Add Claude Code Action-compatible review agents and improve /review-pr

[dceoy]

Summary

• Add three Claude Code Action-compatible core review agents: code-quality-reviewer, test-coverage-reviewer, and documentation-accuracy-reviewer
• Rewrite /review-pr command and skill to deterministically run 5 core reviewers for all, plus pr-review-toolkit specialty agents triggered by diff content
• Add new aspect keywords: quality, coverage, docs, documentation
• Normalize findings across all agents with file/line/severity/source/message structure
• Validate every finding against the PR diff before posting; move unanchorable findings to summary body
• Add retry logic and top-level comment fallback when the reviews API call fails
• Update SKILL.md and README.md to document both reviewer sets, all aspects, and posting policy

Changed files

File	Change
.opencode/agents/code-quality-reviewer.md	New — quality, edge cases, robustness, type safety
.opencode/agents/test-coverage-reviewer.md	New — missing tests, brittle tests, coverage gaps
.opencode/agents/documentation-accuracy-reviewer.md	New — docs, README, API docs accuracy
.opencode/commands/review-pr.md	Rewritten — 8-step orchestration, dual reviewer sets, normalization, anchoring, fallback
.opencode/skills/review-pr/SKILL.md	Rewritten — full documentation of both sets, all aspects, posting policy
README.md	Added Review commands section with aspect table and reviewer descriptions
.github/workflows/opencode.yml	Updated model name to match README example

Acceptance criteria

• .opencode/agents/code-quality-reviewer.md exists
• .opencode/agents/test-coverage-reviewer.md exists
• .opencode/agents/documentation-accuracy-reviewer.md exists
• Existing pr-review-toolkit agents remain unchanged
• /review-pr all deterministically includes 5 Claude Code Action-compatible core reviewers
• /review-pr still supports pr-review-toolkit specialty aspects
• Findings are normalized and deduplicated before posting
• Inline comments posted only when safely anchorable
• Unanchorable findings moved to summary
• API failure falls back to top-level PR comment
• README and skill are updated consistently
• QA passes (prettier, yamllint, zizmor, actionlint, checkov all clean)

Test plan

• Trigger /review-pr all on a PR touching code, tests, and docs — verify all 5 core reviewers run
• Trigger /review-pr security performance — verify only those two agents run
• Trigger /review-pr tests docs — verify test-coverage-reviewer and documentation-accuracy-reviewer run
• Trigger /review-pr simplify — verify code-simplifier runs without posting a review
• Verify inline comments anchor to diff lines and unanchorable findings appear in summary body
• Simulate reviews API failure — verify fallback to top-level gh pr comment

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: bf7bd78544

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

OpenCode PR Review

Reviewed 7 files across 6 areas: code-quality-reviewer, performance-reviewer, test-coverage-reviewer, documentation-accuracy-reviewer, security-code-reviewer, code-reviewer.

General observation

The new orchestration expects every agent to return normalized {file, line, severity, source, message} YAML, but the existing pr-test-analyzer.md and code-reviewer.md agent definitions do not specify that format. Update those agent files to match the normalized contract, or have the orchestrator parse their prose output.

[opencode-agent]

Posted a review on PR #9: #9 (review)

Summary:

• Ran 6 reviewers: code-quality-reviewer, performance-reviewer, test-coverage-reviewer, documentation-accuracy-reviewer, security-code-reviewer, and code-reviewer.
• Security found no issues.
• 10 important findings and 4 suggestions posted as inline comments.
• Key themes:
  • Inconsistencies between README.md, .opencode/commands/review-pr.md, and .opencode/skills/review-pr/SKILL.md on which reviewers run for code, tests, and all.
  • pr-test-analyzer is referenced in the normalized flow but its agent definition still uses a non-normalized output format.
  • code-reviewer is listed as a conditional specialty reviewer without a trigger condition.
  • Filtering rule 3 mixes file-set validation with line anchoring.
  • Suggestions around CI model cost, frontmatter tests, and agent-existence tests.

github run

[dceoy]

Review + triage

CI is green, but I would not merge this yet. The existing important comments are mostly valid and point to the same underlying consistency issues.

Keep / fix before merge

1. Normalize the output contract for every agent that can enter the review flow. The command/skill now require {file, line, severity, source, message}, but several existing specialty agents still define prose or incompatible output formats. This applies not only to code-reviewer and pr-test-analyzer, but also to silent-failure-hunter, comment-analyzer, and type-design-analyzer if they remain callable from /review-pr all or explicit aspects. Either update those agent files or make the orchestrator explicitly translate legacy outputs before filtering/posting.

2. Make aspect-to-agent routing deterministic and document it consistently. Replace and/or with the intended set. If code runs both code-reviewer and code-quality-reviewer, state that in the command file. If tests/coverage runs both test-coverage-reviewer and pr-test-analyzer, make /review-pr all, SKILL.md, and README.md match that behavior.

