Inferoa

An agent harness that treats the loop as an inference workload. Apache-2.0, TypeScript, a terminal UI with slash commands, published to npm and built on the vLLM stack. Its own one-line description: "Inference-native Tokenmaxxing Agent Harness for Loop Engineering."

512★ / 87 forks / 48 open issues / 3 watchers, created 2026-06-08, last pushed 2026-06-18, 132 commits — 128 of them one person's. Read 2026-09-01 from the repository, the GitHub API and the npm registry. T1 for the design; there is nothing to tier as evidence, because the project publishes no measurements at all.

The argument it makes

Every other harness in this spoke treats serving as somebody else's layer. Inferoa's premise is that this stops working once the loop is long:

But 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 it makes those visible to the loop rather than under it — "the loop sees serving, routing, context windows, prefix cache, multimodal endpoints, and self-hosted model paths." In practice that is four mechanisms:

Prefix-cache discipline — prompt epochs, deterministic tool schemas and bounded system sections, so the reusable head of the prompt survives iteration.
Bounded mutable context — compression, summaries, graph-shaped repo context, bounded history and bounded tool output, which is context-rot's mitigation applied per turn.
Routing as a loop decision — model paths chosen on cost, safety, privacy, capability, session pressure, multimodal need, and whether a self-hosted path suffices.
A /tokenmaxxing command that shows token and cost pressure across prefix-cache reuse inside a running session.

The stack it names under itself is entirely vLLM's: vLLM Engine and vLLM Omni for serving, vLLM Semantic Router for model selection, with CodeGraph and RTK for context selection.

Why the framing is credible without a benchmark

The author, xunzhuo, describes themselves on GitHub as "Intelligent Routing @vllm-project | Open Source & AI @AMD." The harness is written by someone who works on the router it sits on. That does not substitute for evidence, and it does explain why this is the first source here where harness design and serving mechanics are argued in the same document instead of across the boundary between two spokes.

Loop engineering with the completion promise formalized

/loop "carries an objective across loop tasks, verification, decisions, recovery, and completion evidence instead of stopping after the next answer." Set against loop-engineering's canonical instance — Ralph, where a completion promise fires when the agent judges itself done — this names the same slot completion evidence and puts independent feedback surfaces beside it: plans, tests, tool results and research metrics. That is the spoke's verification corollary implemented as a runtime, and it is the answer auto-company lacked when it removed the stop condition and had to legislate phases instead.

The word means the opposite of what the corpus recorded

tokenmining (Sam Black, July 2026) coined its name as the deliberate inverse of "tokenmaxxing", which that source glosses as spending tokens freely. Inferoa puts tokenmaxxing in its own tagline and means the opposite: "each turn is shaped to preserve cacheable prefixes, bound mutable context, expose token pressure, and pick the right inference path," with the stack table's targets reading "spend fewer prompt and tool-output tokens" and "avoid one expensive path for every turn."

Two sources, two incompatible readings of one word, and both are recorded. This is the second such collision the hub logged on 2026-09-01 — see vibe-coding, where the same shape appears with four senses. Neither term gets a shared page from a pair; a third instance of the tokenmaxxing sense split earns one.

What is missing, and what the numbers say about it

No benchmarks, no evaluation, no measurements. For a project whose entire pitch is efficiency — cache reuse, token pressure, routing spend — there is not one number in the README. Every claim is architectural.
It stopped. Ten days of work: first commit 2026-06-08, last push 2026-06-18, three GitHub releases on consecutive days (v0.14.17, .18, .19), and 119 npm versions published inside that same window. Nothing in the eleven weeks since.
3 watchers against 512 stars. Stars are cheap and watching is the signal of someone who intends to follow the work; the ratio is unusual enough to record without a reading attached.
The README installs inferoa@dev, saying npm latest is "reserved for stable releases." A latest tag does exist and points at 0.14.19, published the day work stopped — so the stable channel is real, one build old, and the documentation still routes users past it.

loop-engineering · agentic-coding-harness · tokenmining · context-rot · xunzhuo · agent-orchestration · vibe-coding

Cross-spoke: the mechanisms this harness exposes are ../llm-inference-wiki's subject — vllm the engine, kv-cache the thing prefix reuse is protecting. This spoke's synthesis has held that harnesses consume inference and neither spoke owns the other's subject; Inferoa is the first source that argues the boundary itself is the problem.