[lafrancef]

Add multi-turn OpenCode continuation for SWE-Together's interactive coding episodes. Seed with continuation: {mode: resume_harness_session}, then append user input through ordinary /v1/responses requests using the session cookie and existing 1-based x-ng-agent-activation header. Use metadata.agent_timeout_sec for each activation's budget. The adapter retains its saved session and HOME between calls.

Each activation has its own invocation ID and status, with harness_session_id recorded as context. Trajectory turns carry the activation ID, and the plugin forwards activation-scoped X-Session-Id values for root and child invocations. Responses expose new output, usage and observations, with harness_session_id also in metadata; close returns the accumulated evidence. Requested in-harness tool observations use harness_tool_calls.

The shared ledger joins identical retries, rejects conflicting or out-of-order inputs, and fences execution during close. A final negative termination receipt fails the activation and skips verification; close releases the borrowed connection and returns cleanup_confirmed: false. Connection release failures remain retryable. Validation runs before consuming an index. Deadlines cover preparation, queueing and staging; expiry before the first launch allows a later activation to start. Retry guarantees remain process-local.

Continuation timeouts fail on explicit OpenCode error events, including after partial output. When close recovers delayed artifacts, it replaces that activation's fallback evidence while preserving its original outcome and earlier activations.

Existing single-turn response metadata, model headers and close behavior are preserved. Unknown cleanup confirmation is omitted. SandboxSession keeps finalization private and supports explicit rearming after confirmed cleanup with fresh launch files. The adapter README and integration guides document the protocol and directory ownership.

Live validation used OpenCode 1.17.11 and gpt-4.1-mini on borrowed OpenSandbox CPU sandboxes. Single-turn and two-activation file-repair smokes passed their external verifiers, identical replay checks, existing observability conformance checks and rollout health checks. The second activation resumed the same OpenCode session and recalled a label from the first. Agent cleanup disconnected the borrowed connection, leaving task files available for verification; the sandbox owner then destroyed the allocations.

No separate tracking issue is needed: this focused change supplies multi-turn continuation for the SWE-Together harness integration.

The terminology-only follow-up renames the protocol, OpenCode helpers, docs and tests consistently. Its focused checks passed (155 tests):

python -m pytest -q -o addopts= \
  responses_api_agents/opencode_agent/tests/test_continuation.py \
  responses_api_agents/opencode_agent/tests/test_opencode_headers.py \
  responses_api_agents/opencode_agent/tests/test_trajectory.py \
  tests/unit_tests/test_agent_activations.py \
  tests/unit_tests/test_rollout_observability.py
pre-commit run --all-files

Whole-tree pre-commit passed; its guardwords check skipped because the private catalog was unavailable. Full local CI and live smokes were not repeated for this renaming; GitHub CI is requested for the final commit.

Validation before the terminology-only follow-up (02aeba2b):

python -m pytest -q -o addopts= --import-mode=importlib \
  responses_api_agents/opencode_agent/tests \
  tests/unit_tests/test_agent_activations.py \
  tests/unit_tests/test_sandbox_session.py \
  tests/unit_tests/test_agent_sessions.py \
  tests/unit_tests/test_checkpoint_agent.py \
  tests/unit_tests/test_agent_session_retries.py \
  tests/unit_tests/test_base_responses_api_agent.py \
  tests/unit_tests/harness_capabilities \
  tests/unit_tests/test_rollout_observability.py \
  responses_api_agents/simple_agent/tests/test_app.py \
  scripts/harness_conformance/tests
bash scripts/ci/core_unit_tests.sh
source scripts/ci/setup_dev.sh
uv sync --extra dev --extra ray --extra sandbox --extra openshell --extra mlflow --extra wandb
pytest --cov=. --cov-append --cov-report=term-missing \
  --cov-fail-under="$(python3 scripts/ci/cov_fail_under.py)" --durations=10 -m sandbox
for shard in 0 1 2 3; do bash scripts/ci/server_tests.sh "$shard" 4; done
pre-commit run --all-files

The focused regression/checker selection passed: 993 tests, 42 platform skips. Whole-tree pre-commit passed. Fresh single-turn and two-activation live smokes passed verification, replay, observability conformance, rollout health and cleanup.

