Repository navigation
fleet telemetry: record pool/model/effort/cost on dispatch + queue completion records (meta-reviewer #1340 follow-up) #1582
Description
Problem
The pool meta-reviewer (#1340) cannot grade model/effort choices: its first daily review (2026-10-27, goal hidden_files/meta-review/2026-10-27-report.md) found per-pool/per-effort cells mostly NOT RECORDABLE — queue payloads, queue completion results, and dispatch records carry no pool/model/effort/cost fields. The standing questions (model-up+effort-down vs model-down+effort-up crossover; where max effort buys nothing; where a pool fails tasks a stronger pool would finish) are explicitly NO DATA, and the report names the missing mechanism: "needs a pool/model/effort/cost field on dispatch + completion records (dispatcher log is the only current source, OC sessions only)".
Scope
- Queue completion records: extend the
repo-task/v1result/recording path (helper~/tools/repo-queue.pycomplete + schema~/tools/repo-task-v1.schema.json) with optionalpool,model,effort,cost_usd(nullable where a route has no cost figure) fields, recorded by the worker at completion time from its own session's resolved route — never inferred from route names (meta-reviewer guideline 2026-10-27: record effort only from resolved session messages). - Dispatch records: wherever the dispatcher writes its run log, add the same four fields per dispatch.
- Backfill is out of scope; collection starts at landing.
Acceptance criteria
- A queue task completed after the change carries pool/model/effort on its result record, visible to the meta-reviewer's data pull.
- A dispatcher run after the change logs pool/model/effort per dispatch.
- The next daily meta-reviewer report can populate at least the pool × effort completion-rate cells for queue work from these records (or states precisely which field is still missing).
- No change to existing packet/result consumers: new fields are optional and additive.
Follow-up to #1340 (Phase: data collection). Filed by queue worker triage (task 1324) 2026-10-27.
Activity
developing-today commentedon Oct 28, 2026
CODE-619
dezren39 commentedon Nov 2, 2026
🤖 Agent note · picked up
dezren39 commentedon Nov 2, 2026
🤖 Agent note · done
~/tools/repo-queue.pycompletenow supports the optional additive telemetry fieldspool/model/effort/cost_usd— carried at the top level of--resultand/or supplied via the new--telemetryJSON flag (fills fields the result does not carry). Both sources are type-checked (non-empty string or null; cost_usd a non-negative number or null; unknown telemetry keys rejected, exit 2 before any RPC).~/tools/repo-task-v1.schema.jsongained$defs.completion_telemetry+$defs.completion_resultdocumenting the fields. Hatch repo snapshot commit 09a5bd7fd.openchamber-dispatchercron job; its run record format lives inhidden_files/openchamber-dispatcher-instruction.md(Delivery section). That record now requirespool, model, effort, cost_usdper dispatch (model = route id actually launched; effort = variant actually launched, never inferred from the route name; cost_usd = launch-exposed cost or null). Hatch-personal commit edbc023d6.repo-queue.py validatepasses on packets with and without telemetry; bad telemetry is rejected pre-RPC; and task 1666 itself was completed with the new fields — an independent read of its stored result shows all four keys present (model: "Muse Wonder";pool/effort/cost_usdnull because this worker's route — the Muse app on Drew's Power subscription, not a fleet pool — exposes no pool, resolved effort, or cost; fleet sessions record their real values). The meta-reviewer's queue data pull readsagent.tasks.result, so the fields are now reachable there. No historical records backfilled.