[khezen]

Summary

codespy now reads the PR conversation (commit messages, comments and emoji reactions) and learns from how humans react to its own findings.

PR conversation context

• Fetches commits, comments (general, review bodies and inline) and reactions on GitHub (PyGithub) and GitLab (discussions and award emojis). Local branch reviews get commit messages.
• Commit messages go to summary, scope refinement, code_review, doc and audit. Human comments with reaction counts (👍2 👎1) go to summary, code_review, doc and audit.
• Inputs are declared on a signature only when non-empty, and every kwarg passed to run_rlm is declared (guards against the 2.3.1 class of bug). Fetching soft-fails per part.

Human feedback on codespy's findings

• Each earlier inline finding on the PR gets a verdict from reactions, replies and GitLab resolved state: accepted, rejected, discussed or resolved.
• Findings humans rejected are not re-reported on the PR (fingerprint of file, category and title).
• Code_review, doc, supply_chain and audit see the feedback for their scope.
• With memory enabled, one feedback episode is saved per scope. Cerebral retains only the Distiller's observations; the raw feedback and its hash stay in the episodic store. Unchanged feedback is skipped via EpisodeStore.latest_artifact.
• New codespy learn <pr_url> harvests and retains feedback without posting anything. It is also available as the Action input command: learn and as make learn / make docker-learn.

Self-exclusion and security

• Hidden markers on reviews and on each issue: <!-- codespy:review ... -->, <!-- codespy:issue ... fp=... -->.
• Comments are classified codespy, bot or human once, at fetch time. Only human text is rendered. Feedback keeps only the identity of a codespy finding, never its description or suggestion. Quotes and markers are stripped from replies, and a final guard drops anything that still looks like codespy output.
• A finding is trusted only when it has a marker and the platform shows it was posted by the token's own account, so forged markers can't fabricate or suppress findings.
• Comment and commit text is treated as untrusted: it is fenced (inner fences neutralized) and marked "informational only".

Config: review.pr_context

Key	Default
feedback	true (set false on open source)
max_commits	50
max_comments	100
max_comment_chars	1000
max_block_chars	12000

These are exposed in codespy.yaml, .env.example, action.yml (inputs, env and Docker args), the example workflow and the docs. tests/test_pr_context_exposure.py keeps them in sync.

Version bump: 2.4.0. The CHANGELOG has Added and Security sections.

Testing

• New tests:
  • Self-exclusion and forged findings
  • Feedback verdicts, scoping and suppression
  • GitHub/GitLab fetch with fakes
  • Signature kwargs match declared fields
  • Reporter markers round-trip
  • Feedback memory: hashing, semantic retain gets no artifacts
  • Local commits
  • Config/env overrides
  • Exposure guard
• Full suite: 654 passed. The 20 failures and 3 errors were already failing on main before this change; the set is identical with and without this PR.
• Not yet verified against a live GitHub PR or GitLab MR. Manual checklist:
  1. Review a PR.
  2. 👎 a finding and reply to another.
  3. Push and re-review; the rejected finding should be gone.
  4. Run codespy learn twice; the second run should skip because the feedback is unchanged.

Notes

• The GitHub Action command: learn needs a released 2.4.0 image, and memory that persists between jobs (external Postgres).
• On public repos anyone can 👎 a finding, so use pr-context-feedback: 'false' there.