[pjbgf]

https://entire.io/gh/entireio/forgemark/trails/9

Agents keep a per-op latency estimate that rises at once and decays slowly, and stop issuing pushes/clones expected to overrun the level window. In-flight ops get drainGrace (30s) before the hard cancel, so a queueing server no longer sees a burst of cancelled requests at every level end. tokenCovers accounts for the drain.

---

Note

Medium Risk
Changes benchmark timing, cancellation, and JWT token budgeting; behavior under load and long runs shifts, but scope is limited to the forgemark bench harness rather than production forge paths.

Overview
Benchmark agents now track a per-op-kind latency estimate (fast rise on slow samples, slow decay) and refuse to start clones/pushes that are unlikely to finish before the level window ends. Runner.RunLevel passes an explicit end alongside the existing start, extends the level context by drainGrace (30s) only for ops that overrun that estimate, and tokenCovers budgets that drain per remaining level.

Clone, session, and default push loops all gate on fits() for the relevant OpKind. internal/server/runmgr.go comments are updated so live bucket emission still accounts for post-window drain (and session ref cleanup). New agent_test.go covers estimate behavior, fits, and an integration check that the window close does not cancel in-flight HTTP requests.

^{Reviewed by Cursor Bugbot for commit a281030. Configure here.}

[pjbgf]

@cursor review

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit a281030. Configure here.}

[pjbgf]

The changes now report the number of cancelled operations, while trying to reduce them.

Reporting only:

  c=64   push_ok=13360  push/s=111.3    p50=442.8   p95=1403.5   p99=1991.6   max=3950.3   cas=0 err=0 cancelled=64

Avoiding creating new pushes that won't be finished before the end of the session:

  c=64   push_ok=13559  push/s=113.0    p50=433.7   p95=1405.9   p99=2003.3   max=3813.4   cas=0 err=0 cancelled=0