feat: 10KB hard cap on tool output entering context

[samsja]

Summary

Hard-caps any single tool result entering the conversation at 10KB, middle-out (keep head+tail), with a warning header naming the original size and line count:

Warning: truncated output (original token count: 12500)
Total output lines: 10001

<first 5KB>
[... 40000 bytes truncated ...]
<last 5KB>

Applies at the engine tool-result boundary — so it covers all tools (bash, edit, and ipython cell output) uniformly.

Scope: context only — tools yes, skills no

• The session log keeps the full output (audit trail unchanged).
• Skill return values inside the kernel stay uncapped: out = await bash(...) holds the complete string for in-cell filtering; only what the cell prints (the tool result) is subject to the cap. Capture-then-filter workflows keep full fidelity — the cap only protects the context window.

Compatibility

The function is byte-compatible with the truncation in verifiers#2454 (bash harness) and the copy bundled inside #147 (auto-compaction) — extracted standalone here so the cap can ship independently; #147 can relocate/dedupe it when it lands.

Motivation from eval traces: unclipped tool results produced single-turn context blowups (observed up to ~690k tokens from one cat on a large file), which no compaction threshold can save you from after the fact.

Tests: truncation unit test added; 139 passed (6 pre-existing test_acp.py env failures, same as clean main); ruff clean.

---

Note

Medium Risk
Changes what every tool returns to the model mid-run; large outputs lose middle content in context (by design), though logs retain full fidelity.

Overview
Adds a 10KB UTF-8 hard cap on each tool result before it is appended to the conversation. Oversized output is head+tail truncated with a warning that includes estimated original token count and line count.

Truncation runs in RLMEngine when building the tool message (truncate_tool_output); session.log_tool_result still records the full string, so audit trails and in-kernel skill return values are unchanged—only what the model sees in context is limited.

A unit test in test_tools.py checks passthrough for small output and warning/size behavior for large output.

^{Reviewed by Cursor Bugbot for commit 32bb729. Bugbot is set up for automated code reviews on this repo. Configure here.}