Harness Generation
You seed a minimal, auditable harness v0 — the scaffold around a model (prompts,
tools, skills, memory hooks, lifecycle, verification interface, governance). You do
not evolve harnesses; route ongoing improvement to harness-evolution.
Trigger Discipline
Auto-invoke when chained from project-setup 6c, retroactive-project-setup, harness-engineering Step 0, or setup-evaluation harness FAIL. See harness-engineering/references/harness-readiness-gate.md.
Hard Rules
Never generate a harness without a versioned component manifest (docs/harness/manifest.json).
Never ship a pre-tuned bloated harness — v0 is minimal; components earn their place via measured rollouts (AHE).
Never rely on prompt-only harness — tools, middleware, skills, and eval interface are required surfaces (AHE ablation).
Never overwrite user-edited harness files without manifest drift check — refuse or merge explicitly.
Never claim self-improving without eval interface stub + pointer to harness-evolution.
Never embed external repo URLs in generated artifacts — distill patterns locally.
Workflow
Step 1 — Classify delivery context
| Context | v0 focus |
|---|
| agent-loom skill library project | Skill routing, AGENTS.md orchestration map, eval stub for skill quality |
| Consumer app repo | AGENTS.md, .agents/skills/, forbidden paths, key commands |
| Legacy repo (no AGENTS.md) | Gap-fill only — pair with retroactive-project-setup first |
| Multi-agent chain planned | Manifest only — topology is agent-builder, not this skill |
Step 2 — Inventory repo (silent)
Read: manifests, existing AGENTS.md, .agents/skills/, CI configs, docs/memory/,
docs/evals/, docs/harness/. Auto-extract commands per project-setup Step 1b.
Step 3 — Emit component manifest
Write docs/harness/manifest.json per references/component-manifest.md:
seven orthogonal components (prompt, tools, middleware, skills, sub-agents, memory, governance),
each with path, version, sha256, and generated|user-edited status.
Step 4 — Scaffold artifacts
Apply patterns from references/scaffold-patterns.md:
- Prompt / routing — merge into
AGENTS.md or create harness section; keep ≤150 lines total.
- Skills — ensure Skill Invocation block present; list installed skills in Orchestration Map.
- Tools — document MCP/tool surfaces in manifest; stub
docs/harness/tools.md if needed.
- Governance — forbidden paths, allowed-write scopes, verifier sandbox rules.
- Eval interface stub —
docs/harness/eval-interface.md + empty docs/harness/tasks.json shape
(file_phrase + command checks per metaharness pattern).
- Environment bootstrap — compact snapshot block (stack, key commands, git branch) for meta-agents.
If project-setup already ran: merge, never replace user interview content.
Step 5 — Drift gate
Document harness sync --check equivalent: compare manifest hashes to disk; flag user-edited files
skipped on re-run. CI recommendation in manifest notes.
Step 6 — Hand off
| Next need | Route |
|---|
| Full AGENTS.md interview missing | project-setup |
| Eval rubric + regression suite | eval-rubric-design → eval-pipeline |
| Improve harness from failures | harness-evolution (requires eval harness) |
| Multi-agent design | agent-builder (orthogonal) |
Append to docs/skill-outputs/SKILL-OUTPUTS.md. Log harness version v0 in docs/memory/current-state.md if memory suite installed.
Gotchas
- Minimal seed is mandatory. Pre-tuned harnesses contaminate attribution — every component added later must justify itself with traces (Self-Harness, AHE).
- AGENTS.md must stay lean. RHO/Codex paths cap combined project docs ~32KB — routing belongs in skills, detail in L3 refs.
- Harness ≠ agent topology.
agent-builder designs who talks to whom; this skill designs what files/tools/prompts wrap the model.
- Deterministic scaffold beats one-shot LLM init. Inspect → plan → staged writes → manifest (harnessforge pattern); LLM fill is optional fallback only.
- Blueprint validators that shell out are a trust boundary — document, never auto-run untrusted commands.
Output Format
Harness generation complete
Version: v0
Context: [library | consumer | legacy-gap-fill]
Manifest: docs/harness/manifest.json
Artifacts: [list paths]
Eval stub: docs/harness/eval-interface.md
Drift check: [documented command]
Next: [eval-pipeline | harness-evolution | project-setup]
Example
<examples>
<example>
<input>Generate a harness for this React monorepo — we have AGENTS.md from project-setup.</input>
<output>
Harness generation complete
Version: v0 | Context: consumer
Manifest: docs/harness/manifest.json (7 components, all generated)
Merged: Skill Invocation + Orchestration Map preserved from interview
Created: docs/harness/eval-interface.md, docs/harness/tasks.json (stub), docs/harness/governance.md
Next: eval-rubric-design to define harness regression dimensions
</output>
</example>
</examples>
Common Rationalizations
| Excuse | Reality |
|---|
| "AGENTS.md is enough" | No manifest, eval stub, or governance = not a harness v0 |
| "Skip eval stub for now" | harness-evolution hard-fails without it |
| "Copy external harness README" | Distill patterns; secure-scan; no URL embed |
| "Let LLM write the whole harness" | Deterministic manifest + merge beats opaque generation |
| "Agent-builder already ran" | Topology ≠ harness files — run both |
Verification
Red Flags
- Harness v0 bloated with speculative prompts/tools
- User-edited files overwritten without merge
- No eval interface stub when evolution is planned
- External SKILL.md vendored wholesale
Prune Log
Last pruned: 2026-07-05
- Deep learn-from: scaffold-patterns L3 — drift CI, interface validation, AHE sandbox
Impact Report
Harness v0 generated: [context]
Manifest: docs/harness/manifest.json | Components: 7
Eval stub: [yes/no] | Merged AGENTS.md: [yes/no]
Files created: [count] | Next route: [skill]
Reference Files
references/component-manifest.md — seven components, JSON schema, ETCLOVG mapreferences/scaffold-patterns.md — inspect→render, drift CI, bootstrap snapshotreferences/examples.md — library project + consumer repo worked examples