Agents / corpora
AGENTS.md1 starsChanged 58 days agoApache-2.0View on GitHub
Frozen, generated eval data. Protected, because swapping a corpus is how you fake a pass. Four versioned corpora the harness loads directly, each regenerable byte-for-byte from a seeded generator. README.md documents what each corpus contains and the properties it measures rather than asserts. This file is the handling contract: what you may change, what regenerates, and why a human has to look. - Never hand-edit a corpus file. Every item is hashed in its manifest.json, so a hand edit fails…
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 repository39 other files it gives its agents.
- agent-core/AGENTS.mdAGENTS.md
- .agents/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- behavioral-regression/AGENTS.mdAGENTS.md
- claude-foundation/AGENTS.mdAGENTS.md
- config/AGENTS.mdAGENTS.md
- demo/AGENTS.mdAGENTS.md
- docs/AGENTS.mdAGENTS.md
- examples/AGENTS.mdAGENTS.md
- experiments/AGENTS.mdAGENTS.md
- flow-corpus/AGENTS.mdAGENTS.md
- flow-protocol/AGENTS.mdAGENTS.md
- openspec/AGENTS.mdAGENTS.md
- scripts/AGENTS.mdAGENTS.md
- skills/AGENTS.mdAGENTS.md
- tests/AGENTS.mdAGENTS.md
- claude-foundation/CLAUDE.mdCLAUDE.md
- merge-gate-auditorSkill
- update-executive-reportSkill
- update-executive-reportSkill
- architecture-drift-guardSkill
- commonSkill
- corpus-guardianSkill
- dataset-lintSkill
- deploySkill
- e2e-matrix-sentinelSkill
- eval-corpus-forgeSkill
- hierarchical-recursive-brainstormSkill
- model-benchSkill
- openai-judgeSkill
- openspec-implementation-reviewSkill
- openspec-peer-reviewSkill
- openspec-quality-planSkill
- pre-pr-gateSkill
- project-setupSkill
- quality-gateSkill
- refactoring-decomposerSkill
- repo-invariant-reviewSkill
- test-completeness-guardSkill
The file
We split this AGENTS.md into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 57
Never hand-edit a corpus file: Every item is hashed in its manifest.json, so a hand edit fails make corpus-check rather than quietly …
≈ “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.
Files
AGENTS.md — corpora
Frozen, generated eval data. Protected, because swapping a corpus is how you fake a pass.
Four versioned corpora the harness loads directly, each regenerable byte-for-byte from a
seeded generator. README.md documents what each corpus contains and the
properties it measures rather than asserts. This file is the handling contract: what you may
change, what regenerates, and why a human has to look.
Map
| Path | Role |
<name>/<version>/manifest.json | Schema version, generator seed, split counts, a hash per item |
<name>/<version>/items.json | The corpus itself |
<name>/<version>/eval/ | The harness-loadable records a config points its dataset at |
testgen/v1/ | Sixty synthetic focal methods, seeded mutants, gold obligations |
rca/v1/ | Root-cause-analysis corpus (ADR 0046) |
requirements/v1/ | Twenty-five epics with gold acceptance criteria and an evidence store |
answer_quality/v1/ | Fourteen question and answer envelopes replayed by the replay target |
Diagram
flowchart LR
accTitle: corpus lifecycle from generator to gate
accDescr: A seeded generator writes a corpus and its manifest, the corpus check regenerates and compares bytes, an eval config loads it, and any change needs the approval label.
GEN["scripts/gen_*_corpus.py<br/>seeded generator"]
subgraph CO["corpora/ (you are here)"]
MAN["manifest.json<br/>seed + per-item hash"]
ITEMS["items.json"]
EV["eval/ records"]
end
CHECK["make corpus-check<br/>byte-identical or fail"]
CFG["config/*_eval.yaml"]
RUN["eval-harness run"]
LABEL["eval-change-approved"]
GEN --> MAN
GEN --> ITEMS
GEN --> EV
MAN --> CHECK
ITEMS --> CHECK
EV --> CFG
CFG --> RUN
ITEMS --> LABEL
classDef here fill:#e8f0fe,stroke:#1a73e8,stroke-width:2px
class CO hereRules that bite here
- Never hand-edit a corpus file. Every item is hashed in its
manifest.json, so a hand edit failsmake corpus-checkrather than quietly changing a score. Change the generator in../scripts/, regenerate withmake corpus-write, and commit both. - This directory is protected, and the reason is specific. Swapping a corpus for an
easier one moves every number in a matrix without touching a scorer, a threshold or a gate
rule. It is the cheapest possible way to make a failing eval pass, so it requires the
eval-change-approvedlabel and a code owner. - A version directory is frozen. New material goes in
<name>/v2/, never on top ofv1; the configs and the reported numbers name a version and must stay comparable. - The seed is part of the data. Regeneration is only meaningful because the seed and the split keying live in the manifest. Do not change a seed to make a run look better, and do not reshuffle a holdout: the split is keyed by hash, not shuffled.
- Nothing host-specific or scraped belongs here. Real internal source would collide with the charter invariant that nothing host-specific is committed; generation avoids the question and buys reproducible difficulty strata instead.
Verify
make corpus-checkSubagents
| Task in this directory | Agent | Why |
| Find which configs, scorers and docs name a corpus before adding a version | explorer | Read-only; a corpus path is referenced from configs, tests and prose at once |
| Regenerate and confirm the bytes still match | test-runner | Has Bash; make corpus-check runs four generators and names the one that drifted |
| Review a generator change before requesting the label | narrow-critic | A tie-break or seed change silently restates every score and is not lintable |
See also
| Doc | Read it when |
README.md | You need what a corpus contains and which properties it measures rather than asserts |
../config/AGENTS.md | You are pointing an eval config at a corpus, or changing which one it loads |
../docs/decisions/0043-testgen-evaluation-seam.md | You are working on testgen/v1/ and need the seam it was built for |
../docs/decisions/0047-requirements-eval-matrix.md | You are working on requirements/v1/ or its evidence store |
../scripts/AGENTS.md | You are editing a generator and need its gate and coverage obligations |
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
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.

