer to intrinsic context-rot — if context never grows, the attention budget never erodes.

Gap-relevance: this is the seventh measurement to pass growth edge #2 (the harness-vs-model substitution) without testing it — it holds the backbone fixed at Qwen and varies the harness, the same shape as the six before. The edge stands, better posed: the harness arm now has a price at real N (+28.9 pts over 114 WeaveBench tasks; 108 on OSWorld 2.0), so what is missing is one cell, not the whole design. Synthesis gained the split this makes visible — every independent study in the corpus finds process effects and no outcome effect, every vendor result claims a large outcome gain, and the two groups are also testing generic vs purpose-built structure, so the confound is unresolved either way.

[2026-08-05] ingest | Agentic Awesome Skills (sickn33) + vibe-tools (via Dan Kornas on X)

Two sources routed from the hub minutes apart. Ingested together because they answer the same question from opposite ends, and the synthesis section says so.

agentic-awesome-skills (Collection/SoftwareSourceCode, T3, first-party README read in full — 729 lines): 2,003 SKILL.md playbooks, MIT, ~44.5k★, v15.9.0. The catalog is not the story. At 2,003 skills the index stops fitting, and the README documents its own breakage — Antigravity watching ~/.agents/skills can exhaust context and crash-loop, so the installer refuses to run bare against that target, and manifests cap at 128. The project's answer is a local read-only MCP the agent queries (search_skills/get_skill/compose_stack), with the agent choosing exact IDs and Core validating structure while explicitly certifying nothing about semantic fit, compatibility or safety to apply. Also recorded: a full install ships critical and authorized-use-only offensive instruction files.

vibe-tools (SoftwareSourceCode, T3): eastlondoner's Node CLI, MIT, ~4.8k★ — Perplexity research, Gemini whole-codebase context, Stagehand browser control, GitHub/Linear/docs/YouTube integrations. It installs by writing the harness's instruction file. The arriving source was a promotional X thread by dan-kornas (13 likes, newsletter pitch in the reply); I re-derived every specific from the repository and graded on that, noting the thread as the arrival route rather than the evidence.

New synthesis section, "Progressive disclosure has a ceiling, and a catalog found it": the cheap-frontmatter premise of agentskills-spec holds per skill and fails in aggregate; the pair of sources is a spectrum with the context budget as its axis — push capability into the context until it stops fitting, then make the agent go fetch it. Neither source measures whether either works.

Entity nodes: sickn33, dan-kornas — both thin, both paged because who curated this is load-bearing for the grade.

[2026-08-05] ingest | SuperClaude Framework

Routed from the hub (runner-up: none). New page superclaude-framework (source, T3 — project README). 30 slash commands, 20 role-persona agents, 7 behavioral modes, 8 MCP servers, pipx-installed as injected configuration into claude-code. MIT, 23.7k★, v4.3.0; v5 is being rebuilt as a TypeScript plugin system.

The useful thing is the comparison, not the feature list. It teaches nearly the same development arc as superpowers and compound-engineering-plugin and differs on who invokes — operator-typed commands, not agent self-invocation — so the spoke now has three packs agreeing on the methodology and disagreeing on the control point. Folded into synthesis alongside the existing invoke-axis paragraph, and added the cross-link from superpowers.

Recorded as the section's standing hole rather than one entry's caveat: none of the three has been evaluated. cc-thinking-skills is still the only ablation of a comparable pack, and it found nothing.

[2026-08-05] ingest | Ruflo (rUv / ruvnet)

Routed from the hub (runner-up: none). New page ruflo (source, T3 — first-party README). Agent meta-harness over claude-code: swarm topologies, 100+ agent roles, ~210 MCP tools in parallel, AgentDB/HNSW vector memory, SONA self-learning, zero-trust federation across organizations. MIT, 67.1k★, formerly Claude Flow. Not a duplicate of anything held — nothing in the corpus was paged under either name.

Filed as the stack-everything maximum against entries that each defend one claim (iai-pme memory, nono sandboxing, qm multi-operator, orchestrate-100-agents-claude-code fan-out). One conditional number in the whole README (vector search 1.9×–4.7× over brute force above the crossover). Synthesis now states the pattern directly rather than per-entry: adoption and measurement are uncorrelated here.

Federation flagged as the genuinely new claim — no other entry attempts cross-org agent collaboration.

[2026-08-05] ingest | OPC Skills (ReScienceLab)

Routed from the hub (runner-up: ../search-marketing-wiki, on the SEO skill only — the spoke's standing rule files skill packs here). New page opc-skills (source, T3). Apache-2.0, 1.2k★, ~11 skills, 16+ host tools.

Two things it adds. It is the first pack in the corpus whose subject is not software — market research, domain hunting, SEO, logo/banner generation for a one-person company — which means it operates where tests, a compiler and CI do not exist, and nothing in it replaces that verification. And it declares dependencies between skills (domain-hunter needs twitter and reddit), which agentskills-spec does not support and agentic-awesome-skills answers instead with an MCP control plane. Both folded into synthesis; composition now has three documented placements and no argument for any of them.

[2026-08-05] ingest | The Hitchhiker's Guide to Agentic AI (Roitman, arXiv 2606.24937)

Routed from the hub (runner-ups: ../engineering-education-wiki, ../machine-learning-wiki). New page hitchhikers-guide-agentic-ai (source, T1). Book-length practitioner reference, cs.AI, revised 2026-07-27.

First source in this spoke that arrives already assembled rather than as one artifact to be synthesized. Two independent confirmations worth recording: it gives loop-engineering a chapter with named mechanisms (generate-verify-retry, adaptive budget control — the second is a lever nothing here has documented), and it states the in-context/external/episodic/semantic memory taxonomy the spoke had assembled from iai-pme and jcode.

Tension kept rather than resolved: the book treats agent evaluation as teachable while the field ships huge unmeasured packs. Cross-spoke context noted in the page — the substrate/alignment half belongs to ../machine-learning-wiki and ../llm-inference-wiki, and the artifact is also a curriculum of the kind ../engineering-education-wiki studies. Routed here on the second half, which is this spoke's domain.

Not new evidence: a synthesis is only as good as its references, and the page says so.

[2026-08-06] ingest | Claude Code Development Kit — the doctrine shipped as defaults

Routed from the hub (Telegram, curator); no competing spoke, this is squarely agent tooling.

New page claude-code-development-kit (T1, first-party repo). Peter Krueck, MIT, 1.4k★. Four docs/ai-context/ files kept in step by /update-docs, tiered /prime (~4–6k tokens at the light default), /review-work fanning out Bug Hunter / Rules Auditor / Architect subagents, /second-opinion escalating to Codex or Gemini CLI, plus secret-scanner and deny-list hooks.

The reason it is worth a page rather than a line: its foundation is the exact mechanism context-files-ablation-paper ablated to a null across 288 gold-test runs. The kit reports no evaluation, so it adds nothing to that. But all of its own claims are process claims — less re-derivation, smaller loaded floor, cleaner merges — and process effects are precisely what the measured studies did find. So it is not in conflict with the corpus's strongest measurement; it is selling the half that survived. Folded into synthesis under the structure-vs-capability table, with a new edge: nobody has priced the structural layer against the burn it removes, and /update-docs, /review-work and /second-opinion each cost a run.

Two tensions recorded rather than resolved. claude-md's discipline is hand-authored, never auto-generated; /update-docs is machine maintenance, so the kit joins the auto-generation pole next to openwiki and autowiki-factory — with the wrinkle that its files are opt-in behind /prime, which exempts them from the instruction-ceiling argument but not the low-signal-bloat one. And the review-on-stop hook nudges across three stops before a final commit, which by this spoke's own agent-guardrails and constraint-evading-behavior pages is a prompt rather than a boundary; the deny list and secret scanner are the real ones.

Also classified: it is not a harness — no loop, no runtime. A configuration layer over someone else's, and the packaged form of governed-context-claude-code.

No entity node for Peter Krueck. The source gives a name and a LinkedIn link and nothing else; ../ENTITIES.md is evidence-only and the page would have been a name. Same call as staskh earlier today.

[2026-08-06] ingest | Model or Harness? An Interaction-Centric Taxonomy for Localizing Agent Failures (hub-routed, Telegram, URL-only, T1)

Raj, Gupta, Mahmoud, Dumitru, Yi, Sabharwal & He (Scale AI), arXiv:2607.28802v1, 2026-07-30. Created source summary model-or-harness-taxonomy (ScholarlyArticle, url-only, T1 — frozen protocol, published agreement statistics, stated limitations; preprint, 40 examples annotated by the taxonomy's own authors). Created concept page agent-failure-localization: the repair-assignment problem and the edge-plus-fault-side scheme that answers it, with the root-cause rule (label the earliest failure from which execution does not recover) and the note that 36-of-41 model-side is an artifact of the attribution rule, not a measurement.

Threads updated. agent-evals gains "A score does not tell you what to fix" — an eval answers did it regress, never which component to repair; localization presupposes the trace, so the page's tracing argument arrives a second time from the diagnostic side. agentic-coding-harness gains "When the harness is the thing that broke" — agent = model + harness splits the blame too, and the taxonomy's uncomfortable result for this spoke's thesis is that blame still defaults to the model.

Synthesis: growth edge #2 (the harness-vs-model tradeoff) gets its first entry that is not an eighth miss — the paper is upstream of the measurement the edge wants, supplying the classification a frequency study would have to count with. Edge restated to ask for both the controlled comparison and a frequency study. The "does structure substitute for capability" open question gains a third bound: the five non-model failure modes are the cleanest statement yet of where structure is not substitutable.

Entities: scale-ai (Organization/Corporation, sourceOrganization) and harsh-raj (Person, lead author) — both thin, evidence is the one paper. Budget: 2 of 8. Noted on scale-ai that an evaluation-company publishing a failure-attribution scheme validated on its own annotations is adjacent to its commercial interest, which is why it does not satisfy growth edge #3.

[2026-08-06] ingest | Scaffold Effects on GAIA: A Controlled Comparison (research pass, T1)

Starace, arXiv:2606.08529 (2026-06-07), hunted via the hub Research Pass against ## Most wanted #1. This is the design the spoke's oldest edge specified, and it closes it: 5 models × 3 scaffolds × 139 GAIA questions × 3 attempts, pre-registered, 5,907 completed runs — both factors varied in one experiment, where the previous seven sources each held one fixed.

Created scaffold-effects-gaia (ScholarlyArticle, url-only, T1) and jason-starace (Person, thin, no affiliation stated). Entity budget 1 of 8.

The finding, in the form the spoke needed it: scaffold choice moves a single model by up to 23 points, but never across a capability tier — Haiku's best configuration (0.459 L1) stays below Sonnet's worst (0.660). It does carry one step: Sonnet + Planner-Actor-Rater beats Opus + planner-executor on both levels, which is agentsys's README claim confirmed under control for the first time since June, and simultaneously bounded. Two corollaries recorded: the best scaffold is model-specific (Opus tops L1 on plain ReAct; the multi-agent scaffold hurts it, and its advantage never left the Anthropic family), and every single-scaffold benchmark number in this corpus is a scaffold-conditional estimate.

Weaknesses recorded, not smoothed: the scaffold arms vary their tool surface alongside loop structure, so the effect is scaffold-plus-tools; GAIA validation answers are public and the author says L1 may partly reflect memorization; 314 records lost to provider errors; cost is log-derived not billed. Single author, single non-coding benchmark.

Synthesis: the open question "does structure really substitute for capability, and how far?" is marked answered with the history and the earlier bounds kept below it (record-don't-overwrite). Growth edge #2 retired as CLOSED and replaced, ranked lower, by the same design on a coding benchmark.

[2026-08-06] ingest | diagram-design (hub-routed, Telegram, URL-only, T1)

cathryn-lavery's Claude Code skill for editorial diagrams (MIT, ~3k★/220 forks, 22 commits). Created source summary diagram-design (SoftwareSourceCode, url-only, T1 — official repo; every quality claim is first-party and nothing is measured against a generic-diagram baseline) and the author node cathryn-lavery (Person, thin — BestSelf.co founder; a product-design practice, not an engineering one). No new concept page: the substance lands on reasoning-scaffolds, which gains a paragraph — this is the output-shape category in visual form, and the first one whose form rules are checked by a linter rather than by the model's own pre-send inspection (i-have-adhd) or a reader's eye. Second point of interest: a one-time brand-capture step (read the user's website, extract colors + fonts, reuse forever) — claude-md/agents-md durable context applied to visual identity, with the state sourced from outside the repo. Folded into synthesis.md in two places: the output-shape/restraint passage, and the vertical-generalization list, where the designpowers pairing gives one design vertical two opposite structural bets (10-agent pipeline vs one skill over 34 progressively-loaded reference docs). Also noted: self-contained HTML+SVG output (no build/JS/CDN) as text-first-agent-design's instinct pointed at the deliverable, rhyming with beautify-github-readme's Markdown/SVG split. Cross-spoke context recorded on the page: design-system substance is adjacent to ../ui-frameworks-wiki's design-system, which stays the owner of design systems as a building concern.

[2026-08-07] ingest | Claw-SWE-Bench (Zheng et al., arXiv 2606.12344) (via research pass)

T1 — arXiv preprint, 2026-06-10. Closes the coding half of growth edge #2, the successor the 2026-08-06 pass wrote when scaffold-effects-gaia closed the general form: "a T1/T2 crossed model × harness comparison on SWE-bench-class tasks."

350 issue-resolution instances, 43 repos, 8 languages. Prompt template, Docker image, repo state, 3600s timeout and the SWE-bench evaluator are all held fixed, so the deltas are harness or model and nothing else. Harness spread 12.5 pp (GLM 5.1) / 27.4 pp (Qwen 3.6-flash); model spread 29.4 pp across 9 models. Authors: "harness choice is a first-order factor… large enough to reorder leaderboard conclusions if the harness is not specified."

The finding folded into the synthesis is the interaction: harness spread roughly doubles as the model weakens, and the weakest harness (generic) loses 24.5 pp across that step against openclaw's 7.4. So scaffolding substitutes for capability, and is worth most where capability is least — which extends GAIA's "one tier, not two" rather than repeating it.

Recorded limits: two grids rather than a full factorial (only 2 of 9 models meet all 5 harnesses); no frontier model in the harness sweep, so the top of the range is the untested cell — filed as the replacement edge. One ordering artifact in the paper's model table (DeepSeek-V4 Flash listed below MiniMax 2.7 despite scoring higher); every Pass@1 checks out against its resolved count over 350.

Rejected in the same hunt: CodeScout (arXiv 2603.05744). Two fetches of the same paper returned irreconcilable model sets — a search summary said GPT-5-mini / DeepSeek R1 / Qwen3 Coder 480B, a PDF read said Claude 3.5 Sonnet / GPT-4o / Claude 3 Opus with a suspiciously uniform "+8pp for all three" and approximate figures. Not ingested: the numbers could not be pinned down, and it is a problem-statement-augmentation study rather than a harness comparison. Inside the Scaffold (2604.03515) also passed over — a 13-agent source-code taxonomy with no benchmark results. No entity nodes created (Zheng et al. — thin one-off authors; no recurrence yet).

[2026-08-07] ingest | Stanford is running 37,000 AI agents as a virtual biotech (VentureBeat)

Routed from the hub (Telegram). VentureBeat's write-up of James Zou's VB Transform 2026 talk on virtual-biotech — T3: trade press relaying a first-party conference talk, tiered by what it rests on. Every load-bearing number is the speaker's own and the linked bioRxiv preprint (2026.02.23.707551) is not read.

New pages: stanford-virtual-biotech-37000-agents (source summary), virtual-biotech (the system), paperclip (its context layer), agent-environment-design (the environments-not-workflows principle). Updated agent-orchestration — Virtual Biotech is now the largest instance in the wiki by two orders of magnitude, and a different shape: agents differentiated by role and data source, with debate between them as the claimed mechanism rather than coverage or a shorter tail.

Synthesis gained a section ("At 37,000 agents the unit of design changes") and one contradiction bullet — specify more vs specify less, since the spoke's spec-driven thread and Zou's "stop telling agents how to work" cannot both be the general rule. Also recorded: wrapping a legacy database in MCP leaves the interface mismatch untouched, which qualifies how this wiki has treated MCP.

WebFetch was blocked (HTTP 429); the text came from the firecrawl fallback.

Entities (6 created, 0 linked, budget not hit): james-zou, stanford-university, merck, venturebeat, ben-dickson, vb-transform. entity-index returned no match for any of them; the "Stanford University" hits were other universities matching on the common word.

Cross-spoke context noted on the source page rather than paged: the drug-discovery substance (nanobodies, CD276 ADC, FDA breakthrough designation) has no owner in this hub, the "agent school" supervised fine-tuning is ../machine-learning-wiki's subject, and the question of what an autonomously designed therapeutic must clear is ../ai-governance-wiki's.

[2026-08-07] ingest | OpenEdit (veedstudio/open-edit)

Routed from the hub (Telegram), second source of the evening. VEED's Apache-2.0 agent skill that turns a coding agent into a video editor — T1 (official project repo per the QUALITY.md ladder), promotional and unmeasured.

New pages: open-edit (source summary), veed (Organization). Updated video-use — the two are the same job answered differently, so each now points at the other — and text-first-agent-design, which gains the neighbouring move: OpenEdit has the agent author captions in HTML/CSS, choosing the notation the model knows rather than the description the platform maintains. Same instinct as paperclip, routed an hour earlier.

Synthesis gained "Two answers on one task, four days apart" under The surface dictates the representation. The two portable findings: transcription providers are abstracted behind one file while the per-word-timing precondition is a hard refusal (a mechanical gate on an input, rare in this corpus), and the vendor shape is an application company publishing the agent-native replacement for its own subscription, funnel moved to transcription.

Recorded as claimed, not shown: "up to 2.2x faster than Chrome-driven renderers," which the README itself calls preliminary and not reproducible outside VEED. Also noted: Apple Silicon + macOS Tahoe 26.0 only, and V1 scope is captions with motion graphics/charts/brandbook styling admitted to be rough.

Entities: 1 created (veed), 0 linked. Fal, Seedance 2.0 and Figma stayed plain mentions — composed services, not agents of this source.

Tier discrepancy for the next Quality Cycle: video-use carries tier: T2 for the same class of source (a project's own README). The ladder's "official project repo stays T1" rule and the curator's 2026-08-05 call both say T1. Left alone here rather than changed mid-ingest.

[2026-08-08] ingest | Harness-Bench (arXiv 2605.27922) — research pass

Via the hub Quality Cycle's research pass, against this spoke's #2 corpus-wide edge — harness spread at the frontier, written 2026-08-07 as the successor to the coding-benchmark edge. T1, and it closes the edge outright.

harness-bench: 8 model backends × 6 harnesses (+ Codex), 106 tasks, 5,194 trajectories, shared environments/budgets/protocol while preserving each harness's native execution. Frontier models are in the grid (claude-opus-4.6, gpt-5.4, gemini-3.1-pro-preview), which is exactly the cell claw-swe-bench could not test.

The finding: "Stronger model backends tend to achieve higher mean scores while exhibiting lower cross-harness variance." Harness sensitivity falls as capability rises — the trend Claw-SWE-Bench saw between two mid-tier models continues to the top. It does not reach zero: 23.8 pp separates NanoBot (76.2%) from openclaw (52.4%).

