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

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:

1. Prompt / routing — merge into AGENTS.md or create harness section; keep ≤150 lines total.
2. Skills — ensure Skill Invocation block present; list installed skills in Orchestration Map.
3. Tools — document MCP/tool surfaces in manifest; stub docs/harness/tools.md if needed.
4. Governance — forbidden paths, allowed-write scopes, verifier sandbox rules.
5. Eval interface stub — docs/harness/eval-interface.md + empty docs/harness/tasks.json shape (file_phrase + command checks per metaharness pattern).
6. 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

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

Common Rationalizations

Verification

• docs/harness/manifest.json exists with 7 components and hashes
• Eval interface stub + tasks.json shape present
• AGENTS.md ≤150 lines; user content preserved on merge
• Governance forbidden paths documented
• SKILL-OUTPUTS.md appended

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 map
• references/scaffold-patterns.md — inspect→render, drift CI, bootstrap snapshot
• references/examples.md — library project + consumer repo worked examples