compozy

Reconciles a finished workflow's planned decisions (its Accepted ADRs) against the settled reality (git diff, review issues, and task status), then promotes the proven, cross-feature-durable ones into a durable project decision log at .compozy/DECISIONS.md plus .compozy/decisions/AD-NNN.md. Use when a workflow has finished its full pipeline (review round, reviews fix, and final verify) and you want to capture its durable decisions as the final step, or when re-running capture to refresh the log after further changes. Do not use for capturing decisions mid-implementation before review remediation, for PRD or TechSpec authoring, for PR review remediation, or for generic note-taking.

Explains Compozy capabilities, CLI commands, core workflow skills, optional extension skills, configuration, artifact structure, reusable agents, and extensions. Use when the user asks how to use Compozy, what commands are available, how the workflow pipeline works, or how to configure a workspace. Do not use for executing workflow steps — use the specific cy- skills instead.

Executes one PRD task end-to-end uninterrupted — resolves spec conflicts autonomously, implements, validates, and updates tracking without pausing for questions. Use when a prompt includes a task specification that must be implemented, validated, and reflected in task tracking files. Do not use for PR review batches, generic coding tasks without a PRD task file, or standalone verification-only work.

Enforces fresh verification evidence before any completion, fix, or passing claim, and before commits or PR creation. Use when an agent is about to report success, hand off work, or commit code. Do not use for early planning, brainstorming, or tasks that have not yet reached a concrete verification step.

Dispatches scoped-write explorer agents in parallel via `compozy exec` for multi-area research that must produce written artifacts — one seven-section analysis file per slice plus a parent-authored summary. Use when a research question spans several distinct areas and chat output is not enough. Do not use for single-file lookups (use Explore) or edits to existing code.

Independent audit of AI-implemented work — certifies a completed task actually did what it claims, checking files, diffs, tests, and CI rather than the agent's self-report. Flags skipped or weakened tests, mock-hidden integration, snapshot drift, happy-path-only coverage, flaky retries, and status/evidence mismatches. Use when validating completed Compozy tasks, AI-authored PRs, or codex-loop iterations. Not for real-user, persona, or journey QA — use qa-execution for those.

Architectural audit that hunts dead code, duplicated functionality, anti-patterns, type confusion, and code smells across a whole codebase. Use when the user asks for architectural analysis, to find dead or unused code, identify duplication, or assess codebase health. Don't use for style/formatting, performance profiling, security audits, or feature-level code review.

Orchestrate Claude Opus and Codex GPT-5.5 worker TUIs from a controller agent in the current cmux workspace. Use when delegating bounded tasks to worker panes, running plan-first delegations (Claude Code plan mode, Codex Plan mode), monitoring and tracking worker state, or verifying worker reports. Not for end-user cmux control (see cmux, cmux-workspace) — this drives agent workers.

Production-ready CSS transitions for web apps. Use when implementing notification badges, dropdowns, modals, panel reveals, page transitions, card resizes, number pop-ins, text swaps, icon swaps, success checks, avatar group hovers, error state shakes, search/input clear, skeleton loaders, shimmer text, sliding tabs, tooltips, staggered text reveals, card hover tilt, plus-to-menu morph, or accordions. Triggers on "add a transition", "animate the dropdown", "make the modal open smoothly", "swap icon", "page slide", "stagger animation", "open / close transition", "make it animate", "fade between", "success animation", "form error", "shake on invalid", "hover lift", "avatar stack hover", "clear the search", "skeleton loader", "loading shimmer", "shimmer text", "sliding tabs", "segmented control", "tooltip", "reveal text", "tilt card", "3D hover tilt", "cursor glare", "plus to menu", "FAB morph", "accordion", "collapsible", "expand / collapse", "disclosure", "replace this transition", "which transition fits here"

AGH visual-design authority for production UI, static artifacts, prototypes, and reviews. Use when creating or reviewing an AGH surface or changing tokens, typography, spacing, depth, icons, or motion. Do not use for capture-only verification; use agh-ui-screenshot.

Screenshot capture for deterministic PNG evidence and visual-contract bundles for AGH Storybook stories and local UI URLs. Use for visual audits, regression diffs, and design-parity checks. Do not use for interactive E2E flows, remote authenticated sites, or Storybook test execution.

