CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

gamussa/coding-policy

Coding policy for Viktor Gamov's AI agents: language-agnostic quality rules, autonomous shipping discipline, and stack defaults for JVM, Swift, TypeScript, and Python

1.47x
Quality

93%

Does it follow best practices?

Impact

99%

1.47x

Average score across 10 eval scenarios

SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

REVIEW_DETAILS.mdskills/release/

PR Reviewer Mechanics

Reference for Step 4 of the release skill — how the two PR reviewers are wired.

Policy reviewer — one deployment

.github/workflows/review-codex.yml runs the OpenAI Codex CLI authenticated by a ChatGPT subscription (the CODEX_AUTH_JSON secret — no API key) via codex exec with the prompt and schema under skills/onboard-repo/templates/ (this repo) or .github/codex-review/ (a consumer repo scaffolded by onboard-repo). Consumer repos check out gamussa/coding-policy@main into .coding-policy/ and review against those rules; this repo reviews against its own rules/.

  • Trigger: pull_request opened / synchronize / reopened — reviews on open and re-reviews each pushed commit. Fork PRs are skipped (no secret access).
  • Authorship: submitted with the workflow's GITHUB_TOKEN, so the author is github-actions[bot] everywhere.
  • Lane marker: post-review.sh opens the body with <!-- coding-policy:codex-policy-review:v1 -->, so skills/release/review-lanes.sh tells this lane from another reviewer posting as github-actions[bot]. Unmarked older reviews fall back to this lane by login (docs/adr/0010-review-lanes.md).
  • Verdict: derived from per-finding severity by post-review.sh — any blocking finding posts REQUEST_CHANGES, advisory-only posts COMMENT, a clean pass posts APPROVE.
  • APPROVE limitation: github-actions[bot] cannot APPROVE (GitHub returns HTTP 422), so a clean pass falls back to COMMENT. A later COMMENT never supersedes an earlier CHANGES_REQUESTED in GitHub's merge gate, so Step 7 dismisses the stale request via skills/release/dismiss-stale-reviews.sh.
  • Auth / cost: the token is read only from the secret at runtime and deleted before the post step; Codex refreshes the access token in-memory. If the refresh token expires, the review fails loudly — re-seed with gh secret set CODEX_AUTH_JSON --repo github.com/<owner>/<repo> < ~/.codex/auth.json (the target from skills/release/origin-repo.sh).

Claude — advisory code-quality lane

.github/workflows/review-claude.yml (this repo only; onboard-repo does not scaffold it) runs the Claude CLI on a Claude subscription token (CLAUDE_CODE_OAUTH_TOKEN) and reviews for correctness, bugs, security, and test-coverage gaps that no rule targets.

  • Trigger: the same pull_request events as the policy reviewer; no review request is needed.
  • Lane marker: the body opens with <!-- coding-policy:claude-quality-review:v1 -->, so skills/release/review-lanes.sh classifies it claude-quality although it posts as github-actions[bot].
  • Verdict: always COMMENT while the lane is advisory (docs/adr/0014-retire-copilot-lane.md). Its body must be read (rules/reviewer-feedback-reading.md); it never gates the merge.

Copilot is no longer requested. A Copilot review in a repo that still enables it belongs to no lane: its body is read through other_reviews, and its CHANGES_REQUESTED gates until resolved through review (rules/review-severity.md).

skills

README.md

tile.json