[atharvsatpute]

Summary

Adds a Fraud Triage environment: a sequential transaction risk-decisioning
RL task where an agent must approve/flag/escalate/block synthetic
transactions. Fills a gap in the existing finance environments (finrl_env
covers stock trading, finqa_env covers document QA) — neither covers
transaction-level risk decisioning.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• New environment
• Refactoring

Alignment Checklist

Before submitting, verify:

• I have read .claude/docs/PRINCIPLES.md and this PR aligns with our principles
• I have checked .claude/docs/INVARIANTS.md and no invariants are violated
• I have run /pre-submit-pr (or bash .claude/hooks/lint.sh and tests) and addressed all issues

RFC Status

• Not required (bug fix, docs, minor refactoring)
• RFC exists: #___
• RFC needed (will create before merge)

Test Plan

Ran the existing test suite locally:

All 4 tests pass:

• test_reset_returns_valid_observation
• test_approve_and_block_actions_return_reward
• test_episode_terminates_after_episode_length_steps
• test_state_tracks_confusion_matrix_counts

Also manually verified end-to-end with a heuristic policy (amount z-score

• new-device/new-merchant thresholds) achieving 100% fraud recall over a
  200-transaction episode, confirming the reward signal and episode
  termination behave as intended.

Environment follows existing Gym-style conventions (see connect4_env) —
Pydantic wire types, WebSocket-based client, no MCP tool exposure, rewards
computed entirely server-side inside step().

Claude Code Review

N/A

---

Note

Low Risk
Self-contained new environment and documentation with synthetic data only; no changes to core OpenEnv runtime, auth, or existing environments.

Overview
Adds a new fraud_triage_env OpenEnv package for sequential payment fraud triage: agents see one synthetic transaction at a time (risk features only; labels hidden) and choose APPROVE / FLAG / ESCALATE / BLOCK.

Server logic streams transactions from a lightweight TransactionGenerator, scores decisions with asymmetric rewards (heavy penalty for missed fraud, smaller cost for false positives, extra cost for escalate), and tracks episode confusion-matrix stats in state. Wiring follows the usual pattern: Pydantic models, FraudTriageEnv WebSocket client, FastAPI app via create_app, openenv.yaml, Docker image, and pyproject.toml.

Docs add fraud_triage to the environments index and toctree, plus a dedicated guide and env README. Tests cover seeded replay, reward/counter behavior, config validation, episode termination, and WebSocket reset/step/state parity with the client parsers.

^{Reviewed by Cursor Bugbot for commit d9e54c5. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 2ccb2a3. Configure here.}

[atharvsatpute]

Hi everyone, I've addressed the main items and the PR is ready. Could a maintainer please trigger the automated tests? (Note: I already checked the Bugbot output and everything looks good on that end). Thanks for your time!

[burtenshaw]

Thanks! Could you also deploy this space on hugging face for review of the runtime?

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[burtenshaw]

Runtime review on 0062719: the environment's Dockerfile built successfully, and the packaged server passed health, typed-client reset, a 200-step episode, exact termination, state-counter checks, post-terminal rejection without mutation, and deterministic seeded replay. The temporary container was removed and its absence verified. Image ID: sha256:468241ba1670e4efdeb9e7d682768ac4a6f43256781dcbc86175bed98e640873. This is local container evidence; no Hugging Face Space was published.

The 20 focused tests also pass. The docs now state the simplified reward semantics explicitly: FLAG and BLOCK have equal payoff, while ESCALATE adds a cost and provides no extra evidence.

[burtenshaw]

Reviewed 0062719. The post-terminal reward bug and current reset/step API incompatibilities are fixed, action values are validated, and seeded resets replay the full observation stream. All 20 focused tests and the packaged Docker server's 200-step typed-client smoke pass, including terminal rejection, state immutability, replay, and container removal. Required dependency lock and docs/catalog entries are present. The simplified FLAG/BLOCK/ESCALATE economics are now documented accurately. No remaining blocker found; no public Hugging Face deployment was performed.

[burtenshaw]

Refreshed onto main 07f26b0 after #1198 and wired FraudTriageState into the app in both import layouts. All 50 focused environment/State-wiring tests and touched-file lint pass at d9e54c5.

Rebuilt the actual Dockerfile from this head and verified the packaged app's source hash. The live local container passed health, /schema and /state with all eight subclass fields, typed-client reset, all 200 steps, exact termination, rejection of an extra step without state mutation, and deterministic replay. Container removal was verified. Image: sha256:4d90eeb8d103f061996af3fb90166c4a0300260f9e7c6cfdda6d5de5655878c4. Fresh CI is running.

[burtenshaw]

Reviewed the #1198 integration: FraudTriageState is imported in both package and standalone paths and passed to create_app. Independent validation passed 50 tests; HTTP /schema and /state returned 200 and retained all eight State fields. Touched-file lint passed.

[burtenshaw]

Reviewed and validated current head 0c2e23bf68844f3c48ca601f4baf69cb3b029296 against main 9b2c0bea: 20 scoped tests pass. Tests cover seeded replay, all action/reward cases, terminal-step rejection, and WebSocket serialization/client parsing. Touched lint passes. No Docker build was run locally. Refreshed CI and the final merge decision remain with the coordinating maintainer.