framework

Aggregates telemetry into a weekly analytics report: tool usage, error rates, compaction patterns, and day-of-week activity.

Run the Model Change Protocol — 15-item assessment whenever you add, remove, or re-route any AI model in your project. Produces a decision log and assessment artefact. Mandatory before any commit that touches .claude/model-routing.json or agent/skill model declarations.

Shared termination + escalation loop for routine-native autonomy and /watch-pr. Stops on done/blocked_on_human/blocked_on_infra/superseded/gave_up; merges when safe.

Structured invariant verification for active multi-phase projects. Runs four independent checks: phase gates, datastore write health, required files, checkpoint freshness. Always advisory.

Checks config.yaml, fly.toml, and routing JSON diffs for missing sibling parameters — when one entry in a group gets a param, all siblings must too.

Runs the skill lifecycle curator: detects stale skills (30d/90d thresholds), archives inactive ones, identifies overlapping skills, and produces a curation report.

Diagnose guarded-path PR CI failures without pushing commits. Use when CI fails on a PR whose changed files match .ai/guards.json, or when an autofix loop detects a forbidden guarded-path touch.

Checks .md file diffs for prose claims about system behaviour and verifies each claim against the actual code or config.

Audit React/Next.js/Tailwind UI code against accessibility (WCAG), performance, and composition rules (Vercel-style) — for Next.js/Tailwind front-ends.

Detect Supabase migrations that are merged but not applied to live databases

Checks SQL migration diffs for NOT NULL without DEFAULT, missing IF EXISTS guards, irreversible DDL without down-migration, and missing transaction wrappers.

Checks PRs adding scheduled workflows, cron jobs, or new infrastructure for missing alerting, dead man's switch, runbook reference, and explicit ownership.

Checks config.yaml changes for missing capability-matrix entries and code changes for parameters unsupported by the target OSS model (response_format, functions, etc.).

Evaluate phase gate invariants before advancing a multi-phase project. Advisory only; returns PASS/FAIL checklist for one specific phase transition.

Write a structured Codex execution plan from the current Claude Code design discussion and print a copy-paste Codex bootstrap prompt. Does not invoke Codex.

Verify-before-you-code for third-party provider APIs (Sentry, Fly, Vercel, Doppler, Supabase, Neon, OpenAI/Anthropic). Use BEFORE writing or changing code that calls a provider SDK/API surface — fetch the live doc, confirm the call shape, record the verification.

Checks Python diffs for bare-except patterns, Python version compat issues, silent argparse flags, and missing mutually-exclusive groups.

Run static application security testing (Semgrep, plus CodeQL when available) over changed files or a path, triage findings by exploitability, and promote true-positives to the security-deep-dive agent.

Checks if a PR diff stays within the stated scope. Flags files modified that appear unrelated to the PR title/description, new abstractions with only one call site, and promised outcomes the diff does not deliver.

Flags diffs that introduce/modify security boundaries (silo, allowlist, entity isolation, input sanitisation) without at least one adversarial test that attempts to break the boundary.

Checks shell script diffs for missing timeouts, secret-passing anti-patterns, missing errexit guards, and soft-fail dependency checks.

Flags functions in src/ or scripts/ that were changed, have observable side effects, and have no corresponding test file updated in the same PR.

Quiet safety net for a non-engineer product owner (typically Clayton) working across product code. Use judgement to decide whether the user is about to author a genuinely load-bearing change that a solutions architect should see first — the sort of change that, if it goes wrong, is expensive or dangerous to unwind (data model, sign-in / session mechanics, payment / subscription state, a brand-new deployable service). If yes, pause and offer to draft a short Product Requirements Document (PRD) instead of coding. If no — and this is the default — stay silent and let the user work. Bias hard toward silent: it's better to let the user code and course-correct than to interrupt their flow. Do not invoke on ordinary product work (new pages, new endpoints, refactors, config, workflow YAML, styling, copy, bug fixes near sensitive areas). When genuinely uncertain, do not invoke.