guarded-knowledge-closeout

name	guarded-knowledge-closeout
description	Use when working on OmniAuto knowledge closeout, observation emission, manual closeout, or AI-assisted candidate generation. This skill governs both AI behavior and the platform policy boundaries for automatic knowledge growth, and should be used whenever a task touches `platform/src/omniauto/knowledge/`, `knowledge/`, `omni closeout`, or candidate promotion rules.

Guarded Knowledge Closeout

This skill is the governance layer for OmniAuto's automatic knowledge growth.

Use it whenever you:

• change platform/src/omniauto/knowledge/
• add or refine record_knowledge_observation(...)
• adjust omni run / omni closeout closeout behavior
• review or generate AI-assisted knowledge candidates
• update closeout-related docs or tests

Read First

Before changing behavior, open:

• ../../platform/src/omniauto/knowledge/policy.py
• ../../platform/src/omniauto/knowledge/manager.py
• ../../docs/KNOWLEDGE_WORKFLOW.md

Non-Negotiable Rules

1. Treat the existing rule-based closeout path as the authoritative base. Preserve default behavior unless the user explicitly requests a policy change.
2. Automatic writes may only land in knowledge/ and runtime/knowledge_runs/.
3. Automatic writes must never touch skills/, platform/src/, or platform/tests/.
4. AI-assisted output is review-only by default. It may write candidate notes under knowledge/review/ai_candidates/, but it must not directly edit formal knowledge files.
5. skills/ and hard platform/ landings always require explicit user approval.

Controlled Task Rules

• Controlled automatic closeout is for workflows under workflows/.
• External or ad hoc scripts should use omni closeout or OmniAutoService.closeout_task(...).
• If you need to change what counts as controlled, update policy.py, manager.py, docs, and tests together.

Observation Rules

When task code emits observations, keep them structured and minimal:

• kind: pattern, lesson, capability, or proposal
• title
• summary
• domain
• evidence

Prefer the smallest reliable observation over a long narrative. If evidence is weak, do not emit a stronger claim.

AI Candidate Rules

AI-assisted closeout is a strict sidecar, not the main pipeline.

• Default mode should remain conservative.
• AI input must be an evidence pack, not a free-form repository scan.
• AI output must be schema-shaped and evidence-bound.
• Invalid candidates should be discarded, not coerced into formal knowledge.
• Candidate files stay in knowledge/review/ai_candidates/ until a human promotes or rewrites them.

Change Checklist

When changing closeout governance:

1. Update policy.py first if the rule itself changes.
2. Update manager.py only after the policy boundary is clear.
3. Keep AI-assist behavior isolated from the main success/failure path.
4. Add or update unit tests for policy preservation and write boundaries.
5. Run static compile, targeted closeout tests, a real omni run probe, then full pytest.