chore: drop rlm_ prefix from rlm harness metrics

[mikasenghaas]

Summary

Surface rlm's meta.json metrics under their own names instead of prefixing them with rlm_ (verifiers/v1/harnesses/rlm/harness.py). The metric keys are already rlm-specific (num_compactions, ipython_input_*_mean, num_ptc_calls, ...), so the prefix was redundant.

Pairs with PrimeIntellect-ai/rlm-harness#110, which prunes the rlm metric set down to only what verifiers v1 can't derive natively (compaction volume, ipython input size, programmatic tool calls) and renames the survivors.

Breaking

• rlm harness metrics now appear on the trace as e.g. num_compactions instead of rlm_num_compactions. Any dashboard/query keyed on rlm_* for rlm runs must drop the prefix.

Verification

Ran terminal-bench-2-v1 fix-git on the prime runtime with rlm installed from the rlm-harness#110 branch (--harness.version chore/prune-redundant-metrics) and this prefix drop active.

• Task solved (solved: 1.0, agent_completed, 15 turns).
• Metrics logged on the trace (bare names — prefix drop confirmed):
  num_compactions, turns_since_last_compaction, turns_between_compactions_mean, compaction_chars_dropped_mean, compaction_summary_chars_mean, ipython_input_chars_mean (81.5), ipython_input_loc_mean (2.2), num_ptc_calls, num_ptc_calls_bash.
• The removed rlm metrics are covered natively by the v1 trace on the same rollout: num_branches=1, num_turns=15, completion_tokens=2248 (sum of node usage).

Note

Drop rlm_ prefix from RLM harness metric keys

In harness.py, the RLMHarness.rlm coroutine now returns metrics using the raw keys from meta['metrics'] instead of prefixing each key with rlm_. Behavioral Change: any consumer reading prefixed keys like rlm_accuracy must update to use the unprefixed form (e.g. accuracy).

^{Macroscope summarized c396dd6.}

---

Note

Medium Risk
Renaming trace metric keys is a breaking change for any consumer keyed on rlm_*, though runtime behavior is unchanged.

Overview
RLM harness metrics now use the same keys as meta.json (num_compactions, ipython_input_chars_mean, etc.) instead of prefixing every value with rlm_.

The RLMHarness.rlm metric coroutine only changes how keys are named when numeric fields from sessions/*/meta.json are merged onto the trace; filtering of non-numeric metrics is unchanged. Comments were updated to match the slimmer rlm metric set (compactions, ipython input size, programmatic tool calls).

Breaking: dashboards or queries that expect rlm_* keys must switch to the unprefixed names. This pairs with rlm-harness pruning redundant metrics that v1 already derives from the trace.

^{Reviewed by Cursor Bugbot for commit c396dd6. Bugbot is set up for automated code reviews on this repo. Configure here.}

[macroscopeapp]

Approvability

Verdict: Approved

This is a mechanical rename that drops the 'rlm_' prefix from metric keys. The underlying metric collection logic is unchanged - only the key naming format is simplified.

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