KB: review merged PR #301 - fix: embed daemon liveness probe, honest health states, recall fast-skip when not warm

[github-actions]

A merged PR may require knowledge-base updates. Review the change and update library/ as needed.

Source PR

• PR: #301 - fix: embed daemon liveness probe, honest health states, recall fast-skip when not warm
• Author: @thenotoriousllama
• Merged into: main
• Files changed: 19

PR description

Summary

Fixes ISS-007/ISS-008: the embed daemon (:3851) could wedge — accept TCP, never reply (event-loop-blocked inference) — while /health kept reporting embeddings on from a one-shot warm latch in the supervisor. Recall then silently degraded to lexical, every search burned the full bounded embed deadline (fast 1.5s / heavy 3s), and doctor — whose only sensor is /health — stayed blind to the whole chain.

Designed for all three live observations from the dogfood box: (a) a wedge while running, (b) the ~40s post-external-respawn warmup tail (embeds at 10–27s before settling to ~25ms) where a binary alive/dead probe would flap, (c) an externally respawned child never being marked warm by the supervisor's spawn-path-only warmup.

Changes

1. Supervisor liveness (src/daemon/runtime/services/embed-supervisor.ts)

• Periodic bounded liveness probe after warm (default 30s cadence, 2s probe timeout, unref'd-timer scheduler seam) + on-demand checkNow().
• State machine warming → warm → suspect → failed: one missed probe → suspect (no flap), a second consecutive miss confirms the wedge → mark not-warm, kill, respawn via the existing bounded crash-restart machinery; the fresh child re-warms through warming (absorbs the 40s tail — no flap).
• Non-reentrant + coalesced probing; additive optional suspect/state/checkNow on the EmbedSupervisor contract (structural fakes unaffected).

2. /health honesty (src/daemon/runtime/health.ts, assemble.ts)

• reasons.embeddings is no longer "known at assembly": it derives from the supervisor's current state per health call and additively gains warming/suspect/failed — not-warm → never on. embeddingsState gains suspect.
• Additive for both consumers: doctor's health-probe parses reasons.embeddings as a plain string (un-blinded with zero doctor changes); the Hive dashboard prefers embeddingsState (parses warming/failed; zod `.ca

Files touched

• .claude-plugin/marketplace.json (+2/-2)
• .claude-plugin/plugin.json (+1/-1)
• CHANGELOG.md (+4/-0)
• embeddings/src/index.ts (+92/-1)
• harnesses/claude-code/.claude-plugin/plugin.json (+1/-1)
• harnesses/codex/package.json (+1/-1)
• harnesses/openclaw/openclaw.plugin.json (+1/-1)
• harnesses/openclaw/package.json (+1/-1)
• package-lock.json (+2/-2)
• package.json (+1/-1)
• src/daemon/runtime/assemble.ts (+20/-2)
• src/daemon/runtime/health.ts (+67/-28)
• src/daemon/runtime/memories/api.ts (+15/-0)
• src/daemon/runtime/memories/recall.ts (+123/-27)
• src/daemon/runtime/services/embed-supervisor.ts (+210/-0)
• tests/daemon/runtime/health.test.ts (+40/-3)
• tests/daemon/runtime/memories/recall-embed-gate.test.ts (+272/-0)
• tests/daemon/runtime/services/embed-supervisor.test.ts (+272/-0)
• tests/embeddings/embed-daemon.test.ts (+99/-0)

---

Auto-generated by .github/workflows/kb-issue-on-merge.yaml on merge.

[coderabbitai]

🔗 Related PRs

#46 - feat(prd-025): semantic recall on by default [merged]
#60 - fix(ci): de-flake main — compaction-live poll-convergence + exec stdout-drain race [merged]
#112 - fix(daemon): never block boot on the job-queue startup reap (v0.1.1) [merged]
#190 - fix: unblock macOS daemon boot [merged]
#240 - Fix vault-scatter (secrets/settings), the secret CLI, and embeddings health honesty [merged]
#242 - Fix the embed-daemon spawn path (embeddings never ran) + honest /health state [merged]

---

📝 Issue Planner

_{Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.}

• Create Plan

---

🧪 Issue enrichment is currently in open beta.

You can configure auto-planning by selecting labels in the issue_enrichment configuration.

To disable automatic issue enrichment, add the following to your .coderabbit.yaml:

issue_enrichment:
  auto_enrich:
    enabled: false

💬 Have feedback or questions? Drop into our discord!