Folded into the standing contradiction, which had said "watch the harness-substitutes-for-model claim against future neutral benchmarks" — three have now arrived and agree, so the line is rewritten as a finding rather than a watch. The authors' corollary is kept: report capability at the model-harness configuration level, since a score without a named harness says little.

Worth sitting with: openclaw — the most-starred harness in this corpus and the centre of gravity for a dozen of its pages — comes last of six. Adoption and execution visibly disagree here.

Successor edge written: the benchmark scores whole harnesses, so it cannot attribute the spread to a mechanism; this spoke documents mechanisms. Needs a T1/T2 one-mechanism-at-a-time ablation on a fixed model.

Entities: 0 created — twelve authors, no recurrence, abstract-and-tables knowledge only; same call as the claw-swe-bench ingest made on Zheng et al.

[2026-08-08] ingest | DDR-Bench — one mechanism at a time, on a fixed model

Research pass, hub ## Most wanted edge #8. ddr-bench new (T1, arXiv 2602.02039, Liu/Yu/Orini/ Du/He).

The benchmark itself (Deep Data Research over databases) is not why it is here. Its Section 5.2 ablations are: reasoning budget, memory mechanisms, and reactive-vs-proactive, each varied independently on one fixed model under a minimal ReAct scaffold. That is the design the edge asked for, and it is the complement to what the spoke already had — scaffold-effects-gaia crosses models with whole scaffolds, harness-bench scores whole harnesses at scale, and neither can attribute a point of the spread to a component.

Direction reported: memory and reasoning modifications produce inconsistent results rather than clear improvements.

Recorded and worth re-reading later: the paper's stated reason for a minimal scaffold is that harness machinery is a confound when measuring a model. This wiki documents that machinery as the product. Both hold, but the tension is noted on the page rather than smoothed over.

Correction made during the ingest. A search summary attributed to this paper a framework comparison (ReAct vs Plan-and-Execute vs CoALA vs AutoGen, with complex frameworks degrading against ReAct). The paper does not contain it. Checked against the HTML before writing; that claim is not recorded here.

Not held: Tables 3–5 did not extract from PDF or HTML, so there are no per-mechanism deltas. Edge marked partly closed, still open for the numbers.

[2026-08-08] lint | DDR-Bench — correcting my own correction, with the table this time

The ddr-bench entry above records that a search summary claimed the paper compares ReAct against Plan-and-Execute, CoALA and AutoGen, that I checked the HTML, and that the paper "does not contain it."

That was wrong, and the search summary was right. The comparison is Table A1, in the appendix, which neither the HTML nor the PDF summarizer surfaced. Extracting the PDF locally with pypdf found it immediately.

The numbers, ReAct baseline against each single addition, three datasets × three models: 2 of 27 comparisons beat plain ReAct, both +Plan on 10-K. Memory never wins anywhere. Multi-agent is worst or near-worst in every row and takes MIMIC/Qwen3-4B from 16.67 to 4.44. The paper's own words: "complex agents mostly degrade performance relative to the ReAct baseline" and "memory modules often destabilise performance", explained as scaffolding disturbing model confidence into "premature self-termination or over-/under-thinking."

Page updated with the full table. Edge #2's successor is no longer "still open for the numbers" — the numbers are here, and they point away from what this spoke has assumed. What remains open is a coding workload, since DDR-Bench is database analysis.

Method note worth keeping: a web-extraction failure is not evidence of absence. Twice today a paper's real content sat in a PDF that the fetch tool rendered as compressed binary. Extract locally before recording that something is not in a source.

[2026-08-09] ingest | SWE-bench, τ-bench, WebArena — the benchmarks the spoke argued about (via research pass)

Coverage edge 7, mostly closed: the three benchmark papers, all T1 — Jimenez et al. (ICLR 2024), Yao et al. (2024), Zhou et al. (CMU, 2023). BrowserGym, also named by the edge, is still absent.

New pages: swe-bench, tau-bench, webarena.

Why this mattered more than page count. SWE-bench appeared in seven pages here as a bare number. Reading the paper puts those numbers in their place: at publication the best model resolved 1.96% of 2,294 tasks. Nearly everything scored since then was bought by the scaffolding — retrieval, edit tools, test feedback — which is the spoke's own subject. A benchmark the corpus treated as a model ranking is better read as evidence for the harness thesis.

τ-bench supplies the metric the spoke was missing. pass^k — succeed on the same task in all k independent trials — is a reliability measure, and GPT-4o's pass^8 in retail is under 25%. Single-run comparisons between harnesses cannot answer the question shipping-without-human-verification asks; this can.

WebArena covers the third surface: real self-hosted web apps, functional-correctness grading, and a 14.41% agent score against 78.24% for humans in 2023. Stale in absolute terms, and the right citation for the browser-agent claims in ../web-browsers-wiki, where vendors publish their own numbers on their own setups.

Successor edge recorded: all three papers are from 2023–24 while the scores quoted here are current, so contamination/saturation on SWE-bench and an independent re-run are what the corpus now lacks.

Entities: 0 created — Princeton and CMU deferred; no author pages on a single paper each.

[2026-08-09] ingest | Auto mode as the Claude Code default — the spoke's first permission source

Arrived via Telegram. simon-willison, 2026-08-08. T2 — independent commentary, transparent about which claims are the vendor's, wrapped around figures that are all Anthropic's.

New pages: claude-code-auto-mode (source) and agent-permission-model (mechanism). claude-code updated with the change and re-dated. simon-willison linked cross-wiki to ../llm-providers-wiki. Trajectory Labs named once and deferred, per ../ENTITIES.md.

What it establishes. Auto mode becomes the default for new sessions on Pro, Max and Team plans from 2026-08-14: the agent approves its own actions rather than stopping for confirmation. Anthropic reports it blocking 89% of harmful actions against 13.6% caught by the human reviewers it replaces (n=1,053 paid developers), and a commissioned Trajectory Labs evaluation in which none of 720 injection attempts succeeded against Fable 5, Opus 5 or Sonnet 5 under auto mode.

Why it changed the synthesis rather than adding to it. This is the first source here about what a harness forbids rather than what it enables, and the 13.6% figure indicts a design most of the corpus's tools share. Confirmation fatigue means an approval prompt degrades in proportion to how much the agent is used — so the spoke's own enthusiasm for longer, wider agent runs (run-claude-code-agents-24-hours, orchestrate-100-agents-claude-code) was quietly disarming the safety story those tools ship with. Folded in as "Permission is the dial that runs the other way."

Recorded against it. Every figure is vendor-generated, including the third-party test the vendor commissioned. A 720-of-720 result describes the attack set, not the attack surface. Willison accepts the design — "I absolutely buy that auto mode is a better solution than asking humans to constantly approve actions" — and wants "more independent confirmation of this", naming the vector the suite plausibly misses: a malicious package whose instructions read as legitimate work rather than as an attack. His standing prediction of a coding-agent security disaster in 2026 is unretracted.

[2026-08-10] ingest | MimiClaw — an agent harness in C on a $5 microcontroller

Routed here by the hub (runner-up: ../embedded-iot-wiki, which owns esp32, microcontroller and espressif — the substrate half is cross-linked there, not duplicated). Source: https://github.com/memovai/mimiclaw, T1, MIT, 5,608★/829 forks/114 open issues, repository created 2026-02-04, read 2026-08-10.

Pages: mimiclaw. openclaw gained a fourth derivative shape; synthesis gained "The substrate floor".

Why it belongs here rather than in the hardware spoke. The subject is an agent harness; the ESP32 is where it runs. It credits openclaw and Nanobot as the architectures it reimplements, and the spoke already documents that family in depth. The hardware question — what an MCU can host — is real and is the runner-up spoke's.

The finding. It ships four of the six parts of the openclaw-hermes-control-layer anatomy: runtime, gateway, memory, tools. No identity, no policy, no approval prompt. Those two omissions are exactly what nemoclaw and microsoft-scout were built to add to OpenClaw, so the anatomy's claim that they are the control layer and not the working layer now has its demonstration at the bottom of the range.

Autonomy without a gate, and why it is survivable. A 30-minute heartbeat re-prompts the agent from HEARTBEAT.md, and cron_add lets the model schedule its own wake-ups across reboots. Against claude-code-auto-mode's 13.6% figure, there is no control here to degrade. What limits the damage is that the five tools are web search, a clock and three cron calls — containment by capability poverty, a third mechanism beside agent-guardrails sandboxing and the approval prompt. It costs nothing and it ends with the sixth tool.

Correction recorded against the marketing. The README headline says "AI assistant on a $5 chip"; the repository's own GitHub description says "Harness on a $5 chip", and the second is accurate. The board asked for is 16 MB flash / 8 MB PSRAM at ~$10, and every turn is a metered API call to Anthropic or OpenAI with the key on the board (model-or-harness-taxonomy).

Flagged, not editorialised. API key, Telegram token and WiFi credentials sit in a build-time C header or NVS flash, on a board with OTA firmware update and an unauthenticated WebSocket gateway on port 18789. The docs do not discuss authenticating it.

Entities: memovai and the X handle @ssslvky are named once each with no biography evidenced; deferred per ../ENTITIES.md.

[2026-08-10] ingest | Agent SDK Workshop — the layer stack, stated as an order

Routed from the hub (Telegram). github.com/anthropics/agent-sdk-workshop — Apache 2.0, Python, ~56★, one commit. Anthropic's own hands-on course for the claude-agent-sdk. T1 as a first-party artifact; authoritative about what the SDK offers and about how its vendor thinks it should be learned, evidence of nothing about whether agents built this way work.

Dedup: claude-agent-sdk already existed and was refreshed in place with a "How Anthropic teaches it" section rather than duplicated. New page: agent-sdk-workshop.

The finding. The workshop takes one agent through four config booleans — system prompt → ENABLE_TOOLS → ENABLE_SUBAGENTS → ENABLE_MEMORY. The switches are independent, so the order is a judgment, not a constraint: capability first, coordination second, persistence last. This spoke built the same layer list bottom-up from artifacts that each argued for one layer; this is the first time it arrives as a dependency chain from the party that ships the loop. Recorded in synthesis.

The caveat, and one exception. Every stage is presented as an improvement and nothing is measured, which is the shape cc-thinking-skills, context-files-ablation-paper and jetbrains-ponytail-tested have taught the corpus to discount. The exception is stage 1: tools is the layer none of those studies ablated, and the one with a plausible mechanism for a real gain. Nobody has measured it.

Also noted: four of the six breakout scenarios (chief-of-staff, customer-support, sre-agent, account-intelligence) are operations roles rather than coding, against a corpus that is almost all coding harnesses. sre-agent is the seam with ../platform-ops-wiki, as a teaching fixture only. All tools run on mock local data.

[2026-08-10] ingest | Turing Post's 13 frameworks — the layer under the harness, counted

Routed from the hub (Telegram). turingpost.com/p/frameworks-sdks — Alyona Vert, 2026-08-09. T3: a newsletter survey, secondary, unmeasured, no inclusion criteria stated.

New pages: turingpost-13-frameworks (source), agent-framework (the layer), turing-post (publisher; entity-index returned nothing ≥ threshold, so a new node).

Dedup: adk and langchain already existed and are linked, not rewritten. The other eleven frameworks are named in the source page's table and not given stub pages — this is a census, and the awesome-osint-list precedent applies: page the census, not every row.

What it opens. This spoke's corpus is things you run — harnesses, skillpacks, meta-harnesses. This is the layer you build with, and it had never been paged as a class. Three findings went to synthesis: the layer is Python (12 of 13; Mastra alone isn't), it has begun specializing by use case (voice, RAG, simulation, enterprise), and Anthropic is absent from a list holding OpenAI, Google, Microsoft and AWS-via-Strands. That last one is recorded as an observation about the article's selection — with no stated inclusion rule there is no way to separate an omission from a judgment, and the page says so rather than guessing.

A vocabulary collision worth having caught. The article's ladder (model → API → SDK → framework) sorts by how much arrives pre-built. Anthropic's, already on claude-agent-sdk, sorts by who runs the loop and where. Both call the middle rung an SDK and they are not close.

Same hole, one layer down. Nothing measures a framework against building on the API directly. The survey grades ergonomics, community size and one star count. Up a layer, three parties ran controls and found process effects without correctness effects; down here nobody has tried.

[2026-08-10] ingest | LLM Engineer Toolkit — a second census, and it disagrees with the first

Routed from the hub (Telegram). github.com/KalyanKS-NLP/llm-engineer-toolkit — Apache-2.0, 10.7k★, 150+ libraries in 15 categories. T3: per-entry descriptions and a maintained category scheme, but no inclusion criteria, no dates, no per-row licence or maintenance signal — same grade and reason as awesome-agent-evolution.

New pages: llm-engineer-toolkit (source), kalyan-ks (thin Person, on the astrosp rationale). agent-framework and turingpost-13-frameworks both updated in place.

Why it routed here against ../machine-learning-wiki and ../llm-inference-wiki: Agents is the largest category (23), and agents + application development + prompts + structured outputs is about a third of the catalog. The training/fine-tuning (16), data (12), inference+serving (9), RAG (11) and safety (9) facets are recorded as cross-spoke context on the source page rather than split out.

The finding: the category has no agreed membership. Read against turingpost-13-frameworks from one day earlier — 27 agent libraries here against 13 there, 7 on both. The misses are patterned: Turing Post has the vendor SDKs (adk, Microsoft Agent Framework, Strands) and the non-Python entries (Mastra, LiveKit) that this list lacks; this list has a community stratum (AutoGen, Swarms, Composio, Browser Use, AgentOps, Langroid) Turing Post lacks. One census is shaped like the vendor landscape, the other like the Python OSS community, and neither says so.

Anthropic absent, twice. Checked deliberately, because yesterday's page recorded the absence from Turing Post and declined to explain it. There is no Anthropic agent library here either — "Anthropic" occurs once in 150+ entries, inside traceAI's description of what it instruments. Second data point, still not an explanation.

A correction to how this spoke states its own complaint. Evaluation is the catalog's second-largest category — 15 libraries. "Nobody evaluates agent tooling" is too loose. The instruments exist in quantity; what the corpus has never found is anyone using them to publish a result about the tools paged here. Reworded in synthesis.

2026-08-11 — ingest: Confucius Code Agent (research pass, via hub quality cycle)

confucius-code-agent (T1, Wong et al., Meta + Harvard, arXiv 2512.10398v5). Closes the coding half of the mechanism-attribution edge — the one ddr-bench left open because its ablation ran on deep-data-research tasks, not code.

Held: the four-mechanism scaffold (context management, note-taking agent, extensions, meta-agent), 54.3% Resolve@1 on full SWE-Bench-Pro with Claude 4.5 Opus, and Table 2's ablation on a 100-example subset — context management +6.6 pp on Claude 4 Sonnet (42.0 → 48.6) against +0.6 pp on Claude 4.5 Sonnet (51.0 → 51.6); tool sophistication +7.0 pp (44.0 → 51.0).

The finding is about stability, not size. A mechanism worth nearly seven points on one model is worth half a point on its successor. That reframes DDR-Bench's inconsistent results rather than contradicting it: if per-mechanism value collapses as the base model improves, an ablation on a single model will read as inconsistent from outside. Two labs, two workloads, same conclusion — harness gains do not transfer across models, so a reported gain without its model is not information.

Limits recorded on the page: 100-example subset, single runs, no variance, only two of four mechanisms varied (note-taking and meta-agent never removed), and the lab is evaluating its own scaffold. Successor edge written — an ablation with repeated runs and variance, covering persistent memory.

[2026-08-12] ingest | WebMCP — the agent→page slot, and a spec arguing against MCP

Routed here from the hub after a ../webperf-wiki re-read found lighthouse auditing pages for WebMCP conformance while no page in any of the 35 spokes said what WebMCP is. It was hub Quick win 13 for a few hours.

webmcp (T1, the W3C Web Machine Learning Community Group explainer, with Chrome's own docs as a second T1 for shipping detail). A page exposes its JavaScript functions and <form> elements as tools with descriptions and JSON schemas, and an in-browser agent calls them: document.modelContext.registerTool({name, description, inputSchema, execute}, {signal, exposedTo}). getTools(), executeTool() and a toolchange event are named but still TODO. A declarative form synthesizes tools from annotated <form>s; both exist because "some of the web's functionality is only possible with JavaScript."

The problem statement is the find. WebMCP's case against backend integrations names MCP and OpenAPI and calls the cost "UI Disintermediation & Context Loss" — the server-side approach forces a developer to "replicate the user's state, active context, and authentication credentials on a separate server." This spoke has held that objection before, but only from critics. Here it is the opening argument of a specification written inside the same ecosystem, which still "derives direct inspiration and shares a common vocabulary with MCP."

It slots into agentic-resource-discovery's stack and then dents it. The stack was MCP (agent→tool), A2A (agent→agent), Agent Skills, AGENTS.md, ARD (discovery). WebMCP is agent→page — and a capability living inside a page is one an ai-catalog.json entry cannot point at, so ARD's index has a blind spot by construction. Recorded on both pages.

Third governance shape, and the pattern is now properly broken. A2A, AGENTS.md and ARD all ran vendor → neutral foundation → default. pilot-protocol inverted that by being a company operating its own network. WebMCP has no single vendor to invert: Microsoft and Google are co-authors from the first commit (13 August 2025 — Brandon Walderman, Leo Lee, Andrew Nolan; David Bokan, Khushal Sagar, Hannah Van Opstal), in a W3C CG rather than a foundation built around the artifact. Argued in synthesis.md that the three shapes fail differently, which is what makes the distinction worth keeping.

Status recorded honestly: unsettled. 124 commits against 108 open issues, no version number, Chrome 149 origin trial plus a local flag. Neither Apple nor Mozilla appears anywhere in what was read — a gap named on the page rather than glossed, since a proposal adding an agent-facing API surface to every page on the web needs more than one engine.

Security recorded as incomplete, not adequate. Origin isolation by default, Permissions Policy allow="tools" for cross-origin iframes with NotAllowedError on refusal, an exposedTo opt-in array — and the explainer defers the rest to a spec section. A page volunteering executable, natural-language- described entry points is a new attack surface, prompt injection reaching a registered tool being the obvious one, and the source does not work that through. Stated as a limit on what this corpus can claim.

Entities: microsoft and google linked; the six named authors were not paged — authorship alone is not the evidence ../ENTITIES.md asks for. Cross-spoke facets recorded to webperf (lighthouse — the audit shipped ahead of the standard), search-marketing (llms-txt — the other bet on machine-usability), and web-browsers (a CG proposal behind one vendor's origin trial).

[2026-08-13] ingest | DeepSeek Harness — the agent loop as a config row

Routed here by the hub router from a Telegram link (github.com/deepseek-ai/deepseek-harness). T1 primary source: the repo, its architecture doc, and the package READMEs. Fetched through gh after raw.githubusercontent.com/.../main/README.md 404'd — the default branch is master.

Two pages created. deepseek-harness (source summary) and cordis (the plugin meta-framework under it, paged separately because it predates the harness by four years and is the more general claim).

Timing worth recording. The repository was created 2026-08-13 at 11:56 UTC and stood at 37,128 stars / 2,869 forks when read the same evening. That is the largest attention figure in this spoke and it is not evidence of anything; the ingest is filed as architecture. The repo's own BENCHMARK.md is one sentence pointing at the Python SDK, with no result in it.

