ci(claude-review): raise max-turns to 80, make review non-blocking

[giancolombi]

Problem

The Claude Code Review action has been failing across the org on non-trivial PRs:

"subtype": "error_max_turns", "num_turns": 31
error: Claude Code returned an error result: Reached maximum number of turns (30)
##[error]Process completed with exit code 1.

Root cause

The review reads the diff + CLAUDE.md + the changed files and posts one inline comment per finding, so each read/finding is its own turn. On a larger PR the reviewer exhausts the --max-turns 30 budget before it can finish; the SDK returns error_max_turns and claude-code-action surfaces that as a hard failure (exit code 1) → a red X on the PR, even though the review is advisory.

Fix (three guards, all at the step level so continue-on-error actually covers them)

1. --max-turns 30 → 80 — comfortable headroom over what a large review needs. Normal runs already finish well under 30, so cost is unchanged.
2. timeout-minutes moved from the job to the claude step — a job-level timeout cancels the whole job and red-Xes the PR regardless of continue-on-error; a step-level timeout is a step failure that continue-on-error can swallow. 20 min stays the wall-clock/cost bound.
3. continue-on-error: true — a max-turns exhaustion, the step timeout, or a transient API error all leave the posted inline comments + the track_progress sticky comment in place and the check stays green. The sticky comment still surfaces any error, so failures stay visible without blocking.

Context

This workflow is byte-identical across 10 org repos; this PR is part of an org-wide rollout. Canonical change & rationale: prodcycle/prodcycle#1554.

Note: this PR edits the review workflow itself, so claude-code-action's workflow-validation guard (which requires the workflow to match the default branch) will make this PR's own review step exit non-zero — continue-on-error keeps the check green. The fix is exercised normally once merged to main.

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR fixes the Claude Code Review action failing on non-trivial PRs by raising --max-turns from 30 to 80 and restructuring the timeout/error-handling guards so a turn-budget exhaustion or API error leaves the PR check green while still surfacing the failure via the sticky tracking comment.

• --max-turns 80: gives the reviewer enough budget for large diffs; normal runs stay well under 30, so cost is unchanged.
• Step-level timeout-minutes: 20 + continue-on-error: true: the previous job-level timeout cancelled the entire job (bypassing continue-on-error), red-X-ing the PR. Moving it to the step level means any failure is swallowed by continue-on-error while the outer 25-min job timeout acts only as a safety net for hangs in other steps.
• The 25 > 20 minute hierarchy is intentional and correct: if the outer job timeout fired during the review step it would bypass continue-on-error, so it must always be larger than the step timeout.

Confidence Score: 5/5

Safe to merge — the change is scoped entirely to CI configuration and makes the review step non-blocking without removing any visibility into failures.

The three guards (higher turn budget, step-scoped timeout, continue-on-error) are well-reasoned and correctly ordered. The 25-min job ceiling sits above the 20-min step ceiling, preventing the outer timeout from ever firing during a normal review. No application logic is touched.

No files require special attention. The only substantive change is in .github/workflows/claude-review.yml, which is straightforward CI configuration.

Important Files Changed

Filename	Overview
.github/workflows/claude-review.yml	Core change: raises --max-turns 30→80, moves timeout-minutes from job to step level, adds continue-on-error: true. The 25-min job timeout correctly sits above the 20-min step timeout so it never fires on a normal review run.
.github/workflows/publish.yml	Publish pipeline gates both npm and PyPI releases on the shared test.yml suite. Action references use floating tags (v5/v6) rather than SHA pins, but this appears to be a pre-existing pattern not introduced by this PR.
.github/workflows/test.yml	Test matrix for Node (22, 24) and Python (3.12, 3.13) with workflow_call hook so publish.yml can reuse it as a gate. No issues identified.
node/src/api-client.ts	Comprehensive API client with retry logic, chunked sessions, backfill handling, and configurable timeouts via environment variables. No issues identified in this read.
node/src/utils/vcs.ts	Async git wrapper for computing VCS-local-only paths with fail-safe returns and shallow-clone detection. Logic is well-documented and defensively written.

_{Reviews (2): Last reviewed commit: "ci(claude-review): raise max-turns to 80..." | Re-trigger Greptile}