Repository navigation
v1.36.0 — dispatch_count review-cost telemetry
Ports stride's dispatch_count telemetry (W2130), completing the G417 review-convergence port alongside v1.34.0 (the two-round review cap) and v1.35.0 (the cosmetic finding class).
The reviewer entry of workflow_steps may now carry an optional dispatch_count: how many times that subagent was dispatched, counting one re-dispatched after a crash, because a crashed dispatch still spent its tokens. It counts dispatches, not rounds — the round cap excludes a crashed dispatch and this key includes it. Optional, omitting stays valid, and no seventh step name.
The limits are the half that matters. A figure trusted past its accuracy is worse than no figure, because the pair can rank two tasks in the wrong order. All six of stride's limits ship inline, since this port's skills are single-file:
- Wall-clock is not token cost. Measured tok/s varied ~2.1× by dispatch kind; on real records the pair ranked one task 20.3% more expensive when it was in fact 1.4% cheaper.
duration_ms / dispatch_countis not a per-round figure — the duration aggregates folded-in sub-step dispatches while the count covers only the reviewer's own. Both fold-in sites now say: fold the duration, do not increment the count.- A fully compliant
3is not a two-round-cap breach — the count cannot separate a crashed re-dispatch from a genuine extra round, which matters here precisely because v1.34.0 shipped that cap.
Nothing validates the value on the way in — verified against the server's valid_step?/1 rather than assumed — so the contract assigns the guard obligation to the first consumer that reads it.
Suite: 661 passed / 0 failed, up from 640. check-port-canon.sh now reports no missing canon entries for stride-codex.