feat: tool truncation + auto-compaction

[mikasenghaas]

Summary

• make context compaction an explicit ExecutionPolicy.compaction option
• use an explicit threshold or discover one from the model card with 16k tokens of headroom
• compact after threshold crossings, context-bound decodes, and recognized provider overflows
• retry checkpoints from the latest usage-verified conversation when the full checkpoint overflows
• resample empty or tool-calling checkpoints and end cleanly when recovery is exhausted
• apply the same policy to root agents and recursive agents while preserving IPython state
• truncate tool results over 20KB before they enter the conversation
• keep compaction prompts, context detection, and truncation helpers outside the main loop

Companion: verifiers #2454.

Breaking

• ExecutionPolicy.compaction now controls proactive and reactive compaction.
• ExecutionPolicy.summarize_at_tokens now defaults to None instead of 256_000.
• Set compaction=true to enable compaction. Leave summarize_at_tokens unset to use model-card discovery.

Verification

• uv run ruff check .
• uv run ruff format --check .
• UV_PROJECT_ENVIRONMENT=/tmp/nano-rlm-merge-tests.xApDnC/.venv uv sync --group dev
• UV_PROJECT_ENVIRONMENT=/tmp/nano-rlm-merge-tests.xApDnC/.venv uv run pytest tests/ — 157 passed

---

Note

Medium Risk
Changes core LLM turn handling and default context behavior (compaction off unless configured), so long runs and training trajectories may diverge from prior releases; mistakes in overflow detection or checkpoint retry could drop history or end runs early.

Overview
Context compaction is opt-in instead of defaulting to a 256k token threshold. Enable with RLM_COMPACTION=1 or by setting RLM_SUMMARIZE_AT_TOKENS; when compaction is on but no explicit threshold is set, the engine discovers the model window via /models and compacts at window minus 16k (small windows keep at least half). Docs drop RLM_MAX_OUTPUT / RLM_MAX_TOOL_OUTPUT_CHARS in favor of a fixed 20KB head/tail truncation on tool results before they hit the chat.

Compaction logic moves into rlm.compaction (overflow heuristics, checkpoint prompts, truncation). The agent loop gains _complete: proactive compaction after large tool turns, reactive compact-and-retry on provider 400/413 context errors, and on finish_reason=length when usage crosses the threshold. Checkpoints retry up to three times with _last_good fallbacks, plain-text-only summaries (no tool calls / no reasoning channel), and CompactionFailed clean stops instead of treating every oversized body as fatal.

ExecutionPolicy.compaction and ACP/runtime snapshots expose the flag; semantic edge tracking can release and reclaim summary requests on resample. Tests cover overflow recovery, disabled compaction, discovered thresholds, and sub-agents.

^{Reviewed by Cursor Bugbot for commit f300615. 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 7f6a408. Configure here.}