3. Fix the changed-file filtering rule. The current wording says to drop findings when the file or line is not in the changed-file set, but the changed-file set contains paths, not line numbers. Restrict that filter to file membership and leave line validation to the anchoring step; otherwise valid findings can be discarded before fallback-to-summary logic runs.

4. Clarify code-reviewer in /review-pr all. It is listed under conditional specialty reviewers but has no trigger condition. Move it to unconditional reviewers or add a clear condition such as “always run for AGENTS.md/project-guideline compliance.”

Lower-priority / optional

• The model default change to kimi-k2.7-code is a product trade-off, not a correctness blocker; it is already resolved.
• Tests that validate agent frontmatter and command/skill references would be useful guardrails, but they are optional unless this repo wants schema validation for .opencode/ assets.

Suggested implementation direction

I would fix this by making a single source of truth for the aspect matrix in the command/skill text, then updating all review-capable agents to emit the same normalized YAML shape. After that, adjust README.md examples so each row names the exact agents that run.

[dceoy]

Triage complete (commit f1a209f)

All important findings are addressed. Here's what was done:

Fixed:

• Output contract — updated pr-test-analyzer, code-reviewer, silent-failure-hunter, comment-analyzer, and type-design-analyzer to emit normalized {file, line, severity, source, message} YAML, matching the orchestrator contract.
• Routing ambiguity — replaced and/or with and for the code aspect (code-reviewer + code-quality-reviewer) and tests/coverage aspect (test-coverage-reviewer + pr-test-analyzer).
• code-reviewer in all — moved from conditional specialty to unconditional core reviewers (always runs for AGENTS.md/project-guideline compliance).
• pr-test-analyzer in all — added as a conditional specialty reviewer triggered when test files change.
• Filtering rule 3 — changed to drop by file membership only; line validation left to the diff-anchoring step. Fixed in both review-pr.md (command) and SKILL.md.
• README /review-pr tests docs row — updated to name all three agents: test-coverage-reviewer, pr-test-analyzer, and documentation-accuracy-reviewer.

Deferred (suggestions, no test framework currently):

• Frontmatter validation test for new agent markdown files
• Agent-reference existence test from command table
• Agent-reference existence test from SKILL aspect table

QA clean (prettier, yamllint, zizmor, actionlint, checkov all pass).

[dceoy]

On the suggestion threads (no tests for agent frontmatter / agent file references / skill references): deferred. This repo has no test framework for validating .opencode/ assets. All three would be worthwhile additions to qa.sh in a follow-up PR.

[dceoy]

Follow-up: orchestration quality and noisy output

Pushed c85a4ed to address the remaining review comments. The earlier commit f1a209f already normalized code-reviewer/pr-test-analyzer, removed and/or from routing, moved code-reviewer to unconditional core reviewers, and fixed the file or line filtering wording. This commit closes out the rest:

Normalized agent outputs

• performance-reviewer and security-code-reviewer now emit the same file/line/severity/source/message YAML as every other review-capable agent (the impact/vulnerability fields were folded into message). All agents invoked by /review-pr now share one schema.
• code-simplifier gained the missing permission frontmatter key so all agents pass the required-keys check.

Duplicate/noisy output suppression

• New filtering rule "Aggregate by root cause" in review-pr.md and SKILL.md: one representative inline comment per root cause; remaining occurrences collapsed into the summary body. Cross-document/cross-file consistency problems are summarized once instead of repeated at each location.
• Explicit posting policy in both docs: anchored → inline, unanchorable → summary, duplicated → one representative + summary listing affected files.
• Optional guardrail suggestions (e.g. frontmatter/reference tests) are explicitly downgraded to suggestion severity.

Changed-file filtering vs anchoring

• Filtering rule 3 now states it checks file membership only and that the changed-file set contains paths, not line numbers. Line validation is left to the anchoring step.
• Anchoring step clarified that unanchorable findings on a changed file are moved to the summary, not dropped.

Deterministic routing and docs

• README.md aspect table now names the exact agents that run for each keyword (e.g. /review-pr code → code-reviewer, code-quality-reviewer; /review-pr tests docs → test-coverage-reviewer, pr-test-analyzer, documentation-accuracy-reviewer). README, SKILL.md, and review-pr.md now agree on the aspect matrix and /review-pr all.

Guardrails

• Added .agents/skills/local-qa/scripts/validate-opencode.sh (wired into qa.sh): verifies every .opencode/agents/*.md has valid YAML frontmatter with name, description, mode, permission, and that every agent referenced in review-pr.md and SKILL.md exists under .opencode/agents/.

Validation

QA passes: prettier, yamllint, zizmor, actionlint, checkov, and validate-opencode.sh all clean. The PR stays scoped to .opencode/ review orchestration and docs (the unrelated action.yml/workflow formatting that prettier --write . touched was reverted).

The OpenCode review comments on this PR should now be either addressed or obsolete.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1367e629f3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[dceoy]

No noteworthy issues found. Previous feedback appears addressed, and the current CI/CD run is green.