feat: default compaction (256k) and max_depth=1

[samsja]

Summary

Two default changes:

setting	before	after	disable/override
summarize_at_tokens	None (compaction off)	256,000	RLM_SUMMARIZE_AT_TOKENS="" or "0"
max_depth	0 (no sub-agents)	1	RLM_MAX_DEPTH=0

Why

• Compaction on by default: without it, long trajectories die at the model context limit. 256k is a safety-net threshold — typical rollouts never touch it (SWE medians are ~10-30k), but the rare monster trajectory compacts and survives instead of failing. Compaction behavior (handoff quality, post-compaction recovery via the surviving kernel + conversation log) has been validated in the ablations backing feat: terse persona prompt, bash+ipython default tools, actionable compaction handoff #142; at-threshold rollouts solved at 0.75-0.88 with the improved handoff prompt.
• max_depth=1 by default: one level of await rlm('sub-task') is available out of the box, matching how the harness is deployed in practice.

Notes

• Empty/0 env values for RLM_SUMMARIZE_AT_TOKENS now mean "disabled" (previously "" meant unset->off; unset now means the 256k default).
• With max_depth>=1 the engine owns a SessionTreeSupervisor; embedders using the sync close() inside an event loop must switch to await engine.aclose() (the idempotency test now pins max_depth=0 for that reason — it tests retries, not recursion).
• Tests: 130 passed; the 6 test_acp.py failures are pre-existing on clean main in this environment.

---

Note

Medium Risk
Behavior changes for all deployments that relied on implicit defaults (compaction off, no sub-agents); long runs may compact unexpectedly and recursive rlm() is enabled unless env overrides are set.

Overview
Default execution policy now turns on context compaction at 256,000 tokens and allows one level of sub-agent recursion (max_depth=1), including matching env defaults for RLM_MAX_DEPTH.

RLM_SUMMARIZE_AT_TOKENS parsing changes semantics: omitting the variable applies the 256k threshold; "" or "0" explicitly disables compaction (replacing the old “unset/empty means off” behavior). _summarize_at_tokens encodes that logic and invalid positives still raise.

Tests assert the new defaults, cover disable-via-empty/zero, switch async engine teardown from close() to await aclose() (needed when recursion/supervisor is on), and pin max_depth=0 on the compaction/idempotency test so it stays focused on retries rather than sub-agents.

^{Reviewed by Cursor Bugbot for commit c37848a. 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 bc59190. Configure here.}