skill-skill-usage-close-loop

name

skill-skill-usage-close-loop

description

Asynchronously close the loop on skills used in the current thread by routing to PersonalBrain's canonical `skill_skill_usage_close_loop` workflow, spawning one bounded background subagent, and appending per-skill evidence rows for later aggregate analysis. Use when a user explicitly asks to record, review, or retrospectively evaluate the performance of one or more skills used in the conversation without blocking the main task.

Skill Usage Close Loop

Overview

Use this skill as the global wrapper for PersonalBrain's canonical skill_skill_usage_close_loop.

Source of truth:

• <PERSONALBRAIN_ROOT>/30_skills/meta/skill_skill_usage_close_loop.md
• <PERSONALBRAIN_ROOT>/40_governance/skill_evolution/ownership_registry.md
• <PERSONALBRAIN_ROOT>/40_governance/skill_evolution/results.tsv

If the canonical PersonalBrain path is unavailable, stop and report that the durable close-loop target is missing instead of improvising a parallel ledger.

Workflow

1. Resolve the canonical context

• treat the PersonalBrain skill file as the durable contract
• if already working inside PersonalBrain, operate there directly
• otherwise read the canonical file above before proceeding

2. Identify the review roster

• prefer the explicit skill list from the user
• otherwise infer only clearly used skills from the current thread
• skip ambiguous candidates rather than inventing a fake roster

3. Keep the main task unblocked

• spawn at most one bounded background subagent
• pass only the minimal task summary, reviewed skills, failures, touched artifacts, and verification signals
• do not wait by default while the main task still has critical-path work
• wait only if the user explicitly wants the review result now or the thread is already at a natural pause

4. Append the durable review ledger

• append one JSON object per reviewed skill to <PERSONALBRAIN_ROOT>/40_governance/skill_evolution/usage_reviews.jsonl
• create the file lazily on first real append
• use eval_mode: thread_review_dry_run unless stronger replay evidence actually exists

5. Route follow-up actions

• keep the review as passive evidence by default
• escalate recurring issues to the PersonalBrain dashboard when they become pattern-level, not one-off noise
• route canonical improvement work to the scorecard or Darwin optimizer flow instead of editing skills silently

Review Row Shape

Include at least:

• timestamp
• review_batch_id
• thread_scope
• reviewed_skill
• used_with
• task_summary
• observed_strengths
• observed_failures
• recommended_next_action
• evidence_refs
• eval_mode
• review_agent
• blocking_mode

Guardrails

• log meaningful skill usage, not every tool call
• keep private or low-signal session chatter out of durable memory
• do not silently edit canonical 30_skills/*.md
• do not create a second ledger outside PersonalBrain
• record a deferred or partial review when evidence is too weak