TaskEnvironment protocol + SQLite database env (#11, #9)

[askmy-stack]

Summary

• Add shared TaskEnvironment protocol (reset / step / validate_success) under agentfailbench/environments/base.py
• Wire the episode runner to type environments as TaskEnvironment; CustomerApiEnv documents protocol compliance
• Implement deterministic in-memory SQLite investigation environment (list_tables, describe_table, execute_sql read-only, submit_answer)
• Extend TaskSpec with optional expected_answer for investigation validators
• Document the interface + env map in docs/benchmark-specification.md and README repo map
• Unit tests cover protocol satisfaction and the normal (non-failure) DB success path

Closes #11
Closes #9

Test plan

• make ci (ruff, mypy, pytest) — 50 passed
• tests/unit/test_task_environment.py covers clean episode + write rejection + reset
• Spot-check README / benchmark-spec render on GitHub

Made with Cursor