Enforces fresh verification evidence before any completion, fix, or passing claim, and before commits or PR creation. Use when an agent is about to report success, hand off work, or commit code. Do not use for early planning, brainstorming, or tasks that have not yet reached a concrete verification step.

Deep review of branch diffs, working trees, or GitHub PRs at any size. Use when the user asks for CodeRabbit-grade review, an incremental re-review after new pushes, publication of findings to a PR, a cross-LLM peer-review verdict round, or conformance review against spec artifacts. Don't use for applying fixes, reviewing specs or PRDs as documents, or quick single-file feedback.

AGH runtime and contribution guide. Use for sessions, agents, native tools, skills, memory, network, tasks, capabilities, bundles, QA, docs, and repo work. Do not use for unrelated projects.

Dogfoods AGH through an autonomous startup scenario with live providers, cross-surface observation, and strict evidence audit. Use for release or complex-integration QA. Do not use for smoke, static, mock-only, or unit-test work.

Bootstraps isolated AGH QA labs and emits the canonical manifest for downstream QA. Use when local QA needs daemon, workspace, provider, or playbook setup. Do not use for unit tests, planning-only work, or browser-only checks.

Explains Compozy capabilities, CLI commands, core workflow skills, optional extension skills, configuration, artifact structure, reusable agents, and extensions. Use when the user asks how to use Compozy, what commands are available, how the workflow pipeline works, or how to configure a workspace. Do not use for executing workflow steps — use the specific cy- skills instead.

Creates a Product Requirements Document through interactive brainstorming with parallel codebase and web research. Use when starting a new feature or product, building a PRD, or brainstorming requirements. Do not use for technical specifications, task breakdowns, or code implementation.

Decomposes PRDs and TechSpecs into detailed, independently implementable task files with enrichment from codebase exploration. Use when a PRD or TechSpec exists and needs to be broken down into executable tasks, or when task files need enrichment with implementation context. Do not use for PRD creation, TechSpec generation, or direct task execution.

Creates a Technical Specification by translating PRD business requirements into implementation designs through interactive technical clarification. Use when a PRD exists and needs a technical plan, or when technical architecture decisions need documentation. Do not use for PRD creation, task breakdown, or direct code implementation.

Executes one PRD task end-to-end using a provided task file, PRD directory, tracking file paths, and auto-commit mode. Use when a prompt includes a task specification that must be implemented, validated, and reflected in task tracking files. Do not use for PR review batches, generic coding tasks without a PRD task file, or standalone verification-only work.

Performs a comprehensive code review of a PRD implementation and generates a review round directory with issue files compatible with cy-fix-reviews. Use when reviewing implemented PRD tasks, creating a manual review round without an external provider, or performing a quality audit of code changes. Do not use for fetching reviews from external providers, fixing existing review issues, executing PRD tasks, or editing source code.

Runs an optional cross-LLM peer review of an implemented change (the diff) via `compozy exec` and requires the reviewer to write one scoped Markdown findings artifact for user-directed remediation. The reviewer runtime is configurable through --ide / --model / --reasoning (defaults claude / opus / xhigh) and the verification gate through --verify. Project-agnostic: works in any repo and any language, auto-discovers project rule files (CLAUDE.md, AGENTS.md, .cursor/rules, CONTRIBUTING.md), auto-detects the verify command, and writes artifacts to --out, an auto-detected .compozy/tasks/<slug>/qa/ layout, or a .peer-reviews/<timestamp>/ fallback. Use after any implementation pass (feature, bug fix, refactor) when the user explicitly asks for an external review of the diff before commit or PR. Do not use for spec/TechSpec review (use spec-peer-review), automatic remediation, or auto-looped review cycles.

Runs an optional cross-LLM peer review of a spec (TechSpec, design doc, RFC, or detailed PRD) via `compozy exec` and requires the reviewer to write one scoped Markdown findings artifact for user-directed incorporation. The reviewer runtime is configurable through --ide / --model / --reasoning (defaults claude / opus / xhigh). Project-agnostic: works in any repo, auto-discovers project rule files (CLAUDE.md, AGENTS.md, .cursor/rules, CONTRIBUTING.md), and writes artifacts to --out, an auto-detected .compozy/tasks/<slug>/qa/ layout, or a .peer-reviews/<timestamp>/ fallback. Use when a spec draft has already been approved by the user and they want an external review round, especially for autonomy/network/security/migration-impacting designs. Do not use for implementation/diff review (use impl-peer-review), automatic approval gates, or auto-looped review cycles.

