cloudflare

Utility commands and scripts: /cws-submit, /ccgm-sync (reverse-sync local config to CCGM), /user-test, statusline, agent-team launcher.

Security vetting gate for new AI models: before any open-weights or hosted model is integrated into the Claude Code harness or local dev system, verify weights provenance and integrity, file-format safety, license and data terms, the serving path, and stage its agentic access.

Hard PreToolUse gate on AskUserQuestion that blocks questions whose decision context is invisible to the user. Three deterministic gates: deictic references to unseen context ('with that context', 'see above'), identical re-asks after the user pushed back in free text, and mid-workstream questions asked with no visible text emitted this turn. Block messages teach the exact recovery (visible context brief, then a self-contained question). Fail-open on transcript errors; CCGM_ASK_CONTEXT_OFF=1 escape hatch.

Captures permission events, tool failures, and user-correction signals; runs a daily analyzer via direct Anthropic API call; produces a local digest + opt-in Resend email; opt-in real-time security alerts and opt-in confidence-gated auto-apply. Cross-clone-safe via fcntl file locks. Webhook publisher seam is dormant by default.

Nightly, cost-capped dreaming service that mines Claude Code session transcripts for cross-session failure patterns and optimistically auto-integrates evidence-tagged memory-store changes (write-then-review, made safe by per-op-kind postures, a 24h dwell window, per-run blast caps, batch-anomaly detection, a windowed circuit breaker, an eval gate, a daily report, and git-backed rollback). Extends self-improving's learnings store with an out-of-band analyzer -- autoheal's capture-analyze-propose pipeline, retargeted at session transcripts instead of permission events. Provides the deterministic transcript miner (discover/mine/cluster/budget plus a schema-drift canary), the map-reduce analyzer (evidence bundle -> per-change proposals -> digest, with cost caps), the optimistic integration engine (per-op-kind postures, 24h dwell window, blast caps, batch-anomaly detection, windowed self-healing circuit breaker, plus an opt-in composite eligibility gate -- default off, add/supersede only -- scoring each mined add/supersede on four transcript-verified signals through a deterministic no-LLM waterfall before admission) with the /dream, /dream-digest, /dream-review, /dream-apply commands (/dream-review vetoes/reverts post-hoc; /dream-apply is the back-compat human-gated path) plus the nightly LaunchAgent scheduler, and the memory eval harness (with/without-memory A/B plus a saturation third arm, a four-bucket outcome classifier, and an end-to-end task exercising the analyzer's own mined output -- establishing the --gate contract the optimistic engine consumes), and a read-only reconciliation report comparing Claude Code's own harness auto-memory against the learnings store (import candidates + contradictions, appended to the daily digest -- never writes to auto-memory).

Evidence-before-claims methodology. Requires fresh execution of verification commands, reading full output, and confirming results before asserting completion.

Hard PreToolUse gate that blocks Edit/Write/NotebookEdit and mutating git commands (commit, add, stage, apply) while HEAD is on the repo's default branch. Fires before the first edit so no work is ever produced on main. Bypass-proof (exit 2); ALLOW_MAIN_COMMIT=1 escape hatch; exempts in-progress rebase/merge/cherry-pick, unborn HEAD, and direct-to-main allowlisted repos.

/transcript <url> grabs a YouTube transcript via yt-dlp AND dispatches a subagent to produce an opinionated implications doc against your project memory. Saves both files with matching slug+date so the pair correlates by name. --no-analysis flag for extraction only.

Structured resolver for PR review comments. /resolve-pr-feedback fetches unresolved review threads via GraphQL, triages new vs already-handled, and if 3+ new items arrive (or a cross-invocation signal fires) runs cluster analysis - categorizes each into 11 fixed concern categories and groups by category + spatial proximity. Clusters surface systemic issues instead of dispatching 10 one-off fixes. Parallel pr-comment-resolver subagents apply unambiguous fixes, post inline replies via gh api, and resolve threads; taste questions are batched for human decision.

SSH access to a remote server. Adds /onremote command for health checks and remote command execution. Delegates to Haiku.