Classify orchestrator child stderr and throttle skip noise

Context

Operator observation on 2026-05-16 showed npm run ai:orchestrator running a large Knowledge Base sync and logging two separate noise classes:

• repeated Skipping knowledge base sync; task already running (PID: 2082). lines every poll while the existing sync child was still active
• normal KB embedding progress emitted through the daemon as [ERROR] [ProcessSupervisor] knowledge base sync stderr: [LOG] Processed and embedded batch 83 of 237

The embedding model/vector-dimension question is already closed separately: qwen3-8b @ 4096 dimensions is confirmed by healthcheck, direct endpoint smoke, and Chroma sample inspection. This ticket is only about daemon observability quality while that sync runs.

The Problem

ProcessSupervisorService currently treats all child stderr data as daemon-level ERROR, even when the child is using stderr as its normal log channel. That is correct enough for MCP/harness transport visibility, but it is wrong at the daemon supervisor boundary: an informational child line should not become an operator-facing daemon error.

The repeated already-running skip line is also low-value noise during long-running child tasks. It confirms the singleton guard is working, but printing the same message every orchestrator poll obscures real progress and real failures.

The Architectural Reality

• ai/daemons/services/ProcessSupervisorService.mjs:195 detects a task already running and logs an INFO skip every invocation.
• ai/daemons/services/ProcessSupervisorService.mjs:207 spawns child tasks with stderr piped.
• ai/daemons/services/ProcessSupervisorService.mjs:209 forwards every stderr chunk as writeLog('ERROR', ...) regardless of child log prefix.
• ai/daemons/TaskDefinitions.mjs:72 defines kbSync as a child process label knowledge base sync, so buildScripts/ai/syncKnowledgeBase.mjs progress lines flow through this supervisor path.
• ai/daemons/Orchestrator.mjs:295 uses the orchestrator poll cadence, so already-running state is checked repeatedly while a long child task is active.

Existing adjacent tickets are not equivalent:

• #10576 / #10579 handled durable KB sync logging and are closed.
• #10088 is broader post-merge KB sync automation.
• #9919 / #9920 cover daemonization and re-queueing.
• #11075 is deferred config cleanup.

The Fix

Add a narrow ProcessSupervisorService logging policy for child task output:

1. Classify child stderr lines by explicit child log prefix before forwarding to daemon logs. Examples: [LOG] and [INFO] should map to daemon INFO; [WARN] to WARN; [ERROR] or unclassified stderr can remain ERROR unless a stronger local convention exists.
2. Preserve stderr capture. The child process should continue piping stderr so MCP/server/script log channels remain observable.
3. Throttle or dedupe the already-running skip line per task/PID/reason window so a long KB sync does not emit identical skip messages every poll.
4. Keep failure semantics intact: non-zero exits, spawn failures, and real child error lines still surface as ERROR and still record failed task outcomes.
5. Add focused unit coverage in test/playwright/unit/ai/daemons/services/ProcessSupervisorService.spec.mjs for classification and skip dedupe/throttle.

Contract Ledger Matrix

Acceptance Criteria

• Child stderr lines prefixed [LOG] or [INFO] are written as daemon INFO, not ERROR.
• Child stderr lines prefixed [WARN] are written as daemon WARN.
• Child stderr lines prefixed [ERROR] and unclassified stderr continue to be written as daemon ERROR.
• Already-running skip logs are deduped or throttled per task/PID/reason window while still recording a skipped outcome if that is required for health state.
• Existing spawn failure, child error event, and non-zero exit logging remain ERROR.
• Unit tests cover severity classification and repeated already-running skip behavior.
• No changes remove stderr usage from MCP servers or KB sync scripts; this is a daemon supervisor classification fix only.

Out of Scope

• Changing the embedding provider, vector dimensions, or KB content sync semantics.
• Removing stderr from MCP servers or CLI scripts.
• Reworking the full logging framework or introducing a new structured logger.
• Changing orchestrator poll cadence or KB sync interval defaults.
• Implementing post-merge auto-sync; that belongs to existing automation tickets.

Related

• #10576 — KB sync observability: tee MCP child logger output to .neo-ai-data/logs/ for tail-able progress
• #10579 — KB Sync: Implement stderr observability (Shape A)
• #10088 — Automate post-merge knowledge-base sync trigger
• #9919 — Implement fs.watch Daemonization for Autonomous Orchestrator
• #9920 — Agent Error Recovery & Re-Queueing for Orchestrator
• #11075 — Migrate orchestrator magic numbers + retention constants to config

Origin Session ID: 6ec143cb-2e5b-4964-94d6-eb28cb25bde2

Handoff Retrieval Hint: query_raw_memories(query="orchestrator ProcessSupervisor knowledge base sync stderr LOG batch task already running PID skip noise child stderr severity classification")