AGENT SKILL

build-orchestration

Drives an approved IDSD run by delegating deterministic state transitions to the build orchestrator, routing actions to specialist agents/commands, and recording structured evidence.

License 许可
Apache-2.0
最近更新
2026-07-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
build-orchestration
description
Drives an approved IDSD run by delegating deterministic state transitions to the build orchestrator, routing actions to specialist agents/commands, and recording structured evidence.

03 SKILL.md 正文

Build Orchestration

Purpose

Coordinate the approved implementation loop. Do not implement, review, validate, or commit directly. The orchestrator owns state; specialist agents own lifecycle work.

Required inputs

  • AGENTS.md
  • harness.config.json
  • docs/bootstrap/environment.md
  • approved Intent, Context, Expectations, and task contracts
  • tasks/run_manifest.json
  • tasks/run_state.json
  • tasks/feature_list.json
  • active project-pack README and public review criteria when configured

Method

  1. Run python -B scripts/build_orchestrator.py status.
  2. If no approved run exists, stop.
  3. If bootstrap is stale, route REFRESH_BOOTSTRAP.
  4. Ask the deterministic controller for the next action.
  5. Execute exactly that action with the specified agent/command.
  6. Require a structured result artifact for every agent action.
  7. Record the result using record-result.
  8. Repeat until terminal.

Action routing

Action Route
REFRESH_BOOTSTRAP /bootstrap-environment
RUN_PREFLIGHT /start-run
IMPLEMENT_TASK build-agent
REPAIR_TASK build-agent with repair packet
RUN_POSTFLIGHT /postflight
RUN_REVIEW:test-agent test-agent
RUN_REVIEW:principal-engineer-agent principal-engineer-agent
RUN_DOMAIN_REVIEW domain-reviewer-agent
RUN_VALIDATOR validator-agent
COMMIT_TASK /commit-task
FINALIZE_SESSION /finalize-session
COMPLETE stop and summarize

State discipline

  • Conversation text is never durable state.
  • Do not update task status manually unless a controller command explicitly asks for that artifact.
  • Every action receives an action ID.
  • Every action result must reference the same run ID, task ID, and attempt.
  • Do not continue from an action whose result failed schema or consistency checks.

Repair discipline

  • Classify failure before repair.
  • Consolidate findings across test/principal/domain/validator gates.
  • After repair, previous postflight, review, and validator evidence is stale.
  • Stop when the repair budget is exhausted.
  • Update repair memory for repeated defects.

Stop conditions

  • missing required artifact;
  • project-pack config cannot be loaded;
  • required reviewer unavailable;
  • manifest/source hash mismatch;
  • task contract conflict;
  • private eval requested by builder;
  • repair budget exhausted;
  • branch/worktree/preflight failure;
  • command would require remote/prod authority not present.

Verification

  • [ ] Every action came from build_orchestrator.py next.
  • [ ] Every agent used appropriate skills.
  • [ ] Every result was recorded.
  • [ ] No role self-approved.
  • [ ] Build progressed only through durable evidence.

Model routing

The orchestrator must select the model alias from harness.config.json:model_routing for every action. /build-auto itself should use orchestration_large. Implementation actions should use execution_medium unless risk/escalation policy says otherwise. Test and principal review should use review_reasoning; validator should use validation_reasoning; domain review should use domain_reasoning.

Every action packet must include model_routing with alias/provider/model/reasoning metadata. Every structured result should record the actual model used and any fallback.

Do not silently downgrade validation/review on high-risk tasks. Prefer independent model aliases for build, review and validation.

04 同仓库更多 Skills