[ffrujeri]

Stacked on #5465 (simple_agent continuation), which is on #5350, the top of the BaseEnvironmentServerWithTools stack (#5346 → #5347 → #5348 → #5349 → #5350). Review those first; this PR's diff against #5465 is environment_servers/bird_interact only.

What changed

BIRD-Interact runs in one Environment Server for each mode, with no Resources Server. This is approach A of the BIRD-Interact RFC; #5202 is the prototype of approach B, which splits each episode between an Environment Server and a Resources Server.

environment_servers/bird_interact subclasses BaseEnvironmentServerWithTools (#5347) directly, with no user-simulation base class. Its run_episode starts the database sandbox, serves the a-Interact tools, seeds the agent session (agent_session.py), runs the turns and grades them. The server owns the database sandbox, the two-stage user simulator, the grader, the phase and BIRD-COIN rules and the turns.

	#5202 (approach B)	This PR (approach A)
Servers in an episode	Environment Server + Resources Server	Environment Server
c-Interact ask and submit	Calls to /c_interact/ routes	In-process
a-Interact tools	Resources Server routes, MCP auto-exposure	A TypedToolset of the episode, over direct HTTP, MCP or both
Agent's credential	The Environment Server's Resources Server cookie, which reaches every route	A signed episode token, for the tool routes of one episode
Who knows the mode	Both servers	The Environment Server
Several workers	Resources Server session routing	Routing on the episode token (#5350)

• Agent session. Seeded with the episode's grants; above one turn it must resume across turns, so each turn sends only the user's new message. A turn over turn_timeout_seconds fails the episode for a retry, the optional interaction_timeout_seconds stops the turns and still grades, and the agent is closed before grading. A failure keeps the completed turns, the tool calls and the agent's evidence.
• Sandbox. Each episode starts its sandbox and registers its stop, so it stops on every path, including shutdown (feat(environment): finish running episodes' cleanups when a BaseEnvironmentServerWithTools shuts down #5348). A command that fails for infrastructure reasons runs again on a fresh sandbox; then c-Interact fails the episode for a retry and a-Interact masks the sample (bird_interact:sandbox_failed), as does a simulator failure in an a-Interact ask (bird_interact:user_simulator_failed).
• Tools. Over direct_http the first request declares the nine tools as strict function tools; over mcp alone it declares none, because Codex refuses a request that declares tools.
• Data. The example datasets are declared on the agent, one taskset for each mode, and validated by the server's task_data.py, the Harbor layout. grading.py, simulator.py, task_data.py and the example data come from feat(environment): add a user-simulation base class and BIRD-Interact on it #5202 unchanged apart from module paths and wording. data/example_rollouts.jsonl is regenerated from the smoke run below.

Tracking: GYMINT-646 (Linear).

Validation

Command	Result
pytest environment_servers/bird_interact/tests/test_app.py	48 passed: the 37 episode tests from before, unchanged, plus the agent session's (each turn on one session through the capture route, with its activation header and cookie; the tool grant ends with the episode, then 410; an agent that cannot resume fails the seed and is still closed; a turn timeout fails for a retry; the interaction deadline stops and grades; a failed turn, verification or close keeps the turns and evidence; token capture moves the turns to the training path; max_turns must be positive). The episode tests run with a scripted agent and user model and real database sandboxes on the host (local provider)
ng_test +entrypoint=environment_servers/bird_interact +should_validate_data=true	48 passed in the server's own venv (mcp 2.3; the core venv has mcp 1.29, and the MCP test runs under both)
gym dataset collate "+config_paths=[environment_servers/bird_interact/configs/bird_interact.yaml]" --mode example_validation --output-dir <tmp>	Passed; example_metrics.json already up to date
pytest -n 12 tests/unit_tests at this PR's head	15,081 passed, 26 failed. The same 26 fail on main at dda609f86 on the same machine (OpenSandbox compose, sandbox transfer and stop retry, Harbor validation, tasks and run summary, benchmark listing)
pre-commit run --files on every file both stacks change	Passed

Rollout evidence

These runs are from before the rebase onto main at dda609f86 and the move of the turns into this server, which keep the behavior (the 37 episode tests pass unchanged); they have not been rerun since. Docker database sandboxes, num_workers: 4 on each BIRD Environment Server, the five example tasks, two repeats. Policy azure/openai/eccn-gpt-6.1-sol; simulator openai/openai/eccn-gpt-6-luna.

Run	Rollouts	Failed	Masked	Mean reward	#5202's run
c-Interact, simple_agent	10	0	0	0.71	0.40
a-Interact, simple_agent over direct HTTP	10	0	0	0.82	0.82
a-Interact, Codex over MCP (tool_transports: [mcp])	10	0	0	0.82	0.80 (3 tasks)

What I checked in the rollouts:

• No tool call returned an error, and no tool output other than ask contained a task's hidden knowledge entry.
• The one failed c-Interact episode (bookshop_2) is the known simulator issue the RFC records: the simulator calls a fraction a percentage, so the agent's ×100 answer fails.
• Codex's own observations show only BIRD's tools; it never used its shell.
• No BIRD sandbox container was left after the runs.
• The servers log no per-worker access lines, so the runs show four workers and no routing errors, not which worker served each call; feat(environment): route BaseEnvironmentServerWithTools tool calls to the worker that runs the episode #5350's e2e test covers that.

Compatibility

New server; existing servers do not change. #5202 stays open as the approach B prototype until these stacks are reviewed.

Not yet covered: gym env test's data and catalog checks still run only for Resources Servers (GYMINT-643), PostgreSQL (Lite and Full), and rows with custom test_cases.

🤖 Generated with Claude Code

[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.

[ffrujeri]

/ok to test cef7fc5

[ffrujeri]

/ok to test 1ba7db8

[ffrujeri]

/ok to test 8ae6ea8

[ffrujeri]

/ok to test 0bb3a5e

[ffrujeri]

/ok to test b14c4bc

[ffrujeri]

/ok to test 81e2d19