mtk-agent-toolkit

Corrective batch of multiple small independent fixes (apply review findings, several/multiple fixes across >3 files, no new contract or architecture) — broader than fix, lighter than implement.

One-time repo setup that detects tech stack, audits the codebase, pulls coding guidelines, and generates a project-specific CLAUDE.md

Full feature implementation loop orchestrating planning, batching, verification, and review skills

Use when the task is a new feature, breaking change, multi-file change, or any work where approval should happen before coding begins.

Use to create, update, and read durable workflow state under .mtk/workflows/ so orchestration survives compaction, crash, and session handoff.

Use after a spec is approved and before multi-file implementation begins, to break work into verifiable batches with checkpoints.

Use before approving a spec or starting a multi-file implementation to confirm no existing skill, helper, handler, or lesson already covers the proposed work.

Use instead of incremental-implementation for 3+ batches, 6+ non-mechanical files, or non-none security_impact — one fresh implementer subagent per batch with orchestrator drift checks.

Use before reporting any task, batch, or fix as complete — requires fresh execution evidence for every completion claim.

Use when starting a session, switching between planning/implementation/review phases, entering unfamiliar code, or when output drifts from project norms.

Unified entry point that routes natural language requests to the right MTK skill

Use when a change stays within 1-3 files and does not need a formal spec (bug fixes, validation tweaks, small config changes) — self-escalates to implement if scope grows.

One-stop setup entry point that bootstraps a repo or re-runs architecture audit

Fast security-focused review of staged changes before every commit, checking only critical compliance rules

Audit the repo to extract architecture principles, or with --merge unify audits from multiple repos into a single team-wide document

Judges the codebase against the agreed architecture-principles.md/conventions.md and reports drift as graded, read-only work items — never auto-fixes.

Refresh generated setup artifacts (principles, conventions, detected tools, AGENTS.md, configs, indexes) against codebase drift — proposes diffs for engineer-edited files instead of overwriting.

Use when context nears limits, before ending a long session with in-progress work, or handing off to a teammate — captures session state into a recovery artifact so a new session resumes cleanly.

Periodic AI-readiness report — scorecard against 12 named assets plus PR review mining from the last N merged PRs.

Use when the engineer corrects your approach, says "no", "not like that", "stop", or redirects your work — capture the correction as a reusable lesson.

Use when you struggle with the same sub-problem 2+ times then find a working approach — captures the failed attempts and the fix as a reusable lesson, no engineer correction required.

Promote a personal lesson from .claude/lessons/personal.md to the team-wide tasks/lessons.md, optionally rewording for team applicability.

Run an MTK install health check (core files, components, hooks, integrity) with PASS/WARN/FAIL diagnostics, --json for CI, and --fix for safe auto-repairs.

Use when the engineer asks what MTK configuration is active — reports tech stack, references, linter packs, domains, hooks, and rules.

Use when a decision needs current external/web info (library best-practices, version behavior, migration guidance) rather than local sources — runs grounded web research and returns a cited brief.

Use when framework, SDK, or library behavior is unfamiliar or version-sensitive — verify from local authoritative sources (decompiled/SDK code, in-repo docs) before implementing.

Provides TypeScript/JavaScript build/test commands, ORM guidance, framework patterns (React, React Native/Expo, Next.js, Tauri, Node), and reference paths for workflow skills.

Use when creating a new skill for this toolkit — ensures the skill follows the anatomy, CSO principle, and is tested against agent rationalizations before shipping.

Provides .NET-specific build commands, test commands, ORM guidance, framework patterns, and reference file paths for workflow skills.

Use after a feature or fix is verified and passing to reduce complexity, remove dead code, and improve clarity without changing behavior.

Use after implementation is verified and before merge, or when reviewing a PR, to check correctness, security, architecture, and test quality against project standards.

Mine past session transcripts for durable lesson and memory candidates, applying a reject-by-default rubric — suggest-only, never auto-writes.

Use before spec writing when the approach is unclear, multiple designs are plausible, or the engineer wants to explore alternatives before committing.

Provides Python-specific build commands, test commands, ORM guidance, framework patterns, and reference file paths for workflow skills.

Use periodically or when CLAUDE.md feels stale, drifts from the codebase, or commands break — audits CLAUDE.md against a quality rubric and proposes minimal append-only diffs.

Use at the end of a session that surfaced context CLAUDE.md was missing — captures discovered commands, gotchas, and patterns as minimal append-only additions, applied only with approval.

Use when executing an approved multi-file implementation to ensure each batch compiles, tests, and stays within the approved manifest.

Mine recurring reviewer-feedback phrases from the last N merged PRs and surface them as suggest-only [MINED:feedback] candidates.

Use when the change touches auth, secrets, audited state, audit trails, infrastructure, or external inputs — treat security as a design constraint, not final polish.

Use after implementation completes and before review begins, to verify the actual change matches the approved spec — files touched, public contracts added, security impact, and declared scope.