sw-build

name	sw-build
description	TDD implementation of one work unit. Delegates test writing to the tester agent and implementation to the executor agent. Commits per task.
argument-hint	[work-id] [task-id]
allowed-tools	["Read","Write","Edit","Bash","Glob","Grep","Task","TaskCreate","TaskUpdate","TaskList","TaskGet","AskUserQuestion"]

Specwright Build

Goal

Implement the current work unit with TDD. The per-task loop is RED → GREEN → REFACTOR; end-of-unit integration and regression checks live in one optional after-build phase.

Inputs

• {worktreeStateRoot}/session.json -- selected work for this worktree
• {repoStateRoot}/work/{selectedWork.id}/workflow.json, {workDir}/spec.md, {workDir}/plan.md
• {workArtifactsRoot}/{selectedWork.id}/design.md, {workDir}/context.md
• {workArtifactsRoot}/{selectedWork.id}/approvals.md -- durable design and unit approval ledger when present
• {projectArtifactsRoot}/CONSTITUTION.md, {projectArtifactsRoot}/config.json

Outputs

• After each task: failing tests, passing implementation, task commit, workflow progress, updated {workDir}/implementation-rationale.md, refreshed {repoStateRoot}/work/{selectedWork.id}/units/{selectedWork.unitId}/stage-report.md
• After all tasks: as-built notes in plan.md, three-line handoff to /sw-verify, ready-to-verify build state; the handoff points at Artifacts: {repoStateRoot}/work/{selectedWork.id}/units/{selectedWork.unitId}/stage-report.md

Constraints

Execution model (LOW freedom): Run in the foreground in the current turn. "Autonomous" means unattended decisions inside this build, not background execution.

Stage boundary (LOW freedom): Follow protocols/stage-boundary.md. Implement only the active unit; never create pull requests, run gh pr create, or invoke /sw-ship. Before the terminal handoff, write {repoStateRoot}/work/{selectedWork.id}/units/{selectedWork.unitId}/stage-report.md; the handoff points at it and the Next line is Next: /sw-verify.

Branch setup (LOW freedom): First action before coding: resolve the session-selected work from the current worktree, verify that no other live top-level worktree owns it, then check out the feature branch from config.git.branchPrefix and sync it per protocols/git.md. The selected work's recorded targetRef, when present, is the first branch-resolution input via protocols/git.md; repo config defaults and the baseBranch compatibility alias are fallbacks only. Use {git.branchPrefix}{selectedWork.unitId} for multi-unit work and never commit to the base branch. If the selected work is already owned elsewhere, STOP with explicit /sw-adopt guidance instead of mutating it silently; do not fall back to implicit adopt/takeover behavior.

Build freshness checkpoint (LOW freedom) — after branch setup: Evaluate the build checkpoint via protocols/git-freshness.md using the selected work's recorded targetRef and freshness. require blocks stale, diverged, and blocked freshness results; warn surfaces advisory drift; queue-managed results stay distinct and do not trigger implicit local rewrites. When branch-head validation is blocked and rebase or merge reconcile is configured, run protocols/git-reconcile.md in the owning worktree and continue in the same stage after a successful reconcile. manual remains an explicit fallback: stop with manual reconcile guidance, reconcile the current branch against the recorded target in the owning worktree, or run /sw-adopt first if a linked-worktree ownership conflict exists, then rerun /sw-build. Do not clear the block by silently rewriting targetRef or freshness metadata.

Approval checkpoint (LOW freedom) — before task loop: Use protocols/approvals.md and the shared helper with the current unit artifact set (spec.md, plan.md, context.md). When sw-pivot or replanning regenerated the current unit artifacts, that regenerated artifact set becomes the current approval surface. Interactive /sw-build runs may record an APPROVED unit-spec entry in {workArtifactsRoot}/{selectedWork.id}/approvals.md with source classification command, and they must refresh or record that approval for the regenerated surface before any task executes. Headless runs must validate existing human approval instead. Approval refresh does not replace branch reconciliation and must not be used to clear a separate freshness block. Never move approval truth into workflow.json.

Task loop (MEDIUM freedom): Work one task at a time. Finish it before starting the next and emit a status card after each task commit.

