Keep duplicate /skill-doctor rows, refuse zero-row tables, report unused plugins; record the 2026-10-15 plugin dispositions - #1153

Scope

• What this PR changes, in one or two sentences: tools/skill-usage/skill_usage.py keeps every /skill-doctor row, so a skill listed from two sources loses neither row. It refuses a table that parses to zero rows as "not measured", and it reports each plugin whose listed skills all show no use in 7 days. The practice skill's skills-library row records the 2026-10-15 plugin dispositions and the claude-native-practice double row (option (a), accepted by the command center).
• Base commit: f165498aa
• Lane: lane:foundation
• Owned paths touched:
  • tools/skill-usage/skill_usage.py, its README.md and plugin-exceptions.json (new and empty);
  • tests/test_skill_usage.py, tests/test_native_practice_skills_listing.py (new) and tests/fixtures/skill_usage/skill-doctor-duplicate.txt (new);
  • the practice skill's slots.json and instructions-skills.md, plus the slot manifest's sha256_when_labeled;
  • evidence/artifacts/skills-listing-20261015/ (new);
  • docs/decisions/2026-10-04-pwsh7-windows-guidance.md: a line reference only;
  • manifests/evidence.json.

Double listing, measured cost (a correction to the accepted basis): option (a) was accepted on a cost of about 130 tokens/turn for the second row. The model is in fact shown the skill once, so the second row costs 0:

• Model listing: in 2.1.300 the model's skill listing (T5r, byte 228551628) passes loaded skills through Oi(…,"name"), where var Oi=T (byte 216279796) is lodash uniqBy.
• /skill-doctor table: its skill set (Hfr, byte 227068501) uses the same dedupe. The table (tkn, byte 238412509) prints that name's estimate on every row, and the use cells are keyed by name too.
• Transcripts: all 182 initial skill-listing attachments in this repository's sessions since the personal copy appeared (2026-10-11T23:38:30Z; clients 2.1.296, 2.1.299 and 2.1.300) name it once.

The report gives table_context_tokens (260: the two rows added together, not a cost) and listing_context_tokens (130). The row stays option (a): no setting change, with the same two checks re-run each release.

Items:

1. Rows kept: duplicate rows keep their cells, and each name gains a listings list.
2. Not measured: zero parsed rows give "not measured", and the message says whether the text was empty and how many skills skillUsage lists. skillUsage is read in process, names and counts only, and only when no row parsed. An empty table counts as a broken capture, because the client prints "(no skills loaded)" rather than empty text.
3. Unused plugins: listing_sources and plugins_unused_7d, minus the exceptions in plugin-exceptions.json, which starts empty. Owner-direction exceptions are added there by pull request.
4. Skills-library row:
   • the built-in claude-api skill supersedes the disabled claude-api@anthropic-agent-skills plugin;
   • the typesafe plugin duplicate is disabled, and the preloaded typesafe-ai is kept;
   • the synced plugins are disabled (Record the claude.ai synced plugins as an owner-synced exception; correct the sync row (#27) #1138);
   • the double row is recorded as one listing.
5. ci: maintain native research automation and attest catalog archives #26: the syncClaudeAiSkills rejection is reworded. Unset and true both sync. A false in .claude/settings.local.json is honored, and the shared .claude/settings.json is not.

Fresh /skill-doctor run (2026-10-15T06:13:14Z, 2.1.300, 0 turns, $0): 72 rows and 71 names, so 1 duplicate.

• Unused 7 days: 5 plugins (claude-hud, codex, context-mode, socraticode, worktrunk), 0 excepted. The flags are measurements, not a recommendation.
• Row drop: from 140 to 72 rows, matching the 68 rows of the six disabled plugins.

SOTA sources

• Claude Code 2.1.300 binary (sha256 674deba0bba9aee479f4503c438b49303b6fdd304b9a471b14977aaec5215453). Byte offsets:
  • T5r at 228551628;
  • Hfr at 227068501;
  • Oi=T at 216279796;
  • tkn at 238412509;
  • tcs at 227067812.
• anthropics/claude-code CHANGELOG.md at 94506ff0d3bd9185f1752da36e968775021fb48f:
  • L7009 (2.1.69, built-in claude-api);
  • L3528 (2.1.234).
• Claude Code docs:
  • https://code.claude.com/docs/en/settings-reference.md (fetched 2026-10-15T05:52:18Z, sha256 8ba90434…), syncClaudeAiSkills;
  • https://code.claude.com/docs/en/skills.md "Resolve skills that share a name" (05:52:28Z, sha256 4d090296…).
• Earlier measurement: Record the claude.ai synced plugins as an owner-synced exception; correct the sync row (#27) #1138 listing-cost.json at head 9f63fc311acd14ca4551d62e125517f4f563c356.

Evidence-class table

Local commands run

$ python3 -m unittest tests.test_skill_usage tests.test_native_practice_skills_listing   # base code
KeyError: 'listings' / KeyError: 'duplicate_listings' / AttributeError: no attribute 'claude_config_path' /
AssertionError: 'not measured' not found in '' / TypeError: unexpected keyword argument 'skill_usage' /
TypeError: unexpected keyword argument 'plugin_exceptions' / FileNotFoundError plugin-exceptions.json /
AssertionError on the (a)-(c) row text / FileNotFoundError listing-attachments.json / AssertionError on #26 wording
$ python3 -m unittest tests.test_skill_usage tests.test_native_practice_skills_listing
Ran 250 tests in 90.146s
OK (skipped=8)
$ python3 -m unittest tests.test_native_practice_* tests.test_local_pages_* tests.test_skill_usage tests.test_skill_usage_daily \
    tests.test_skills_manifest tests.test_paired_runs_results tests.test_primary_source_reading_counts
917 tests OK (skipped=13)
$ python3 scripts/validate.py
rc 0; {"components": 71, "hashed_files": 12052, "profiles": 4, "receipts": 240, "status": "passed"}

Notes for the reads

• Follow-up, daily run timeout: tools/skill-usage/daily_skill_usage.py:103 passes no --claude-timeout. One of three headless /skill-doctor runs hit the 30 s default, and the retry took about 32 s.
• Follow-up, manifest: adoption/skills/manifest.json still names the two disabled plugins.
• Follow-up, stale headers: the slots.json and page headers still say client 2.1.295.
• Manifest rebase: every open practice-row PR rewrites the slot manifest's sha256_when_labeled, so later ones need a manifest rebase.

Checklist

• No GitHub Actions changed.
• No workflows changed.
• No secrets are printed, logged or committed. skillUsage is read names and counts only.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code