Enhance multi agent review

[theMickster]

🎟️ Tracking

📔 Objective

Making a number of minor changes to the skill.

1. Main purpose; the documentation about the skill stunk; now much better
2. There were a couple minor, but important callouts from the /skill-creator in-terms of wording around the prompt sent to the subagents performing the code review.
3. Added a suggested expansion of the why we don't want subagents running async in the background.
4. Took advice from both the /skill-creator and the plugin validation pipeline to really trim up duplicate or overly fluffy wording.
5. Enhanced the finding shape and added a sample report as recommended from the /skill-creator. Although I don't think Claude had any major problems performing these duties; I think they are reasonable enough additions and show future maintainers of this and other skills what to look for.
6. Enhance the model selection with more fine-grained control. Idea came about because of fable.
7. Removing the tech-lead subagent because there's minimal, if any gains using it. Instead, going to K.I.S.S. by crafting a stronger prompt and use a general subagent.

[github-actions]

Plugin Validation Summary — bitwarden-code-review

Overall: ✅ PASS (with minor warnings) — No critical issues, no blocking errors. All structural, skill-quality, and security checks pass. The findings below are best-practice/convention improvements only.

Validated against PR #146 changed files using the plugin-validator and skill-reviewer agents (plugin-dev) and the reviewing-claude-config security skill.

---

1. Plugin Structure Validation (plugin-validator)

Result: PASS with warnings. Manifest is valid JSON with all required fields; versions are fully consistent across all four tracking files; CHANGELOG is well-formed; all referenced component files resolve.

✅ Passing checks

• Version consistency (1.12.0) across all four sources:
  • .claude-plugin/plugin.json:3, .claude-plugin/marketplace.json:27, agents/bitwarden-code-reviewer/AGENT.md:3, CHANGELOG.md:8
• Agent frontmatter (AGENT.md): name bitwarden-code-reviewer is 23 chars, lowercase-with-hyphens (valid 3–50); model: opus valid; system prompt substantial (~160 lines). All 5 skills in the skills: field exist.
• Skill frontmatter (SKILL.md): correct name/description; name matches directory.
• CHANGELOG.md: Keep a Changelog format; ## [1.12.0] - 2026-06-18 entry present with populated ### Changed section.
• File organization: README.md present; no stray/unnecessary artifacts.
• Marketplace source points to ./plugins/bitwarden-code-review; name/description match the manifest.

⚠️ Warnings — should fix (non-blocking)

1. agents/bitwarden-code-reviewer/AGENT.md:4 — MINOR: Agent description uses a prose trigger list but contains no <example> blocks. The plugin-dev agent-development standard recommends <example> blocks demonstrating trigger scenarios.
   • Remediation: Add one or more <example> blocks, e.g. <example>user: "review this PR" → launches bitwarden-code-reviewer</example>. Functional without it.
2. agents/bitwarden-code-reviewer/AGENT.md (frontmatter) — MINOR: No color field. Optional but recommended by the standard.
   • Remediation: Add color: <blue|cyan|green|yellow|magenta|red>.
3. .claude-plugin/plugin.json:21 — MINOR: "agents" is a single string, while the sibling "commands" field uses an array. Agents are also auto-discovered from agents/, so the explicit registration is redundant.
   • Remediation: Convert to an array (["./agents/bitwarden-code-reviewer/AGENT.md"]) for consistency, or remove the field and rely on auto-discovery (consistent with the prior removal of the skills field in CHANGELOG 1.8.1).

ℹ️ Informational

• CHANGELOG.md:172 — the [1.3.3] entry lists its change as a bare bullet without a category subheading. Cosmetic; does not affect the current version's validity.

---

2. Skill Review (skill-reviewer) — performing-multi-agent-code-review

Result: PASS. No errors, no blocking warnings. Exemplary progressive disclosure; all references resolve.

✅ Passing checks

• Frontmatter: valid name + description; description is 490 chars (within <500 guideline) with concrete trigger phrases ("structured", "deep", "thorough", "multi-pass", "multi-agent", "commit range", "time window", "N most recent commits"). No "Load this skill" anti-pattern.
• Content quality: SKILL.md body ~2,630 words (within the 1,000–3,000 target); consistent imperative/infinitive writing style.
• Progressive disclosure: lean core delegating schemas/standards/templates to references/ and a rendered example in examples/. Strong DRY discipline (the "Review Rules" bundle is defined once and referenced by name).
• Reference integrity — all resolve. SKILL.md references six files (the brief listed four); both extras exist:
  • examples/sample-report.md ✅
  • references/discovery-standards.md ✅, references/finding-shape.md ✅, references/report-template.md ✅
  • references/evaluation-standards.md ✅ (extra), references/modes.md ✅ (extra)
  • All cited intra-file section anchors (Hygiene Sweep, Line Number Accuracy, Severity Levels, Do Not Flag, Confidence Scoring, Finding Shape, Source-Agent Friendly Names, Model Header) resolve.

