[giuliolovisotto]

Automatic compaction produces a model call for the summary, but the collector currently turns that helper response into an extra policy turn. Exclude uniquely resolved compaction helper references from fallback turn construction while retaining the calls, ownership and token usage. Keep helpers out of turns even when agent ownership claims conflict, preserving those claims for validation. This matches the verifier exclusion rule in #4630. Preserve producer-provided trajectories and the HTTP-error exclusion introduced in #4045.

Stack: this PR → shared support #5307 → Kilo/Codex/Stirrup/Hermes fixes → #5439, which registers the scenario last.

Earlier validation:

• 32 focused collector tests passed, including all reference forms, ownership conflicts in both invocation orders, ambiguous/mismatched references, retained usage and HTTP errors. Six new regression cases fail on the prior implementation and pass with the fix; ownership validation still rejects duplicate claims.
• python -m pytest -q responses_api_agents/terminus_2_sandboxed_agent/tests --disable-warnings: 135 passed, 4 skipped with Harbor 0.22.0. Align the existing saved-projection test's fallback expectation with helper exclusion while preserving its assertions for producer decisions and command-batch counts (original contract). The earlier CI failure expected six rebuilt turns; only three are retained because three referenced compaction helpers are excluded.
• All-file pre-commit passed.
• Real OpenCode 1.17.11 with a controlled local provider: compaction changed from 4 calls / 4 turns to 4 calls / 3 turns; the helper's 511 tokens remain included in the 27,738-token total. The no-compaction control retained 3 calls / 3 turns.

Prepared with Codex. Full repository CI is not yet validated.

Rebase validation — 2026-11-02

Rebased onto main 0c7ac5119c99e0aa0dfeda7be59157c589680ffd, preserving the existing stack order. The assembled original stack passed 1,719 shared conformance, collector, Codex and Kilo tests (16 skipped, 76 configuration-test cases deselected). The separate Terminus saved-projection suite passed all 30 tests. Fresh compaction rollouts retain 16 calls and 15 policy turns on OpenCode, Kilo, Codex, Stirrup and Hermes, including helper usage. All-file pre-commit passed on both stack tips; commits have DCO sign-off and SSH signatures. These are local pinned-runtime runs with a controlled provider, not remote deployment or full repository CI qualification.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[svcnvidia-nemo-ci]

SHIP

LGTM — no concerns found.

Reviewed the collector changes and regression tests, tracing reference resolution, ownership validation, token accounting, and health consumers. The exclusion preserves captured calls and ownership claims without changing producer-provided trajectories or existing HTTP-error handling.

Static review only; tests were not executed.

Note: formal approval is not available for this request, so this review is posted as a comment only.

Review generated by codex. This was a full review. Request a full review with /review reset=true.

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/41797691599