title: Harness Implementation Plan category: projects tags: [harness, ultimate-pi, implementation, architecture] status: developing created: 2026年04月28日 updated: 2026年04月28日 sources:

• "harness-implementation-plan" related:
• "agentic-harness"
• "spec-hardening"
• "structured-planning"
• "grounding-checkpoints"
• "adversarial-verification"
• "persistent-memory"
• "automated-observability" path: "ultimate-pi" purpose: "Full implementation plan for the 8-layer agentic harness" summary: Project page for the ultimate-pi agentic harness implementation. Covers 9 build phases, shared schemas, key ADRs, token budgets, and risk surface. provenance: extracted: 0.9 inferred: 0.1 ambiguous: 0.0

The agentic harness is an 8-layer mandatory pipeline for the ultimate-pi project. Every task flows through all 8 layers. Build order: schemas → memory → spec hardening → planner → execution+grounding → QA → critics → observability → Archon workflow.

For the full concept, see agentic-harness.

• ADR-008 (Black-Box QA): Tests from spec only, never from implementation — see adversarial-verification
• ADR-009 (claude-obsidian Mode B): Replaces Vectra + embeddings with LLM-native search — see persistent-memory

All inter-layer data contracts in lib/harness-schemas.ts:

• HardenedSpec — Layer 1 output (success criteria, anti-criteria, ambiguities, scope)
• ExecutionPlan / PlanNode — Layer 2 output (DAG with dependencies, risk, verification)
• GroundingCheckpoint — Layer 3 checkpoints (spec version, drift, state hash)
• SubtaskResult — Layer 3 output
• TestSuite / QAResult — Layer 4 output (spec-only test cases, coverage, verdict)
• CriticReview / CriticFailure — Layer 5 output (verdict, failures, severity)
• ObservabilitySpec — Layer 6 output (metrics, health checks, alerts)
• WikiEntryType — Layer 7+8 wiki categories
• AgentCapability / TaskRouting — Layer 7 orchestration

lib/harness-config.ts — all tunable parameters. No enable/disable for layers. The harness is always on when installed.

Default config in .pi/harness.json.

Typical 5-subtask plan: ~83,500 tokens overhead + coding tokens.

Major risks and mitigations:

• AI hallucinates ambiguity → max_ambiguity_retries cap; human force-approve via approve-spec
• Invalid plan DAG → Automated cycle detection + validation
• QA tests don't compile → Fallback to schema-only validation if runner fails
• QA test writer leaks implementation → Architectural enforcement: prompt never includes implementation; spec_only is immutable
• Critic false positives → conditional_pass doesn't block; max_attack_rounds caps rework
• Memory grows unbounded → max_entries config; wiki-lint detects orphans; failures never evicted
• Archon runtime dependency → Harness normalizes terminal states; compensation/resume in own state machine
• claude-obsidian is external → Pin skill versions; skills are ~50KB text, vendorable
• LLM-native search cost → hot.md short-circuits most queries at ~500 tokens; deep queries optional

Each phase complete when:

1. TypeScript compiles (npm run check:ts)
2. Extension loads in pi.dev
3. Unit tests pass for core logic
4. Integration test: minimal task through the layer in isolation
5. Phase 5: generated tests pass; spec-only constraint verified
6. Phase 8: archon run harness-pipeline completes end-to-end
7. Workflow status round-trips through Archon JSON