verify-learning-staleness

name: verify-learning-staleness description: "Scans verify-learning/SKILL.md for stale Check:/Bash: assertions whose targets (paths, phase headers, grep patterns) no longer exist in the codebase. Four lanes: L1 path references, L2 Phase/Step references inside SKILL.md targets, L3 grep-target patterns, L3b Grep Phase X for pattern`` variant. Skips negative assertions ("MUST NOT exist", "test ! -f", "absent") and external paths (meta/, world/). Use whenever the agent suspects verification-checklist drift — after major SKILL.md refactors, after extracting inline pseudocode into scripts, after renaming phases, OR via the recurring goal under asp-115 (g-115-219) that fires weekly. Sister to felt-sense-checkin Phase 5b (which calls the same scanner at 75-goal cadence) — this skill is the on-demand entry point." user-invocable: false triggers: [verify-learning-staleness, staleness-scan, check-drift, audit-verify-learning, stale-checks, verify-learning-audit] tools_used: [Bash] companion_scripts: [core/scripts/verify-learning-staleness.py] conventions: [] minimum_mode: assistant revision_id: "skill-bootstrap-verify-learning-staleness-1a0562" previous_revision_id: null

/verify-learning-staleness — Verify-Learning Staleness Scanner

Detect stale Check: and Bash: assertions in verify-learning/SKILL.md whose underlying targets (paths, phase headers, grep patterns) have been moved, renamed, or extracted out of the codebase without the check being updated.

Why This Exists

verify-learning/SKILL.md accumulates ~2000 verification assertions over time. Each assertion references a concrete artifact: a script path, a SKILL.md phase number, a grep pattern in a named file. When refactors happen — extracting inline pseudocode into Python, renumbering phases, deleting retired scripts — the verification check goes stale silently. Future /verify-learning runs report PASS for stale checks because the assertion can't even be evaluated against current code, producing false confidence.

This skill is the structural complement: it scans every assertion and reports which ones reference artifacts the codebase no longer contains.

Companion Script

• core/scripts/verify-learning-staleness.py — the canonical scanner. Modes:
  • Default (no flag) — JSON to stdout, exit 0 if clean / exit 1 if stale.
  • --text — human summary on stdout, JSON to stderr.
  • --skill-md <path> — override the target file (default: framework path).

Procedure

1. Bash: py -3 core/scripts/verify-learning-staleness.py --text — run the scanner. Output lists each stale finding as [L<lane>] line N: <stale_ref> with detail explaining why the target couldn't be resolved.

2. For each finding, decide the disposition:

   • Relocated — the underlying contract still exists, but moved to a different file/phase/script. Update the Check: text to reference the new location.
   • Retired — the contract was removed entirely. Delete the Check: line.
   • Future-tense — the check was added speculatively for upcoming work. Leave alone if work is genuinely imminent, otherwise convert to a comment or delete.

3. After applying fixes, re-run the scanner — stale_count should ratchet down (or stay constant only with documented future-tense exemptions).

Lane Reference

Negative-Assertion Guard

Checks that EXPECT artifacts to be missing (e.g., test ! -f X, X does NOT exist, Phase X absent, MUST NOT contain Y) are exempt from staleness reporting. The negative-phrase list lives in _NEGATIVE_PHRASES in the scanner script — extend there if a new negative idiom appears.

Expected Output

Clean state:

verify-learning-staleness: scanned <N> assertions, found 0 stale

Drift state:

verify-learning-staleness: scanned <N> assertions, found <M> stale
  [L2_phase] line 162: Phase 0.5.0a in .claude/skills/aspirations-precheck/SKILL.md
    Phase '0.5.0a' not found in ... — phase/step was renumbered, removed, or extracted
  [L1_path] line 3892: core/scripts/validate-loop-digest.sh
    path 'core/scripts/validate-loop-digest.sh' does not exist
  ...

Failure Modes

• Scanner returns 0 stale but verify-learning still drifts — the scanner regex didn't catch the pattern. File an Idea goal to extend detection (e.g., add a new lane or loosen _PATH_RE).
• Scanner returns false positives — likely a new negative idiom not in _NEGATIVE_PHRASES. Add the phrase, re-run.
• Scanner can't find a referenced grep target — an L3 finding for grep target 'X' could not be resolved. Either the basename collides with multiple files (rare, scanner takes first match) or the file truly doesn't exist (true positive).

Chaining

• Called by:
  • The aspirations loop via recurring goal g-115-219 under asp-115 Recurring Infrastructure Monitoring (interval 168h / weekly). Pulls drift detection into the regular hygiene rotation.
  • felt-sense-checkin Phase 5b at the 75-completed-goals cadence.
  • The agent on-demand whenever it suspects drift — e.g., right after editing verify-learning/SKILL.md, after a SKILL.md extraction refactor, or when a /verify-learning run reports suspicious 100% PASS.
• Calls: core/scripts/verify-learning-staleness.py.
• Does NOT call: /verify-learning itself (that's a much heavier diagnostic — this skill is the structural pre-check, not a replacement).
• Modifies: nothing. This skill is read-only. Stale-check repair is done via the standard Edit tool against verify-learning/SKILL.md after the scan reports findings.

Return Protocol

See .claude/rules/return-protocol.md — last action must be a tool call, not text. The terminal Bash call is the scanner invocation; never end the skill with a text summary paragraph.