What it adds. "Everything is a plugin" stated at full strength — the model adapter, tool registry, session log and the agent loop itself are rows in a Cordis tree, replaceable by a patch file. This is agent-middleware past the point where the framework keeps its own loop. Also: the runtime invariant "model-visible means logged", which makes model-or-harness-taxonomy's worked ambiguity (did compaction drop the instruction, or did the model ignore it?) decidable from the log; capability seams with three required roles; scope-layered skill and tool registries that implement agent-scope-isolation at the registry level.

Pages updated. agentic-coding-harness gains it as the fully decomposed end of the instance spectrum. agent-framework gains a heavyweight TypeScript entry against turingpost-13-frameworks's "TS-native is a single entry" (three days old at the time) plus a note that the harness/framework boundary blurs from this side — mount only dsh-base and it is a library. synthesis.md gains "The loop becomes a config row".

Cross-spoke context, not fragmented. The vendor angle (DeepSeek the lab, its models, its API) belongs to ../llm-providers-wiki, which already holds deepseek — linked, not duplicated. Worth noting there if it is revisited: the harness is not a lock-in play, since llm-pi-ai makes any OpenAI-compatible gateway a config entry. Runner-up spoke was ../llm-providers-wiki; the dominant substance here is harness architecture, so it routed to this spoke whole.

Entities: deepseek linked cross-wiki (../llm-providers-wiki). No individual authors paged — the repo is org-published with no named authorship, and authorship alone would not meet ../ENTITIES.md's evidence bar anyway.

Open, deliberately. Whether "replaceable from configuration" survives third-party plugins is untested; the dsh-plugin GitHub topic is where that will show. Developer preview, breaking changes promised — this page will need a refresh, and it is marked freshness: volatile.

[2026-08-13] ingest | Sheets canvas — generative UI grows a third pole, and a vendor

Routed here by the hub router from a Telegram link (blog.google, Workspace). Both spokes' boundary notes sent it here: ../ui-frameworks-wiki's scope section assigns "agents emitting UI (generative UI)" to this spoke, and agent-ready-ui states the same seam from its side. Runner-up recorded as ../ui-frameworks-wiki.

T2, justified. First-party vendor announcement — authoritative on what shipped, silent on everything else. No model named, no artifact format, no limits, no failure cases, no numbers. Rare to ingest a source this thin; it earns the slot by being the first shipped consumer product in a corner that until today held two repos.

Two pages created. sheets-canvas (source summary) and generative-ui — the corner is now paged as a concept, which synthesis.md had explicitly flagged as a spin-out candidate "if generative-UI keeps growing." Three entries is when a design axis becomes visible rather than asserted.

What it adds to the axis. The corner's question is what does the model actually emit. arrow-js emits real TypeScript and bounds its reach with a WASM sandbox; openui emits a restricted DSL and bounds its vocabulary; canvas emits something Google declines to name, inside Google's own app, where containment is moot because nothing is exposed. Enclosure instead of containment — and no export, inspection or diff. "No coding required" has an unstated converse: no code is offered either.

The property that is actually new is binding, not generation. Canvas is "a dynamic, read-write layer that sits directly on top of your spreadsheet data," always in sync and shared like any sheet. The other two render output for a viewer; this one is a durable multi-user surface over a live datastore. Generated UI becomes an application with state — and inherits concurrent edits and schema drift, none of which the announcement addresses. Recorded as an open question, not as a criticism the source answered.

Updated: synthesis.md (third pole folded into the rendered-output passage), index.md, and cross-wiki agent-ready-ui in ../ui-frameworks-wiki — one paragraph noting the boundary is sharpened by a product that removes the toolkit from view. No page created there; the seam is a link.

Entities: google and gemini linked cross-wiki to ../llm-providers-wiki. No new entity pages — the post carries no byline.

Evidence hole, unchanged. Nobody has measured generated UI against a hand-built one on correctness, usability, or re-prompting rate. Two of the three entries are now consumer products, which makes the absence louder rather than quieter.

[2026-08-14] lint | Quality Cycle — tier gap fixed

cordis (added 2026-08-13 with the DeepSeek Harness route) carried no tier — the ladder's own rule ("an official project repo... stays T1, whoever links to it") wasn't stated on the page itself. Added tier: T1 frontmatter and a one-line justification. No other integrity issues found: all four new pages from the DeepSeek Harness and Sheets canvas routes are linked (3–5 inbound refs each), no orphans, no dupes. Zero new pages this pass.

[2026-08-14] ingest | Sayna (SaynaAI)

Arrived via Telegram, routed by the hub (runner-up: ../speech-audio-wiki, which owns the STT/TTS providers Sayna wraps but not the aggregation layer itself). sayna: Rust unified voice-layer server for agentic frameworks — one WebSocket/REST API over Deepgram/ElevenLabs/Google/Azure STT+TTS, LiveKit rooms, Silero-VAD turn detection (T3, ~303★). Cross-linked against claude-to-speech (the corner's earlier, single-provider entry) and synthesis.md's cross-spoke-adjacency section updated with a new ../speech-audio-wiki line. index.md updated; no contradictions. No entity paged for SaynaAI (org name only, no further detail in the fetched metadata).

[2026-08-14] ingest | ECC (affaan-m)

Arrived via Telegram, routed by the hub (no runner-up — the domain match is unambiguous). ecc: "the agent harness operating system" — 68 agents, 284 skills, AgentShield security scanner (1,282 tests/102 rules), Memory Vault, installed separately across 13 harnesses (Claude Code, Codex, Cursor, and more). Single maintainer, MIT, created 2026-01-18. GitHub API confirms 240,127★ / 36,445 forks — within 1,000 stars of superpowers's already-flagged ≈242k, a second instance of the same anomaly in this spoke rather than an isolated one. Added to agentic-coding-harness's instance list; synthesis.md's existing 242k/67.1k adoption-vs-measurement thread extended with the new convergence. index.md updated. No entity paged (single-maintainer repo; affaan-m not otherwise documented in the README).

[2026-08-14] lint | ECC/superpowers star-count correction

Follow-up to the same-day ECC ingest. Refreshed superpowers against GitHub's live API while checking whether other repos share its flagged star count: superpowers has grown to 272,113 stars (from the ≈242k recorded 2026-06), not the ≈242k the ECC page had compared it to. That made the "ECC and superpowers converge on the same number" claim wrong — it compared ECC's live figure to superpowers' two-month-old snapshot. Corrected on both pages, agentic-coding-harness, index.md, and synthesis.md: the two star counts are ~32k apart, not near-identical. What survives the correction — both remain individually extraordinary counts for single/small-maintainer packs this young — is kept; the false-coincidence framing is not. Swept */wiki/*.md + */index.md + */synthesis.md hub-wide for any star count above 90k as a sanity check: the only other entries in that range (openclaw ~380k/272k-adjacent-but-unrelated corner aside, coding-interview-university ~352k, developer-roadmap ~363.7k, Next.js ~133.9k, gstack 126.3k) are long-established, widely-known projects years old — a different category, none close to 240k. No third repo in this spoke's corpus sits near the ECC/superpowers range.

[2026-08-14] ingest | byLLM + Meaning-Typed Programming paper

Two sources arrived via Telegram in the same batch: byllm.jaseci.org (WebFetch returned an empty JS shell; fell back to firecrawl per HUB.md edge handling) and the ACM paper it implements (dl.acm.org/doi/10.1145/3763092 — the PDF reader route 403'd, the abstract/body route via firecrawl worked). Created byllm (SoftwareSourceCode, T1): the by operator, the shipped module inside jaseci's jac repo (github.com/jaseci-labs/jac, 686★, not the 240k-star kind of number), model-agnostic across OpenAI/Anthropic/Google/self-hosted. Created meaning-typed-programming (ScholarlyArticle, T1, PACMPL OOPSLA2 2025, Dantanarayana et al.): the paper's own benchmark suite against DSPy and LMQL — 4.0×/5.3× less code modified, N=20 blind user study (3.2× faster, 45% fewer LOC, LMQL zero success on the hardest task), accuracy 90.85%/98.23%/98.92% for LMQL/DSPy/MTP, token/cost savings, and an identifier-degradation robustness test. Updated jaseci's MTP paragraph to point at both and folded the paper into synthesis.md's standing "adoption vs. measurement uncorrelated" thread as the corpus's counter-case: a framework claim this corner actually measured. index.md updated (SoftwareSourceCode + ScholarlyArticle sections). No entity paged for the paper's nine academic co-authors, consistent with how swe-bench/tau-bench/webarena handle academic author lists elsewhere in this spoke.

[2026-08-15] ingest | Agent Skills Can Be Harmful (arXiv 2608.11888)

Routed from the hub (Telegram, PDF URL). Downloaded and extracted locally with pypdf — 12 pages, full text read. Created agent-skills-can-be-harmful (ScholarlyArticle, T1): Dong, Gao, Li, Xu, Hua, Yang — microsoft Research + Huazhong University of Science and Technology + UIUC, 12 Aug 2026. Differential-testing design: paired target/reference runs with task, verifier, repository state, harness and model held fixed and only the skill setup varied; the skill pool widened from smithery.ai/skillsmp.com by embedding metadata with all-MiniLM-L6-v2 (top-5, cosine ≥0.7), taking the comparison space from 826 to 20,664 potential pairs. OpenCode 1.15.1 + Claude Opus 4.6. 665 labelled candidates refined to 307 confirmed cases — 125 functional failures, 182 efficiency regressions, each hand-labelled by consensus.

Created three Thing pages. skill-induced-failure (DefinedTerm) holds both taxonomies and the result that reframes the spoke's skills thread: Applicability Mismatch — the wrong skill loaded — is 2 of 125 functional failures (1.6%), while Task-Implementation Fault is 86 (68.8%), i.e. on-topic skills whose examples/defaults/templates get followed as task requirements, plus 37 cases where the skill moved a verifier-visible surface (artifact path 24, environment state 13). On cost: Excessive Procedure 114/182 (62.6%) against Context Bloat's 46, and within Context Bloat 43 of 46 are the mandatory body vs 3 for lazy-loaded material. skilltriage (SoftwareApplication) records the attribution tool's measured agreement (88.8%/72.5% exact subcategory; 93.6%/79.7% category) and the limit that it triages only pairs already confirmed. skillsbench and swe-skills-bench (Dataset) page the two benchmark subjects — both at second hand, from this paper's summary; the primaries (arXiv 2602.12670, 2603.15401) are unread and now named as what growth edge #1 needs.

Updated agent-evals (a pass rate scores 182 cost regressions green and averages the functional failures away; the paired design as the constructive alternative), agent-failure-localization (localization by re-execution rather than trace labelling — buys certainty, costs a second run, reaches only toggleable configuration), and agentskills-spec (progressive disclosure prices text and not the procedure the text prescribes; the 43-vs-3 split says authors put the wrong material above the L2/L3 line; the spec has no way to mark a requirement apart from an example, which is the mechanism behind 86 of 125 failures).

synthesis.md extended in four places, per the standing no-new-dated-sections rule: the skills layer of the thesis (the packs all ship the exact material that causes the dominant failure mode and none marks what is binding), Progressive disclosure has a ceiling (the catalog found the aggregate ceiling; this finds the per-skill one, and it is made of steps not tokens), Verification is where the bottleneck went (excessive verification is the single largest cost bucket at 67/182 — the corpus's own verification-loop discipline, priced), and growth edge #1 (the two utility benchmarks exist at the scale the edge specified; edge stays open until the primaries are read). No entity paged for the six academic authors, consistent with swe-bench/tau-bench/webarena and the 2026-08-14 MTP precedent; microsoft linked as the affiliation rather than duplicated as "Microsoft Research". index.md updated, including a new ## Dataset (benchmarks) section.

[2026-08-16] ingest | Stop burning tokens on code review (Swizec Teller) — routed from Telegram

Arrived mid-quality-cycle via Telegram; routed here by the hub over ../dev-tooling-wiki, which owns linters as build infrastructure. The dominant substance is not the linter, it is where the model sits in a loop: used once to compile prose guidelines into rules, then removed from the per-commit path that agent-authored code goes through.

Source page stop-burning-tokens-on-code-review (T2 — a practitioner's measured figures from an operation he runs, per QUALITY.md's "someone tested, measured or used the thing" rule): 2,900 PRs last quarter, ~10 reviews/day, ~$1,000/week for Cursor's BugBot, $1,000 in one day for Claude review skills in GitHub Actions, 27-second full linter sweep against a 10–30 minute model review.

New concept deterministic-review-gate — the mechanism the post argues for, placed against the spoke's existing verification thread: agent-loops-verification wants better environments to verify in, shipping-without-human-verification wants the human out, this wants the model out of the hot path and keeps the human. constraint-evading-behavior is carried on the page as the standing caveat, since a compiled rule is easier to game than a reviewer's reading of one.

Entity swizec-teller (Person). claude-fable-5 cross-linked — it is the model that compiled the rules.

What the source cannot support, recorded on both pages: no A/B, no defect-detection rate for either approach, and no statement of what the linters miss. Cost and latency are measured; quality is asserted. synthesis gains Review gets a price tag.

[2026-08-16] lint | Freshness pass — three repo pages re-verified, one failed

The 60-day volatile window came due on three pages from 2026-06-16; all three checked against the GitHub API.

  • zouroboros — re-verification failed. The repo at its URL was created 2026-08-14, two months after this page was written from it, and carries 0 stars, 0 forks, no tags. The v2.1.0 / ~19★ / 757-test artifact the page describes is gone from that address. Design description kept, all quantities marked unverifiable, volatile retained deliberately — the subject is a live artifact and the next check belongs at the same address.
  • autonovel — 1,487★ / 286 forks, last pushed 2026-03-20. Not archived, not maintained. The page called it the most concrete deployment evidence in the Hermes orbit; it is now a static one.
  • awesome-hermes-usecases — 188★, pushed 2026-08-11, actively curated. Its primary-source rule is still being applied, which is the only reason this spoke holds it.

Corpus freshness queue 4 → 0 for the batch.

[2026-08-17] ingest | ai-marketing-claude (AI Marketing Suite for Claude Code)

Arrived via Telegram, routed by the hub (runner-up: ../search-marketing-wiki, declined on the standing precedent its own registry note sets — marketing/SEO skill packs are tools and file here; the marketing strategy files there). The third domain instance of agent-skills in this vertical after agentic-seo-skill and claude-skills-ppc.

Read the tree, not the README. 37 files, 132 KB: skills/ holds 14 SKILL.md files and the advertised 15th is market/SKILL.md, the dispatcher; 5 subagents; 4 Python scripts; 6 templates; install.sh/uninstall.sh. So the pack is almost entirely prompt text. agentic-seo-skill, the same form factor one vertical over, ships 89 scripts it calls evidence collectors. That contrast is the finding, and it lands on the verification thread: this pack's deliverable is a scored, client-facing PDF resting on no checkable observation.

The metadata is the other finding. Created 2026-03-02 01:09 UTC, last pushed 01:18 — nine minutes — and untouched in the five and a half months since, holding 2,451★ / 692 forks / 7 open issues / 35 watchers. A 28% fork-to-star ratio is copy-taking, not dependence. agentic-awesome-skills established that stars measure attention; it had 2,464 commits under the number. This has one push. The corpus now has the clean case.

Also recorded: the documented install is curl … | bash from an unmaintained repo, installing instructions an agent then executes with the user's tools — two trust decisions in one command, in a spoke that holds agent-skills-can-be-harmful. No bad intent implied; the mechanism just leaves no moment to look.

Tiered T1 as the primary artifact, with the page stating plainly that this is provenance and not quality. Nothing has been run: no output sample, no eval, no third-party review, no verifier of the kind skillsbench defines, and no published rubric for the six-dimension score it advertises. synthesis.md and index.md updated.

[2026-08-17] ingest | Data Formulator (microsoft/data-formulator)

Arrived via Telegram, routed by the hub (runner-up: ../analytical-databases-wiki). Routed here on the generative-ui precedent — an agent that emits an interface files in this spoke, the toolkit it targets belongs to ../ui-frameworks-wiki — and by subject rather than mechanism: DuckDB underneath is an implementation detail, not the subject. Metadata read from the GitHub API: 16,279★ / 1,543 forks / 93 open issues / 104 watchers, MIT, created 2024-06-07, pushed 2026-08-15, v0.8 beta.

data-formulator is the corpus's fourth generative-ui entry and answers that page's central question — what does the model actually emit — with the narrowest answer yet: Flint chart specs across 30+ types. Not framework code bounded at runtime (arrow-js), not a component DSL (openui), not an undisclosed artifact (sheets-canvas) — a chart algebra, in which the model cannot invent an interaction at all.

The finding is the comparison. sheets-canvas shipped four days earlier with the same product idea (an agent over tabular data emitting an interactive artifact) and is opposite on every disclosure question: Google names no model, never says what the artifact is, and runs it inside Sheets; Microsoft publishes the artifact language, lets the user pick the model through LiteLLM, ships MIT source, and runs on the user's machine. The generative-ui page's enclosure complaint stops being an inference from one vendor's silence — the same capability exists in the open, so the silence is a choice rather than a property of the form. Folded into that page along with a third position on the output-vs-application axis: Data Formulator binds to live data but is a read path, so it inherits schema drift and not concurrent edits — the first evidence here that the two hazards separate.

Second idea worth keeping: data threads. Work branches rather than accumulating as a transcript, which is a claim about the shape of the interaction rather than about the model, and the first non-linear unit of work in this corpus. Unmeasured — no study, no task set, no chat baseline.

Recorded weaknesses: nothing run, no accuracy figure, no benchmark, no user study, no paper identified behind a Microsoft Research artifact, and no statement of how often a generated chart misreads its data — which is the gap that matters most here, since a wrong chart is more persuasive than wrong prose. T1 as the primary artifact, provenance not quality, stated on the page as it was for ai-marketing-claude earlier the same day. Also noted: unsigned, un-notarized desktop preview builds — the second unsigned distribution channel the hub has recorded in two weeks.

[2026-08-20] ingest | NanoClaw (github.com/nanocoai/nanoclaw)

Routed from the hub (Telegram). T1 — the project's own repository, read from the raw README plus the GitHub API rather than through a summary. MIT, TypeScript, 30,570★ / 12,841 forks / 926 open issues, created 2026-01-31, pushed the day it was read.

Why it matters beyond being another clone. mimiclaw ported this architecture to an ESP32 and credited OpenClaw for it. NanoClaw rewrites OpenClaw as an argument against it: its author would not give "complex software I didn't understand full access to my life," and charges that OpenClaw's security is application-level (allowlists, pairing codes) rather than OS-level, in one Node process with shared memory. The remedy is a Linux container per agent group, mounts you name, and credentials the agent never holds (OneCLI Agent Vault injects them at request time).

That is NVIDIA's OpenShell argument from the opposite direction. The spoke had recorded external containment as the thing platform vendors bolted onto OpenClaw because the project would not do it; this is the case where containment is the entire reason a competing open-source implementation exists, from an independent developer with no runtime to sell. Recorded on agent-scope-isolation's thread and folded into openclaw.

The finding worth keeping. The codebase is deliberately small for a model's comprehension budget, not a human's — "just ask Claude Code to walk you through it" — configuration is rejected in favour of editing code, and the intended workflow is to fork and have Claude Code modify the fork. Legibility-to-an-LLM as a design constraint on a source tree is new in this spoke. Distribution follows the same logic: trunk is registry and infrastructure only, channels and providers live on separate branches and are installed per-fork by /add-<name> skills, and the contribution rule is "Don't add features. Add skills." — the agent-skills thesis applied to the harness's own repository.

