feat: per-rollout token limits (max_input/output/total_tokens)

[mikasenghaas]

Summary

• Add framework-enforced per-rollout token budgets to EnvConfig, alongside the existing max_turns:
  • max_input_tokens — caps the trace's prompt_len
  • max_output_tokens — caps the trace's completion_len
  • max_total_tokens — caps the trace's total_tokens
• Introduce RolloutLimits (a small frozen bundle in interception.py) that holds all framework limits — turns + the three token budgets — and exposes reached(trace), returning the name of the first limit reached (or None). This subsumes the old standalone max_turns check so all framework limits go through one place and thread as a single object (Environment → Rollout → InterceptionServer).
• The interception server checks limits.reached(trace) before serving each turn (same mechanism as a @stop): the first limit reached refuses the turn, halting any harness, and is recorded as the trace's stop_condition.
• Trace.is_truncated now treats the token-limit stop conditions (max_input_tokens / max_output_tokens / max_total_tokens) as truncation, like max_turns / harness_timeout.

The checks read the trace's existing computed properties (prompt_len / completion_len / total_tokens), so they apply uniformly to any harness. Token caps are soft by one turn: they're checked between turns, so the turn that crosses a cap still completes.

Breaking

• InterceptionServer(...) and Rollout(...) no longer take a max_turns keyword — they take limits: RolloutLimits | None instead. Environment.max_turns is replaced by Environment.limits. These are internal constructors (only Environment.episode / Rollout.run call them); the public EnvConfig.max_turns config field and its CLI flag are unchanged.

Verification

• ruff check --isolated clean on all changed files.
• Smoke-tested RolloutLimits.reached: >= boundary semantics, the input/output/total → prompt_len/completion_len/total_tokens mapping, and limit priority (turns before tokens when both are reached). Imports of env / rollout / trace / interception confirmed (no circular import from the new env → interception edge).

Follow-up (separate PR, prime-rl side): surface these fields on the orchestrator's env config so they can be set from training TOMLs.

Note

Add per-rollout token limits (max_input/output/total_tokens) to rollout execution

• Introduces a RolloutLimits frozen dataclass in interception.py that holds optional caps for turns, input tokens, output tokens, and total tokens, with a reached() method that checks a Trace against all limits.
• Adds max_input_tokens, max_output_tokens, and max_total_tokens fields to EnvConfig in env.py, which are aggregated into a RolloutLimits object and passed down to each Rollout.
• The InterceptionServer now calls limits.reached() before serving each turn; if any limit is hit, it returns HTTP 400 with rollout stopped: {limit} and sets the trace stop condition to the limit name.
• Trace.is_truncated in trace.py now returns True for the three new token-limit stop conditions in addition to max_turns and harness_timeout.

^{Macroscope summarized 5ae2372.}

---

Note

Medium Risk
Changes rollout halt behavior and internal Rollout/InterceptionServer APIs; misconfigured limits could truncate eval rollouts unexpectedly, but enforcement mirrors existing max_turns.

Overview
Adds framework-enforced per-rollout token budgets on EnvConfig (max_input_tokens, max_output_tokens, max_total_tokens) alongside existing max_turns, exposed as top-level eval/serve CLI flags like other env limits.

Introduces RolloutLimits with reached(trace) so turn and token caps are checked in one place before each intercepted chat turn (same halt path as @stop). Environment builds limits from config and passes them through Rollout to InterceptionServer instead of a standalone max_turns argument.

Trace.is_truncated now treats token-limit stop conditions (max_input_tokens, max_output_tokens, max_total_tokens) as truncation, consistent with max_turns and harness_timeout. Token caps are evaluated between turns (soft by one turn).

^{Reviewed by Cursor Bugbot for commit 5ae2372. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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.}

^{Reviewed by Cursor Bugbot for commit 5ae2372. Configure here.}

[cursor]

Missing docs for token limits

Medium Severity

This PR adds framework max_input_tokens, max_output_tokens, and max_total_tokens on EnvConfig (top-level eval/TOML flags like max_turns), but none of the core docs under docs/ describe them. Users have no reference for configuring or interpreting these limits and truncation stop conditions.

 

^{Triggered by project rule: BugBot Instructions}

^{Reviewed by Cursor Bugbot for commit 5ae2372. Configure here.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR introduces new framework-enforced token limits that change rollout termination behavior - a new feature capability. While the implementation is straightforward and well-structured, new features introducing user-facing behavior changes warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}