| name | health-harness-orchestrator |
| description | Codex adapter for implementing, debugging, verifying, reviewing, and delivering an already-defined Reva change. Invoke only when reva-workflow-router selects implementation or incident mode, or when product-pipeline delegates its S5 child phases. |
复元 Health Harness — Codex Orchestrator
This thin Codex adapter executes the project-managed delivery harness. The
agent-neutral rules live in:
docs/governance/agent-skill-registry.json
docs/governance/agent-skill-governance.md
docs/design-agent-operating-harness.md
AGENTS.md and docs/governance/*
Read relevant source files, nearby tests, and the active Dossier before planning
implementation.
Controller boundary
- Run as
primary_controller only when Router selects
health-harness-orchestrator for implementation or incident mode.
- When invoked by
product-pipeline, remain a delegated S5 child phase in the
parent's run. Do not create a second Dossier, ledger, plan, or finish state.
- Capabilities such as systematic debugging, TDD, and verification add
discipline; overlays such as safety or database review can block progress;
neither category owns the workflow.
- Activate
system-map only when its task trigger is met and Router exposes it
in deferred_by_phase.on_demand; otherwise inspect known source and tests directly.
- A release terminal is selected by target and occurs only after its Gates pass.
Codex-native workflow
- Scope: inspect
AGENTS.md, the active Dossier, relevant source and tests,
dirty files, open PRs, and exact acceptance evidence.
- Plan: define the smallest coherent tasks and cross-surface contracts.
Keep one writer per shared file set.
- Implement: use Codex subagents only for independent bounded work.
Read-only investigators and reviewers may run in parallel; stateful writers
need disjoint ownership or isolated worktrees.
- G3 incremental QA: test each completed slice. For a real defect, establish
the cause and a failing test before changing production behavior. LLM-facing
changes also pass
scripts/harness_llm_change_gate.py and any required live
regression gate.
- G4 safety review: sensitive health data, medication, genetics, safety rules,
authentication, notifications, and write paths require the selected safety
overlay. BLOCK returns to implementation and requires re-review.
- G5 release health: hand off to exactly one target-specific release workflow only
after tests and review pass. Native changes are not mobile OTA changes.
- G6 validate: prove the actual production/user path and return evidence to the
parent controller or close this run when operating independently.
For multi-agent, interruptible, or adversarial-review work, keep one trace:
python3 scripts/harness_workflow_trace.py init \
--kind health-harness \
--dossier docs/dossiers/<date>-<slug>.md \
--budget-tokens <hard-limit> \
--label "<short label>"
When delegated, reuse the parent run instead of calling init. Record subagent
starts with spawn, Gate decisions with verdict, checkpoints with event, and
resume from summary. Exit code 2 means stop and reduce scope or seek a human
decision.
Completion discipline
- Do not claim success from edits alone; report fresh command output and user-path evidence.
- Do not suppress exceptions, test failures, security blocks, or missing production proof.
- Preserve unrelated changes; stage and publish only the files owned by this run.
- If the user says pause, immediately preserve state and stop all mutating or release actions.