kekzlkekzl/imp/AGENTS.md

imp

AGENTS.md77 starsChanged 5 days agoMITView on GitHub

Focused agent roles for working on imp; the cross-tool companion to CLAUDE.md. Every rule in CLAUDE.md applies to every role: single architecture, gated performance, GPU-free check before a GPU job, make dev to iterate and make build to measure, no bare make format, branch off main and never stack, no busy-polling, the file-size gate, no sudo. This file adds only what is role-specific: scope, tools, MUST and MAY NOT. Rules with no other home: - Performance comparisons are single-session only.…

Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,774 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository16 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- tests/CLAUDE.mdCLAUDE.md
- add-model-archSkill
- benchmark-cudaSkill
- building-and-testingSkill
- check-degenerationSkill
- codebase-auditSkill
- code-graphSkill
- docs-layersSkill
- docs-syncSkill
- find-stubsSkill
- numerics-bisectSkill
- quant-formatsSkill
- server-apiSkill
- shipping-prsSkill
- sm120-cuda-expertSkill

The file

We split this AGENTS.md into the rules it gives an agent

6 rules. About 1 is common, 3 are distinctive, and 2 are very specific to this project.

≈1
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.

Common
Distinctive
Very specific

Common:
Verify every finding against the real source: before acting on it (fan-out sweeps over-flag).
line 23
Distinctive:
Allowed tools: read/search tools, read-only sub-agents.
line 30
Very specific:
MUST: keep the build green; clean-reconfigure after a build-system change.
line 37
Distinctive:
MUST: run a before/after benchmark in the same session for any perf-affecting change (warm clocks,
line 48
Distinctive:
MUST: add an independent oracle with a justified, inline-documented tolerance.
line 56
Very specific:
MUST: keep docs in sync with code.
line 72

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.

Files

AGENTS.md

<!--
layer: L3
audience: agents
verified: 2026-09-11
commit: b5de0dd7
-->

AGENTS.md - subagent roles and guardrails for imp

Focused agent roles for working on imp; the cross-tool companion to CLAUDE.md. Every rule in CLAUDE.md applies to every role: single architecture, gated performance, GPU-free check before a GPU job, make dev to iterate and make build to measure, no bare make format, branch off main and never stack, no busy-polling, the file-size gate, no sudo. This file adds only what is role-specific: scope, tools, MUST and MAY NOT.

Rules with no other home:

Performance comparisons are single-session only. Compiler/cuBLAS autotuning makes cross-session numbers unreliable; compare results captured within one run. Decode tg128 is the headline signal; the gate is 8 % decode / 8 % prefill / 10 % own_peak_mb.
No version strings in markdown/configs: versions live in CMake and lockfiles only.
Verify every finding against the real source before acting on it (fan-out sweeps over-flag). Secrets via env only.

Roles

auditor

Scope: whole repo, read-only assessment.
Allowed tools: read/search tools, read-only sub-agents.
MUST: read docs/audit/SETTLED.md before forming hypotheses and generate against it (eight of the 2026-07-29 audit's thirteen hypotheses described duplication that earlier campaigns had already collapsed); verify each finding against source before reporting; write only a dated report under docs/audit/, or append to an existing running findings log where one owns the area (AUDIT.md for the memory subsystem, which records REFUTED results too); append new refutations to SETTLED.md with their anchors; rank by severity+effort.
MAY NOT: edit any code or config; act on an unverified sweep result; report a finding that contradicts a SETTLED.md entry without first disproving that entry's anchor; propose multi-arch or speculative rewrites.

build-engineer

Scope: CMakeLists.txt, cmake/, CMakePresets.json, Dockerfile, Makefile, dependency pins, .github/workflows/.
Allowed tools: edit (build/CI files), bash (configure/build).
MUST: keep the build green; clean-reconfigure after a build-system change; bump both dep-pin sites (CMake + Dockerfile) and both halves of a pin (tag + commit SHA) together; keep every workflow uses: on a 40-hex SHA; keep the single-arch gencode block intact.
MAY NOT: touch kernel/algorithm logic; add multi-arch paths; rename the Build CI job (branch-ruleset required check); put a --mount=type=cache on the build dir in the Docker build (the ccache mount is content-addressed and stays); collapse the Dockerfile's toolchain/builder split (make dev compiles in the toolchain stage) or let the two build paths diverge on compiler flags: a -march difference between them would silently confound every A/B.

kernel-optimizer

Scope: src/compute/** and src/quant/** only.
Allowed tools: edit (those dirs), bash (build, benchmark).
MUST: run a before/after benchmark in the same session for any perf-affecting change (warm clocks,
=3 trials, decode tg128); coherence-check after hot-path edits; stay single-arch.
MAY NOT: edit the build system, public API, or runtime orchestration; commit a perf change without an in-session A/B; insert temperature cooldown waits (water-cooled GPU, no throttle).

test-writer

Scope: tests/**.
Allowed tools: edit (tests), bash (build, run tests).
MUST: add an independent oracle with a justified, inline-documented tolerance; tests adapt to the engine, not the reverse.
MAY NOT: change src/ to make a test pass; assert exact-equality on known-nondeterministic paths (e.g. NVFP4 MoE atomic-scatter); commit a conflated/unsound golden.

benchmark-runner

Scope: scripts/bench_gate.sh, scripts/gen_perf_baseline.sh, the scripts/verify.sh gates, tools/imp-bench/, tests/perf_baseline*.json.
Allowed tools: bash (benchmark), edit (baselines + bench scripts).
MUST: warm the clocks, run >=3 trials single-session, enforce the 8%/8%/10% gate, sample nvidia-smi clocks during the run to rule out depressed host state.
MAY NOT: compare across sessions/days as if equal; refresh a baseline silently (must be stated in the PR).

docs

Scope: *.md, docs/, imp.conf.example.
Allowed tools: edit (docs only).
MUST: keep docs in sync with code; cite perf numbers from tests/perf_baseline.json rather than re-typing them; English.
MAY NOT: edit any .cpp/.cu/.h/.cmake; introduce version strings; invent benchmark numbers.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

It worked
Partly
It didn't

What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free