Implementation rationale (LOW freedom) — at each task commit boundary: Maintain {workDir}/implementation-rationale.md as an append-only curated artifact with one section per completed task. In tracked work-artifact mode, stage the task's rationale content before creating the task commit so the tracked tree stays clean; if the resulting commit SHA is only known after the commit is created, capture that SHA in the next synchronized rationale update instead of leaving tracked artifacts dirty between tasks. Each task entry must record the task ID, relevant AC references, changed files, tests added or updated, why this approach was chosen, any deviation from the approved unit artifacts, the execution path (executor or build-fixer), and the task commit SHA once it is known. This artifact captures rationale, not transcript excerpts.

TDD cycle (LOW freedom for sequence):

1. RED: delegate to specwright-tester, write hard-to-pass tests, and confirm they fail.
2. GREEN: delegate to specwright-executor, pass the failing tests, and stop on any plan mismatch or pre-existing type/signature discrepancy.
3. REFACTOR: simplify only code written for the current task; keep behavior unchanged. Per-task integration and regression runs do not happen inside this loop.

Mid-build checks (MEDIUM freedom): Follow protocols/decision.md#late-discovery-lifecycle and protocols/build-quality.md for late discoveries, behavior capture, and as-built notes.

Build failures (MEDIUM freedom): If RED tests pass, the tests are wrong. If GREEN or after-build checks fail, delegate to specwright-build-fixer for at most 2 attempts and follow protocols/headless.md for persistent failures. Treat executor-reported signature/type mismatches as a plan mismatch, not a build-fixer case.

Commits (LOW freedom): One commit per completed task. Stage only the files for that task, never use git add -A, and run configured format/lint commands before committing when present.

After-build (MEDIUM freedom): Optional end-of-unit phase only. Delegate post-build review, then run commands.test and commands.test:integration when configured; integration now runs here once per unit, not per task. On failure, use specwright-build-fixer (max 2 attempts); if it still fails, surface it to the user in interactive mode and skip with a recorded note in headless mode.

Task tracking (LOW freedom): When Claude Code task tools are available, create and update task records, but keep the selected work's workflow.json as the source of truth. Tracking failures never block the build.

State updates (LOW freedom): Follow protocols/state.md: acquire the per-work lock on the selected work before mutations, update the selected work's tasksCompleted after each committed task, refresh currentTask, and append as-built notes before handoff. Mutate only the selected work's workflow.json and the current worktree's session state.

Parallel execution (MEDIUM freedom): Only use protocols/parallel-build.md when the experimental config flag enables it; otherwise ignore it and stay sequential. When mutable concurrency needs more than subordinate helpers or read-only lanes, split the effort into separate works with integration criteria rather than sharing one mutable workflow across top-level worktrees.

Protocol References

• protocols/stage-boundary.md -- scope and final handoff
• protocols/git.md -- branch lifecycle and commit discipline
• protocols/git-freshness.md -- build-entry freshness checkpoint
• protocols/git-reconcile.md -- lifecycle-owned branch reconcile
• protocols/approvals.md -- spec approval capture and validation
• protocols/review-packet.md -- rationale artifact contract consumed by the reviewer packet
• protocols/delegation.md -- tester/executor/build-fixer context handoff
• protocols/build-quality.md -- post-build review and as-built notes
• protocols/decision.md -- late discoveries and error handling
• protocols/headless.md -- non-interactive failure handling
• protocols/parallel-build.md -- config-gated parallel mode
• protocols/state.md -- workflow locking and task progress

Failure Modes

Condition	Action
No active work unit	STOP: "Run /sw-design and /sw-plan first"
Selected work owned by another live top-level worktree	STOP with explicit /sw-adopt guidance
Build/test command not configured	STOP: "Configure commands in config.json or run /sw-init"
Build freshness reconcile under branch-head require + rebase or merge fails	STOP and surface the reconcile failure; keep the recorded target/freshness metadata intact.
Build freshness checkpoint is blocked under branch-head require + manual	STOP with manual reconcile guidance, keep the recorded target/freshness metadata, then rerun /sw-build.
Tester writes tests that pass immediately	Re-delegate RED with stronger failing cases
Executor reports a pre-existing type/signature mismatch	STOP and surface the plan mismatch
Build-fixer exhausts 2 attempts	STOP and show the failure
Per-work lock held by another skill	STOP with lock info