Skill Candidate Harvester

Skill Candidate Harvester

Use this skill when repeated workflows need to move beyond task-local notes and become structured skill candidates or drafts.

Preconditions

• The target environment should have:
  • tracked repos listed in $CODEX_HOME/memories/durable-knowledge/repos.yaml
  • candidate storage under $CODEX_HOME/memories/durable-knowledge/skill-candidates/
  • draft storage under $CODEX_HOME/memories/durable-knowledge/skill-drafts/
• If those paths are missing, seed them first from this skill's assets/ templates.

Read first

1. $CODEX_HOME/memories/durable-knowledge/repos.yaml
2. the candidate schema in assets/skill-candidate.template.yaml
3. each tracked repo's:
   • AGENTS.md
   • docs/work/.current
   • recent task meta.yaml files
   • recent task learning-log.md files
   • recent task status.md files, especially Potential Skills
   • knowledge/changelog.md
   • relevant pitfalls, checklist, or verification docs when evidence needs confirmation

Workflow

1. Load the tracked repo list.
   • use assets/repos.template.yaml only as a seed, not as the live registry
2. Collect candidate evidence.
   • triaged or promoted entries from learning-log.md
   • explicit skill_candidates from task metadata
   • Potential Skills sections in task status files
   • repeated durable findings in changelogs, pitfalls, or checklists
3. Merge by workflow identity.
   • prefer a stable pattern_key first when present
   • otherwise merge by stable skill-style name plus normalized summary similarity
   • merge evidence instead of creating near-duplicates
4. Update or create candidate records under $CODEX_HOME/memories/durable-knowledge/skill-candidates/.
   • use assets/skill-candidate.template.yaml
   • keep pattern_keys, trigger_types, source_entries, recurrence_count, source_repos, source_tasks, validation, and failure modes current
5. Decide whether a draft should be created.
   • require at least two independent tasks or two repositories of evidence
   • require at least one validation path
   • require at least one pitfalls or edge-case record
   • require stable steps that no longer depend on large amounts of one-off context
6. When the threshold is met:
   • create or refresh a draft under $CODEX_HOME/memories/durable-knowledge/skill-drafts/<skill-name>/
   • link back to source repos, source tasks, and verification state
   • hand off to skill-creator for the final installed skill
7. If the workflow is already an installed skill:
   • mark the candidate installed
   • update the global skills changelog instead of generating a duplicate draft

Hard rules

1. Do not create multiple candidates for the same workflow just because wording differs.
2. Do not promote a candidate to a draft without evidence, validation, and at least one failure-mode record.
3. Do not treat manual enthusiasm alone as enough for draft promotion.
4. Do not overwrite source-repo nuance; keep blocked or partial validation states intact.
5. Do not install a skill directly from this skill; draft first, then use skill-creator.

Bundled assets

• assets/skill-candidate.template.yaml
• assets/repos.template.yaml
• assets/global-memory/

Deliverables

1. Tracked repos are aggregated into candidate evidence.
2. Candidate files stay normalized and deduplicated.
3. Mature candidates become drafts with clear provenance.
4. Installed skills remain linked back to their source evidence.