Give board-claim-guard the dumb-fast treatment before it goes standing (no per-write node cold-start)

task-board-claim-guard-dumb-fast

task confidence inferred status backlog 2026-06-16 owner platform-engineer
source log-auditor — surfaced recording DEC-0041 (the cold-start removal); the decision and the hooks README both name this as the one place the per-write node boot could silently return

Give board-claim-guard the dumb-fast treatment before it goes standing

Dumb-fast trace capture + off-hot-path distill/prune removed the per-call node cold-start (~120 ms) from the PostToolUse capture hook — by narrowing the matcher to mutating calls and replacing the node spawn with a dumb-fast .cmd append. That fix has a mirror-image risk on the other hook: the PreToolUse board-claim-guard (board-claim-guard.mjs, Agentic "sprint board" architecture — a git-resident OKF task board worked by bounded, hook-governed Agent SDK loops §5) is a Node script and is currently opt-in / not wired. Wiring it standing as-is would re-introduce the exact ~120 ms node boot DEC-0041 just removed — on every mutating call, on top of capture.

The risk (one place the cold-start could silently come back)

The .claude/hooks/README.md already flags this verbatim (the "Do not pay a second cold-start" callout, DEC-0041): the guard is acceptable while opt-in/local, but before it goes standing it needs the same treatment. This task makes that durable and owned — a doc note evaporates; a board task does not.

Shape (decide, then build)

Two viable approaches, to be chosen and recorded:

  1. A dumb-fast .cmd/shell front that fast-paths the common no-opinion cases (non-task paths, reads, unclaimed or expired-lease) with ~0 interpreter boot, and only escalates to the full node verdict when a live foreign lease might actually be at stake — so the cold-start is paid on the rare contested write, not every write.
  2. Fold the claim check into the runtime board worker (DefaultBoardWorker / the reserved AgentSdkBoardWorker, Build the runtime BoardWorker seam (DefaultBoardWorker + reserved AgentSdkBoardWorker)) so no per-write hook boot is paid at all — the protocol around the claim is already load-bearing there (claim-commit-then-verify-HEAD; GH-Actions concurrency serialization).

Either way, the latency fix must not weaken the guard's deterministic-verdict, fail-open, KB-agnostic (DOSSIER_KB), and isolation guarantees (Agentic "sprint board" architecture — a git-resident OKF task board worked by bounded, hook-governed Agent SDK loops §5).

Notes

Depends on Build the PreToolUse claim/lease governance hook (the guard's verdict logic) as an ordering hint, not a hold — this is the "before you wire it standing" gate on that work, not a separate hook. Owned by Platform / Runtime Engineer (it changes a governance hook's transport). Provenance: surfaced by the log-auditor recording Dumb-fast trace capture + off-hot-path distill/prune, confidence: inferred (agent-judged real, not human-curated). Board globbed before filing — task-pretooluse-claim-lease-hook covers building the verdict, not the dumb-fast transport concern, so this is distinct and not a duplicate.