fix(agents): enforce command-driven REPL state

[FluffyAIcode]

Summary

• replace first-line artifact heuristics with explicit WAITING_FOR_GOAL / READY / RUNNING states
• accept only /continue, /steer <text>, /new <goal>, and /quit after goal anchoring
• reject all bare text after the initial goal, including copied model-output fragments
• never read stdin during Generator/Critic inference
• atomically persist complete goal, Generator, Critic, and run ID to a private 0600 checkpoint
• recover an exact completed turn from the timestamped transcript and optionally auto-continue after restart

Truncation diagnosis

The model output was not truncated. Run br_2a37518138bfab7c completed with Generator and Critic EOS. The authoritative log shows a complete accepted input, while Terminal displayed an asynchronously overwritten Operator ... \\mat fragment. This PR removes the implicit-input path that allowed automation to race the prompt boundary.

Recovery validation

Recovered from the real pre-incident run:

• run: br_2a37518138bfab7c
• Generator: 4,223 characters
• Critic: 3,762 characters
• immutable goal: recovered from the anchored transcript

Tests

• command whitelist and phase rejection
• private atomic checkpoint round trip
• timestamped complete-run recovery
• targeted suite — 19 passed
• full platform-neutral suite — 879 passed, 1 skipped
• Python compile and diff checks pass

Made with Cursor

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.