fix(ci): reset stale PR review context after history changes

[Sun-sunshine06]

Summary

• Reuse prior bot-review context only when the reviewed head is a verified ancestor on a complete, merge-free linear extension.
• Discard prior review IDs and run a fresh full review after rebase, force-push, rollback, merge commits, incomplete comparisons, or ancestry-check failures.
• Treat the current Files Changed list and current PR diff as authoritative in both Codex and DeepSeek review paths, with paginated PR file loading.
• Prevent an older matching review from suppressing a new run in an A -> B -> A history rollback.
• Clarify that bot reviews are advisory; a green workflow run only reflects automation completion.

Why

PR #298 exposed the failure mode: the bot compared diverged heads (04bcc817...e91a849, 13 ahead / 7 behind) and inherited findings for files that were not in the PR's current seven-file diff. The workflow previously treated any different reviewed SHA as a valid follow-up.

The new checks fail closed: old review text and old-head compare data are available only for a verified safe linear update. All other cases reset review context and review the current PR from scratch.

Validation

• npx vitest run tests/deepseek-common.test.ts tests/codex-pr-review-context.test.ts (11 tests passed)
• npm run typecheck
• npm run lint (0 errors; 8 existing warnings)
• node --check for both changed DeepSeek scripts
• Workflow YAML parse, Prettier check for tracked formatted file types, and git diff --check
• Live GitHub compare replay for PR Fix/wsl sandbox bash tool wiring #298: diverged range rejected; linear range accepted

The review event remains COMMENT; this PR does not turn advisory findings into a merge gate.

[github-actions]

Review mode: initial

Findings

• [MAJOR] listPullRequestFiles will return an empty array for any PR with files due to incorrect flatMap usage with --paginate --slurp – The gh api --paginate --slurp flag returns a single flattened JSON array of all items, not an array of per-page arrays. The added flatMap with Array.isArray check discards every item because each element is an object, not an array.

  Evidence: .github/scripts/deepseek-common.mjs:420-424

  const pages = JSON.parse(raw);
  return pages.flatMap((page) => (Array.isArray(page) ? page : []));

  This breaks the DeepSeek PR review path whenever the PR has at least one file — the file list will be [], so no file excerpts or patches are included.

  Suggested fix: Remove the flatMap and return JSON.parse(raw) directly:

  - const pages = JSON.parse(raw);
  - return pages.flatMap((page) => (Array.isArray(page) ? page : []));
  + const pages = JSON.parse(raw);
  + return Array.isArray(pages) ? pages : [];

• [MAJOR] Test mock for listPullRequestFiles does not reflect actual gh api --paginate --slurp output – The test mocks the command to return '[[{"filename":"first.ts"}],[{"filename":"second.ts"}]]' (an array of arrays), but the real output is '[{"filename":"first.ts"},{"filename":"second.ts"}]' (a flat array). The test passes only because the implementation accidentally flattens array-of-arrays, which is the opposite of real behavior.

  Evidence: tests/deepseek-common.test.ts:116-122
  Mock returns a nested array, but actual gh api --paginate --slurp returns a single array.

  Suggested fix: Update the mock to return a flat array:

  - return '[[{"filename":"first.ts"}],[{"filename":"second.ts"}]]';
  + return '[{"filename":"first.ts"},{"filename":"second.ts"}]';

  And adjust the expectation or the implementation accordingly (the implementation should not use flatMap — see above).

• [MINOR] gh pr diff --name-only result is loaded but never used – In .github/prompts/codex-pr-review.md line 55, the command gh pr diff "$pr_number" -R "$repo" --name-only is run, but the output is not stored in a variable nor referenced later. It wastes an API call.

  Suggested fix: Remove the line unless it is needed for another purpose.

• [NIT] In isLinearReviewUpdate, has_merge_commit field could be undefined for empty commit arrays – If comparison is provided but commits is empty (e.g., ahead_by=0), the function accesses comparison?.has_merge_commit which may be undefined. It will still compare === false, so it works, but explicit handling would be clearer.

  Suggested fix: No change needed, but consider adding a guard for clarity.

Summary

• Review mode: initial
• Review policy: advisory — the check reflects automation health/completion only; it does not approve the PR or resolve findings.
• Major issues: The listPullRequestFiles pagination implementation is incorrect and will break the DeepSeek review path. The accompanying test is also wrong. These must be fixed before merging.
• Minor: Unused command in prompt.
• Other logic: Ancestry check and context-reset logic appear sound and correctly handle failures by discarding context.

Testing

• New tests for isLinearReviewUpdate and PR file pagination were added, but the pagination test is invalid due to a mismatched mock. The existing tests for linear update correctness are good.
• Suggested tests: Once the pagination bug is fixed, verify that listPullRequestFiles works with real gh output (e.g., by running against a PR with >100 files). Also add a test for the prior_context_discarded and is_follow_up_review output logic in the workflow.

Open Cowork Bot

[Sun-sunshine06]

I verified the pagination findings against the real GitHub CLI/API before changing the implementation, and they are false positives.

Using this PR with per_page=2 to force three pages:

{"outerIsArray":true,"outerLength":3,"firstIsArray":true,"pageLengths":[2,2,2],"firstPath":".github/prompts/codex-pr-review.md"}

gh help api also states that --slurp wraps all pages of JSON arrays or objects in an outer JSON array. Running the committed helper directly against PR #301 returns all six changed paths, so the nested mock and flatMap are intentional and match real output.

The gh pr diff --name-only command is also intentional: this is an agent prompt, so its stdout is supplied directly to the reviewing model as the authoritative current Files Changed list; it does not need a shell variable to be consumed.

No code change is needed for these findings. The live commands used were:

gh api --paginate --slurp "repos/OpenCoworkAI/open-cowork/pulls/301/files?per_page=2"
node --input-type=module -e "import {listPullRequestFiles} from './.github/scripts/deepseek-common.mjs'; console.log(listPullRequestFiles('OpenCoworkAI/open-cowork', '301'))"