| name | cc-audit |
| description | SOTA-2026 deepresearch-backed audit of new Claude Code CLI releases. Triggered on every CC update (or when a third-party signal — tweet, blog, changelog mirror — claims a new feature). Fetches the official changelog + Anthropic docs + GitHub release notes, verifies claims against primary sources, decides for each change IGNORE / WRAP / INTEGRATE / DEFER. Records the verdict so the harness never ships a conflicting or redundant tool. |
| allowed-tools | WebFetch, WebSearch, Read, Grep, Bash, Edit, Write |
/cc-audit — audit new Claude Code releases before integrating
The harness's defense against tweet-bot hallucinations, vendor over-promises, and feature-overlap. Every new Claude Code version is suspected of conflicting with claudemax's existing skills/runtime until proven otherwise.
Trigger
- Manual:
/cc-audit <version> (e.g. /cc-audit 2.1.147) when a third-party signal claims a new feature
- Automatic (recommended): on every
claude --version bump detected by cmax doctor, queue a /cc-audit <new-version> task in the daily list
Pipeline
- Anchor the temporal context. Resolve "today" against the local time-anchor (
local_date from the SessionStart hook). Every web fetch records its accessedAt date.
- Pull primary sources, in parallel.
WebFetch https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md (the canonical changelog)
WebFetch https://code.claude.com/docs/en/release-notes/claude-code (Anthropic docs redirect)
WebFetch https://github.com/anthropics/claude-code/releases/tag/v<version> (GitHub release notes)
- Cross-check third-party claim (the tweet / blog / mirror that triggered the audit) against the primary sources. If primary sources do NOT corroborate, treat the third-party signal as
unverified.
- For each new entry in the changelog version range:
- Detect overlap with existing claudemax skills/runtime/hooks via
grep -rn <feature-key> /home/fer/Documents/minmaxingv2 --include="*.ts" --include="*.md" --include="*.json" --include="*.sh".
- Decide one of:
- IGNORE — irrelevant to claudemax's surface (e.g. an MCP integration the harness doesn't expose).
- WRAP — expose the new primitive behind an existing claudemax skill (e.g.
cmax doctor gains a --hooks flag when CC exposes a hook-introspection API).
- INTEGRATE — replace claudemax's own path with the new primitive (e.g. if CC ships a native bounded-concurrent subagent dispatcher, claudemax's
agent-teams.ts should delegate to it).
- DEFER — track in
CHANGELOG.md [Unreleased] under ## Deferred with reason + revisit date.
- Record the audit in
docs/CC_AUDIT_LOG.md as a dated row: version, source URLs, per-entry verdict, who/when/why.
Anti-patterns the audit catches
- Tweet-bot hallucinations. Example: @ClaudeCodeLog (2026-05-21) claimed 2.1.147 added a "Workflow tool for deterministic multi-agent orchestration (CLAUDE_CODE_WORKFLOWS=1)". Primary changelog had no such entry; the audit ruled
unverified and the harness skipped a fictitious integration.
- Feature creep without subtraction. When CC ships a native version of something claudemax already does (e.g. native goal-mode, native rate-limit-reset surfacing), the audit forces a
keep-ours / delete-ours / wrap-theirs decision rather than maintaining both paths indefinitely.
- Silent overlap. A new CC flag (
--dangerously-skip-permissions, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS) might already be relied on by claudemax. The audit makes that explicit.
Output shape (paste into CHANGELOG / docs/CC_AUDIT_LOG.md)
## CC v<version> audit — <date>
Primary sources verified:
- https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md (accessed <date>)
- https://github.com/anthropics/claude-code/releases/tag/v<version> (accessed <date>)
Per-entry verdict:
| Change | Source | Verdict | Action |
|---|---|---|---|
| <feature> | changelog L<n> | INTEGRATE | edit packages/runtime/...; PR #N |
| <feature> | changelog L<n> | WRAP | new skill /skills/<name>; PR #N |
| <feature> | changelog L<n> | IGNORE | irrelevant — no harness surface |
| <feature> | changelog L<n> | DEFER | revisit after <condition> |
Third-party claims investigated:
| Claim | Source | Corroborated? | Verdict |
|---|---|---|---|
| <claim> | <url> | NO | unverified — do not act |
When NOT to invoke /cc-audit
- Patch-level CC releases (e.g. 2.1.148 fixing a bash exit-127 regression from 2.1.147): cmax doctor's version check is enough; no audit needed.
- Pre-release or beta builds: not stable enough to integrate; the audit produces noise.
Coordination with other skills
/harness-audit is the periodic broader review of claudemax itself (skill catalog, hook coverage, rule wording). /cc-audit is the narrower upstream-change review.
- When
/cc-audit decides INTEGRATE, follow up with /cmax to run the actual deepresearch → multispec → /goal → verify pipeline on the integration work.
- When
/cc-audit decides WRAP, follow up with /agentfactory to scaffold the wrapper skill.