| name | ai-build |
| description | Canonical implementation gateway: reads approved plan.md, resolves stack from manifest, deterministic-routes each task to its adapter, dispatches the build agent in an isolated worktree, runs TDD self-validation per task, then a single final quality loop with one bounded quality-remediation pass on the full changeset before /ai-pr. Trigger for 'go', 'start building', 'execute the plan', 'implement it', 'lets do this', 'build the plan', 'resume', 'continue'. Not without an approved plan; run /ai-plan first. Not for multi-concern specs needing decomposition; use /ai-autopilot instead. Not for a single function or subcomponent; use /ai-code. |
| effort | cheap |
| argument-hint | [spec-NNN | --resume | --no-hitl] |
| model_tier | haiku |
| mirror_family | antigravity-skills |
| generated_by | ai-eng sync |
| canonical_source | .claude/skills/ai-build/SKILL.md |
| edit_policy | generated-do-not-edit |
Build
Execution engine for approved plans: reads plan.md, dispatches one subagent per task (fresh context, TDD self-validation), then a single final quality loop with one bounded quality-remediation pass on the full changeset before /ai-pr. Requires an approved plan — run /ai-plan first; resume with /ai-build --resume. Route multi-concern plans (≥ 3 concerns or ≥ 10 files) to /ai-autopilot, not /ai-build. If stuck: STOP and re-plan.
Workflow
Principles: §10.5 TDD (per-task RED/GREEN/REFACTOR self-validation), §10.4 DRY (plan patch-hunks unlock cheap-tier execution downstream).
- Preflight — verify
.ai-engineering/specs/plan.md, .agents/skills/_shared/execution-kernel.md, handlers/quality.md, handlers/deliver.md exist. Any missing → STOP and report the exact path(s). Never improvise missing orchestration.
0b. Executor route gate (spec-145) — read plan frontmatter status + execution_route.executor. status ≠ approved → STOP for approval. executor: autopilot → refuse this run, print execution_route.safe_next_command (normally /ai-autopilot), STOP. Route metadata absent → legacy behavior.
- Board sync (in_progress) — read
.ai-engineering/specs/spec.md frontmatter refs; for each ref whose hierarchy rule ≠ never_close (user_stories/tasks/bugs/issues), /ai-board sync in_progress <ref> (fail-open). Then read <spec_id> from spec.md spec: and python .ai-engineering/scripts/spec_lifecycle.py start <spec_id> (APPROVED→IN_PROGRESS; mirrors status: in-progress; fail-open — non-zero exit does NOT block DAG construction).
- Advise advisory — invoke the
ai-advise agent in gate mode before dispatching any build task (fail-open: log + continue, never block dispatch).
2b. Stack routing (D-127-06 / sub-008, spec-128 D-128-01) — per task, call tools.skill_app.deterministic_router.resolve_adapter(task_path, spec_stack) for the overrides dir .ai-engineering/overrides/<stack>/; pass it to the build agent to load conventions.md, tdd_harness.md, security_floor.md, examples/ before writing code. spec_stack precedence: spec frontmatter stack: over file-extension inference. UnknownStackError halts the task ("no overrides for this stack").
2c. Model dispatch (D-131-08 / sub-003, §10.4 DRY) — per task, inspect the plan.md block:
Patch (deterministic): present AND no synthesis hint → model_tier=haiku, effort=cheap (mechanical).- patch absent OR synthesis hint present →
model_tier=sonnet, effort=mid (judgment).
- operator
--max-effort OR task tagged architecture: true → model_tier=opus, effort=high (deep-architecture override).
Log via emit_agent_dispatched(..., metadata={"model_tier": <tier>, "effort": <effort>, "patch_present": <bool>}). Investing in /ai-plan (exhaustive patch-ready output) is what unlocks cheap-tier execution downstream (D-131-08).
2d. No-HITL contract (D-134-03) — if --no-hitl is set, read handlers/no-hitl.md and apply: single-concern gate, NEEDS_CONTEXT → BLOCKED promotion, quality_loop_blocked → exit 78, --no-watch implied for delivery, no auto-retry. Absent flag → default behavior unchanged.
- Execute kernel —
.agents/skills/_shared/execution-kernel.md. Wrap each task (Sub-flow 1 dispatch → 2 build self-validation TDD RED/GREEN/REFACTOR → 3 artifact collection → 4 board sync). As each task reaches a terminal state, update .ai-engineering/specs/plan.md immediately before dispatching the next task. Do not defer checkbox/status writes to the end of the phase or the end of the spec.
- Quality check — read
handlers/quality.md: Verify+Review on the full changeset, single round, fail-loud with ONE bounded quality-remediation pass. Blocker/critical/high findings may be fixed once when scoped to quality-loop evidence; remaining blocker/critical/high findings → STOP + escalate (no second remediation pass).
- Deliver — read
handlers/deliver.md: PR via /ai-pr with quality report. The deliver/quality handlers render the operator report through the value block (value-lens.md).
Resume Protocol
--resume uses observable evidence only — never guess hidden state. The build skill is the canonical implementation gateway (D-127-11).
- Missing/placeholder plan:
plan.md missing or still # No active plan → STOP, run /ai-plan.
- Incomplete tasks: unchecked checkboxes → resume at the first incomplete phase; skip completed tasks.
- Quality evidence missing: all tasks checked but no
## Quality Outcome section → resume at Quality Check (handlers/quality.md).
- Quality evidence present: resume at Deliver;
handlers/deliver.md detects an existing open PR and enters watch-and-fix or creates/updates the PR.
- Conflicting evidence: choose the earliest safe step and log why. Safety over convenience.
Handler Dispatch Table
| Phase | Handler | Agent Pattern |
|---|
| Quality Check | handlers/quality.md | Verify + Review parallel |
| Deliver | handlers/deliver.md | PR pipeline + cleanup |
| No-HITL | handlers/no-hitl.md | deterministic / single-concern only |
Common Mistakes
- Giving subagents the whole codebase — scope tightly.
- Continuing past a BLOCKED task without user input.
- Batch-updating
plan.md at the end instead of per task close.
- Modifying RED-phase test files during a GREEN-phase task.
- Invoking
--no-hitl on a multi-concern plan — use /ai-autopilot.
Examples
User: "the plan is approved, go"
/ai-build
Reads plan.md, dispatches one agent per task with fresh context; each task self-validates via TDD; the final quality loop verifies the full changeset and may consume one bounded quality-remediation pass before handing off to /ai-pr.
Integration
Called by: user directly post-/ai-plan approval. Calls: ai-build agent, ai-verify, ai-review, /ai-pr, /ai-board sync. Reads: _shared/execution-kernel.md. Transitions to: PR merge, or back to /ai-plan. See also: /ai-autopilot (multi-concern + --backlog), /ai-resolve-conflicts.
$ARGUMENTS
