[deve1993]

Why

Production routing has been degraded for ~10 days. skill_stats reported a healthy 293-skill catalog, but only 27 were active — and only active skills are ever listed or routed:

type	total	active
domain	129	19
process	114	0
command	23	0
agent	22	3
lifecycle	5	5

The 19 surviving domain skills were the 17 marketing skills from #9 plus nextjs and _routing-index, so skill_route answered every query from those:

"add stripe payments checkout"   → attribution, marketing-plan, sms, marketing-loops, prospecting
"setup Payload CMS multi-tenant" → lead-magnets, attribution, ads, marketing-plan, prospecting

The same queries against a healthy catalog return payments + stripe-subscriptions-webhooks and payload + mongodb + cms. The ranking algorithm was never at fault — the data was.

skill_read kept working for the missing skills, which is why nothing looked broken.

Root cause

A --full prune whose discovery set covered only a slice of the bundle. The guard added in 5438d7a only refused on skills.length === 0, so a catastrophically partial discovery passed straight through and deprecated everything else. The survivors match the prune's own protection clause exactly (category NOT IN ('System','Lifecycle')) — _routing-index and the 5 lifecycle skills.

Changes

Prevention — prune now refuses when it would deprecate >25% of the eligible active catalog and ≥10 skills. The absolute floor keeps small catalogs working (1 of 2 is a legitimate 50% prune). --force is the explicit override.

Recovery — new --reactivate, the exact inverse of --full: restores deprecated skills that ARE in the bundle. Genuinely retired skills never appear in the discovery set so they stay deprecated; pending security-gate quarantines are never auto-approved.

Detection — stats() counted every status while list/route serve only active. That contradiction is how this stayed invisible. It now reports active / byStatus / activeBy*, and skill_stats prints a warning banner when the active catalog drops below half the total.

Deploy gap — 29 domain skills existed only in the .claude/skill workspace mirror and were never COPYed into the image. Confirmed against prod: skill_read("stripe-subscriptions-webhooks") → not found. Now in the bundle.

CRLF frontmatter — seo-page and seo-programmatic used CRLF, so ^---\n failed and both lost their entire frontmatter, landing in the DB as "Domain skill: <name>" with none of the trigger keywords routing ranks on. Parser now tolerates CRLF; both files normalised.

Cross-zone duplicates — the bundle ships 29 names in both data/skill/ and data/lifecycle-skills/; whichever copy was walked last silently decided the skill's type. Now collapsed deterministically and reported.

Minor — expo-ui-* renamed to kebab-case (they were keyed under "Expo UI SwiftUI", unreachable by the obvious name); typecheck script added to codegraph (pnpm -r typecheck skipped it); /build/ gitignored.

Verification

Build 3/3, typecheck 3/3 (now including codegraph), 363 tests pass. Four new regression tests cover the partial-discovery guard, --reactivate semantics, CRLF frontmatter, and cross-zone collapse.

End-to-end, against a workspace built exactly the way entrypoint.sh builds one in the container:

1. Imported the bundle → 315 active
2. Recreated the incident (deprecated all but the marketing skills) → 24 active
3. Ran --reactivate → 315 active, the 2 security-gate quarantines correctly untouched
4. Routing verified correct on payments / Payload / 3D / Coolify queries

Deploying this

Merging + redeploying shows the 29 missing skills and installs the guard, but does not by itself repair the catalog — the boot import passes status: NULL and the upsert deliberately keeps existing status rather than resurrecting deprecated rows. Recovery is an explicit step, in the container:

sqlite3 /data/.codegraph/graph.db ".dump" | gzip > /data/.codegraph/pre-reactivate.sql.gz

DATA_DIR=/data
mkdir -p "$DATA_DIR/.opencode" "$DATA_DIR/.agents/skills"
ln -sf /app/data/skill   "$DATA_DIR/.opencode/skill"
ln -sf /app/data/agents  "$DATA_DIR/.opencode/agents"
ln -sf /app/data/command "$DATA_DIR/.opencode/command"
for d in /app/data/lifecycle-skills/*/; do
  name=$(basename "$d"); mkdir -p "$DATA_DIR/.agents/skills/$name"
  ln -sf "$d/SKILL.md" "$DATA_DIR/.agents/skills/$name/SKILL.md"
done

node dist/cli.js import-skills "$DATA_DIR" --reactivate

rm -rf "$DATA_DIR/.opencode" "$DATA_DIR/.agents"

The symlinks are required — without them discovery is empty and --reactivate restores 0 (harmless, but useless). Docs updated in DEPLOY-SKILLS.md.

Left out deliberately

• .claude/skill/pixarts/ holds 6 real skills that are never imported (walkDir doesn't recurse into nested dirs). They self-name pixarts/*, which the project's naming rule disallows — renaming them is a content decision, not a bug fix.
• webxr-spatial, shopify-expert sit in pending on a skill-guard BLOCK verdict. By design those need human review.

🤖 Generated with Claude Code