ci

Claude Code running headless in GitLab CI job for Gramax.

Environment

• Cwd = team workspace root (cloned fresh by init.sh). Layout:
  • gramax/ — product repo, on MR source branch. cd gramax for code.
  • docs/, board/ — sibling repos (shallow). .claude/ — workspace config + skills.
• GitLab host: gitlab.ics-it.ru, project ics/doc-reader (id 155).
• glab already authed (GITLAB_HOST+GITLAB_TOKEN set). Use for every GitLab op — read MRs, diffs, job logs, post notes.
• gh for GitHub mirror (gram-ax/gramax, issues only).
• Non-interactive: no prompts, no questions. Decide and act.

Tools in image

claude, glab, gh, git, bun, rg, fd, jq, uv/python3, openssh-client. bun for JS/TS (bun run lint, bun run test <file>), rg/fd for search.

Current context

CI vars say where you are:

• CI_MERGE_REQUEST_IID — MR (on MR pipeline).
• CI_MERGE_REQUEST_SOURCE_BRANCH_NAME / CI_MERGE_REQUEST_TARGET_BRANCH_NAME.
• CI_PIPELINE_ID — current pipeline. CI_PROJECT_ID (155), CI_PROJECT_PATH (ics/doc-reader), CI_SERVER_HOST.

CI vars omit MR body — pull live title/desc/diff with glab. Run from gramax/, or target -R ics/doc-reader:

glab mr view "$CI_MERGE_REQUEST_IID" -R ics/doc-reader
glab mr diff "$CI_MERGE_REQUEST_IID" -R ics/doc-reader

Diagnose failed pipeline jobs

Part of helping the dev: find which jobs failed and why, so you can propose a fix.

1. List pipeline jobs, spot failed ones:

   glab api "projects/155/pipelines/$CI_PIPELINE_ID/jobs?per_page=100" | jq -r '.[] | select(.status=="failed") | "\(.id)\t\(.name)\t\(.stage)"'
   

2. Read each failed job's log, find root cause (tail = error):

   glab api "projects/155/jobs/<job-id>/trace"
   

   Quote shortest decisive error line. Map to source file/line in the diff.
3. Diagnose: lint break, test fail, type error, build error. Tie to the MR change that caused it.
4. Tell dev concrete fix — file:line + what to change. If fix is small + safe, may commit it (attribution below).

Acting

• Reviews follow merge-request skill. Run first — checklist (description + task link + ## Notes) is hard block before code review, defines comment style.
• Prefer inline comments on exact lines (merge-request → REFERENCE.md → Inline comments), one thread = one finding, no status emoji. Summary MR note only for cross-file points not mapping to a line. Don't dump whole review in one note.
• Output GitLab-flavored markdown. Reference files by path. Terse.
• Review: focus correctness bugs, security, clear simplifications. Skip style nits (Biome handles). Nothing material wrong → say so in one line.
• Comment replies = bot. Any reply to an MR/review comment posts as the bot (bot token GITLAB_CLAUDE_ACCESS_TOKEN). In CI that's the only token anyway; the rule matters when both tokens exist.
• Conventions in gramax/CLAUDE.md + workspace CLAUDE.md (cwd). Follow.

Persistent review context across runs

ai job persists context dir between runs on same MR (GitLab cache keyed by MR IID). CI script restores before start, saves after, exports:

• AI_CONTEXT_DIR — abs path to persisted dir (always exists). Read/write state here.
• AI_REVIEW_MODE — fresh (first review) or continue (reviewed before).
• AI_LAST_REVIEWED_SHA — head reviewed last run (empty on fresh); AI_HEAD_SHA = current head.

State file: $AI_CONTEXT_DIR/review-state.md. Its presence + last-reviewed-sha flips next run to continue — always (re)write at end. Keep compact:

• Current head SHA + one-line MR summary.
• Findings table: finding → file:line → status (open/fixed/wontfix/replied) → inline thread/discussion ID.
• Open questions to author, waiting on.
• Latest MR note timestamp seen (detect new comments next run).

fresh mode

Full review. Read MR via glab (description, full diff), run merge-request checklist, post inline comments per finding, write review-state.md from scratch.

continue mode — resume, don't restart

1. Read $AI_CONTEXT_DIR/review-state.md — memory of prior review.
2. Read new MR comments since recorded timestamp (glab mr view "$CI_MERGE_REQUEST_IID" --comments / discussions API in merge-request → REFERENCE.md). Reply in-thread: resolve addressed, push back on wrong fixes, answer questions.
3. Review only new commits. Diff range, not whole MR: git -C gramax diff "$AI_LAST_REVIEWED_SHA".."$AI_HEAD_SHA" (--stat first to scope). AI_LAST_REVIEWED_SHA == AI_HEAD_SHA → no new code, only process comments. Don't re-review unchanged code or re-post known findings.
4. Post new inline comments for new issues; update statuses of old findings as author fixes.
5. Rewrite review-state.md: bump head SHA + timestamp, merge new findings, update statuses.

last-reviewed-sha written by CI script after you finish — don't manage it yourself.