/feature-workflow

description: Run the gated, agent-driven workflow end-to-end.
argument-hint: "[work-name]"

Feature Workflow (Gated Orchestrator)

Goal: take a high-level plan and drive it to completion using specialized subagents, TDD gates, and explicit acceptance.

Inputs

• `work-name`: short slug (e.g. `file-mgmt-rebuild-phase0`)
• Optional: an existing plan doc to refine.

Workflow (always in this order)

1) **Plan (Planner agent)**

• Create or refine a modular plan in the project docs directory.
• Break into Work Items with explicit acceptance criteria, tests, and rollback notes.

2) **Spec Check (Spec Guardian agent)**

• Validate plan vs specs and project rules in `AGENTS.md` and `.claude/rules/*.md`.
• Stop if specs conflict or constraints are violated.

3) **Impact (Impact Analyst agent)**

• For each Work Item, map the minimal file set, dependency edges, and risks.
• Propose the smallest correct change boundaries.

4) **Implement (Implementer agent)**

• For each Work Item, do a **preflight investigation** before writing tests:
• Reproduce/describe current behavior and expected behavior.
• Trace the exact call chain and identify the smallest test seam.
• Confirm the minimal file impact; update the Impact map if needed.
• Subagent delegation is encouraged for this "map and report" step.
• For each Work Item: write the test first (RED), implement (GREEN), refactor safely.
• Encourage subagent delegation for large/mechanical diffs; always review + run gates after.
• Keep side effects isolated; keep changes local; keep files under ~300 lines.

5) **Test (Test Runner agent)**

• Run `pnpm check:all` (and `cargo test` when Rust changes).
• If UI flows are impacted, request the user to run the app and use Tauri MCP for E2E.

6) **Audit (Auditor agent)**

• Review diffs for correctness, architecture drift, and rule violations.
• If issues found: loop back to Implement for fixes.

7) **Manual Test Guide (Manual Test Author agent)**

• Update manual testing docs incrementally per Work Item.
• Ensure final “end-to-end” guide exists and is coherent.

8) **Verify (Verifier agent)**

• Re-run gates as needed; produce a final checklist.

9) **Release (Release Steward agent)**

• Propose one commit per Work Item with clear messages.
• Commit only after explicit user “accept + commit”.

Acceptance Contract

• “Accept” means: tests green, rules satisfied, scope minimal, no known data-loss path introduced.
• If uncertain: stop and ask rather than guessing.