capability-discovery

name	capability-discovery
description	Discover reusable operational capabilities from command learnings, repeated mistakes, and session evidence, then draft skill candidates for later maintainer promotion without auto-writing real skills.

Capability Discovery

Use when Project Governor should turn repeated operational recovery lessons into reusable workflow candidates instead of leaving them as one-off memory notes.

Purpose

Project self-evolution starts with discovering repeatable capabilities from governed evidence:

• failed commands
• repeated mistakes
• corrective lessons
• restart or recovery procedures
• missing script or CLI contract lookups
• local install/cache refresh lessons

This skill is advisory. It drafts skill candidates and reuse triggers, but it does not create, delete, rename, or overwrite skills/<skill>/ directories.

Process

1. Query governed memory first:

   python3 skills/context-indexer/scripts/query_context_index.py --project . --request "command learning repeated mistake reusable capability skill candidate restart recovery" --memory-search --auto-build --format text
   

2. Run capability discovery:

   python3 skills/capability-discovery/scripts/discover_capability_candidates.py --project . --format text
   

3. If candidates should be available to later sessions, write the state file:

   python3 skills/capability-discovery/scripts/discover_capability_candidates.py --project . --write
   

4. Promote a candidate into a real skill only after maintainer review. Promotion requires:

   • evidence from at least one command learning or repeated mistake
   • a clear trigger
   • a bounded workflow
   • tests or smoke checks for the deterministic helper if one is added
   • skills/CATALOG.json, README, and contract docs updates

Output

The deterministic helper returns:

• candidate capability type
• confidence
• source learning IDs
• suggested skill name, description, triggers, and workflow steps
• recommended next action

With --write, the helper writes:

.project-governor/state/CAPABILITY_CANDIDATES.json

Safety Rules

• Do not mine raw chat logs directly.
• Do not promote a single weak signal into a mandatory rule.
• Do not auto-create skill directories.
• Do not store secrets from command output.
• Prefer updating an existing skill when a matching workflow already exists.