Executes full-project QA like a real user by discovering the repository verification and E2E contracts, running build, lint, test, and startup commands, exercising core workflows end-to-end through CLI, HTTP, and browser interfaces, requiring automated regression coverage for supported critical flows, fixing root-cause regressions, and rerunning the full gate. Uses the agent-browser companion skill for Web UI validation when a web surface exists. Use when validating a branch, release candidate, migration, refactor, or risky commit. Do not use for static code review only, one-off unit test edits, planning test cases, or architecture brainstorming without execution — use qa-report for planning and documentation.

Generate comprehensive test plans, test cases, regression test suites, automation annotations, and bug reports for QA engineers. Includes Figma MCP integration for design validation. Use when planning QA before execution, documenting test strategies, marking which flows require E2E follow-up, or creating structured bug reports. Do not use for executing tests against a live repository or running verification gates — use qa-execution for that.

Explains Compozy capabilities, CLI commands, core workflow skills, optional extension skills, configuration, artifact structure, reusable agents, and extensions. Use when the user asks how to use Compozy, what commands are available, how the workflow pipeline works, or how to configure a workspace. Do not use for executing workflow steps — use the specific cy- skills instead.

Creates a Product Requirements Document through interactive brainstorming with parallel codebase and web research. Use when starting a new feature or product, building a PRD, or brainstorming requirements. Do not use for technical specifications, task breakdowns, or code implementation.

Runs an optional cross-LLM peer review of a TechSpec via compozy exec --ide claude --model opus --reasoning-effort xhigh and packages findings for user-directed incorporation. Use when a TechSpec draft has already been approved by the user and they want an external review round, especially for autonomy/network/memory-impacting designs. Do not use for PRDs, automatic approval gates, code review batches, or auto-looped review cycles.

Decomposes PRDs and TechSpecs into detailed, independently implementable task files with enrichment from codebase exploration. Use when a PRD or TechSpec exists and needs to be broken down into executable tasks, or when task files need enrichment with implementation context. Do not use for PRD creation, TechSpec generation, or direct task execution.

Creates a Technical Specification by translating PRD business requirements into implementation designs through interactive technical clarification. Use when a PRD exists and needs a technical plan, or when technical architecture decisions need documentation. Do not use for PRD creation, task breakdown, or direct code implementation.

Executes one PRD task end-to-end using a provided task file, PRD directory, tracking file paths, and auto-commit mode. Use when a prompt includes a task specification that must be implemented, validated, and reflected in task tracking files. Do not use for PR review batches, generic coding tasks without a PRD task file, or standalone verification-only work.

Dispatches scoped-write explorer subagents in parallel for general research and exploration of any codebase, topic, or domain. The operator passes --path (output directory), --agents (parallel count), and --prompt (research question). The parent scouts the territory first to divide work, dispatches N explorer subagents that each write one analysis file at <path>/analysis/NN_analysis_<slug>.md following a seven-section schema, then synthesizes <path>/analysis/summary.md. The skill self-contains the explorer subagent definition and prompts the operator to install it at .claude/agents/explorer.md when absent. Use when running parallel multi-area research that must produce written artifacts. Do not use for competitor-only research already covered by cy-research-competitors, single-file lookups answerable by Explore, or edits to existing code.

Deep architectural audit focused on finding dead code, duplicated functionality, architectural anti-patterns, type confusion, and code smells. Use when user asks for architectural analysis, find dead code, identify duplication, or assess codebase health. Don't use for style/formatting issues, performance profiling, security audits, or feature-level code review.

Enforce root-cause fixes over workarounds, hacks, and symptom patches in all software engineering tasks. Use when debugging issues, fixing bugs, resolving test failures, planning solutions, making architectural decisions, or reviewing code changes. Activates gate functions that detect and reject common workaround patterns such as type assertions, lint suppressions, error swallowing, timing hacks, and monkey patches. Don't use for trivial formatting changes or documentation-only edits.

