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.mdharness.config.jsondocs/bootstrap/environment.md- approved Intent, Context, Expectations, and task contracts
tasks/run_manifest.jsontasks/run_state.jsontasks/feature_list.json- active project-pack README and public review criteria when configured
Method
- Run
python -B scripts/build_orchestrator.py status. - If no approved run exists, stop.
- If bootstrap is stale, route
REFRESH_BOOTSTRAP. - Ask the deterministic controller for the next action.
- Execute exactly that action with the specified agent/command.
- Require a structured result artifact for every agent action.
- Record the result using
record-result. - 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
invoice-domain-review
Domain review for contractor invoice governance, ingestion, reconciliation, attendance attribution and billing evidence.
python-domain-review
Reviews generic Python implementation against task intent, API expectations, simplicity and maintainability.
verified-commit
Creates a task-scoped commit only after validator and exact-tree postflight pass.
using-agent-skills
Selects and applies the smallest relevant set of skills for a task, preventing generic prompting and unused specialist guidance. Use at the start of every command, task, review, or repair cycle.
testing-patterns
Designs focused, hierarchical, regression, boundary, integration, and end-to-end tests for vertical-slice implementation. Use before and after implementation, and during test review.
test-hierarchy
Classifies and enforces the required unit, integration, local E2E, and cloud E2E test tier for a task or run.