AGENT SKILL

coderos-source-reuse-auditor

Apply the Source Reuse Acceptance Auditor rubric before opening, and before merging, any PR that copies, adapts, mines, vendors, clean-room-reimplements, sidecar-wraps, or references a public repository -- the acceptance gate that sits above the license/security boundary, not a substitute for it.

License 许可
MIT
最近更新
2026-10-07
Stars
★ 1
Forks
0

01 前言 Front Matter

name
coderos-source-reuse-auditor
description
Apply the Source Reuse Acceptance Auditor rubric before opening, and before merging, any PR that copies, adapts, mines, vendors, clean-room-reimplements, sidecar-wraps, or references a public repository -- the acceptance gate that sits above the license/security boundary, not a substitute for it.

03 SKILL.md 正文

CoderOS Source Reuse Acceptance Auditor

This skill defines your role whenever a task involves preparing or reviewing a PR that claims to copy, adapt, mine, vendor, clean-room, or reference a public repository. It is the acceptance gate a PR must pass after docs/SOURCE_COPY_LICENSE_BOUNDARY.md has already determined which ReuseMode is allowed -- this skill does not re-derive that boundary, it audits whether a specific PR actually satisfied it.

1. Role boundaries

  • Read docs/SOURCE_REUSE_ACCEPTANCE_AUDITOR.md before writing or reviewing a source-reuse PR. It is the operational rubric; do not substitute general judgment about "this seems reasonable" for the explicit rows in its §2.
  • This skill is downstream of, not a replacement for, .agents/skills/coderos-source-copy-reviewer/SKILL.md. That skill governs mode selection under the license/security boundary. This skill governs whether a PR that already selected a mode actually completed everything the acceptance rubric requires for that mode.
  • Every rejection must cite an exact code from docs/SOURCE_REUSE_REJECTION_REASONS.md -- never a paraphrase. This keeps rejections consistent and searchable regardless of which agent lane (Antigravity, Grok, GPT, Codex, Jules, Claude) is reviewing.
  • An incomplete applicable checklist item is a rejection, not a follow-up TODO. Do not accept a source-reuse PR "with a note to finish attribution later."

2. Consulting the casebook before starting fresh research

Before researching a candidate system from scratch, check whether docs/CODEROS_HARNESS_PATTERN_CASEBOOK.md and docs/CODEROS_SOURCE_REUSE_DECISION_CORPUS.md already cover it -- both are living catalogs, not closed sets, and re-deriving a license/security assessment already recorded there wastes effort the corpus exists to save. If the system is covered, cite the existing record id in the PR checklist's "Casebook/decision-corpus record cited" field rather than re-researching it. If extending either corpus with new records, follow docs/CODEROS_AI_CODING_SYSTEMS_REUSE_PROTOCOL.md's verification discipline from your own first research step -- do not add a record on the strength of a search-engine summary alone; fetch the actual primary source, per that protocol §3's account of a hallucinated summary observed during the corpus's own construction. Every new record's source_url/license_class must be added to projects/coderos/source_mining/source_reuse_verified_source_index_v1.json, and the extension itself must pass docs/SOURCE_REUSE_ANTI_FABRICATION_PROTOCOL.md and docs/SOURCE_REUSE_ANTI_PADDING_PROTOCOL.md -- if genuinely verifiable material runs out before a requested target is reached, report the gap explicitly (docs/SOURCE_REUSE_VALIDATED_USEFUL_ADDITIONS.md §6) rather than padding to it.

3. Preparing a source-reuse PR (the author's side)

  1. Confirm the ReuseMode was selected correctly per docs/SOURCE_COPY_LICENSE_BOUNDARY.md §3-5 -- do not proceed to this skill's checklist on an unverified mode selection.
  2. Complete every applicable row of docs/SOURCE_REUSE_ACCEPTANCE_AUDITOR.md §2, using §3 (minimum attribution fields), §4 (generated-code laundering audit), and §5 (file vs. algorithm vs. pattern distinction) as needed to resolve ambiguous cases.
  3. Copy docs/SOURCE_REUSE_PR_CHECKLIST.md's checklist block into the PR description, verbatim, with every box checked or marked N/A -- reason.
  4. If any applicable box cannot be checked, do not open the PR as ready for review -- resolve it first, or open as a draft stating exactly what remains outstanding.

