[fullsend-ai-coder]

Summary

This changes the review agent's instructions so a review run always leaves a schema-valid agent-result.json, even when the session ends early.

• agents/review.md: as soon as the PR is identified, the orchestrator now writes a provisional failure result (reason: tool-failure). The final result overwrites it. The contextual-labels step now runs before the final write, and the text says outright that it never ends the run, with or without label_actions.
• skills/pr-review/SKILL.md: step 1 points to the provisional write. Step 7 writes the result after the label step, even when there are no labels. The skill stays under its skillsaw context-budget ceiling (17,101 tokens; the ceiling is 17,111).
• docs/review.md: explains the provisional result and adds a troubleshooting row for the Reason: tool-failure notice it produces.
• scripts/pr-review-provisional-result-test.sh (added to make test): runs the documented jq template, checks the output against schemas/review-result.schema.json with validate-output-schema.sh, checks that a final result overwrites it, and pins the guidance wording.

Why

Before this change, agent-result.json was only written in the last pr-review step. Every early exit reported in the issue left no file behind: a crash before synthesis, a kill, or a clean stop right after issue-labels returned "No label_actions to emit". Validation then failed and the post-script was skipped, so nothing was posted on the PR. With this change, those runs post a "PR was NOT reviewed" failure notice instead of failing silently.

Design notes

• The provisional result is a failure notice, not a partial review. Posting raw, unadjudicated sub-agent findings (option 1 in the issue) would conflict with the agent's constraint to report failure rather than a partial review, and could trigger the fix agent on false positives.
• Trade-off: a validated failure result counts as a passed validation. Today validation_loop.max_iterations is 1, so this changes nothing. Once run: a timed-out validation-loop iteration must not retry from scratch; export the budget and send a wrap-up signal fullsend#7042 brings back retries, a session that dies after step 1 will post the notice instead of retrying. Revisit this then.
• Out of scope: runner-side diagnostics for sessions that end without output (the issue's option 3) and the harness-level fallback (option 2). The runner owns those (run: a timed-out validation-loop iteration must not retry from scratch; export the budget and send a wrap-up signal fullsend#7042 and related issues).

Testing

• bash scripts/pr-review-provisional-result-test.sh: passes. Also passing: pr-review-challenger-contract-test.sh, pr-review-remediation-test.sh, pr-review-dependency-scope-test.sh, post-review-test.sh, review-body-fallback-test.sh.
• skillsaw 0.18.0 --strict: 0 errors.
• pre-commit couldn't fetch its hook repos in the sandbox (TLS error), so I ran the hooks directly: the pre-commit-hooks 6.0.0 checks, shellcheck 0.10.0.1, hack/lint-agent-docs, eval/lint-release-cases.py, and gitleaks via scan-secrets.
• make couldn't spawn /bin/sh in the sandbox, so I ran the test scripts above directly rather than through make test.

These paths are protected (agents/, skills/, scripts/), so a human needs to approve.

🤖 Generated with Claude Code

---

Closes #1169

Post-script verification

• Branch is not main/master (agent/1169-review-provisional-result)
• Secret scan passed (gitleaks — e41be7b41a17e17686425892a2f4968031f7a92b..HEAD)
• PR body secret scan passed (gitleaks — no-git)
• Final PR description secret scan passed (gitleaks — no-git)

[fullsend-ai-review]

🤖 Finished Review · ✅ Success · Started 4:15 AM UTC · Completed 4:23 AM UTC

Commit: dd50e9e · View workflow run →

Runtime: pi · Model: sonnet → claude-sonnet-5-5 · Effort: high · Cost: $1.81

[fullsend-ai-review]

Risk Assessment: elevated (3/5)

Details

Touches multiple protected paths and CI configuration in core review orchestrator files with high git churn and regression history, modifying existing behavior without a feature flag despite moderate change size and well-scoped issue alignment.

[fullsend-ai-review]

Review

Findings

Medium

• [protected-path] agents/review.md — This PR modifies protected paths: agents/review.md, skills/pr-review/SKILL.md, and scripts/pr-review-provisional-result-test.sh (new). The PR links issue Review orchestrator fails to synthesize sub-agent results into output file #1169 and its description explains the rationale (write a provisional failure result so early-terminated review sessions still leave a schema-valid agent-result.json), so the context is sufficient. Human approval is always required for protected-path changes, regardless of context.

---

Labels: The PR fixes a reliability bug in the review agent (issue #1169).