Measured, not asserted: forks against stars, read the same day — OpenClaw 81,273/386,899 = 0.21, NanoClaw 12,841/30,570 = 0.42, mimiclaw 889/5,710 = 0.16. A project that tells every user to fork it should show exactly that signature. Held loosely: forks also come from drive-by contributors, and nothing here separates a fork that runs from one that never did.

Flagged, not repeated as fact: the "nearly half a million lines of code, 53 config files, 70+ dependencies" charge is a competitor's count of a rival and nobody in this corpus has verified it. Nor is the isolation claim evaluated — no threat model against a hostile page or a malicious skill, and ClawHub's 341 malicious skills are the standing evidence that skill supply is where this class of system is actually attacked. Docker on macOS runs in a VM and on Linux does not; the README does not distinguish the two boundaries.

[2026-08-20] ingest | DSPy + awesome-prompts (hub route)

Two sources from the hub in one pass, arriving within two minutes of each other and turning out to be the same subject seen from two angles.

dspy (stanfordnlp/dspy + dspy.ai, both read 2026-08-20) — T1 for the artifact's own facts, with the adoption paragraph flagged T3. 37,456★ / 3,241 forks, MIT, 406 contributors, v3.3.0 (2026-08-03). Signatures, modules, and thirteen optimizers. The embarrassing part of this ingest is that the spoke had cited DSPy fourteen times across five pages — meaning-typed-programming's benchmark rival, half of hermes-agent's "GEPA + DSPy" loop — without ever paging it. A dangling concept, found only because a source arrived that made it the subject.

awesome-prompts (ai-boost, 8,727★ / 840 forks, GPL-3.0) — T3: real curation, no stated inclusion criteria, nothing measured. Read for volume rather than for its thesis, since the two disagree: it argues prompts-as-engineering and spends 43% of the README on copy-paste templates.

Two new Thing pages. prompt-optimization — the mechanism both sources are about, five families sorted by what the search is allowed to move. omar-khattab — MIT EECS/CSAIL, the DSP→DSPy→MIPRO→GEPA line, entity discovery off the DSPy ingest (stanford-university already existed and was reused).

Three findings folded into synthesis.

  1. Prompt optimization is the second mechanism for self-improving-agents, which had exactly one (an agent authoring its own skill files) — and the first with peer-reviewed algorithms behind it.
  2. It is the one corner that cannot skip evaluation, since an optimizer without a metric does not run — and that buys less than it looks like. The gain is measured against the optimizer's own seed prompt, never against a human writing the prompt by hand. New open question, and the sharpest form of the spoke's standing benchmark gap: the harness exists, the baseline arm does not.
  3. Where the agent-framework boundary actually runs. DSPy is in neither census — Turing Post omits it, llm-engineer-toolkit files it under LLM Prompts beside LLMLingua and the prompt compressors. Both censuses sort by what you assemble; DSPy sorts by what you optimize, and a compile step has no slot in a taxonomy of assemblies. Third data point on contested membership, and the first that explains the mechanism rather than the vantage point.

And one observation the directories are better at than the papers. awesome-prompts' Coding & Development prompts are distillations of this spoke's own corpus, each citing its origin artifact: Anthropic's auto-mode write-up as a permission classifier, harness guidance as an "Agent Harness Designer", three separate encodings of loop-engineering, skill authors for Codex and Roo Code. The system prompt is acting as the transport layer for agent-tooling technique — and the same repo catalogues six collections of leaked production prompts as curriculum, which is claude-fable-system-prompt-leak's material grown into a genre.

Flagged, not repeated as fact: DSPy's 6.6M monthly downloads and Shopify's "~550× cost reduction" (vendor-relayed, no method or baseline); awesome-prompts' claim that promptfoo was acquired by OpenAI in March 2026; and the long tail of 2026 star counts it relays for repos nobody here has checked. Khattab's own "first and most widely used declarative programming model" is a priority claim this corpus has not tested against LMQL.

[2026-08-21] ingest | agtx (hub route)

agtx (fynnfluegge/agtx, read 2026-08-21) — T1 for the artifact's own facts, benchmark numbers flagged T3. 1,251★ / 111 forks, Apache-2.0, Rust, v0.2.6 (2026-06-28). A kanban TUI that runs several coding agents at once, git worktree and tmux window per task, ten swappable workflow plugins, and an experimental orchestrator agent that drives the board over MCP.

Dedup went cross-wiki. The author was already paged: fynn-fluegge in ../research-wiki, discovered through rocketnotes and left a thin node in August. Updated in place rather than duplicated here, and thickened with what the repository list evidences — agtx is his second-most-starred project, the LLM work starts in 2023 with codeqai and doc-comments-ai, and a 2017–2019 3D-engine line stops. No profession asserted; none is stated anywhere.

The find is in benchmark/, not the README. agtx ships a SWE-bench Lite runner whose TOML config names the workflow plugin, so superpowers, spec-kit, get-shit-done, OpenSpec and skill stacks including ponytail become swappable arms over a fixed task, model and sandbox — duration, tokens, dollar cost and a three-way outcome grade per run. Eight configs published with results.

The results barely support anything: n=1 per arm on one instance of 300, one model, no repeats, the tool's author running it, and every config sets skip_worktree = true, which switches off the parallel-worktree feature agtx exists for. What the table measures is the plugin, with agtx as driver. Recorded as an instrument, not a result.

What it does show is worth carrying: seven of eight arms produced the correct fix and wrote no test (🟡); the one ✅ was superpowers at $2.64 vs the cheapest partial pass's $0.40, 15m46s vs 4m48s. A cost effect with a correctness effect attached is exactly what cc-thinking-skills, context-files-ablation-paper and reasoning-scaffolds all looked for and did not find. One task is not evidence; it is a reason to run the rig properly. Folded into agent-evals and synthesis, and it narrows the llm-engineer-toolkit complaint — somebody published a result, nobody has gone past n=1.

Two smaller findings. Orchestration here is heterogeneous — Gemini research → Claude implement → Codex review, with automatic switching, rather than one model fanning out to copies of itself; a shape agent-orchestration had not carried. And the stuck-task detector (idle 1 minute, no phase artifact) diagnoses by calling read_pane_content on the agent's tmux pane, then nudges or escalates to the human as an MCP tool call — text-first-agent-design applied to supervision, and agent-guardrails' escalation argument as an API instead of a crash.

Conflict recorded, not resolved: spec-driven-ai-tools (T4) scored OpenSpec highest of the tools it hand-tested; here OpenSpec is the leanest arm by tokens and lands 🟡, correct fix with no test. Different criteria, so not strictly a contradiction — but cheap, and does not finish the job survives both readings.

Errata carried onto the page: agtx's plugin table links gsd to github.com/fynnfluegge/get-shit-done-cc, which 404s and exists nowhere on that account (checked 2026-08-21). The GSD this spoke pages is gsd-build/get-shit-done by TÂCHES, 64.7k★, last pushed 2026-05-31. Stale link or deleted repository is not evidenced either way, so both facts are recorded and neither inferred.

Also updated: agent-kanban (the sibling comparison — both converged on worktree-per-worker and on MCP for the orchestrating agent independently; they diverge on where the human sits).

[2026-08-21] quality | cycle — orphans cleared, research pass closes the variance edge

Orphans 6 → 0. The cluster the hub's ## Outstanding flagged on 2026-08-18 as needing "a pass of its own". Each linked from a page where the connection carries an argument: genkit into agent-framework's membership problem (Google is in the censuses through adk and its library rung is in neither), sayna into the specialization axis (a voice layer you mount rather than a voice framework), confucius-code-agent into the measurement hole as the nearest anyone has come, grok-build into awesome-prompts, webarena into agent-evals as the environment-shaped alternative to a patch-shaped suite, hitchhikers-guide-agentic-ai into agentic-tooling.

Research pass — beyond-pass-at-1 ingested (T1), closing the confucius-code-agent successor edge. 396 tasks × 4 duration buckets × 3 domains, 10 open-weight models, two scaffolds, k=3 repeats, 23,392 episodes, 95% CIs. The corpus's first agent study that repeats runs and reports the spread.

Three things it changes here. It names the standard suites' shared blindness in a table — swe-bench, webarena and tau-bench all score one attempt. It gives the spoke a fourth negative result on added structure, and the first on memory: long-horizon GDS falls for 5 of 9 models, worst on the strongest. And it prices the whole thing at ≈$0.004 per episode, which removes budget as the reason nobody measures variance.

Read the PDF locally rather than trusting the fetch. WebFetch returned wrong expansions for two of the four metric acronyms (MOP as "Mean Observed Performance"; it is Meltdown Onset Point) and reported the repetition count as unstated when the abstract states it plainly. pypdf extracted all 23 pages cleanly. Same lesson as 2026-08-08, and it would have put two errors on the page.

Contradiction flagged, not resolved: the paper's abstract says memory scaffolds "universally hurt … all 10 models"; its body says 5 of 9 with one positive. Both on the page, body cited.

Also this cycle: three of the spoke's pages (swe-bench, tau-bench, webarena) were among 33 corpus-wide carrying source: true with no tier: in frontmatter — graded in prose, invisible to every measurement. Lifted as authored, all three T1. Detail in the hub quality-log.md.

[2026-08-21] ingest | Anthropic's Browser Use toolset (The New Stack)

Routed from the hub. anthropic-browser-use-tool — amanda-caswell for the-new-stack, 2026-08-21, T2 — on the Browser Use toolset anthropic announced the day before, next to general availability for Computer Use, the Skills API and the Files API.

New pages: claude-browser-use (WebAPI), anthropic-browser-use-tool (source), amanda-caswell (Person, thin node — the one new entity this ingest discovered; the-new-stack and anthropic already exist cross-wiki).

The finding is in the title. Anthropic ships a protocol and 27 tool definitions and executes none of them. The developer opens the browser, runs each action, keeps the session alive and returns the result. Claude reads the page as an accessibility tree and acts on element references (ref_3) instead of screen coordinates.

Homonym handled, not merged. The spoke already had browser-use — the GitHub org behind the open-source agent framework that drives Chromium over CDP. No relation to Anthropic's tool; the article says so itself. Both pages now carry a disambiguation note naming the other, per the hub's homonym rule.

Three pages absorbed the substance. text-first-agent-design gains a fifth row and its first vendor instance — plus two things the open-source instances never stated: a fixed entry cost (~6,600 input tokens for the default toolset) and a named failure mode (references go stale when the page moves, and the API does not detect it). agent-permission-model gains the batching problem — several actions per model turn means the approval check moves from the protocol boundary into the developer's executor loop, mid-sequence. synthesis files it as the vendor layer's sixth shape, the inverse of gemini-managed-agents-hooks: sell the tool vocabulary, keep none of the runtime.

Weakness recorded. T2 trade reporting, no independent testing, every figure restated from Anthropic's own documentation, and the outlet discloses that its owner invests in Anthropic. No first-party document ingested. The missing number is a Browser-Use-vs-Computer-Use cost or accuracy comparison — the thing that would show whether the tree pays for itself the way mobilerun claims it does on a phone.

[2026-08-21] ingest | The framework layer, measured for the first time — via research pass

agent-framework has carried the same complaint since it was written: the layer is crowded, Python-shaped and specializing, and nobody measures it. Both censuses it rests on grade on ergonomics and community size. agent-storage-footprint is the first source here that puts a number on the layer, and it does it by measuring something nobody had proposed — what each framework leaves on disk when the work is held identical.

The control is what makes it a comparison rather than another census. One scripted trajectory (three reads, 182.7 KB of observations, one answer) replayed through each framework's own persistence adapter: 6.7× spread. Drop the scripting and compare only configurations that all reach 100% on the same suite: 15.7×, LangGraph 5.10 MB against CrewAI 0.32 MB. Same answers, sixteen times the residue.

Three findings worth keeping. Full-history designs grow superlinearly (α up to 1.95 — a 200-round monitoring loop over a single 2 KB status file leaves 323 MB). On a minimal write task, framework residue exceeds the delivered output by up to 24,033×. And a content-addressed store removes 4.8–32.7× of it with every restoration test passing and every reconstructability score reproduced — so this is duplication nobody had reason to look at, not eight teams pricing recovery differently.

The finding that generalizes is the one measured outside the authors' own harness: across 108 instance-normalized swe-bench Verified submissions, per-instance trajectory volume spans three orders of magnitude with no detectable correlation with resolve rate. Whatever storage buys, the benchmark this spoke's evaluation thread is built on cannot see it.

Tier T1, and the provenance check mattered here. InfiAgent is in the evaluated roster and scores well (D=1.8, α=1.15), so the obvious question was whether the authors ship it. They do not — Hong Kong Polytechnic, Zhejiang Normal, and a supply-chain company, none of them shipping any of the eight.

Not closed, and the edge says so. Accuracy in this study is a control, pinned so storage can vary, so it compares frameworks to each other and never to writing the loop yourself — which is the thing agent-framework actually wants to know. Successor edges 9 and 10 written: a comparison where correctness or cost is the dependent variable, and a re-run once these version-pinned results decay.

One methodological note recorded on the page for its own sake: naïve byte-level measurement understates duplication by an order of magnitude, because database paging and JSON escaping break up repeated content past what a byte-level differ can see. A metric that silently under-reports because its subject is stored in a form the instrument cannot parse is the same failure the hub found in its own tier counts the same day.

[2026-08-22] ingest | The SWE-Bench Illusion (Liang, Garg & Zilouchian Moghaddam) — via research pass

swe-bench-illusion (arXiv 2506.12286, T1). Closes the contamination half of growth edge 7, which has been open since swe-bench arrived and the spoke started quoting 2026 scores against a 2023 task set.

The instrument is the contribution. Instead of re-running agents, ask models two things a real solver could not do: identify the buggy file from the issue description alone (no repository access), and reproduce the ground-truth function from the issue plus current file context.

Result across ten GPT and Claude models: 60–76% file-path accuracy on SWE-Bench Verified (500 instances, 12 repos averaging 763.5 files), 57–71% on Full SWE-Bench, and below 53% on repositories outside the benchmark. Verbatim recall agrees — up to 35% consecutive 5-gram accuracy on Verified against at most 18% elsewhere. Claude Opus: 92.11% when the path is mentioned in the issue text, 67.41% when it is not.

Kept honest about scope. This does not make the harness argument wrong — knowing where a bug lives is not fixing it, and the ~2%-to-70% arc that argument rests on is unchanged in shape. It makes small differences between reported resolve rates uninterpretable. Noted on swe-bench and in synthesis, next to agent-storage-footprint's independent finding that trajectory volume spanning three orders of magnitude does not correlate with resolve rate.

Authors' own confounder recorded: RefactorBench's issue descriptions are 14.6 tokens against Verified's 451.2, so the pandas/PyTorch comparison is the cleaner half of the evidence. The independent-re-run half of edge 7 stays open — needs a decontaminated task set with harness results on it.

[2026-08-22] quality | Cross-spoke: a second harness-ablation result on memory

No new page here. vibesearchbench (arXiv:2605.27882, T1, ingested in ../osint-wiki by the hub research pass) ablates a production harness's sub-agent collaboration, local memory and life-long memory on 200 multi-turn search tasks and finds no significant improvement from any. Recorded in synthesis.md beside beyond-pass-at-1: different task, different metric, different harness, same direction. The naive-versus-curated-memory successor edge is unaffected and stays open.

[2026-08-22] ingest | Hermes Agent on Android / Termux (routed from the hub)

hermes-agent-termux — Nous Research's own Termux installation guide (hermes-agent.nousresearch.com/docs/getting-started/termux), arriving by Telegram. T1: a first-party statement of the artifact's platform support, dependency set and self-declared limitations, all checkable against the software. The single non-artifact claim — that a third party's package repo is unaudited — is a disclaimer against interest.

Why it is worth a page rather than a line on hermes-agent. It is the first source in this wiki where a framework enumerates the features it cannot ship on a platform it supports. Termux is labelled Tier 2, with the support level written down: "maintained on a best-effort basis only. Commits to main may break these packages at any point in time."

Surviving on Android: CLI, cron, PTY/background terminals, the Telegram gateway, MCP, Honcho memory, ACP. Absent: voice transcription — blocked because ctranslate2 publishes no Android wheels, so a missing binary artifact three levels down the dependency graph removes a modality — automatic browser/Playwright bootstrap, the .[all] bundle, and Docker-based terminal isolation. That last one matters: hermes-agent lists "command approval, container isolation" as its security posture, and on this platform only the first survives. Plus an OS-level limit: Android suspends background jobs, so "gateway persistence is best-effort rather than a normal managed service."

Two smaller things kept. The pkg install line is mostly a compiler toolchain (clang, rust, make, pkg-config, libffi, openssl) because the wheels do not exist for this architecture — and two entries, ripgrep and ffmpeg, are the agent's tools, so the harness anatomy openclaw-hermes-control-layer describes is visible in an apt line. And the docs publish a community curl | bash APT repo while denying all four of build, sign, host and audit — a documented instance of the convenient install path leaving the vendor's trust boundary, noted for ../defensive-security-wiki rather than paged here.

It also tests the spoke's own framing. openclaw-hermes-control-layer puts Hermes at the infrastructure-you-own pole against openclaw's gateway breadth. A phone is that pole taken to its limit, and what survives is exactly the memory-and-CLI core while the breadth falls away — which is better evidence for the positioning than a README.

hermes-agent updated with a section rather than a rewrite; its standing caveat ("the feature list is README claims") now has a worked example from the vendor. Dedup: no page in any spoke covered Hermes on Android; termux is paged in ../osint-wiki and linked cross-wiki, not duplicated.

[2026-08-22] ingest | Agents that remember: introducing Agent Memory (hub-routed, Telegram, URL-only, T1)

Cloudflare blog, tyson-trautmann + rob-sutter, published 2026-04-17 (Agents Week 2026), revised 2026-07-15. Created source summary cloudflare-agent-memory (WebAPI, url-only, T1 — first-party engineering post with the pipeline described stage by stage).

What it adds. A fourth pole for agent-memory, and the first that is not self-hosted: a managed service on Durable Objects (SQLite per profile) + Vectorize + Workers AI, reached by Worker binding or REST through five verbs. Detail worth keeping — content-addressed message IDs make re-ingest idempotent; an eight-check verifier passes/corrects/drops each extracted memory; classification into facts/events/instructions/tasks with tasks kept out of the vector index; supersession by normalized topic key leaving a version chain; five parallel retrieval channels (full-text, exact fact-key, raw-message safety net, direct vector, HyDE) fused by RRF; date arithmetic computed deterministically and injected, never asked of the model. Model selection is a structure-over-model data point: Llama 4 Scout (17B MoE) for extraction/verification/classification/query-analysis, Nemotron 3 (120B MoE) for synthesis only — "the synthesizer is the only stage where throwing more parameters at the problem consistently helped."

The weakness, recorded. The post names LoCoMo, LongMemEval and BEAM and reports no score against any of them. T1 for architecture, unmeasured on outcome — noted on the page and in synthesis.

Updated agent-memory (new "managed pole" section + hosted/self-hosted axis), synthesis §6, and index.md. Entities: tyson-trautmann and rob-sutter created (thin, byline-only); cloudflare linked cross-wiki to ../cloud-wiki, not duplicated — entity-index reported "create new" because that page is typed WebPage/source, not Organization. "Agents Week 2026" left as prose, not an Event node (single source, tangential).