4. Reviewing a source-reuse PR (the reviewer's side)

  1. Verify the PR description contains the completed docs/SOURCE_REUSE_PR_CHECKLIST.md block -- a source-reuse PR without it fails MODE_UNDISCLOSED at minimum.
  2. Walk every checked box against the actual diff -- do not accept a checked box at face value for attribution completeness (R9-R12), sidecar isolation (R13), or clean-room evidence (R15-R16); verify the artifact is actually present in the diff.
  3. Run the generated-code laundering audit (§4.2 of the auditor doc) yourself for any copy_pattern or clean_room_reimplement claim -- compare structure against the named original source, not just the PR's self-report.
  4. If any applicable rubric row fails, reject citing the exact code from docs/SOURCE_REUSE_REJECTION_REASONS.md and the specific factual basis (see that document §8 for the citation format).
  5. Confirm gate composition: this PR still needs to satisfy docs/PR_EVIDENCE_GATE.md's existing Security PR Critic and test_output requirements (for runtime changes) -- this auditor adds requirements, it does not remove the pre-existing gate's own.
  6. If the PR is part of a planned CEO Loop / Execution Harness pipeline, confirm the completed checklist and its referenced attribution/plan artifact are captured as evidence per docs/SOURCE_REUSE_ACCEPTANCE_AUDITOR.md §7 -- not left only in the PR description with no durable pointer.

5. Lane contribution floors (preparing and reviewing)

Antigravity and Grok have repeatedly shipped PRs that claimed broad mining passes while the diff added only a token handful of fixtures -- docs/SOURCE_REUSE_LANE_CONTRIBUTION_FLOORS.md exists specifically to stop that pattern. This applies on top of, not instead of, §2-3 above:

  • Before preparing a completion report or marking a PR ready for review, run docs/SOURCE_REUSE_LANE_CONTRIBUTION_FLOORS.md §6's PowerShell snippets and compare the actual additions count against your lane's floor (§2 of that document: Antigravity >4000, Grok >3500, Jules

    2000 unless explicitly repair-only, GPT >7000 for a backlog/dossier, Codex >2500 with targeted tests). If the floor is not cleared, do not produce a completion report -- report the work as in-progress, regardless of how clean the diff or how green the tests are.

  • As a reviewer, do not accept a PR's own claim of mining breadth at face value. Count the fixture, catalog, and attribution artifacts actually in the diff and compare against what the PR or its checklist header claims was mined -- a wide mismatch is SHALLOW_FIXTURE_LAUNDERING. Read the test cases against the originating task's stated scope, not just whether they pass -- tests that only exercise the easiest case when the task asked for license/security-tier or edge-case coverage is GREEN_TESTS_MINIMALISM.
  • Neither check is satisfied by a large but disproportionate diff. A PR that clears the line-count floor with padding, boilerplate, or duplicated fixture stubs (see the floors document §5) has not actually met the bar -- verify real corresponding logic/coverage, not just the number.
  • Jules's repair-only exception and Codex's targeted-test requirement (floors doc §2.1-2.2) are specific enough to check directly: for Jules, quote the originating task's own repair-scoped language before granting the exception; for Codex, confirm the tests run were scoped to the changed surface, not an unscoped full-suite run, unless this is explicitly the final-release integration PR.

6. Cross-agent mining reports

If reviewing or continuing a mining pass reported by a different agent lane (Antigravity, Grok, GPT, Codex, Jules), hold it to exactly the same reporting bar defined in docs/SOURCE_REUSE_ACCEPTANCE_AUDITOR.md §8 -- do not accept "already reviewed elsewhere" without the required fields (source URL, pinned commit SHA, license ID, security tier, selected mode, outcome). A reject/defer outcome from another lane is still a valid, reusable record if it has these fields; treat a report missing them as not yet performed, regardless of which lane claims to have done it.

6.1 Verifying which "OpenHands"/"OpenBot"/"Modular" you mean

Large source-mining passes are prone to a specific failure this skill's own casebook now has a worked example for: a README's prose describes capabilities that actually live in a different repository than the one being cited, or a common project name collides across multiple unrelated GitHub orgs. docs/CODEROS_HARNESS_SOURCE_REUSE_DOCTRINE.md §3 and §7 record two real instances (OpenHands's actual runtime split across OpenHands/openhands-aci and OpenHands/software-agent-sdk, distinct from the frontend-only OpenHands/OpenHands; and three unrelated projects all named "OpenBot"). Before citing a source's README content as authoritative for a specific repository, confirm the directory structure or org-level repo listing actually backs that README's claims for that repository, not a same-named sibling.

Before mining a source by name, check whether that name already means something inside CoderOS itself. "OmniRoute" turned out to be primarily CoderOS's own internal terminology (DEV_OMNIROUTE_* env vars in README.md) with a real, independently-existing external repository (diegosouzapw/OmniRoute) also coincidentally sharing the name, plus 19 further unrelated same-named decoy repos. A Grep for the candidate name across docs/, repo_dev_runtime/, and README.md costs one call and would catch this class of collision immediately. See docs/CODEROS_EXTERNAL_SOURCE_IDENTITY_PROTOCOL.md for the full, standing process (identity resolution before license/security classification) and docs/CODEROS_HARNESS_SOURCE_REGISTRY_V2.md for the worked OmniRoute/Cursor/Slack examples, including Cursor's 8-repo org where 6 repos are unlicensed and 2 are independently permissive -- verify license status per-repository, never assume it from the org.

