Summary

Adds gbrain smoke-test — a post-container-restart health check + auto-fix tool for gbrain + OpenClaw. Runs 8 built-in checks (Bun, gbrain CLI, doctor, worker, OpenClaw Codex Zod CJS, gateway, embedding API key, brain repo) with auto-fix, plus a user-extensible drop-in mechanism at ~/.gbrain/smoke-tests.d/*.sh.

Motivating bug: OpenClaw's Codex plugin ships a broken zod@4 on every container restart (missing core.cjs → every Codex ACP session fails). The workaround (npm install zod@4 --force) now runs automatically on every bootstrap via gbrain smoke-test.

Shipping as a patch (v0.19.1), not a minor, because the scope is one new CLI command + one skill — honest to size. Broader doctor-consolidation plans (v0.20 candidate) are tracked separately.

What's in the diff

smoke-test feature (commit 12a395b):

• scripts/smoke-test.sh — 204-line bash runner with test-first/fix-after/re-test loop
• skills/smoke-test/SKILL.md — contract, built-in test catalog, drop-in authoring pattern, known-issue database
• src/cli.ts — gbrain smoke-test dispatch (added to CLI_ONLY)
• skills/manifest.json — skillpack registration

Merge + CI fixes (commits 67fc285, 89a567e, a67e159):

• Merge master (resolved src/cli.ts conflict keeping all CLI_ONLY additions from both sides: smoke-test + master's skillpack/routing-eval/skillify)
• skills/RESOLVER.md — smoke-test registered under Operational so resolver_health check passes
• skills/smoke-test/SKILL.md — added ## Anti-Patterns and ## Output Format sections required by test/skills-conformance.test.ts
• llms-full.txt — regenerated via bun run build:llms to match the new RESOLVER row (drift guard now passes)

Test Coverage

[+] scripts/smoke-test.sh (NEW, 204 lines bash, 8 checks)
  └── [GAP] No automated test — bash script, external side effects (bun install, npm install zod, gbrain jobs work).
            Project convention does not unit-test shell scripts. Verified manually + by CI import.

[+] src/cli.ts (smoke-test dispatch)
  └── [★ TESTED] Dispatch pattern covered by existing CLI_ONLY smoke tests.

[+] skills/manifest.json, skills/smoke-test/SKILL.md, skills/RESOLVER.md
  └── [★★ TESTED] test/skills-conformance.test.ts + test/resolver.test.ts + test/check-resolvable.test.ts

[+] llms-full.txt
  └── [★★ TESTED] test/build-llms.test.ts drift guard

COVERAGE: 4/6 paths tested (67%) — 2 gaps are both in the bash script (intentional, shell side effects)

Tests: 136 test files (unchanged — this PR adds no new test files; existing conformance + resolver + build-llms + skillpack-check tests exercise the skill file additions).

CI Status

Last CI run surfaced 8 failures, all addressed in commits 89a567e + a67e159:

• 3× gbrain skillpack-check (doctor exit 1 cascade) → fixed by RESOLVER wiring
• 1× RESOLVER.md > every manifest skill is reachable → fixed
• 1× checkResolvable — all manifest skills reachable → fixed
• 2× skills conformance > smoke-test > Anti-Patterns + Output Format → fixed
• 1× Tier 1 E2E gbrain doctor exits 0 on healthy DB → fixed (same RESOLVER root cause)
• 1× build-llms > committed files match generator output → fixed by regen

Pre-Landing Review

Documented concerns for a future v0.20 consolidation pass (not blockers for this patch):

• Shell script output uses emoji (✅❌🔧⏭️) — deviates from CLAUDE.md no-emoji rule
• curl | bash auto-install of Bun on missing-runtime path — supply-chain pattern inside a health check
• Hardcoded container paths (/data/gbrain, /app/node_modules/openclaw/...) — brittle to OpenClaw layout changes
• Duplicates gbrain doctor for test 3 (shells out + greps)
• No tests on 204 lines of bash

These are tracked in the v0.20 consolidation plan (doctor-as-single-health-surface, discussed separately).

Plan Completion

This PR ships the original smoke-test feature as-is (v0.19.1). The broader consolidation plan at ~/.claude/plans/lets-look-at-what-swirling-finch.md (folding smoke-test + skillpack-check + check-backlinks + lint + orphans + check-resolvable into gbrain doctor with severity + category tags) was not approved and is deferred to a later release.

TODOS

No TODO items completed by this PR. (Open items in TODOS.md are BrainBench Cat 5 + Cat 6, unrelated.)

Test plan

• bun test test/skills-conformance.test.ts test/resolver.test.ts test/check-resolvable.test.ts test/skillpack-check.test.ts test/build-llms.test.ts — all 191 tests pass
• bun test test/e2e/doctor-progress.test.ts test/progress.test.ts — all 17 run, 5 env-skipped, 0 fail
• gbrain doctor exits 0 locally (resolver_health: OK, all 29 skills conformant)
• gbrain skillpack-check returns healthy: true with empty actions

Other locally-failing tests (dream, orphans, progress CLI — all 5-10s subprocess timeouts) are env-dependent on DATABASE_URL; they skip gracefully in CI.

🤖 Generated with Claude Code