Linux core passed 8877 tests (28 skips), followed by 921 sandbox tests (5 skips). Combined coverage passed the configured 95% gate at 95.09%.

The full server run selected all 233 servers: 210 passed; 23 were blocked by GitHub HTTP 503 responses or PyTorch download connection resets. Both changed servers passed their full isolated Linux suites: OpenCode 472 passed/6 skips; NOOA 226 passed. There were no data-validation failures. Scripted OpenCode conformance statuses match the previous revision, including its existing sandbox.identity and model_error failures.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[yaoyu-33]

Thanks for this, @lafrancef. The ledger and per-activation rearm design is clear, and the README and fern updates make the protocol easy to follow. We ran the PR's tests locally (test_agent_activations: 12 passed; opencode_agent tests: 405 passed, 42 skipped on macOS) and reproduced the inline items with scratch probes against both base and head.

The main item is in opencode_agent/sandbox.py: the single-turn path's failure-time evidence changed, which the description says is preserved. The rest are smaller edges in the new continuation path, test coverage for the shared SandboxSession changes, and docs and contract clarity.

PR hygiene:

• Could you link the tracking issue, or say why none is needed (AGENTS.md "Pull Requests")? Naming the intended consumer of the continuation fields would also help, since nothing in the repo sends continuation yet.
• Could you list the exact local checks you ran (pytest targets, pre-commit run --all-files) next to the live-smoke description?
• CI still needs a vetter's /ok to test.

[yaoyu-33]

Not blocking for this PR, but a suggestion to make the next harness cheaper to add (codex, hermes, openclaw and pi all use SandboxSession too).

Only three things look OpenCode-specific: the resume flag, where native state lives, and how to find new messages in the export. The rest is generic but lives in OpenCode behind ~20 is_continuation branches:

• per-activation reset;
• evidence accumulation;
• activation-scoped IDs;
• deadline handling;
• the close bundle.

Could we split session state from per-activation state and move the generic part into nemo_gym/agent_utils?

@dataclass
class Activation:          # replaced each activation, no field-by-field reset
    id: int
    log: str = ""
    observations: AgentObservationBundle | None = None
    trajectory: TrajectoryRecord | None = None
    runtime_info: HarnessProcessInfo | None = None

@dataclass
class ContinuationState:   # retained across activations
    home_dir: str
    native_id: str | None = None
    seen_message_ids: set[str] = field(default_factory=set)
    evidence: EvidenceAccumulator = field(default_factory=EvidenceAccumulator)  # merge + ID scoping

Two related ideas:

• One path for single-turn and continuation. Treat single-turn as a single activation on the same path, so most is_continuation branches go away. The single-turn evidence-ordering change came from one of them.
• Fewer adapter hooks. Let the base route opted-in sessions to activate_agent_session(), and give _validate_agent_activation / _continuation_close_timeout defaults. Then an adapter mostly implements one method: run one activation with resume.

A follow-up PR is fine too.

[lafrancef]

Ack, agreed splitting off the generic part makes sense, will do once we add multiple-activation support to a second harness

[lafrancef]

/ok to test 9d4caf5

[github-actions]

🌿 Preview your docs: https://nvidia-preview-pr-4747-b3cc21c46cd4.docs.buildwithfern.com/nemo/gym

[lafrancef]

/ok to test 2d711fd

[lafrancef]

/ok to test 221c064

[lafrancef]

/ok to test ea17c8b

[lafrancef]

/ok to test b3cc21c

[github-actions]

🔄 Merge queue validation started! You can track the progress here: https://github.com/NVIDIA-NeMo/Gym/actions/runs/39649856807