[giuliolovisotto]

What does this PR do?

Kilo emits compaction events, but its captured helper calls lack the assistant-message correlation needed to resolve those events. Reuse Gym's existing OpenCode assistant-message header plugin in Kilo's task-environment hooks, declare the captured header, and scope compaction observations to the configured model server. Use matching source-path and destination-filename constants for bundled plugins and installers in both OpenCode and Kilo. Kilo explicitly imports the shared OpenCode correlation hook; existing filenames and runtime behavior are preserved. The collector can then retain the helper and exclude it from policy turns using the existing join.

Validation

The constants cleanup passed 481 OpenCode/Kilo tests (12 skipped), all-file pre-commit, and fresh compaction runs on both pinned runtimes. Each run retained 16 model calls and 15 policy turns, reused the summary, and retained helper usage.

Earlier validation:

python -m pytest -q scripts/harness_conformance/tests tests/unit_tests/harness_capabilities responses_api_agents/kilocode_agent/tests --disable-warnings: 658 passed, 9 skipped on the assembled stack. Plugin-staging tests exercise the reused module. Fresh sandbox tests: 60 passed; fresh assembled-stack shared/Kilo/collector run: 1,194 passed, 9 skipped, 73 deselected. The fresh Kilo compaction run still passes with 16 calls and 15 turns.

Rollouts

With Kilo 7.4.15 and the final shared scenario, python scripts/run_harness_conformance.py --harness kilocode --scenario context_compaction --gate-tier P1 --output results/kilocode-compaction passes: 16 captured calls, 15 policy turns, one compaction event; the summary is reused and helper usage retained.

The eight pre-existing scenarios pass 7/8. model_error still raises an execution error instead of returning a failed response. The same failure and failed checks reproduce on the shared-support branch before any Kilo changes; it is not introduced by this PR.

Compatibility and benchmark impact

The added header carries correlation metadata. Kilo's execution and compaction policy are unchanged. Production artifacts gain resolvable helper references; no benchmark configuration changes. The shared module is existing Apache-2.0 Gym code.

The stack starts with #5122; #5439 registers the scenario only after the shared support and harness fixes.

All-file uvx pre-commit run --all-files passed; all commits have DCO sign-off. Full repository CI has not been validated. Smoke runs use real pinned harness runtimes and a controlled local model provider; they do not qualify remote sandbox infrastructure or model quality.

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). Fresh Kilo compaction passes all five behavioral checks: 16 calls, 15 turns, one event and retained 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.

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/41802880941