fix(compaction): forward tools= with tool_choice="none" to keep exten…

[samsja]

Summary

The compaction summary call drops tools= from the request entirely so the model can only respond with text. Side effect we missed: vLLM's
chat-completions layer injects a # Tools\n<tools>{...}</tools> block into the rendered system message only when tools= is in the request.
Dropping tools= therefore makes the compaction call's system prompt diverge from every regular turn.

Downstream, prime-rl's RL trajectory walker (interleave_rollout in src/prime_rl/orchestrator/trajectories.py) opens a new training sample
whenever a step's prompt_ids doesn't prefix-match any active sample. One compaction event creates two prefix-incompatible boundaries:

1. Compaction summary turn (no tools= in request → no # Tools block in rendered system) does not extend the prior pre-compaction sample.
2. Post-compaction continuation turn (tools back in request → # Tools block back) does not extend the summary turn.

So one compaction event splits the rollout into three samples where structurally two would be enough.

Symptoms in production

RLM-GLM5.1 run (TITO, use_token_client=true), all four envs:

env	samples/rollout	compactions/rollout	breaks (s−1)	breaks/compaction
rlm-swe	3.16	1.07	2.16	2.01
rlm-deepdive	1.99	0.49	0.99	2.02
rlm-swerebench-v2	2.62	0.81	1.62	2.00
general-agent-rlm	2.23	0.61	1.23	2.02

Ratio is 2.0 across all envs — too clean to be tokenization noise; structural.

Fix

Forward the active tool list to the compaction call as tools=active_tools with tool_choice="none". The system prompt now renders identically to
regular turns (extension property holds at the summary turn → 1 break per compaction instead of 2), while tool_choice="none" preserves the
original "text-only summary" behaviour by forbidding tool calls.

request_kwargs = {"model": self.model, "messages": messages}                                                                                       
if active_tools:                                                    
    request_kwargs["tools"] = active_tools
    request_kwargs["tool_choice"] = "none"
response = await call_with_retries(self.client.chat.completions.create, **request_kwargs)                                                          
                                            
Why tool_choice="none" over dropping tools=                                                                                                        
                                                                                                                                                   
tool_choice="none" is the OpenAI-spec way to say "you have tools, I just don't want you to call them this turn." It keeps the request              
schema-consistent with regular turns (same system-prompt rendering on the server), so prime-rl's interleave_rollout keeps merging steps into one   
sample across the compaction boundary instead of opening a third one.                                                                              
                                                                                                                                                   
When this regressed
                                                                                                                                                   
It hasn't — _compact_branch has called chat.completions.create without tools= since compaction was first introduced in #54 (2026-04-23, commit
f7cda58). The bug has been silent the whole time. It only surfaced now because we noticed samples_per_rollout / rlm_compactions_count == 2.0 and
traced the prefix break from the prime-rl orchestrator side.
                                                                                                                                                   
Scope
                                                                                                                                                   
- One file changed (src/rlm/engine.py), 23 insertions / 6 deletions.
- _compact_branch signature gains an active_tools argument; the only caller is updated.
- No test changes — tests/test_metrics.py exercises metrics shape, not the compaction request.