pruning-view

Produce a prune report: stale candidates, stale promoted learnings, and suspected regressions. Use weekly or when startup context feels noisy. Read-only — writes a report but never removes anything on

SKILL.md

GitHub

pruning-view

When to use

  • Weekly hygiene pass.
  • After a noticeable increase in friction records.
  • When stable/global.md approaches the configured budget.

Protocol

  1. Run: bash ~/.claude/evolution/bin/evolve-prune.sh
  2. Run: bash ~/.claude/evolution/bin/evolve-regression.sh
  3. Read the resulting reports/YYYY-MM-DD-prune.md and reports/YYYY-MM-DD-regression.md.
  4. Summarize to the user: N stale candidates, M stale promotions, K suspected regressions.
  5. For each finding, propose the safest action:
    • Stale candidate → move to rejected/ (operator decides).
    • Stale stable learning → suggest /evolution demote <id>.
    • Suspected regression → suggest /evolution demote <id> with the contradicting record IDs as evidence.
  6. Do not execute any mutation. The operator confirms each action.

Staleness definition

From config.yaml: stale_after_days (default 90). A candidate or promoted learning is stale if last_confirmed is older than that cutoff.

Regression definition

config.yaml: regression_failure_count (default 3). A promoted learning is suspected-regression if ≥ N records since its promoted_at share token overlap with its text and are of type correction or friction.

Output

Summary of the two reports inline + the paths to the full files. No writes to stable/.