⚠️ Warnings — optional polish (non-blocking)

1. SKILL.md:3 — MINOR: Description opens with an imperative ("Perform a...") rather than the canonical third-person "This skill should be used when..." form. The "Use when the user asks..." sentences already supply effective triggering, so this is stylistic only.
2. SKILL.md:24, :127 — MINOR (doc-only): Uses skill-runtime variables ${CLAUDE_PLUGIN_DATA}, ${CLAUDE_SKILL_DIR}, and $ARGUMENTS. Correct for skill context; consider a one-line maintainer note distinguishing these from ${CLAUDE_PLUGIN_ROOT} used elsewhere.

---

3. Security Validation (reviewing-claude-config)

Result: PASS — no security issues. Strong security posture.

✅ Passing checks

• No committed secrets / hardcoded credentials across all changed plugin files (api keys, tokens, passwords, private keys, GitHub PATs, AWS keys — none found).
• No settings.local.json committed.
• Least-privilege permissions: .claude/settings.json defines a comprehensive deny list (lines 3–45) blocking all PR/issue/repo/release/secret/org/workflow mutations and every write-method gh api call (POST/PUT/PATCH/DELETE). The single secret-scan hit — Bash(gh secret:*) at settings.json:30 — is a deny entry (a control that blocks secret access), not a credential.
• No dangerous auto-approvals: there is no broad allow list; the config is deny-only.
• Scoped tool access: agent tools and command allowed-tools are scoped to read-only gh/git patterns with :* suffixes; code-review-local correctly scopes Write to two named output files.
• Skill security design: the performing-multi-agent-code-review skill specifies an untrusted-input boundary (CWE-1427 prompt-injection handling), a WebFetch/WebSearch ban with rationale, context partitioning, and a zero-knowledge/threat-model security floor — all propagated verbatim to subagents.

---

Recommendations (priority order)

1. Add <example> trigger blocks to the agent description (AGENT.md:4).
2. Add a color field to the agent frontmatter.
3. Convert plugin.json:21 "agents" to an array, or remove it to rely on auto-discovery.
4. (Optional) Align the skill description's opening clause with the canonical third-person form.

None of the above block merge. The plugin is structurally sound, version-consistent, secure, and all referenced files resolve.

[github-actions]

🤖 Bitwarden Claude Code Review

Overall Assessment: APPROVE

Reviewed the documentation and skill-content changes to the performing-multi-agent-code-review skill: per-stage model flags with a security floor, replacing the bitwarden-tech-lead architecture subagent with a general-purpose agent, a new examples/sample-report.md, finding-shape JSON examples, a Model Header rule in the report template, and a rewritten README section. Cross-checked internal consistency across SKILL.md, the reference files, the README, and the CHANGELOG — the confidence threshold (≥80), security floor wording, and -mixed filename vs. Model Header rules all reconcile, and the intentional divergence between those two rules is explicitly cross-referenced. Confirmed all tech-lead references are removed from the skill (only the historical CHANGELOG entry documenting the removal remains) and no stale fable references survive. Version bump (1.12.0) and changelog entry are present and consistent across marketplace.json, plugin.json, and AGENT.md.

No findings at or above the confidence bar.

Code Review Details

No blocking, important, or refactor-level findings.

[theMickster]

🔍 Plugin Validation Report — bitwarden-code-review

PR #146 · Validated by plugin-validator, skill-reviewer, and claude-config-validator (reviewing-claude-config).

Overall: ✅ PASS — No critical or blocking errors. One major recommendation (agent <example> blocks) and a few minor advisories.

Recommended Actions (priority order)

1. (Major) Add <example> blocks to AGENT.md:4 and narrow the bare "review" trigger — then bump version + add CHANGELOG entry.
2. (Minor) Verify or replace ${CLAUDE_PLUGIN_DATA} / ${CLAUDE_SKILL_DIR} in SKILL.md.
3. (Minor) Document the gh api graphql write-path constraint in settings.json.
4. (Minor) Refresh the CHANGELOG "Version Format" note.

My responses to @claude

1. No, we are not making that change. That is way outside the scope. Just because I touch a file does not mean you get to go wild.
2. Yes, verified this twice. Your training data is old, and you keep screwing up this recommendation every 3rd run
3. No, we are not making that change. Outside the scope as well. That works as-is.
4. No, huh? Terrible, random advice.

[claude]

Claude finished @theMickster's task in 36s —— View job

---

Acknowledged — no changes to make

Understood on all four. These are your calls as the maintainer, and you've scoped this PR deliberately. Nothing to implement here.

1. <example> blocks / narrow "review" trigger — Agreed, out of scope. The AGENT.md change in this PR is a single-line version bump; touching the description to add example blocks would be unrelated scope creep. Skipping.
2. ${CLAUDE_PLUGIN_DATA} / ${CLAUDE_SKILL_DIR} — Confirmed. Both are in active use (SKILL.md:24 for the default output path, SKILL.md:127 for reference resolution) and you've verified they populate at runtime. Leaving as-is. That recommendation was wrong.
3. gh api graphql write-path note in settings.json — Agreed, out of scope and works as-is. The token scope is the real enforcement boundary; no doc change needed in this PR.
4. CHANGELOG "Version Format" note — Skipping.

No version bump or changelog entry required, since nothing is changing. The existing 1.12.0 bump and CHANGELOG entry already on this branch cover the documentation/skill work.
· enhance-multi-agent-review-q2-2026