[2026-08-22] ingest | Claude Managed Agents overview + Dreams (hub-routed, Telegram, URL-only, T1)

Anthropic platform docs, read 2026-08-22. Re-seen subject — refreshed claude-managed-agents in place (its url: now points at the overview rather than the quickstart) and paged one new thing.

From the overview. Anthropic's own framing: Messages API for custom loops, Managed Agents for "long-running tasks and asynchronous work." Four objects — Agent (model/prompt/tools/MCP/skills), Environment (Anthropic cloud sandbox or self-hosted on your own infrastructure), Session, Events. Separating config from placement is what lets one versioned agent satisfy data-residency without being redefined; recorded as the agent-scope-isolation instinct applied to infrastructure. Built-in tools: bash, file ops, web search/fetch with domain allow/blocklists, MCP. Scheduled deployments fire sessions on cron with no client-side scheduler. Public beta, managed-agents-2026-04-01, on by default. Not eligible for Zero Data Retention or a HIPAA BAA because sessions are "stateful by design" — the managed pole's price, stated by the vendor.

The find: claude-dreams. New page. A dream reads a memory store plus 1–100 past sessions and writes a new store with duplicates merged, stale/contradicted entries replaced, patterns surfaced. Own beta header dreaming-2026-04-21; pending → running → completed|failed|canceled; minutes to hours; billed at standard token rates; instructions capped at 4,096 chars. The input is never modified — review the output or discard it. The pipeline is a session you can stream, archived afterwards. And it is process without a store, which makes consolidation a separable stage rather than an engine property.

Why that matters. This morning's cloudflare-agent-memory ingest described exactly this as a roadmap item, with the sleep metaphor. Anthropic has shipped it, under that name. Two vendors converging on consolidation in one day moves it out of iai-pme's single-instance column. Neither publishes evidence that a consolidated store retrieves better than its input — Anthropic's guidance hands evaluation to the customer — so beyond-pass-at-1 still stands unanswered. Recorded on both pages.

Updated agent-memory (new consolidation section), synthesis §6, index.md. No new entities: anthropic already a node; the docs carry no byline.

[2026-08-22] ingest | claude-seo — Universal SEO skill for Claude Code (hub-routed, Telegram, URL-only, T1)

agrici-daniel/claude-seo, MIT, 14.8k★ / 2.2k forks, 203 commits, v2.2.4 (July 2026), single maintainer. Created claude-seo (SoftwareSourceCode, T1) and the Person node agrici-daniel.

Not a duplicate. agentic-seo-skill (2026-05-29, Bhanunamikaze) is a different repo — 16 sub-skills / 10 agents / 89 scripts against this one's 25 / 18 / 32 commands. Checked before creating. The two are now cross-linked as a portability-vs-depth pair: the earlier pack installs identical content in nine harnesses' native formats, this one is Claude-Code-only and spends the difference on auto-discovery, 15-way parallel dispatch, a test suite, and a four-tier Google credential ladder. Same domain, same form factor, opposite bets — the cleanest instance of that trade the spoke holds.

Two firsts, both recorded in synthesis. (1) 410 tests, 91 of them SSRF/DNS-rebinding bypass cases, in CI. Every other skill pack here is markdown with no executable contract; this one treats its fetcher as attack surface, correctly — a tool that takes a user URL and fetches it from the agent's own machine is SSRF with extra steps, and two releases are credential-injection and SSRF fixes. Filed against agent-guardrails as containment showing up in what a helper script may fetch rather than what the agent may do. (2) Its 10-principle framework is a cognitive skill by reasoning-scaffolds' test — the category with negative measured evidence — but its output contract forces every recommendation to carry a "how would we know this failed?" check and a leading indicator. The scaffold is still unverified; its outputs are falsifiable. Kept those two claims apart on both pages, because blurring them would read as evidence the framework works.

Weaknesses recorded, deliberately, because 14.8k stars invite the opposite. The "10–15 min vs 4–8 hours" figure is the repo's own with no method. The case study is one screenshot of one three-month-old site's GSC impressions — no control, no attribution, not a benchmark. The 410 tests measure the code; nothing measures whether the audits improve rankings. Well-engineered and unevaluated at once.

Cross-spoke, corrected mid-ingest. First draft said this "hardens" ../search-marketing-wiki's llms-txt position from unmeasured. Reading that page showed it already holds the stronger, first-party finding — Google's own "Google Search itself doesn't use them". So the source adds no evidence there; what it adds is propagation, a corpus conclusion encoded as a practitioner default. Rewrote the note and added one pointer line to llms-txt rather than a second page.

Open, not resolved: the README names an interoperating skill "AI Marketing Claude"; this spoke holds ai-marketing-claude by a different author (zubair-trabzada). Nothing read settles whether they are the same project. Flagged on the page, not merged.

Also updated agentic-seo-skill, reasoning-scaffolds, index.md, synthesis.

[2026-08-23] ingest | KOT / agentkot (hub-routed, Telegram, URL-only, T1)

GitHub repo Loqira-Labs/agentkot, created 2026-08-20, v1.0.0 2026-08-21, read 2026-08-23 at 4★ / 0 forks / 0 issues. Created source summary agentkot (SoftwareSourceCode, url-only, T1 as a first-party publication — but unverifiable: the repository holds three READMEs, tools/ docs in EN/RU/ZH and prebuilt binaries, no source, no licence, GitHub language: null). Created entity loqira-labs (Organization, Kazakhstan, org created 2026-07-14, also publishes dsp-codegen; entity-index: no match, new node — k-kolomeitsev cut as too thin to page as a Person).

The angle that made it worth a full page: the design is three refusals — no permission gate ("There is no security, and there will not be"), no skills subsystem, no MCP — against a corpus that treats all three as infrastructure. Folded accordingly: agent-permission-model gains a third position (remove the gate, not move it) sitting beside per-action confirmation and model auto-approval, and reading against nanoclaw shows the shared premise underdetermines the design — same argument, opposite conclusion. agent-orchestration gains the widest child taxonomy here (spawn · 12h delegate · mailed teammate · restart-surviving TOML pipeline), plus two shapes new to the page: an until gate with a cheap supervisor agent, and peer-to-peer mail that breaks the strict fan-out tree. durable-agents gains pipeline-resume-by-run_id and a virtual file layer rebuilt from the transcript. agentic-coding-harness gains the refusal end of its spectrum. agent-memory gains the project-scope-as-agent-output note.

Synthesis: new section "The refusals become the product" (the skills refusal answers the registry argument, not the progressive-disclosure one; the MCP refusal concedes the need and rejects the transport); a paragraph added to the permission section; and a new tension — "self-hosted" is doing two jobs, since the security model asks the operator to own the boundary while the binary-only distribution denies them any way to inspect what is inside it. Also recorded: KOT's Agent API mirrors claude-code's field for field (subagent_type, general-purpose, effort off→max, isolation: worktree|remote with remote unavailable, <task-notification>), a dead parameter carried across — noted as an observation about the docs, not a claim about code nobody can read.

Two facts checked and kept because they cut against the source: homepage agentkot.ai fails TLS (serves a *.timeweb.ru certificate), and the publisher's own bio says "no cloud, no frontier API" while the product routes across a dozen hosted APIs. Runner-up spoke: none — squarely agentic-tooling.

[2026-08-23] ingest | @skills: Attention Is All You Have (hub-routed, Telegram, PDF, T2)

arXiv 2608.12610 (2026-08-12), Li Yin + 4 at SylphAI with Zhangyang Wang (UT Austin). PDF pulled and extracted locally with pypdf (26pp). Created source summary attention-is-all-you-have (ScholarlyArticle, url-only, T2 — original full-corpus crawl, but a preprint whose authors sell the remedy; classification shares come from a 100-skill sample at ±10pp and are cited as sampled). Created Thing page atskills-protocol (DefinedTerm/standard) and entities sylphai (Organization) + li-yin (Person); entity-index found no match for either. Zhangyang Wang / UT Austin left as a plain mention — evidenced but tangential, and the budget was better spent on the two nodes the protocol hangs off.

What it adds that nothing here had: counts. 56,804 SKILL.md directories / 1,133 repositories, 56,245 bodies fetched. 21 providers publishing 958 first-party skills (AWS 138, Anthropic 379, Google 92), 79% of provider-product skills naming their own CLI. 75 agents, 54 distinct project-level skill directories, 58 home locations. Median body 921 words. Two aggregators hold 21% of the corpus, top fifteen repos hold 51%. Installed description 50–280 tok per message.

Folded: agentskills-spec gains the directory-fragmentation count against its own portability claim, plus the L1 description as a bidding surface (Stripe's flagship: ~150 words of trigger conditions, ~20× a quiet description). agents-md gains the paper's read of why it won — always loaded, no trigger lottery — which sits uncomfortably beside context-files-ablation-paper's null. context-rot gains the ecosystem-scale form: the channel degrades exactly as adoption succeeds.

Synthesis: new dated section "Somebody counted the skills ecosystem", and — the reason this ingest mattered beyond the numbers — it lands the same week as agentkot, which read the same broken install layer and deleted skills, while this paper deletes the install. Recorded as the spoke's sharpest live disagreement, and a testable one. New open question added: how many skills can an agent keep installed before triggering degrades — the paper bounds it at <100 slots from the long-context literature and states that no controlled measurement exists, which makes it the cheapest high-value experiment on the skills thread.

Honest weakness kept on the page: the headline ratio has a measured numerator and an argued denominator, and the authors say so themselves rather than burying it.

[2026-08-23] ingest | Towards a Science of AI Agent Reliability (research pass, T1)

arXiv 2602.16666, ICML 2026, Rabanser / Kapoor / Kirgis / Liu / Utpala / Narayanan (Princeton). Hunted for hub ## Most wanted #8's successor — a k≥3 repeated-run reliability study covering at least one frontier proprietary model — and it is a clean close: 14 proprietary frontier models and no open-weight ones, K=5 runs per task at temperature zero, GAIA (165 tasks) plus the 26 τ-bench tasks left after Cuadron et al.'s grading errors, run on Princeton's HAL harness. Created science-of-agent-reliability (ScholarlyArticle, T1 — peer-reviewed, neutral party with no product in the race) and entities sayash-kapoor, arvind-narayanan, princeton-university (entity-index: no match for any; the "MATCH" it returned for Princeton was fuzzy noise on the word University).

What it settles: beyond-pass-at-1's own limitations section guessed the closed models were "likely more reliable than the models studied here." They are not — 18 months of releases, modest reliability gain overall and barely any on open-ended GAIA. The plateau is not an artifact of studying open weights, which is the single most useful thing the pair of studies establishes together.

Three findings folded into the Running it once was never a measurement thread rather than a new dated section (standing rule): consistency runs backwards against model size inside a vendor family, which lines up with this corpus's "high variance is a capability signature" from a different metric; the instability is in action ordering and not selection ("what but not when"), which is a sharper mechanism for agent nondeterminism than the spoke had; and resource consistency is poor and degrades with difficulty, so every single-run cost figure here carries an unmeasured spread. agent-evals gains the benchmark-design charge and the twelve-metric menu; beyond-pass-at-1 gains the complement note.

Edges: the two successors from hub ## Most wanted #8 had never been written into this spoke's own ## Growth edges, which is the source of truth — fixed. Edge 11 written and closed the same day; edge 12 is new and is the residue this ingest creates: both repeated-run studies hold the harness fixed, so the spoke's central claim that structure substitutes for capability has never been measured with variance on the structure side. Edge 13 carries the non-naive-memory arm across from the shortlist.

[2026-08-25] ingest | prompts.chat (hub-routed, Telegram, URL-only, T3)

Model: claude-opus-5 · effort: low.

prompts-chat — the website form of f/awesome-chatgpt-prompts, 167.9k★, started 5 Dec 2022, MIT code and CC0 prompt data. Arrived from the curator over Telegram; the hub routed it here over ../engineering-education-wiki (runner-up), which owns the prompt-engineering curricula.

The reason it belongs in this spoke and not in a prompting one: the directory has grown 66 Skills and 13 Workflows, ships an MCP server, and installs into claude-code as a plugin. Its definition of a skill — multi-file bundle of instructions, configuration and supporting files — is agentskills-spec's, reached from the prompt-directory side. The skills are indexed by the system they drive (Supabase, Stripe, GitHub, Obsidian, Termux), not by the persona they ask for, which is the whole difference from the 2022 corpus sitting beside them.

The number worth keeping: 66 skills here against agentic-awesome-skills's 2,003, at roughly thirty times the audience. Either the format is much harder to author than a prompt, or the submission flow still belongs to the old artifact. The source cannot separate those, and the log records both readings rather than picking. What it does establish is direction of travel — nothing is converting skills back into persona prompts.

Recorded as framing, not capability: "workflows are prompts with sequential flows and connections" describes no graph and no handoff. Each is one long prompt with ordered sections and prose prerequisites — the same gap agent-orchestration keeps finding between a claimed orchestration layer and a document about one.

T3: community-submitted, upvote-ranked, no inclusion criteria, no evaluation on any entry. The credibility list (Harvard/Columbia references, 40+ citations, #1 liked dataset on Hugging Face) is the maintainer's own and was not checked. The CC0 licence on the prompt data is the durable fact: it explains why this corpus appears inside other tools unattributed, and makes it a plausible upstream for training and eval sets — unmeasured here, and a growth edge.

New entity: fatih-kadir-akin (thin). No build — check:wikilinks and check:freshness only, per the per-route rule.

[2026-08-28] ingest | How Many Tools Should an LLM Agent See? (Meta) — research pass, T1

Repantis, Gawde, Singh & Blackwell (Meta Platforms), arXiv:2605.24660, 13pp — read from the PDF. Ingested with claude-opus-5 at high effort, hunting edge 14 / the standing open question about installed-skill count.

Created how-many-tools-should-an-agent-see (T1). Updated synthesis (edge 14 half closed with a successor, open question sharpened) and index.md. Not a duplicate of attention-is-all-you-have — that crawls a skill registry and argues from the long-context literature; this measures selection against depth on three benchmarks.

Edge 14 is HALF closed, and the useful part is the reframing. Nine conditions, three benchmarks (BFCL, MetaTool, ToolBench), registries 20 → 3,251 tools, scored with Bits-over-Random, a chance-corrected metric asking whether success at a depth beats random selection at that depth.

  • BFCL, 370 tools pooled so every other function is a hard distractor: 90.3 ± 2.4% at K = 7.4 ± 2.5 against 90.8% for a fixed shortlist of fifty.
  • ToolBench, 3,251 tools: a fixed five wins aggregate coverage (64.7% vs 61.9%) and finds nothing where the right tool ranks 6th–20th; the adaptive agent finds 16.7% of those.
  • Downstream with Claude Sonnet 4.6: 93.1% vs 87.1% selection accuracy against a fixed five, and 76.8% vs 60.9% on medium-difficulty queries.

So the variable is what the model is shown, not what exists. A 3,251-tool registry is fine at K = 7. That moves this spoke's question from "how many skills can I install" to "how many reach the prompt."

And it is why the edge is only half closed. Installed skills have no retrieval layer — their trigger descriptions share the prompt, so installed count is shown count, the K = N corner this paper never visits. The attention-is-all-you-have experiment is still the missing one.

Recorded as the authors' citation, not our reading: they report RAG-MCP showing "clear degradation as N grows from 1 to 11,100 MCP tools," and LongFuncEval and Rabinovich & Anaby-Tavor finding function-calling accuracy degrades as catalogues grow or as semantically similar tools are added.

Limits on the page: the RL agent is "deliberately simple, serving as a probe of the metric rather than a proposed system"; BFCL retrieval is BM25 over descriptions and parameter names; registries are synthetic rather than one team's installed set; one downstream model supplies the selection figures.

Entity note: no new nodes — four authors at one company, and meta is not the subject of any claim here.

[2026-08-30] ingest | Pi — pi.dev

Routed here by the hub router from Telegram (runner-up: none). Model: claude-opus-5 · effort: high. Read from pi.dev (WebFetch, then firecrawl for the full copy) plus the GitHub API for repository facts, since the website states none.

pi-coding-agent (T1). Mario Zechner's Pi, published by Earendil Works: "a minimal agent harness… Adapt Pi to your workflows, not the other way around." MIT, TypeScript, 99,163★ / 12,303 forks, 138 open issues, created 2025-08-09, pushed 2026-08-29 — the largest repository this spoke holds, by an order of magnitude.

The dedup finding came first. The corpus has cited Pi since June without paging it: oh-my-pi was ingested 2026-06-01 as "a fork of Mario Zechner's Pi," and agentic-coding-harness names "configurable harnesses like Pi" in passing. The fork was paged and the upstream was not.

Why it earns a page rather than a line. Pi is the second harness defined by what it removes — no MCP, no sub-agents, no permission popups, no plan mode, no built-in to-dos, no background bash — after KOT on 2026-08-23, and it overlaps KOT on two of them. The bet behind them is opposite. KOT keeps ten fixed built-in tools and no registry; Pi keeps nothing but surface ("Primitives, not features") and ships subagent/, plan-mode/, permission-gate.ts and sandbox/ among 50+ example extensions, with a package registry over npm and git. The synthesis paragraph that generalized "what the refusals share is a bet that built-in beats extensible" was written from one instance and does not survive the second; it now carries a subsection saying so.

A dating correction. This spoke recorded KOT on 2026-08-23 as "the first source here to argue the transport is the accidental part of MCP." Pi's "No MCP" links Zechner's post "What if you don't need MCP?" dated 2025-11-02 — nine months earlier. KOT was the first the spoke read, not the first to say it. Corrected in place.

Also recorded: four modes (TUI / print-JSON / RPC / SDK, with openclaw as the named SDK integration), 15+ providers switchable mid-session, sessions stored as branchable trees exportable to HTML or a gist, AGENTS.md/SYSTEM.md context control with replaceable compaction, and self-modification as the documented way to get a feature ("ask Pi to build it… hit /reload").

No benchmark, no evaluation, no comparison anywhere on the site — unusual enough here to state. Stars are adoption, not quality.

New pages: pi-coding-agent, mario-zechner (same-as github.com/badlogic + mariozechner.at, both verified via the GitHub API; no Wikipedia or Wikidata entry exists), earendil-works (same-as the GitHub org + earendil.com). Updated: agentic-coding-harness (Pi and KOT added to the instance list; the passing Pi mention now links), oh-my-pi (a note that the upstream reframes the fork as the same codebase taken the way its author declined to go), index.md, synthesis.md.

[2026-08-30] ingest | LongHorizon-Harness — re-seen source, refreshed in place

The same URL arrived again through Telegram, 25 days after the 2026-08-05 ingest. Refreshed in place per the re-seen-source rule; no new page. Model: claude-opus-5 · effort: high (the original ingest recorded no effort).

The refresh found a stale conclusion, not a stale number. The page's closing section is titled "Why it does not close growth edge #2" and ends "the edge stands… someone still has to run the fourth cell." It was written 2026-08-05. Growth edge 2 was closed by the research pass on 2026-08-06, and twice more after that:

  • scaffold-effects-gaia (08-06) — crossed 5 models × 3 scaffolds, pre-registered.
  • claw-swe-bench (08-07) — 350 instances, harness spread 12.5–27.4 pp vs model spread 29.4 pp.
  • harness-bench (08-08) — 8 backends × 6 harnesses, 106 tasks, 5,194 trajectories.

