# Agent Harness Status

This status map follows the Agent Harness diagram: Workflow is cognitive
orchestration, Infra is runtime governance.

## Related Documentation
- `README.md`: top-level quick start and Harness Runtime overview.
- `docs/repo-index.md`: low-token repo navigation and runtime surface index.
- `docs/HARNESS_RUNTIME.md`: lifecycle, evidence, checkpoint, permission, and subagent contracts.
- `docs/LIFECYCLE_SKILL_ROUTING.md`: lifecycle stage, workflow, skill, and helper routing.
- `docs/index.html`: Chinese public Delivery Harness Framework docs entry for GitHub Pages.
- `docs/index-en.html`: English public Delivery Harness Framework docs entry for GitHub Pages.
- `docs/delivery-harness-beginner-guide-cn.html`: beginner-oriented Chinese Delivery Harness Framework guide.
- `docs/delivery-harness-beginner-guide-en.html`: beginner-oriented English Delivery Harness Framework guide.
- `docs/CODEX_ENV_REPRODUCTION.md`: Codex + Claude environment reproduction guide.
- `docs/project-lifecycle-harness-flow-cn.html`: Chinese vertical lifecycle flow.
- `docs/project-lifecycle-harness-flow-en.html`: English vertical lifecycle flow.
- `docs/project-lifecycle-harness-flow-skills.html`: Chinese lifecycle skill/helper routing visual guide.
- `docs/project-lifecycle-harness-flow-skills-zh-status-style.html`: current styled Chinese Delivery Harness Framework skill/helper routing visual guide.
- `docs/project-lifecycle-harness-flow-skills-en-status-style.html`: current styled English Delivery Harness Framework skill/helper routing visual guide.

| Diagram Module | Current Status | Implemented Evidence | Remaining Gap | Next Step |
| --- | --- | --- | --- | --- |
| Research | done | `docs/repo-index.md`, startup probes in `delivery-harness-framework`, `AGENTS.md` read-first rules | automatic source freshness scoring | add optional research evidence events per source |
| Requirements | done | lifecycle stage table, `docs/templates/harness-requirements.md`, `scripts/harness_requirements.py` | template must be filled per task before becoming source of truth | validate requirements artifacts before planning or development |
| Planning | done | planning stage policy, lifecycle router, `docs/HARNESS_RUNTIME.md` | no graphical plan report | keep plan output text-first; add visual report only when requested |
| Development | done | scoped write policy for `development`, repo change rules, `scripts/harness_agent_team.py` write-set validator, `--emit-evidence` decision receipt, configured dispatch ask-gate | only configured dispatch tool names/patterns are intercepted; runtime-owned shapes may still need explicit support | validate with `--emit-evidence` before multi-worker dispatch and extend configured shapes as the runtime exposes them |
| Validation | done | `verify_codex_env.sh`, `test_runner.py`, evidence schema, verification gate, skip-counted runner summary, hook/schema runtime content checks | no automatic final-answer gate in the model runtime; runtime activation still requires explicit sync approval | keep AGENTS gate and evidence helper; add completion hook if Codex exposes one |
| Sandbox | done | Codex sandbox/approval model, tool policy guard categories, `scripts/harness_env_probe.py` | global sandbox can be observed and reported, not forced by repo | keep probe output in verification and document runtime limits |
| Memory | done | `docs/harness-state.md`, `codex_subconscious.py`, repo index, `scripts/harness_recover.py` | recovery is an explicit smoke output, not automatic model memory | run recovery smoke at session start or before handoff-sensitive work |
| Skills | done | `codex/skills/*`, sync tests, generic lifecycle skill boundary test | skill quality varies by imported upstream content | add targeted validation for critical local skills |
| Session State | done | `docs/harness-state.md`, local evidence JSONL schema, `scripts/harness_checkpoint.py` | runtime state updates still require explicit helper invocation | use checkpoint helper at phase transitions and handoff |
| Permissions | done | `codex/runtime/tool-policy.json`, `harness_guard.py`, state snapshot phase fallback, unknown-phase read-only fallback, handoff repo-write ask-gate | phase inference is still category-level, not path-scoped | add explicit phase marker support in future Codex hook payloads |
| Hooks | done | `SessionStart`, `UserPromptSubmit`, `PreToolUse`, `PostToolUse`, hook sync tests | completion hook not available in current config | keep final verification gate in AGENTS and skill |
| Observability | done | `evidence.schema.json`, `harness_evidence.py`, `harness_observer.py`, `scripts/harness_report.py`, verify PASS/FAIL/SKIP evidence for missing `codex` without early shell exit | no browser dashboard | generate optional visual report from `harness_report.py --json` |
| Docs Drift | done | `docs/surfaces.json`, `scripts/check_surfaces.py --check-public-nav`, `test_check_surfaces_validates_public_nav()` | public HTML nav generation is still manual | optionally generate HTML nav blocks from the manifest |
| Tool Router | done | phase-based policy, guard classifier, payload/env/state phase resolution order, and configured `agent_dispatch` category | not all Codex tools expose identical payload shape | keep payload parser permissive and test common forms |
| Model Router | done | `codex/hooks/model_router.py`, prompt/subtask complexity tests | hook can recommend and emit JSON; actual model switching depends on runtime or wrapper support | keep deterministic, non-blocking, and quality-floor first |
| Checkpoints | done | `docs/HARNESS_RUNTIME.md` checkpoint contract, `docs/harness-state.md`, `scripts/harness_checkpoint.py` | helper records state but deliberately does not commit | create commits only on explicit user request; otherwise append state checkpoints |
| Guardrails | done | guard hook blocks or asks on destructive, secret, remote, dynamic-exec, phase write violations, unreceipted configured agent dispatch, and verify drift checks for live hook/schema mismatches | high-risk patterns and dispatch tool names need ongoing tuning; live `~/.codex` enforcement is not activated until the approved sync runs | extend policy with observed false positives/negatives |