belt

Experimental (mattpocock/skills wayfinder port, MIT). Charts a foggy effort as a shared Linear decision map, resolves its decision tickets across bounded sessions, and hands one graduated effort to /belt:requirements. Upstream of /belt:feature-dev; one run graduates exactly one effort.

Runs the shared build stage: TDD implementation from the plan document's task list and two-agent code review with autonomous triage. Use standalone with an existing plan.md/fix-plan.md, or composed as the build stage of /belt:feature-dev and /belt:bug-fix. QA runs as its own stage (/belt:qa) after build; integration happens at the orchestrator's integrate phase.

Runs the feature design stage: goal-sheet intake (/belt:goal) followed by a design document (architecture + key decisions) reviewed by belt:spec-reviewer. Use standalone for design-only work, or composed as the first stage of /belt:feature-dev. Task breakdown lives in the plan stage (/belt:plan), not here. --codex enables adversarial spec review.

Runs the bug diagnosis stage: root-cause analysis with a failing reproduction test and reproduction scenarios, fix planning, and adversarial plan review. Use standalone for diagnosis-only work, or composed as the upstream stage of /belt:bug-fix. --codex enables adversarial plan review.

Quality-gated feature pipeline from ticket to integration: goal-sheet intake, design document with spec review, implementation plan with QA scenarios, context-reset checkpoint, TDD implementation, autonomous code review, mandatory QA with human-readable evidence (screenshots / transcripts), and integration with evidence publishing. Accepts a Linear id, URL, free text, or a requirements.md path. --codex enables adversarial review.

Runs the implementation planning stage: writes plan.md (test strategy + task list) and scenarios.yml (QA replay scenarios) from an approved design.md, then reviews the plan via belt:spec-reviewer. Use standalone when a design already exists, or composed between the design and build stages of /belt:feature-dev. --codex enables adversarial plan review.

Interview-driven requirements definition. Resolves a Linear ticket, URL, or free-text request, investigates the codebase, asks only human-decidable questions in dependency-ordered rounds until the frontier is empty, and writes docs/requirements/<YYYY-MM-DD-topic>/requirements.md reviewed by belt:spec-reviewer. The result feeds /belt:feature-dev as input.

General-purpose documentation writing. Takes a topic (feature, module, or theme), investigates the code, and creates or updates documents under docs/ following the repository's existing documentation conventions. Use for architecture overviews, usage guides, and reference pages.

Batched-question intake that turns a Linear ticket, URL, or free-text task into a goal sheet (goal / scope / acceptance criteria). Lightweight replacement for grilling/brainstorming dialogues: investigates the codebase first, then asks only human-decidable questions in dependency-ordered rounds (frontier interview). Use standalone before any feature work, or as the intake phase of /belt:feature-dev.

Two-agent parallel code review (belt:code-reviewer + belt:quality-reviewer) with deterministic merge. --codex adds an adversarial pass via /codex:rescue. In a belt run, critical/high findings are fixed autonomously; standalone runs triage findings in one batched selection.

Runs the QA verification stage: an independent belt:qa-verifier subagent replays docs/features/<topic>/scenarios.yml (browser scenarios via agent-browser with screenshots, cli scenarios by executing the real commands with transcripts), runs an exploratory pass, and writes qa-report.md. Evidence goes to the run directory and is published to PR/Linear per the [qa] evidence config. Use standalone after a build, or composed as the qa stage of /belt:feature-dev and /belt:bug-fix.

Spec review via the consolidated belt:spec-reviewer agent. Reviews any spec-family document (requirements.md, goal-sheet.md, design.md, plan.md). Findings are triaged in one batched selection. --codex adds an adversarial pass via /codex:rescue in the same parallel batch.

Belt Protocol for driving belt-agent CLI. Defines command loop, response interpretation, invoke/artifact/validate semantics, and safety constraints for LLM agents driving belt pipelines.

Runs a quality-gated debugging pipeline with root-cause analysis, fix planning, TDD repair, autonomous code review, mandatory QA replay of the reproduction scenarios with human-readable evidence, and integration. Use when a bug needs structured diagnosis and verified repair. --codex enables adversarial review.

Resumes a previously handed-over belt pipeline run by reading handover.md and state.json from the current worktree, verifying preconditions, and invoking the owning pipeline skill in resume mode via Skill tool args. Use when continuing belt pipeline work after /belt:handover and /clear, when the user invokes /belt:resume, or after a session restart to pick up an in-progress run in the current worktree.

Writes a handover note (Resume hint) under the current belt run directory so a later session can pick up where the pipeline was paused. Use when pausing a multi-phase belt pipeline (feature-dev, bug-fix, debug-flow) before /clear or session end, or when the user invokes /belt:handover.