So the fourth cell was run three days after this page invited someone to run it, and the comparison this page set up resolves against the harness arm: the model gap is the larger of the two. The section is now kept with a retraction block rather than rewritten, because the reasoning was sound on 2026-08-05 evidence and the record of it going stale is itself the useful thing. Nothing about the measurements is retracted — only the framing that read a single-arm result as evidence for the founding bet.

Repository facts the first ingest did not record at all. GitHub API, 2026-08-30: 1,396 stars / 144 forks, 37 open issues, MIT, Python, homepage lh-harness.pages.dev, created 2026-08-04 — so 1,396 stars in 26 days. Five tagged releases since the first ingest (v0.1.3 on 08-07 through v0.1.7 on 08-20) and 38 commits, whose titles are the useful part: "Guard-only failures fail the round, not the run," "Recover from isolated EBADF in control-bus reads; keep the watcher alive," "Recover from local role timeouts." Those are the failure modes a Manager/Executor/Auditor split produces once it runs for dozens of hours — evidence the design is meeting real operating conditions rather than only benchmark ones.

Updated: longhorizon-harness (retraction block, adoption and release cadence, see-also, attribution), index.md.

Second re-sent source in a row today, after ruaccent in ../speech-audio-wiki. Both re-reads corrected a claim rather than adding a fact, and in both cases the error was a sentence the first pass did not check — an assumed absence there, a conclusion overtaken by later ingests here. Worth naming as a pattern: pages go stale at their arguments before they go stale at their numbers, and nothing in the cycle currently re-reads an argument.

[2026-08-30] ingest | SeekerClaw (routed by hub, arrived via Telegram)

seekerclaw — MIT, Android 14+, 299★/71 forks, by sepivip; winner of the Solana Mobile Hackathon, April 2026. T1 for what it is: architecture, module and tool inventory, licence, platform, key handling and the cap policy are the project's own checkable facts. The hackathon win and "first on-device agent that can transact on the open web" are maker claims and are attributed on the page rather than asserted.

Why it matters here, in two places.

It inverts mobile-device-agent's device/brain split. That page was written from one source (mobilerun) and had the runtime on an operator machine with a helper on the phone. SeekerClaw puts the runtime on the phone — Node.js (~24k lines, 50+ modules) inside a Kotlin/Compose app (~27k lines), running as a foreground service — and still calls a remote model. The part that made those frameworks model-agnostic holds; the architecture around it does not. Also inverts what the phone is for: mobilerun's phone is the target being tapped, SeekerClaw's is the host, with device control one tool module of thirteen.

It supplies the guardrail case agent-guardrails was missing. An irreversible Solana transaction is where the reversibility bound and the budget bound meet, and the design refuses both obvious answers. Not "confirm everything" and not "trust it": per-tx and daily caps in USDC and SOL, preflighted before any send, burner key signing silently under cap and falling back to a main-wallet popup over. Plus a split by effect — for x402 payments GET settles silently under cap, POST always asks. Reads bounded by money, writes bounded by consent.

Pages: new seekerclaw, sepivip. Updated mobile-device-agent (the on-device counterexample; its Verification open item partly answered, and its Corroboration item now reads two sources with still no independent measurement), agent-guardrails (a budget bound in real currency), synthesis.md, index.md.

Smaller things kept on the page: skills as Markdown + YAML with a published SKILL-FORMAT.md (another independent agent-skills implementation, installed by sending the agent a link in chat); MCP over Streamable HTTP with rug-pull detection; a loop detector that warns at 3 identical calls and breaks at 5; reasoning handling as its own subsystem across four providers; a bounded flush on Stop. No benchmark of any kind is published — noted, not held against the tier, since none of the T1 claims depend on one.

[2026-08-30] ingest | MAST + the 20,574-session study — the failure frequencies existed all along

Found by the absence-claim drain, not by a hunt. agent-failure-localization said "nobody has published frequencies" for agent failures. One search returned two papers, and the older predates that sentence by seventeen months.

mast-multi-agent-failure-taxonomy (Cemri, Pan, Yang et al., arXiv 2503.13657, NeurIPS 2025, T1) — 14 failure modes in 3 categories; 150 traces hand-annotated at κ = 0.88, then 1,600+ via an LLM judge validated against those humans, across 7 frameworks and 4 models. System design 44.2%, inter-agent misalignment 32.3%. Per-mode: step repetition 15.7%, unawareness of termination 12.4%, disobey task spec 11.8%, loss of history 2.8%, disobey role spec 1.5%.

coding-agent-misalignment (Tang et al., arXiv 2605.29442, May 2026, T1) — 20,574 sessions across 1,639 repositories, misalignment operationalized as developer pushback. 90.50% cost effort and trust rather than destroying anything; 91.49% need explicit user correction; overall rates fall over time while constraint violations and inaccurate self-reporting rise proportionally.

Why it matters beyond closing the claim. Step repetition plus not knowing when to stop is 28% of failures and neither is a reasoning deficit — the losses sit in the scaffold. That is this spoke's structure over capability finding arriving as a defect distribution instead of a benchmark spread. Folded into synthesis.md.

Edge 12 stays open in a narrower form: neither study varies the harness as the independent variable.

[2026-08-30] research | Stop Comparing LLM Agents Without Disclosing the Harness (research pass, edge 12 / Most wanted #29)

Hunted ## Most wanted #29 — "reliability with the harness as the variable," needing K≥5 repeats per configuration with reported spread. Half close.

harness-disclosure (Zhang, Wang, Ge, Xu, Hamm & Reddy, arXiv 2605.23950v1, 7 May 2026, T1, PDF read with pypdf) is the formal statement this spoke had been approximating from five separate measurements. The Binding Constraint Thesis defines HV(M) as variance over harnesses and MV(H) as variance over models and asserts HV > MV for long-horizon tasks among comparable frontier models.

New numbers: a harness swap alone moving Terminal-Bench 2 pass@1 69.7% → 77.0%; 11 pp scaffold-only variation for GPT-5 and 15 pp for Kimi K2 Thinking on SWE-bench Verified; and ~48 pp single-model swing on HAL's SWE-bench Verified Mini — the largest harness effect the corpus holds, against harness-bench's 23.8 pp. Plus model ranking reversal, which is what turns a noisy leaderboard into a misleading one.

It does not close the edge. It is a position paper: it specifies the factorial protocol with reported variance components and aggregates others' single-run numbers to argue for it. Nobody has run K≥5. Successor: the edge can now name the protocol instead of describing a wish.

[2026-08-31] ingest | geo-seo-claude — the marketing pack revises itself toward measuring

Routed from the hub (Telegram). https://github.com/zubair-trabzada/geo-seo-claude — MIT, Python, 9,877★ / 1,531 forks, created 2026-02-18, pushed 2026-08-31 (the day of the ingest). T1 as the primary artifact; the README's headline statistics are relayed third-party numbers and are read as T3 inside it. Runner-up spoke: search-marketing-wiki, declined under the standing rule that SEO/PPC skill packs are tools here (agentic-seo-skill, claude-skills-ppc precedent).

Pages: new geo-seo-claude, zubair-trabzada. Updated ai-marketing-claude (the same author's second pack answers three of its four recorded weaknesses), claude-seo and agrici-daniel (the Skool funnel is a pattern now, not one maintainer's arrangement), synthesis.md, index.md.

What is new to the spoke. A skill pack moved along the verification axis on its own. The same author's ai-marketing-claude had 4 scripts, none collecting evidence, and this synthesis said its scored client PDF rested on no checkable observation. This pack has six scripts of which four fetch facts — page retrieval, a citability score computed from word counts / pronoun density / sentence length / counts of percentages and dollar figures, robots.txt read for 14+ named AI crawlers, and llms.txt validation. It also publishes a 16 KB scoring rubric and states its own limit unprompted: "these weights are not derived from a controlled study." Third posture toward the unverifiable score, beside silence (ai-marketing-claude) and a forced falsification line (claude-seo).

The ceiling is unchanged: every input is a property of the site, none is a measurement of whether an AI cited it.

A defect found by reading the tree. /geo report-pdf is a headline command with a sample PDF in examples/. Its skill (v2.0.0, pandoc + Chrome) needs ~/.claude/skills/geo/templates/; the files exist in the repo and install.sh never copies that directory. It also hardcodes /Applications/Google Chrome.app/ while shipping a Windows installer, and the README still documents a ReportLab script the skill itself replaced. Three files' worth of reading; nothing in 9.9k stars shows it. Not run — this is a claim about the code, not a test result.

Corrections carried. "Nine minutes, then nothing" on ai-marketing-claude is about that repository, not its author. And the README undercounts its own contents: 13 sub-skills claimed, 14 listed, 15 present, plus an undocumented Flask CRM dashboard under scripts/webapp/.

[2026-08-31] ingest | Towards Long-Horizon Agents: A Survey (+ reading list)

Routed from the hub (Telegram). https://github.com/RUC-NLPIR/Awesome-Long-Horizon-Agents — MIT, 1,015★ / 38 forks, created 2026-07-09, pushed 2026-08-15; the paper is doi:10.20944/preprints202607.1328.v1 (Dong et al., 20 authors, ruc-nlpir + NUS), July 2026, with an OpenReview PDF and ~730 catalogued papers. T1 on this corpus's ladder, with two qualifications that belong beside the number: preprint, not peer-reviewed, and a survey measures nothing of its own. Runner-up spoke: machine-learning-wiki (Pillar II is train-time work), noted as cross-spoke context.

Pages: new long-horizon-agents-survey, long-horizon-agency (concept), ruc-nlpir (entity). Updated harness-disclosure (its thesis is dated by construction), synthesis.md, index.md.

Why it earns a page despite measuring nothing. It writes this spoke's central question as a decomposition, Agent = π_θ ⊕ ℋ — the third independent formalization in about two months, after harness-disclosure (measures the split) and model-or-harness-taxonomy (diagnoses across it), citing neither. Three groups, three institutions, three purposes, no shared ancestry. Worth naming on the same day the hub corrected the opposite case in research-wiki, where four apparent independent arrivals turned out to descend from one gist.

The finding with teeth. The survey's organizing claim is that harness engineering and model optimization co-evolve — harness-implemented capability gets internalized into the policy. That puts a clock on the Binding Constraint Thesis: HV > MV is largest exactly where the harness compensates for what the model cannot yet do. ~48 pp is a 2026 reading, not a constant. Not a contradiction (one is a measurement, one a trajectory) but a condition the harness thread now carries. Edge 12 gains a second dimension: run the specified factorial protocol at two points in time.

A structural correction. The nested H1 ⊂ H2 ⊂ H3 levels make agent-memory the definition of H2 rather than a feature of interest, and self-evolution the definition of H3. Two threads filed here as separate enthusiasms are two rungs of one requirement.

New edge: ingest METR (arXiv 2503.14499). The whole horizon axis rests on it and the wiki holds it only as a row label in beyond-pass-at-1.

[2026-09-01] ingest | Vibe-Coding Prompt Template — the method under its own antonym

github.com/KhazP/vibe-coding-prompt-template — MIT, 2,984★ / 377 forks / 0 open issues, created 2025-04-14, pushed 2026-08-21, read today from the repository and the GitHub API. Routed here by the hub; runner-up ../research-wiki, which owns the durable-markdown-as-memory lineage the workflow leans on. T1 for what the workflow is; every efficacy claim is the author's. New pages: vibe-coding-prompt-template, vibe-coding, alp-yalay.

The finding is about a word. The repository is named for vibe coding and prescribes the discipline CDD was defined against — Google's post puts vibe coding at the pole where you "jump straight to code, intent improvised in chat and lost when the session ends." This project's advanced-practices section says the opposite in CDD's own terms: durable facts in files rather than one long chat, repo-owned MEMORY.md over tool-side memory, plan before edit. The author's GitHub bio reads "Vibe Coder."

That made a concept page worth writing. vibe-coding collects four incompatible senses the hub already held across four spokes: CDD's artifact-free pole; termux-launcher's author disclosing "this project is entirely vibe-coded" as a caveat about review (../foss-applications-wiki); the Godot Foundation banning it from contributions (../game-engines-wiki); and now the name of a workflow with approval gates and a review checklist. The consequence is practical, not lexical: a ban written on that word has no agreed test, and a disclosure written on it no longer narrows anything.

What it adds to the method thread. Every earlier SDD instance here bundles the ordering with machinery — spec-kit's slash commands, OpenSpec's delta specs, conductor's tracks, corbell's code graph. This one is four prompts pasted into a chat tool, priced at 2–4 hours, for an audience that "does not need to be an experienced developer" — the sequence without the tooling, which is the cheapest available test of which half carries the benefit. Nobody has run it: no evaluation of any kind, and the evidence offered is four sites built by the author and his associates.

Also recorded: the agents-md hub-and-spoke rule stated outright (one contract, vendor files as "concise tool-specific adapters"), permissions/evals/cost ceiling moved into the planning document with an agent-permissions.example.json emitted at step 4 — agent-permission-model has been a harness-config subject here, not a planning one — and a published freshness policy with a monthly cadence and model families instead of pinned versions.

Pages refreshed: spec-driven-development, context-driven-development, agents-md, synthesis.md (new section + a new cross-spoke seam with foss-applications and game-engines), index.md. Entity: alp-yalay created, same-as GitHub only — the personal domain in the README returned HTTP 500 today, so it is not recorded as an identifier.

[2026-09-01] ingest | Inferoa — the harness that looks down at the serving layer

github.com/agentic-in/inferoa — Apache-2.0, TypeScript, 512★ / 87 forks / 48 open issues / 3 watchers, 132 commits, created 2026-06-08, last pushed 2026-06-18. Read today from the repository, the GitHub API and the npm registry. Routed here by the hub; runner-up ../llm-inference-wiki. T1 for the design — there is nothing to tier as evidence. New pages: inferoa, xunzhuo.

The argument. "Every loop is also an inference workload. As turns accumulate, prompt prefixes drift, cache reuse collapses, stale evidence fills context, model routing gets harder, and serving choices start to matter." So the harness exposes serving to the loop instead of hiding it: prompt epochs and deterministic tool schemas to keep a cacheable prefix alive, bounded history and tool output, routing by cost/safety/privacy/capability, a /tokenmaxxing view of token and cost pressure in-session. Built on vLLM Engine, vLLM Omni and vLLM Semantic Router, with CodeGraph and RTK for context selection.

Why it matters to loop-engineering. The spoke has written that discipline as a correctness problem — how the loop knows it is done, how truthful the feedback is. This is the first source to write it as a cost problem, and it answers with kv-cache mechanics rather than a token budget. It also renames Ralph's completion promise completion evidence and gives it independent feedback surfaces (plans, tests, tool results, research metrics) — the slot auto-company left empty.

The evidence is provenance, not data. xunzhuo describes themselves as "Intelligent Routing @vllm-project | Open Source & AI @AMD", so the harness is written by someone who works on the router it sits on. That is why an unbenchmarked design argument is worth holding. It is not a substitute for one, and the README contains no measurements at all for a project whose whole pitch is efficiency.

Health, recorded because it cuts against the pitch. 119 npm versions published inside the ten days to 2026-06-18, three GitHub releases on consecutive days, then eleven weeks of silence. 3 watchers against 512 stars. The README installs inferoa@dev on the grounds that npm latest is "reserved for stable releases" — latest exists and points at 0.14.19, published the day the work stopped, so the docs route users past a stable channel that is one build old.

Second word to flip in one day. tokenmining is defined as the inverse of "tokenmaxxing", glossed there as spending tokens freely; this source's tagline uses the word for token discipline. Both readings kept, on that page and this one. Same shape as vibe-coding earlier today. No shared page from a pair — a third instance earns one.

Pages refreshed: loop-engineering, tokenmining, synthesis.md (new section, plus the llm-inference seam now marked contested from this side), index.md.

[2026-09-03] ingest | HarnessDev — can a model build its own harness?

Source: https://arxiv.org/abs/2609.01437 (T1). Arrived via Telegram, hub-routed here — harness construction is this spoke's densest thread and there was no competing destination. WebFetch returned PDF binary; pypdf extracted 41 pages cleanly.

Why it is not another harness comparison. scaffold-effects-gaia, claw-swe-bench, harness-bench and confucius-code-agent all vary harnesses humans built. harnessdev varies the builder: six creator LLMs each construct a complete execution system from an identical weak seed — "a runnable compatibility layer, not a task-solving agent" — across four domains and 2,207 instances, with hidden tasks withheld.

Findings recorded: overall 86.2 for the human reference against Opus 4.8's 67.8 and Seed 2.0 Pro's 22.8; but per-domain the picture inverts — model-built harnesses take MLE-bench 32.9 against 24.0 and EQ-Bench3 84.6 against 83.7, and lose BrowseComp 52.4 against 92.2. The domains models win are the ones where the harness has least to do. 77.8% of failed Data tasks are attributed to harness defects, not executor capability.

Two retractions and a cross-link. autoharness carried, since ingest, that harness self-improvement was "still unmeasured here". It is measured now: +4.44 held-out points for the best of six models, regression for three of four lineages once the runtime model is fixed. Retracted in place with the number, and its unverified 42% → 78% headline now has something to sit against. self-improving-agents gained the same finding as a section — the compounding claim is sized rather than refuted, and the size is small enough to require care with the word. harness-bench gained a note on the fourth variable this paper exposes: a Self-Eval score conflates harness design, executor capability, and the compatibility between them, which no factorial in this spoke has separated.

Two caveats on the page rather than buried. The human baselines are "verified public system results rather than paired controls under one executor" — five different published systems on five different models — so the 86.2 is a best-available comparison, not a controlled arm. And the lead affiliation is ByteDance Seed while Seed 2.0 Pro places last, which is the inverse of the usual vendor-benchmark problem: it does not make the work independent, and it removes the obvious motive to distrust the ranking.

avoid-ai-writing run. Verify at the session's end-of-run gate.

[2026-09-04] ingest | claude-batch-toolkit (S2)

Routed from the hub (Telegram), runner-up llm-providers-wiki. A four-part Claude Code add-on — MCP server, /batchy skill, status line, jobs registry — that posts slow work to Anthropic's batch endpoint at half price. T1 primary artifact; nothing in it is measured.

Why it matters to this spoke: it is the fifth token-budget lever and the first that isn't about tokens. Route / compress / modality-arbitrage / prune all shrink what gets billed; this one leaves the prompt alone and buys the provider's asynchronous discount, paying in latency instead of fidelity. Folded into the synthesis under that corner.

Two findings recorded against it. The README is six months stale (created 2026-02-09, last push 2026-03-09): its cost table quotes Opus 4 at $15/$75 and the default model is claude-opus-4-6, so the 50% ratio holds while every dollar figure is wrong — batched Opus 5 is $2.50/$12.50 per claude-message-batches-api. And it never mentions that batched requests get best-effort caching only, which matters more for a harness than for a generic batch user.

Also noted: the daemon-free status-line poller (forked, ≤1/60s, nothing when idle), the Vertex backend that can submit but is not background-polled, and the /batch → /batchy rename forced by a Claude Code built-in — command-namespace collision as the slash-command form of the trigger-slot crowding attention-is-all-you-have measures on descriptions.

New: claude-batch-toolkit, s2-stream-store. Updated: synthesis.md, index.md.

