| name | director-build |
| description | Runs the Codex Director build lane for bounded implementation packets. Use when a Director worker brief explicitly invokes $codex-director:director-build for code, docs, configuration, tests, commits, or other implementation work that one worker can plan, edit, verify, and report. |
| metadata | {"short-description":"Run bounded implementation lane"} |
Director Build
Quick Start
Use only when explicitly invoked by a Director brief or by the user.
This split workflow skill is worker-side. If its activation marker, skill body, callback text, or readback evidence appears in a parent Director thread, the parent treats it only as routing input or child evidence. It does not load this workflow into the parent and does not permit parent inline browser, repo, provider, or project execution.
- Read local instructions and the owning Director brief.
- Load Build Workflow.
- Report activation:
workflow-skill-loaded:$codex-director:director-build, selected workflow/playbook director-build, top-level control loop, scope, user-outcome fit, model/thinking rationale, helper/subagent lane plan, and evidence contract.
- Gather only the context needed for the bounded change.
- Plan, edit, verify, run the required review/self-check, and return concise evidence.
Native Helper Mapping
During activation, report native helper runtime surface, V2 helper policy, planned RP-style helper profiles, and helper cleanup. multi_agent_v2 is the worker-internal equivalent of RP agent_run/agent_manage, not a substitute for Director-created top-level worker threads.
When V2 is available:
- Treat
spawn_agent as RP agent_run op=start: choose a stable lowercase task_name, set agent_type, model, reasoning_effort, service_tier, and fork_turns when exposed, and otherwise state those choices in the helper prompt.
- Treat
wait_agent as RP wait/poll: it is only a mailbox wake signal. Use send_message for queued context, followup_task to steer or continue a helper turn, list_agents for status/path checks, and close_agent for cleanup after evidence is consumed.
- Helper briefs must include role/profile, goal, exact scope, leave-alone files/modules, sibling lanes, plan/artifact paths, model/thinking/fork rationale, expected output, evidence standard, and cleanup expectation.
- Use
explore helpers for context/test mapping, engineer helpers for clear bounded sub-edits after the plan is known, pair helpers for complex implementation reasoning, and design helpers only for bounded UX/copy/design critique.
- For parallel helpers, require disjoint files/modules and sibling-awareness. For sequential helpers, use
followup_task only inside the same bounded implementation decision path.
- V2 evidence counts only after this worker reads the helper output, spot-checks material claims, summarizes verified evidence into final build evidence, and records
close_agent or close_blocked:<reason>.
Workflow
- Activate: restate the assigned outcome, repo/path, files likely involved, constraints, selected workflow, top-level loop, commit authority, helper policy, done criteria, and whether this build proves the final user outcome or only a prerequisite checkpoint.
- Baseline: read local instructions, check git status, identify existing patterns, and run or inspect the narrowest relevant baseline when useful.
- Plan: outline the implementation steps, verification commands, review gate, rollback/recovery notes when relevant, any files to avoid, and the remaining proof still needed if the user asked for end-to-end behavior.
- Helper lane: for non-trivial work, use a real helper/subagent lane for context mapping, implementation sketch, risky-area review, or verification. If unavailable, record
blocked:<reason> before proceeding.
- Implement: make the smallest scoped edits that satisfy the done criteria. Preserve unrelated dirty changes and local style.
- Verify: run focused tests/checks and any required smoke or static validation. If validation cannot run, capture the exact blocker and residual risk.
- Review: perform an adversarial self-check or delegated review against correctness, scope, tests, security/data risk, and user-facing behavior.
- Finish: record changed files, validation evidence, review verdict, commit status when authorized, unresolved risks, and cleanup/archive state.
Escalate Or Stop
Ask the Director to reroute to $codex-director:director-orchestrate or $codex-director:director-dynamic-workflow if implementation splits into independent lanes, crosses repos, needs approval gates, touches production/external writes, hosted provider config, direct deploy/alias work, secrets/auth/security, browser/desktop proof, or requires long-lived monitoring. Stop for secrets, destructive operations, unclear commit authority, or ownership ambiguity.
Required Invariants
- Stay inside the assigned repo/path, done criteria, commit authority, and constraints.
- Use orchestration as the top-level control loop for non-trivial build work unless the brief records a direct-leaf exception.
- Non-trivial work needs a real helper/subagent lane or a clear blocked helper capability. Direct leaf is allowed only with separate tiny, mechanical, and low-risk rationale.
- If the task grows into multiple independent lanes, risky writes, cross-repo work, or packetization, stop and ask the Director to reroute to
$codex-director:director-orchestrate or $codex-director:director-dynamic-workflow.
- Do not call the user goal complete from passing tests, a deployed preview, a provider setting save, or a health check unless that is the exact final verification surface. Report checkpoint evidence and remaining proof separately.
- If authorized to perform a direct preview deploy or alias move, record the deployment id, alias target, no-production proof when relevant, validation URL, residual gap, and the durable git/PR follow-up. Direct deploy evidence does not replace a requested commit/PR.
- Do not create nested top-level Codex workers unless the Director explicitly delegates that authority.
- Do not treat
wait_agent, list_agents, or helper final-status notifications as implementation evidence.
- Treat final output as candidate evidence until the Director reads the worker thread and reconciles it.
Output
Report changed files, commands/tests, review verdict, commit status when authorized, native helper surface, V2 helper paths/evidence/owner verification/cleanup when used, unresolved risks, cleanup/archive state, and exact blockers.