Repository navigation
Per-task turn cap and cost/token budget #236
Description
Goal
Add cheap hard limits per task: a maximum number of turns and a maximum cost/token budget, instead of complex stuck-detection (see Harness Engineering, Part 4: "skip stuck-detection machinery, but ship cheap caps").
Behavior
- First check what already exists on main (step/turn limits, usage tracking, pricing) and extend rather than duplicate.
max_turns: stop the loop cleanly after N model turns.- Budget: max total tokens and/or max cost in USD per task from provider-reported usage; cost uses a configurable per-model price table (unknown model → token budget only, warn once).
- Warn the model once at ~80% of a budget so it can wrap up; at 100%, stop cleanly with a clear message and defined exit status, recorded in output, JSON events, trajectory and OpenTelemetry if enabled.
- Subagents/workers count against the parent's budget.
- Flags + config keys consistent with existing patterns; sensible defaults (turn cap on, cost cap off unless set).
Tests
- Turn cap stops at N.
- Token budget stops the run; 80% warning injected once.
- Cost computed from usage + price table.
- Worker usage counts toward the parent.
- Unknown model price → token budget fallback.
Do not refactor unrelated code. No CHANGELOG edits.
Activity
vicgupta commentedon Oct 30, 2026
Slice in progress: #246 adds the token budget (
--token-limit/token_limit, statusTokenLimitExceeded), the one-time 80% budget warning (message +budget.warningevent), and delegate/worker token accounting against the parent's budget. Step limit, cost limit with a price table, and the unpriced-model warning were already on main.Still left for this issue:
step_limitapart from an explicit 0, and changes behaviour for existing configs).--checklimits line and ininit-configoutput.--token-limitthrough eval, attempts and workflow stages where the other limits are forwarded.Merge pull request #246 from vicgupta/amf/issue-236
vicgupta commentedon Oct 30, 2026
Decision from Vic: keep the turn cap off by default (no default-on turn cap). Remaining work for this issue: show the token limit in
--checkoutput andinit-config, and pass--token-limitthrough eval, attempts and workflow stages.Show the token limit in --check and init-config; forward it through e…
Merge pull request #247 from vicgupta/amf/issue-236