[2026-09-05] route | Environment Evolution for Terminal Agents (arXiv 2609.04128) → agentic-tooling-wiki

Arrived by Telegram. Classified here rather than to ../machine-learning-wiki on the dominant substance: the paper's mechanism is a "loop-engineered multi-agent harness" generating training environments for terminal agents, scored on Terminal-Bench 2.1. The RL half is real and is the runner-up spoke's; the harness half is this spoke's subject, and the interesting claim lives there.

New page: environment-evolution (T1, published 2026-09-03). Dedup: rg found no page for this paper, and no page named "environment evolution", anywhere in the corpus.

Why it earns a page rather than a note. This spoke tracks the harness as a runtime for doing work. Here the same construct is a runtime for manufacturing the difficulty another model trains against, which is a use of the object the spoke had no instance of. It also inverts self-improving-agents: what improves generation over generation is the environment, not the agent — and the authors argue that letting the model generate its own curriculum from its own rollouts is precisely what limits the existing co-evolution approach. Folded into that page.

Numbers recorded with their caveats on the page: +14.4 (Qwen3.6-27B) and +18.0 points (Qwen3.6-35B-A3B) on Terminal-Bench 2.1, plus rollout difficulty checks against Hy4 preview, claude-opus-5 and GPT-5.6 Sol. All the authors' own, v1 preprint three days old, no affiliations on the listing, no baseline score stated, and "simple long-horizon RL" left unspecified at abstract level — so neither number should be carried elsewhere before the body is read.

Cross-spoke: ../machine-learning-wiki owns curriculum learning and RL training method; if that spoke later wants the training half, link this page rather than re-summarising the paper.

[2026-09-06] ingest | omp product site (omp.sh) — re-seen, refreshed in place; one citation chased

Arrived via Telegram as https://omp.sh/ ("omp — a coding agent with the IDE wired in"). Routed here by the hub. WebFetch returned 403; Firecrawl read the site, the docs tree and the README, and the GitHub API supplied the repository counts. Runner-up spoke: ../dev-tooling-wiki (LSP, DAP and an in-process toolchain are its vocabulary) — declined, the subject is an agent harness.

Re-seen, not new. oh-my-pi has been paged since 2026-06-01 under the repository URL, so this refreshed that page rather than creating a second one. The site is a different artifact from the README, though: omp.sh now carries a full documentation tree (quickstart, providers, sessions, plan, subagents, collab, skills, MCP, CLI, RPC, SDK, extension authoring) that did not exist at either earlier reading.

Citation chased, and it is the interesting half. The README's benchmark table links to Can Bölük's post The Harness Problem (blog.can.ac, 2026-02-12). New page harness-problem-hashline — T3, because it is the tool's author benchmarking his own edit format against his competitors'. It answers, partially, a caveat this spoke has carried on oh-my-pi since June ("the 61% figure — unbenchmarked here"): 16 models × 3 formats, 3 runs × 180 React-mutation tasks, +15 points average over apply_patch, ahead in 14 of 16, Grok Code Fast 1 6.7% → 68.3%. Fixtures and per-run reports are published; that makes it reproducible, not independent.

Two more new pages. agent-edit-format (DefinedTerm) — the four designs in circulation (apply_patch, str_replace, Cursor's fine-tuned apply model, hashline) and why every one but the last makes the model retype content it already read. And two entities under ../ENTITIES.md: can-boluk (Person, same-as GitHub + can.ac, both verified — a Windows-kernel and reverse-engineering background that shows in the tool) and stencil-labs (Organization, thin, the copyright line on the MIT LICENSE).

What the numbers moved to: 21.8k★/2.1k forks on 2026-08-04 → 29,778 / 3,035 today, with 2,391 open issues, 21,374 commits, and a repository created 2025-12-31 — eight months old. Also new since August: native Windows with no WSL (the in-process coreutils are why), curated cross-session memory with swappable backends, an advisor model reading every turn of the main agent, /collab live session sharing over a relay, 16 internal URL schemes, and desktop control via the OS accessibility tree. The front page teases "omp², soon!" with nothing attached.

Synthesis: new section "The harness effect gets isolated to one tool, by an interested party" — this is the finest-grained harness measurement the spoke holds (harness-disclosure and harness-bench vary whole harnesses; this varies one tool inside one), and the gradient contradicts the slogan: the gain concentrates in weak models and reverses at the frontier. Two new growth edges follow — 17 (read Diff-XYZ arXiv:2510.12487 and EDIT-Bench arXiv:2511.04486, the independent check on a mechanism the corpus has started to lean on) and 18 (where the harness effect goes at the frontier).

Also recorded, without endorsement: the post reports Google disabling the author's Gemini account while the benchmark ran, with a screenshot as the only evidence. One party's account of an enforcement action; the page says so.

[2026-09-07] ingest | oh-my-pi repository — re-seen a day later, page refreshed in place

Same subject as yesterday, different artifact: the curator sent https://github.com/can1357/oh-my-pi, which is the URL oh-my-pi has carried since 2026-06-01, one day after the omp.sh reading. Refreshed that page rather than creating anything; no new pages, no synthesis change.

Re-checked against the GitHub API instead of assumed: 29,934★ / 3,047 forks / 2,389 open issues, pushed the same hour. Against yesterday's numbers that is +156 stars and +12 forks in a day, with open issues down two. The point of recording it is the rate — yesterday's entry called this the fastest star accumulation in the spoke after Pi, and a day of daily-resolution data says that was not an artifact of the site launch.

Two README counts no earlier reading had: 23 web-search backends and 50+ tree-sitter grammars. Both are the project's own figures and joined the "still unmeasured" list on the page. The README moved no claim otherwise, and "omp², soon!" still has nothing behind it.

[2026-09-10] ingest | Claude 101 (Anthropic Academy)

Routed in from the hub (Telegram). The course page is a Skilljar catalogue entry; fetched with curl and read as stripped text. Lessons sit behind registration — only the syllabus was read, and the page says so twice.

Source: claude-101-course, T1 — first-party, and the primary record of how Anthropic sequences its own concepts for a beginner. freshness: volatile; no published:, the page carries no date or revision marker.

Why it earned a synthesis section: unit two is projects, artifacts, skills. This spoke has held skills as an engineering object throughout — spec, catalogues, benchmarks, failure analyses — and this is the first source showing the vendor teaching them to non-developers, in an introductory course with no code in it. A concept crossing from tooling into product literacy is a dateable event and this is the date.

Two smaller reads recorded: the unit order (chat → containers → reach → role use cases) restates agentic-adoption-ladder's sequence from the vendor's side rather than the practitioner's; and the surface taught is the desktop app, a reminder that the agentic-coding corner this spoke documents in depth is the minority product.

Not paged: Skilljar. It is the LMS the course is hosted on, evidenced only by its own FAQ block on this page, and it says nothing about agent tooling. Named in the source page instead.

Runner-up logged at the hub: ../engineering-education-wiki, on shape — this is a curriculum with a certificate, which is that spoke's subject. Its own boundary rule sends it here: a course teaching a tool routes to the tool's spoke; a course teaching a discipline stays there.

[2026-09-14] ingest | Measuring Time Horizon using Claude Code and Codex (METR) — via quality cycle

Found while draining the absence-claim queue (hub cycle), not by a hunt: long-horizon-agents-survey said the factorial model × harness experiment had never been run, and it has — once. Source: metr-scaffold-comparison, T1, published: 2026-02-13. Retracted in place: the survey page's "nobody has run" sentence (struck, dated, corrected to "run once, one time point"); long-horizon-agency's longitudinal claim re-checked and holds. Edge 12 half closed (K≥5 half stands). New entity metr paged the same cycle from the hub's Outstanding list. Synthesis section The product harness does nothing at the top of the range.

[2026-09-16] ingest | Dream-RSI: Recursive Self-Improvement through Evolving Worlds

Source: dream-rsi, T1, published: 2026-09-14, via Telegram (HF papers page → arXiv 2609.14858). Routed by the hub (runner-up optimization-algorithms). New pages: dream-rsi, google-deepmind (thin Organization). self-improving-agents gains a fourth loop site (exploration policy, replay feedback). Synthesis section The loop moves to the policy over the agent. ingest_effort: low.

[2026-09-17] ingest | Hypit (hypit-ai README + LICENSE)

Source: hypit, T3, via Telegram (msg 2134), routed by the hub (runner-up llm-providers). No new entity — hypit-ai is an org with no evidence beyond the repo. Licence read from the file: modified Apache-2.0, fails OSI. Synthesis section Three video skills, three licences.

[2026-09-17] ingest | Stax (google-labs-code README + developers.google.com docs)

Source: stax, T3, via Telegram (msg 2136), routed by the hub (runner-up machine-learning). New entity google-labs (thin, Wikipedia same-as). Product page 403; docs page read instead. Synthesis section The eval platform arrives without the eval.

[2026-09-17] ingest | ScienceIDE (arXiv 2609.19134)

Source: scienceide, T1, published: 2026-09-16, via Telegram (msg 2138), routed by the hub (runner-up machine-learning for the SFT/RL half). qwen reused cross-wiki; AItonomy / PhAI-Labs not paged (no evidence beyond the byline). Synthesis section The environment is a real repository this time; self-improving-agents gains a line.

[2026-09-20] ingest | Omega — SingularityNET neuro-symbolic agent framework (README + docs/, GitHub API)

Source: omega-framework, T1 (first-party README + docs/; every number self-measured on the maintainers' own agent), via Telegram (msg 2183), routed by the hub (runner-up research-wiki, formal methods; second adjacency knowledge-representation-wiki for AtomSpace). No published: — the repo states none. Dedup: zero hits corpus-wide for SingularityNET / Hyperon / MeTTa / OpenCog / NAL. New pages: omega-framework (source), neuro-symbolic-agent (mechanism, first instance), singularitynet (Organization; same-as = official site + GitHub org, Wikipedia/Wikidata have no entry). Touched: self-improving-agents (negative instance — three failed self-patches), agent-memory (three-tier store with a logic tier). Two claims checked against the tree: "~200 lines" is src/loop.metta (119), core ~594; Docker image named omegaclaw. Synthesis section The engine the model cannot overrule, and where the errors went; adjacency note under research-wiki. ingest_effort: high.

[2026-09-23] ingest | agent-workspace-linux (agent-sh README + docs/permission-model.md + SECURITY.md, GitHub API)

Source: agent-workspace-linux, T1 (first-party design; nothing measured), via Telegram (msg 2201), routed by the hub (runner-up defensive-security for the sandbox boundary). No published:: the repo states none (latest release v0.3.3, 2026-09-19). Dedup: agentsys is the same org; no other hit. New entities agent-sh (Organization; same-as agent-sh.dev) and avi-fenesh (Person; same-as GitHub; own site as url:). Sibling computer-use-linux mentioned, not paged. Touched: agent-permission-model (new section: one approval, then a two-layer ceiling), agentsys and claude-browser-use (links). Two findings recorded on the page: SECURITY.md supports only 0.1.x at v0.3.3; README ad for the author's own company, not disclosed. Synthesis section The agent gets its own desktop. ingest_effort: high.

[2026-09-24] ingest | Claude Code model router (self-authored build notes)

Source: claude-code-model-router, T3 (self-run measurements on one device, a few dozen requests, not independent — the design is checkable, the numbers aren't), published: 2026-09-24, raw file, no URL, routed by the hub (runner-up llm-providers-wiki for the pricing/tier angle, recorded in the source page's own cross-spoke section only). Dedup: no existing page on prompt-cache economics or model-tier routing; claude-apps-gateway and tokenmining cover adjacent but distinct ground (fleet governance; token-budget levers without a switching-cost account). New page cache-aware-model-routing (mechanism: a provider's prompt cache is per-model, so a mid-session downgrade only pays through output tokens and the loop, never through the context switch itself). Touched: claude-code (new "wire format is undocumented" note — system-role folding, cache breakpoint on the last message, per-model request shape, claude -c resume-model behaviour), agent-orchestration (subagents sidestep the switch cost structurally), tokenmining (one link added, claim untouched). Synthesis: flagged as a qualification of tokenmining's "~60% cost cut from routing" — both figures hold, they describe different points in a session (before vs. after a session has built up cached context). ingest_effort: medium.

[2026-09-27] ingest | ai-marketing-claude — refresh (GitHub API re-check)

Source: ai-marketing-claude, T1 (primary artifact), already held since 2026-08-17 (last refreshed 2026-08-31 for the geo-seo-claude comparison). Routed by the hub again (runner-up search-marketing-wiki), re-ingested in place per the edge rule. Dedup: same repo, no new entity. Re-checked commit history and counts via the GitHub API: still exactly one commit (e5aa0ea4, 2026-03-02), confirming "nine minutes, then nothing" as a stable property rather than a stale read. Counts moved 2,451★/692 forks/7 issues/35 watchers → 2,686★/712 forks/6 issues/37 watchers; the fork-to-star ratio held at 27–28% across the six-week gap, which is new evidence against the ratio being an early-life artifact. Updated the source page, zubair-trabzada's cross-reference count, the index.md entries, and the synthesis paragraph in the popularity-signal thread. No structural change to the repo itself. ingest_effort: low.

[2026-09-27] ingest | ai-marketing-claude — reconfirmed, no change (duplicate Telegram send)

Same URL arrived again via Telegram a second time today, after the refresh logged above. Re-checked the GitHub API before touching anything: still one commit (e5aa0ea4, pushed 2026-03-02T01:18:40Z), still 2,686★/712 forks/6 open issues/37 watchers — identical to the numbers just written into ai-marketing-claude and this log. Nothing material changed in the few hours between sends, so no edit was made to the page or index.md; updated: was already today's date. Recorded here only so the duplicate send doesn't look unhandled. ingest_effort: low.

[2026-09-27] ingest | StarNet — pixel-art multi-agent desktop

Source: starnet, T1 (primary artifact — README/repo speaking for itself, a statement about provenance not verified quality), url: https://github.com/androoAGI/starnet, no published: (repo carries no self-dated publication). Routed by the hub (runner-up foss-applications-wiki for the local-installed-software angle; recorded in the source page's own framing and appended to the existing vibe-coding cross-spoke adjacency bullet rather than opening a new one). Dedup: rg for "starnet", "andrew sims", "androoagi" across the hub found no existing page; entity-index on "Andrew Sims" / "StarNet" / "androoAGI" returned no match ≥ threshold — new pages confirmed clean. New page starnet: local-first Rust/Tauri desktop app running several agents in a pixel-art space station, where room = capability-scoped team, hallway = authorized handoff lane, placed object = real capability grant — a spatial answer to the question agent-permission-model already tracks from the approval-UX side. Stated design principle ("the interface must never assert state the harness cannot prove") recorded as the project's own claim, not verified independently. Also notes: BYO-model (OpenRouter/Anthropic/OpenAI/Google, keys in OS keychain, or local Ollama), chat channels (Telegram/Discord/Slack/Signal/Matrix), Night Shift unattended mode, recipes/skills/cron, Task Briefs, MCP connectors, voice I/O, spend/budget ledgers, import from openclaw/hermes-agent personas. MIT code; name/logo/artwork proprietary to the creator. 641★/109 forks/14 watchers, 2026-09-27 snapshot. Entity discovery: new Person page andrew-sims (GitHub handle androoAGI) — no personal site or social profile linked from the repo, so no same-as: recorded, matching the zubair-trabzada precedent for a thinly-documented solo maintainer. Entity budget: 1 of 8 used. Touched: agent-permission-model (linked, no claim edit), openhuman (comparison paragraph on the source page — closest prior local-first Tauri multi-agent desktop, differentiated by memory-tree vs. spatial scoping), agent-workspace-linux and oh-my-pi/pi-coding-agent (linked as the other answers to the same permission question). index.md gained the source entry and the Person entry. Synthesis section The permission model becomes a floor plan added after The agent gets its own desktop — folds StarNet in as a third answer (spatial/visible) beside per-action approval (degrades with use) and the sandboxed-workspace ceiling, none of the three measured against each other. avoid-ai-writing pass (detect): prose checked against house style already established in this spoke's comparable pages (oh-my-pi.md, ai-marketing-claude.md); no P0/P1 flags beyond the existing em-dash-heavy convention already consistent across the corpus. ingest_effort: medium.

[2026-09-27] quality | quality cycle

Refreshed openmanus, perplexity-api-platform-skills, a2a-protocol (GitHub and spec re-checks). Drain: six pages hold. pilot-protocol and kimi-claw stay stale (not re-read). Scorecard in the hub quality-log.md.

[2026-09-27] ingest | Jev-Code-Reviewer

Routed by the hub (Telegram, egma-ai/jev-code-reviewer: "Review behavior, not just diffs. Jev prioritizes human attention; OpenAI explains the changes."). Fetched README, docs/ARCHITECTURE.md and the org page directly (WebFetch); no transient errors. Dedup: rg for "jev-code-reviewer", "egma-ai", "jev-reviewer" across the hub found no existing page; entity-index on "egma-ai" and "jev-code-reviewer" returned nothing ≥ threshold — new pages confirmed clean.

New page jev-code-reviewer: local CLI + agent skill + unpacked Chrome extension for reviewing a coding agent's own PRs by behavior rather than diff. TypeSafe's Jev model classifies each change unit P0/P1/P2; OpenAI writes the old/new-logic prose reviewers read; a local policy file and hash gate what's shown, and only P0 shows by default. No hosted backend, no GitHub App, posts nothing to GitHub. Recorded a connection the dedup sweep turned up rather than missed: the tool is named after TypeSafe's Jev model, which ../llm-providers-wiki already documents on its own terms (typesafe-ai, system-one-models-and-jev, T3, the vendor's self-graded 67.8% aggregate accuracy) — cross-linked both directions on this page rather than duplicated; not touched, per hub convention, in the other spoke's own files. Runner-up llm-providers-wiki logged on the source page itself (CodeRabbit angle and the Jev-model angle both declined — this page is about the harness, not the model). 81★/1 fork/10 commits, updated 2026-09-25. Tier T1: everything load-bearing is the project describing its own design (commands, architecture, the P0/P1/P2 mechanism, stated limitations); the tagline is marketing framing, not a measured claim, and no source measures reviewer time saved or what the priority split misses — recorded as an open gap, not folded into the tier. No published: — the source states no date beyond a © 2026 footer.

Entity discovery: new Organization page egma-ai (GitHub org, 3 followers, one public member @nischalj10 with no bio — not paged separately, plain mention only per the relevance gate). same-as: recorded as the org's own linked X account (@egma_ai), shown directly on its GitHub profile page — no Wikipedia/Wikidata entry found. Entity budget: 1 of 8 used.

Touched: stop-burning-tokens-on-code-review and deterministic-review-gate (Related-list cross-link only, no claim edit — Jev is a third position in the thread those two pages already hold). index.md gained the source entry and the Organization entry. Synthesis section Review keeps the model, and moves the queue local added after The permission model becomes a floor plan — folds Jev in as a third review-loop position (triage locally, post nothing to GitHub) beside "remove the model" (deterministic-review-gate) and the GitHub-Actions setup stop-burning-tokens-on-code-review priced, plus the Jev-model naming connection to llm-providers-wiki. avoid-ai-writing pass (detect): flagged and fixed two double-em-dash sentences; left the corpus's established em-dash/bold convention alone, per the starnet precedent. ingest_effort: low.