Executes real-user QA sessions through public interfaces using personas, journeys, exploratory charters, test tours, edge-case probes, CFR checks, and browser evidence. Reads qa-report artifacts from <qa-output-path>/qa/ when present, captures issues/screenshots/reports under the same output tree, and classifies bugs by user impact. Use when validating a release candidate, migration, refactor, or user-facing change against production-like behavior. Do not use for AI implementation audits, task-status reconciliation, CI gate runs, integration/security/performance templates, or flaky-test triage; use agent-output-audit for those.

Plans real-user QA deliverables: personas, journey maps, exploratory charters, persona/journey/tour/CFR test cases, regression suites, Figma validation checks, automation intent, and user-impact bug reports. Writes artifacts under <qa-output-path>/qa/ for qa-execution to consume. Use when planning QA before execution, documenting journey-driven test strategy, marking flows that need E2E follow-up, or filing structured bug reports. Do not use for live execution, AI implementation audits, CI gate ownership, or technical integration/security/performance suites; use qa-execution or agent-output-audit instead.

Analyzes codebases to identify refactoring opportunities based on Martin Fowler's catalog of code smells and refactoring techniques. Detects duplicated code, high coupling, complex conditionals, primitive obsession, long functions, and other structural issues. Produces a structured refactoring report with prioritized findings saved to docs/_refacs/. Use when auditing code quality, preparing for a refactoring sprint, or reviewing architectural health. Don't use for style/formatting issues, performance optimization, or security audits.

Comprehensive testing doctrine for software and AI systems — covers positive patterns, anti-patterns, gates for coding agents writing tests, CI discipline, and an LLM/agent evaluation primer. Use when authoring or reviewing tests, adding mocks, deciding test placement, generating tests via agents, debugging flaky CI, designing eval suites for LLM features, or rebuilding a brittle test suite. Contains 12 positive patterns (selector hierarchy, table-driven, builders, real-system gates), 25 anti-patterns across Brittleness, Flakiness, Mock-misuse, Process, and AI-specific families, 7 mandatory gates for agents writing tests, flaky-test taxonomy with quarantine workflow, contract / property / mutation testing patterns, and an oracle-ladder primer for LLM-as-judge and agent eval. Language-agnostic — pseudo-code only. Don't use for general code review, library-specific debugging unrelated to tests, non-testing CI pipeline design, or production observability.

Guides Codex Loop setup and long-running tracked execution by installing or refreshing the runtime, explaining activation headers, bootstrapping durable tracking artifacts, decomposing broad requests into tasks, and maintaining per-iteration memory and state. Use for Codex Loop installation, status, activation, or tracked restart-safe implementation, QA, review, and remediation loops. Do not use for one-off coding tasks that do not need Codex Loop or persistent tracking.

Free AI search via Exa MCP. Web search for news/info, code search for docs/examples from GitHub/StackOverflow, company research for business intel. No API key needed.

Executes full-project QA like a real user by discovering the repository verification and E2E contracts, running build, lint, test, and startup commands, exercising core workflows end-to-end through CLI, HTTP, and browser interfaces, requiring automated regression coverage for supported critical flows, fixing root-cause regressions, and rerunning the full gate. Uses the agent-browser companion skill for Web UI validation when a web surface exists. Use when validating a branch, release candidate, migration, refactor, or risky commit. Do not use for static code review only, one-off unit test edits, planning test cases, or architecture brainstorming without execution — use qa-report for planning and documentation.

Install or refresh the CC Loop runtime for structured [[CC_LOOP ...]] activation with min="...", rounds="...", or goal="...". Use only for cc-loop setup, status, uninstall, or activation guidance.

Perform comprehensive research and analysis using multiple discovery and analysis tools to provide a solid foundation for complex tasks.

Use when asking about Rust code style or best practices. Keywords: naming, formatting, comment, clippy, rustfmt, lint, code style, best practice, P.NAM, G.FMT, code review, naming convention, variable naming, function naming, type naming, 命名规范, 代码风格, 格式化, 最佳实践, 代码审查, 怎么命名

Authors and structures professional-grade agent skills following the agentskills.io spec. Use when creating new skill directories, drafting procedural instructions, or optimizing metadata for discoverability. Don't use for general documentation, non-agentic library code, or README files.