Work Skill

Overview

Execute implementation through an AgentTeam using adaptive TDD. The lead coordinates and delegates -- agents write code.

Methodology

1. Pick tasks from bd ready or $ARGUMENTS
2. Mark tasks in progress: bd update <id> --status=in_progress
3. Read the epic description (bd show <epic>) for spec context -- EARS requirements guide what "done" looks like
4. Read Acceptance Criteria: Locate the ## Acceptance Criteria table in the parent epic description. Each AC row defines a testable criterion that the implementation must satisfy. If no AC section exists, flag as a process gap and proceed using EARS requirements directly.
5. Read Verification Contract: Locate the ## Verification Contract section in the parent epic description.
   • Treat Required evidence as the epic's definition of done.
   • Use Profile, Surfaces, and Risks to decide what evidence must be produced during implementation, not just at review time.
   • If the contract is missing: flag it as a process gap and fall back to the acceptance criteria, EARS requirements, and legacy test + lint gates.
6. Run ca search per agent/subtask for targeted context. Display results.
7. Assess parallelization: identify independent tasks that can be worked simultaneously
8. Deploy an AgentTeam (TeamCreate + Task with team_name) with MULTIPLE test-writers and implementers:
   • Role skills: .claude/skills/compound/agents/{test-writer,implementer}/SKILL.md
   • Scale teammate count to independent tasks; pairs coordinate via SendMessage on shared interfaces
9. Agents communicate via SendMessage when working on overlapping areas.
10. Lead coordinates: review agent outputs, resolve conflicts, verify tests pass. Do not write code directly.
11. If implementation diverges from spec requirements, acceptance criteria, or the Verification Contract, stop and discuss with user via AskUserQuestion before proceeding.
12. If blocked, use AskUserQuestion to get user direction.
13. Shut down the team when done: send shutdown_request to all teammates.
14. Commit incrementally as tests pass.
15. Run full test suite for regressions.
16. Close tasks: bd close <id>

Memory Integration

• Run ca search per delegated subtask with the subtask's specific description
• Each agent receives memory items tailored to their assigned task, not a shared blob
• Run ca learn after corrections or novel discoveries

MANDATORY VERIFICATION -- DO NOT CLOSE TASK WITHOUT THIS

Before bd close, you MUST:

1. Read the parent epic's Required evidence list and satisfy every listed item. The minimum legacy floor is test and lint.
2. Run pnpm test then pnpm lint
3. If Required evidence includes build, run pnpm build
4. Produce the contract-specific evidence needed for the touched surfaces:
   • browser_evidence / responsive_check / edge_states_check / a11y_smoke: capture runtime/browser proof and summarize it in the review handoff
   • contract_checks / contract_examples: run representative requests/examples and record outcomes
   • command_transcript: record happy-path and bad-input CLI flows
   • examples_run / docs_examples_sync: run or update public examples alongside the code
5. Run /implementation-reviewer on changed code -- wait for APPROVED If REJECTED: fix ALL issues, re-run required evidence, resubmit. INVIOLABLE per CLAUDE.md.

The full 8-step pipeline (invariant-designer through implementation-reviewer) is recommended for complex changes. For all changes, /implementation-reviewer is the minimum required gate.

Beads Lifecycle

• bd ready to find available tasks
• bd update <id> --status=in_progress when starting
• bd close <id> when all tests pass

Parallelization Strategy

• Always prefer parallel work: independent tasks should be assigned to different teammate pairs simultaneously
• Scale the team adaptively: deploy multiple test-writer + implementer pairs proportional to independent task count
• Subagent spawning within teammates: each teammate should spawn opus subagents for independent subtasks (e.g., a test-writer spawning subagents to write tests for multiple modules in parallel)
• Coordinate on shared interfaces: teammates working on overlapping APIs must communicate via SendMessage before implementing

Literature

• Consult docs/compound/research/tdd/ for TDD methodology, test-first development evidence, and best practices
• Consult docs/compound/research/property-testing/ for property-based testing theory and invariant design
• Run ca knowledge "TDD test-first" for indexed knowledge on testing methodology
• Run ca search "testing" for lessons from past TDD cycles

Common Pitfalls

• Lead writing code instead of delegating to agents
• Not injecting memory context into agent prompts
• Modifying tests to make them pass instead of fixing implementation
• Not running the full test suite after agent work completes
• Ignoring acceptance criteria from the parent epic
• Ignoring the Verification Contract and closing work with only test/lint evidence

Quality Criteria

• Tests existed before implementation code
• Agents received relevant memory context
• Lead coordinated without writing implementation code
• Incremental commits made as tests pass
• All tests pass after refactoring
• Task lifecycle tracked via beads (bd)
• Implementation aligns with spec requirements from epic
• Implementation satisfies acceptance criteria from parent epic
• Implementation produced the evidence required by the parent epic's Verification Contract

PHASE GATE 3 -- MANDATORY

Before starting Review, verify ALL work tasks are closed:

• bd list --status=in_progress must return empty
• bd list --status=open should only have Review and Compound tasks remaining If any work tasks remain open, DO NOT proceed. Complete them first.