[2026-09-28] ingest | Agensh: Scaling Organizational Intelligence to 1,024 Agents

Routed by the hub (Telegram, from Lucsw; no runner-up spoke named). Fetched the arXiv abstract page and the v1 PDF (extracted locally with pypdf per the standing PDF-extraction note — the WebFetch PDF read came back metadata-only). Dedup: rg for "agensh", the arXiv id, and all seven author names across the hub found no existing page. entity-index on the authors and "Microsoft Research" found nothing ≥ threshold for the people; "Agensh" fuzzy-matched agent-sh (0.75) but that's an unrelated GitHub org (publisher of agentsys/agent-workspace-linux) — checked and ruled out, not a dedup hit. "Microsoft Research" matched the existing microsoft Organization node (0.67) — reused, no duplicate.

New page agensh (T1 — design/mechanism is the authors' own checkable artifact; the ProgramBench numbers are the authors' own run on their own five hand-picked hardest tasks with no third-party replication, flagged in the tier note rather than folded into it; published: 2026-09-22 from the v1 timestamp). Zhan, Song, Dong, Huang, Lian, Xia, Wei, Microsoft Research General AI group. A multi-agent harness that removes the central orchestrator: identical workers loop gather-context → claim-subtask → act → verify → merge, coordinating only through a shared Git workspace, a Mattermost message interface, and an append-only shared-context log (credited to DeLM). Scaling 1→128 agents on the five hardest ProgramBench tasks (GPT-5.6-sol, Copilot as executor) raises mean pass rate 19.31%→28.78%; scaling 1→1,024 on pandoc alone raises it 33.89%→55.06%. Cooperation forms are not specified in the (identical, worker-ID-only-varying) prompt — trajectories show peer negotiation at 8 agents, review chains at 32, standardized merge protocols at 128, duplicate specialist roles with failover at 1,024. Own stated limits recorded on the page: no orchestrator baseline run at matched agent count, no cost accounting for a 1,024-agent run against a 6h budget, one model/one harness/one benchmark family.

Entity discovery: no new Person or Organization pages. The seven authors are plain-mentioned rather than paged — MSR General AI researchers with no independent connections in this corpus beyond this one paper, judged sub-threshold for the relevance gate at ingest_effort: low (settings.json effortLevel: low for the active model, checked directly rather than trusted from memory). Microsoft Research reused the existing microsoft node: added a "Microsoft Research's own multi-agent research" paragraph there distinguishing the paper (orchestrator-free, research-side) from Foundry's product-side "multi-agent orchestration" APIs the page already documented, same company, different team and claim. Entity budget: 0 of 8 used (by choice, not by exhaustion).

Touched: agent-orchestration — new "Instances in this wiki" bullet framing Agensh as the sharpest counter-example to the page's own orchestrator-worker-fan-out framing, with the two scaling numbers; added to Related. jcode and virtual-biotech — Related-list link only, no claim edit (both already discussed at length in the synthesis sections this ingest extends). index.md gained the source entry and an addition to the microsoft org blurb.

Synthesis: two edits. In "Shared state is the next boundary," added a new paragraph after the jcode swarm discussion — Agensh as a third topology (removes the orchestrator rather than relocating coordination, as agent-orchestration vs. jcode's peer-to-peer cache-invalidation both do), noting it's the first source here to put a controlled number on scaling past the orchestrator bottleneck. In "At 37,000 agents the unit of design changes," noted agensh as the new numeric midpoint between orchestrate-100-agents-claude-code (100+, unmeasured) and virtual-biotech (37,000, role-differentiated, unmeasured against a baseline) — homogeneous like the former, measured unlike either — and flagged that its pass-rate curve is the closest thing to numbers on virtual-biotech's undemonstrated "debate improves outcomes" claim, while noting the two populations (converging-on-roles vs. pre-differentiated-by-data-source) aren't a clean comparison.

avoid-ai-writing pass (detect, on the new source page and the synthesis/microsoft additions): 14 em-dashes / 919 words on the source page, above the stated 1-per-1,000 target but consistent with this corpus's established convention (documented precedent: starnet, jev-code-reviewer entries above); no word-list (Tier 1/2) hits beyond "harness" used in its literal, corpus-standard technical sense, not the verb; no other P0/P1 patterns found. Left as-is per the skill's own instruction not to mangle accurate phrasing to satisfy a convention this spoke has already settled otherwise.

Checks: cd site && npm run check:wikilinks — clean, no unresolved links. npm run check:freshness run (frontmatter written: freshness: stable on the new page) — clean, no new stale/malformed entries. ingest_effort: low.

[2026-09-28] ingest | supermemoryai/company-brain: Open-sourcing our company brain

Routed by the hub (Telegram). Runner-up spoke ../research-wiki named (memory/PKM angle — team knowledge tracking, cross-link to gbrain if relevant). Fetched the GitHub README via WebFetch and a WebSearch to confirm supermemory's own official site/docs for same-as. Dedup: rg -il for "supermemory", "company-brain", "company brain", "slack bot", "slack agent" across the hub found no existing page — no collision, no re-ingest.

New page company-brain (T1 — official project repo/README, primary source; no published:, the README only says "discontinued a few weeks ago" and the footer shows © 2026, left off rather than guessed). An open-sourced (Apache-2.0) Slack "teammate" agent from supermemory, formerly a paid product with thousands of users, now a one-click deploy to the user's own Cloudflare account (Workers/Durable Objects/D1) while memory stays on supermemory's hosted service. Reads team Slack conversations into scoped memory (public-channel/private-channel/DM), answers grounded questions, and acts via MCP (GitHub, Linear, Notion, Google Workspace). Tool writes run under the asking user's own tool connections; a request needing a teammate's tool sends that teammate an approve/deny card first. 717★/105 forks/42 commits at read time — no independent usage report or audit found, every claim is the maker's own README.

Entity discovery: entity-index -- --query supermemory and -- "supermemory" both returned nothing ≥ threshold (best fuzzy hits 0.36-0.40, unrelated orgs) → new entity. Created supermemory (Organization, same-as: https://github.com/supermemoryai — verified via WebSearch/WebFetch against the official supermemory.ai docs, distinct from the page's own url:). Entity budget: 1 of 8 used (Layer 1: SoftwareSourceCode → copyrightHolder/owner). No Person entities — the README names no individual author.

Touched three Thing pages, all with substantive new sections (not link-only):

  • agent-permission-model — new "fifth position" section: Company Brain gates on whose authority a write uses (the asking user's own connections) rather than how often the agent is asked, a different axis from the four existing positions (per-action, batched, none, one-ceiling).
  • agent-scope-isolation — new "second instance" section: Company Brain independently converges on qm's public/private/DM partition and asker-scoped writes; contrasted qm's written-down SECURITY.md limitations against Company Brain's README stating the same scoping as a finished feature with no comparable gaps list.
  • agent-memory — new section: Company Brain decouples the hosted/self-hosted axis cloudflare-agent-memory and the self-run poles assumed travelled together — a self-hosted agent still depending on a hosted third-party memory store (supermemory).

index.md: new entries under SoftwareSourceCode (company-brain) and Organization (supermemory); updated the one-line summaries for agent-permission-model, agent-scope-isolation and agent-memory to reflect the above.

synthesis.md: new dated section "Two Slack agents converge on the same shape, and the permission dial gets a fifth position (2026-09-28)" tying the three Thing-page edits together as one finding (convergent design under independent authorship, still resting on one maker's README each). Added a runner-up note under the ../research-wiki cross-spoke-adjacency bullet: gbrain was the second-best fit on the memory/PKM angle, but the dominant substance (Slack bot, MCP tool actions, permissions graph) is this spoke's per the 2026-06-01 split — noted on the source page rather than split.

avoid-ai-writing pass (detect, on company-brain and supermemory plus the new sections in the three touched pages and synthesis.md): em-dash density ran 9.9-20.3 per 1,000 words across the new prose, above the stated 1-per-1,000 target but consistent with this corpus's established convention (same finding and same disposition as the agensh entry above); no word-list (Tier 1/2) hits anywhere in the new text; bold used at most once per section, as labels on genuinely list-shaped feature bullets (MCP/sandbox/skills/web UI), not decoration. Left as-is per the skill's instruction not to mangle accurate phrasing to satisfy a convention this spoke has already settled otherwise.

Checks: cd site && npm run check:wikilinks — clean, no unresolved links. npm run check:freshness (frontmatter written: freshness: volatile on company-brain, entity page supermemory left at the Organization default per ENTITIES.md) — clean, no new stale/malformed entries. npm run check:arguments (three existing pages refreshed) — clean, no contradictions. ingest_effort: low (settings.json effortLevel: low for the active model, checked directly per standing note).

[2026-09-28] ingest | auto-browser (LvcidPsyche)

Source: https://github.com/LvcidPsyche/auto-browser — an MIT-licensed, single-maintainer MCP browser control plane. One shared Playwright/Chromium session behind a 20-tool curated MCP profile (full profile adds diagnostics/audit/admin), noVNC human takeover, auth profiles, and model-agnostic adapters spanning OpenAI/Claude/Gemini/OpenRouter/Ollama/vLLM plus LangChain/LangGraph/CrewAI. Safety layer: approval-gate dashboard, Ed25519-signed "witness receipt" audit chains independently verifiable without trusting the controller, host allowlists, production-mode hard requirements, and a published adversarial self-audit (docs/audits/2026-08-execution-audit.md) disclosing controls that "reported success while doing nothing" before being fixed. Also ships an experimental "Convergence harness" (Stage 0) — Agent Skill Induction: verified task traces become staged, provenance-signed skill candidates requiring explicit manual promotion. Explicit non-goals: CAPTCHA solving, unauthorized scraping, deceptive bypass tooling. No runner-up spoke was named for this source.

Created auto-browser (SoftwareSourceCode, source: true, tier T1 — architecture, tool profile, licensing and the allowlist/production-mode mechanics are checkable artifact facts; the star count, the Glama badge and the self-audit's account of its own fixes are the maker's own claims, flagged as such rather than folded into the tier). Dedup: rg -il found no existing page for auto-browser or LvcidPsyche anywhere in the corpus; this is a new node, not a refresh.

Cross-linked into three existing Thing pages rather than left as a dangling source page:

  • agent-permission-model — new closing section, "A sixth axis: prove the gate held, after the fact." The five existing positions all vary how much oversight the agent gets before acting; auto-browser's witness-receipt chain and its own published self-audit answer a different question — whether a gate that looked like it held can be checked after the fact — which is a record layered on top of an authorization design, not a rival position on the same scale.
  • self-improving-agents — new bullet under Instances for the Convergence harness (Stage 0), a fifth design point on the drift-control question: signed provenance plus mandatory manual promotion, closer to zouroboros's consensus vote than Hermes's off-by-default review queue.
  • browser-use — one paragraph distinguishing this MCP-based, human-takeover tool from both the browser-use org's CDP-driven framework and Anthropic's own Browser Use API toolset, which the page already tracked as a two-way homonym collision; now a three-way adjacency.

synthesis.md: added "A browser control plane adds a sixth axis to the permission dial, and a fifth drift control (2026-09-28)" tying the two Thing-page findings together (witness-receipt verification as a new axis; Convergence harness as a fifth self-improvement instance), both flagged as resting on one maintainer's README plus a self-reported audit, with no independent verification of either claim. No runner-up spoke was logged for this source, so no cross-spoke-adjacency bullet was added.

Entity discovery: cd site && npm run entity-index -- "LvcidPsyche" returned nothing at or near threshold. The GitHub page gives no name, bio or other repos beyond this one project, so per ENTITIES.md's evidence-only rule no Person node was created for a bare pseudonymous handle — same disposition as corbell and ecc's solo maintainers, mentioned in prose rather than paged.

avoid-ai-writing pass (detect, then hand-fixed): the first draft of auto-browser ran 10 em dashes and 10 bold spans in ~900 words, well over the skill's targets. Rewrote to zero em dashes and one bold-free page; the new sections added to agent-permission-model, self-improving-agents and synthesis.md were written clean from the start (0 em dashes each, checked by grep). No Tier 1/2 word-list hits in any of the new prose.

Checks: cd site && npm run check:wikilinks — clean, no unresolved links. npm run check:freshness (frontmatter written: freshness: volatile on auto-browser) — clean. npm run check:arguments (three existing pages refreshed: agent-permission-model, self-improving-agents, browser-use) — clean, no contradictions flagged. ingest_effort: low (settings.json effortLevel: low, checked directly).

[2026-10-04] ingest | cc-thinking-skills refresh (github.com/tjboudreaux/cc-thinking-skills, hub-routed)

Already paged 2026-07-22 (cc-thinking-skills, T2, tj-boudreaux); refreshed in place, no new pages. Read the README, analysis/AUDIT.md and repo metadata via the GitHub API. The pack is now 28 skills, not 39: 11 were deleted after their mechanisms were absorbed (the August cutover), the survivors were lean-rewritten and are all manual-only. Audit verdict is unchanged in kind (Elevate 0, AUTO-RETAIN 0, validity provisional) but the numbers moved: best row thinking-scientific-method is now +4.0pp, under the repo's own +5pp margin, and the July +5.3pp / +8.0pp estimates are flagged provisional for scoring, placebo and archive defects. The follow-up portfolio-v1 and workflow-v1 studies recorded zero model calls (not a null). A recovery ledger admits 14 deleted result artifacts. Metadata: 1,471 stars, 172 forks, 4 open issues, MIT, v1.0.0 2026-08-04, last push 2026-08-07. Updated counts and added the caveat on tj-boudreaux, reasoning-scaffolds, index.md and synthesis.md. Gap-relevance: bears on the standing Benchmarks question; the spoke's one comparable ablation is now weaker, not stronger. Runner-up spokes: none. published: left off (no source-stated date for the repo page).

[2026-10-04] ingest | Subpool (github.com/gesta-run/subpool, hub-routed)

New source page subpool (T1 for design and metadata, nothing measured): self-hosted Go/React/Postgres gateway pooling Codex, GitHub Copilot and OpenAI-compatible accounts behind one OpenAI-compatible endpoint with per-employee keys, limits, expiry/revocation and subscription-first fallback to API accounts. Read the README and repo metadata via the GitHub API (112 stars, 12 forks, 0 issues, created 2026-09-02, v0.2.0 2026-09-29, Apache-2.0). Dedup: rg across all spokes found no prior subpool/gesta-run page; related gateway pages omniroute and ai-gateway live in ../llm-providers-wiki and are linked, not duplicated. Added index line and a synthesis section ("The seat becomes a pool"). No org page: gesta-run and its sponsor Ubang have no evidence beyond the README banner. published: left off (README states no date). Runner-up spoke: llm-providers-wiki (API access across providers; weaker, gateway not provider), recorded in the source page only.

[2026-10-04] ingest | FunctionGemma formatting and best practices (ai.google.dev, hub-routed)

New source page functiongemma-formatting-and-best-practices (T1, documentation, nothing measured, freshness: volatile). Covers the developer-turn trigger phrase, six control tokens and the <escape> token, the extra stop sequence, supported single-turn and parallel calls, untrained chaining and multi-turn, and the description-enrichment fix. Dedup: rg across all spokes found FunctionGemma only as one table row in ../llm-providers-wiki gemma-variants; no tool-calling protocol page here, so no Thing pages created. Added index line and a synthesis section. published: left off (live docs, no date). Runner-up spoke: llm-providers-wiki (Gemma variant), recorded in the source page only.

[2026-10-07] quality | Pilot Protocol re-read

Most-stale volatile page refreshed. pilot-protocol: both IETF drafts (draft-teodor-pilot-*-01, 2026-04-06) are individual submissions with no working group and are about to expire. Requests/s rose from ~33,157 to 52,734 while "agents online" stayed at ~219,000; both recorded as the company's own counters. The $4.5M figure is no longer on the site. Tier stays T3.

[2026-10-07] research | edge 14 closed — skill shadowing (Databricks)

New page skill-shadowing (T1, arXiv 2605.24050). It runs the sweep attention-is-all-you-have proposed: all skills in context, 52 / 102 / 202 skills, pass-rate drop 0.08 / 0.14 / 0.21. Edge 14 and hub Most wanted #28 closed; successor 14a asks for a mitigation arm across providers. Two inconsistencies inside the paper recorded on the page.

[2026-10-07] route | Personal Jarvis — GitHub

Routed from the hub. New page personal-jarvis (T3): a community multi-harness desktop app (Claude Code, Codex + 6 others in parallel panes), voice, computer use, MCP/skills marketplace. Folded into the vendor-runtime-governance thread beside nemoclaw (NVIDIA), Scout (Microsoft) and kimi-claw (Moonshot) as a fourth, non-vendor shape: governance by UI convention, no sandbox, no routed inference. No independent review or security audit found; macOS build unnotarized.

[2026-10-07] ingest | Moyai (BerriAI, GitHub)

Source: moyai, T3 (source), via Telegram. Self-hosted multi-harness background coding agent — sixth owner-type in the multi-harness-ownership thread (gateway company, gating by credential isolation rather than sandbox/identity/hosting/UI-convention). Carries the spoke's first head-to-head cost claim between two autonomous-coding products: ~79% cheaper than devin per BerriAI's own 31-day internal switch, self-reported. devin updated with the comparison; synthesis multi-harness thread extended. No new entities (BerriAI evidenced only by the repo; LiteLLM already noted in llm-providers-wiki's ai-gateway page, not duplicated here).

[2026-10-07] ingest | claude-marketing-research-skill (ishwarjha, GitHub)

Source: claude-marketing-research-skill, T3 (source), via Telegram. Fifth marketing- vertical instance of agent-skills, after agentic-seo-skill, claude-skills-ppc, ai-marketing-claude and geo-seo-claude. Adds a true zero to the evidence-collector-scripts spectrum ai-marketing-claude set up (0/4/6/89) — one SKILL.md, no scripts, competitor data from "web search." Small enough (55★, 3 commits) that rough edges are still visible: install commands name the wrong repo, credits still read "Created by [Your Name]." ai-marketing-claude and synthesis updated with the comparison. Runner-up: search-marketing-wiki (the content is general positioning/copywriting methodology, not search/SEO specifically — the agent-skill packaging is the dominant substance here).

[2026-10-08] ingest | From Spec-First to Enterprise-Ready: Extending GitHub Spec Kit (Microsoft for Developers)

Refreshed spec-kit in place with the enterprise extension model: presets/extensions/bundles/ workflows as four composable layers, a governed-vs-community catalog split, and a layered constitution (org → domain → team). Updated spec-driven-development with a new "from one repo to an org" note. No new pages — this is governance scaffolding on an already-held tool, not a new subject. First-party (Microsoft, GitHub's owner), conceptual/architectural, no benchmark.

[2026-10-11] lint | absence-claim drain (step 3a), 8 claims on 7 pages

Two claims retracted in place on agent-evals. The first said no prose-quality metric exists, but EQ-Bench Creative Writing v3 and LitBench (EACL 2026) both do; LitBench's best off-the-shelf judge agrees with human preference on 73% of story pairs. The second said nobody had published a run through i-have-adhd's release gate, but the repo's evals/RESULTS.md (2026-08-02) reports weighted score 4.045 → 4.473 and a failed gate. The i-have-adhd source page now carries the same correction. Still holding: claude-browser-use, diagram-design, dream-rsi (v2 posted 2026-10-06, same authors), harness-problem-hashline (Google forum acknowledges third-party-tool bans in general, but has said nothing about this account), and mast-multi-agent-failure-taxonomy.

Related pages

Links to