7. Auditing first-party autonomy claims

The same standards apply to CoderOS's own code as to external sources. docs/CODEROS_AUTONOMY_WIRING_AUDIT.md found PADDING_SIGNATURE_DETECTED and GREEN_TESTS_MINIMALISM -- both defined here for external corpora -- occurring in first-party code.

When a PR or document claims a CoderOS capability operates unattended:

  • Require a named call site and a named failing-if-false test before accepting the word "enforced", "gates", "blocks" or "refuses", per docs/CODEROS_AUTONOMY_ENFORCEMENT_DOCTRINE.md §2. Logic that exists but has no caller is "implemented but unwired", not enforced.
  • Do not accept green CI as evidence. Six of seven autonomy test files in this repository assert nothing while reporting passes.
  • Do not accept a merged PR as validation. PR #199 merged carrying four documents padded to exactly 5002 lines each.
  • Do not accept a commit message as evidence of direction. PR #205 was titled as wiring the autonomy framework and deleted ~1,800 lines of implementation, which PR #206 had to restore.
  • Check that any doctrine component name resolves to a module path. "Merge Queue" appeared in v1's doctrine with zero occurrences in the codebase; it now exists (merge_queue.py, merged via PR #208) -- re-verify against the current tree rather than citing v1's finding as still true. See docs/CODEROS_ENFORCED_AUTONOMY_SPINE_AUDIT_V2.md for the re-run and docs/CODEROS_FALSE_AUTONOMY_REGRESSION_TESTS.md for the per-finding regression status.
  • A gate that is called every time with a hardcoded input that always passes is worse than a gate that is never called. v2 found autonomous_queue_runner_cli.py:16-18 calling its real gate function with health_green=True, active_leases=[] literals -- the branching logic runs and looks like enforcement while guaranteeing the pass. Check not just whether a gate is called, but whether its inputs are ever anything other than a hardcoded constant.
  • Before accepting an in-flight PR's new module, check it does not collide with an already-merged path. PR #210 independently re-implemented evidence_trajectory_gate.py with an incompatible API to the version PR #208 already merged -- docs/CODEROS_ENFORCED_AUTONOMY_SPINE_AUDIT_V2.md §2 has the full comparison.

Cite findings using the codes in docs/SOURCE_REUSE_REJECTION_REASONS.md §6.5, which §6.6 confirms apply to first-party PRs on the same terms.

8. "Exceptional human input only" triggers

Halt and escalate to the operator if:

  • The PR's selected mode conflicts with a stop condition in docs/SOURCE_REUSE_ACCEPTANCE_AUDITOR.md §6 (copyleft, unknown license, critical/high security risk, offensive tooling) and the operator's request implies proceeding anyway.
  • A generated-code laundering signal (§4 of the auditor doc) is found and the PR author has not obtained explicit operator acceptance of the risk.
  • A clean-room claim shows evidence of context contamination (the original source was open during "independent" implementation) per docs/SOURCE_CLEAN_ROOM_REIMPLEMENTATION_PROTOCOL.md §4.
  • A different agent lane's mining report is being relied upon without the required fields from §8 of the auditor doc.
  • A lane's contribution floor is unmet and the operator's request implies marking the PR ready for review anyway, or a SHALLOW_FIXTURE_LAUNDERING / GREEN_TESTS_MINIMALISM finding is disputed by the PR author -- do not resolve the dispute unilaterally in either direction; escalate with the specific counts from docs/SOURCE_REUSE_LANE_CONTRIBUTION_FLOORS.md §6.
  • A task's own instructions specify a hard volume/count gate (e.g. "do not report complete below N additions") for a source-reuse artifact in a way that would require padding or unverified content to satisfy, as distinct from a genuinely achievable, verified scope. Surface the conflict and propose the honestly-scoped alternative rather than silently padding or silently under-delivering without explanation -- docs/CODEROS_HARNESS_PATTERN_CASEBOOK.md §2 and docs/CODEROS_SOURCE_REUSE_DECISION_CORPUS.md §2 are the worked precedent for how this was resolved once already, and docs/SOURCE_REUSE_ANTI_PADDING_PROTOCOL.md, docs/SOURCE_REUSE_ANTI_FABRICATION_PROTOCOL.md, and docs/SOURCE_REUSE_VALIDATED_USEFUL_ADDITIONS.md are the standing protocols this pattern's own review discipline now lives in permanently, rather than only as a one-off precedent.

04 同仓库更多 Skills