Skip to content

brain_search ~20% first-call timeout for Cursor MCP clients — infra bug, not a retry-tax to keep paying #458

@EtanHey

Description

@EtanHey

From gen-11 weave E04 (7 evidence spots; lived experience across tonight's judge fleet)

First brain_search call from Cursor CLI clients times out ~20% of the time under enrichment/drain load; retry succeeds. The fleet-wide workaround is a 'retry-once' rule baked into every dispatch prompt — a standing tax on every worker and a silent-degradation source (tonight: 529/673 judge verdicts ran evidence_degraded because search was DB-locked through the batch).

Ask

Chase the root: likely first-call cold path (connection/setup racing the writer lock) in the BrainBar socket path or the busy_timeout handling for read queries. Goal: first call as reliable as the retry, so the retry-once prompt rule can be deleted from fleet dispatch templates.

Evidence: orchestrator docs.local/weave-2026-06-06/ (S4-edits E04) + 2026-06-05-kg-judge-verdicts-report.md (79% degraded batch).

🤖 Filed by brainlayerClaude per weave dispatch

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions