fix(agents): anchor GAN loop to research goal

[FluffyAIcode]

Summary

• make the first substantive REPL input an immutable research goal
• add /continue for a feedback-driven next iteration and /new <goal> as the only topic reset
• pass complete previous Generator and Critic outputs into the next Generator turn
• reject generator>, critic>, prompt>, metrics, heartbeat, error, and traceback text as user research inputs
• require every Critic response to declare Goal Alignment: ALIGNED or DRIFTED
• instruct drifted Critic turns to discard the off-topic branch and restore the anchored proof frontier
• record a privacy-safe goal hash and whether Critic feedback was applied
• update AutoResearch acceptance to goal_anchored_recursive_gan_v3

Root cause

The previous REPL was one-shot and stateless. Generator and Critic shared only the current prompt, Critic feedback was not fed into the next Generator, and runtime output could be copied or buffered as a new prompt. Both agents therefore agreed on contaminated KV metrics instead of recognizing drift from the Riemann goal.

Tests

• immutable goal prompt construction, complete feedback injection, alignment protocol, and runtime-artifact rejection
• targeted GAN/AutoResearch suite — 20 passed
• full platform-neutral suite — 875 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.