fix: stop repeated identical tool doom loops

[kshitijk4poor]

Summary

• Detect and interrupt repeated identical tool calls to prevent doom loops.
• Add recovery behavior that skips the repeated call and prompts the model to change approach or summarize.
• Add regression coverage for repeated-call handling, CLI continue/stop paths, and per-turn state reset behavior.

Scope

• Includes: run_agent.py, tests/test_run_agent.py
• Excludes: adapter/CLI/browser cleanup work (moved to refactor: split adapter and cli cleanup #939)
• Issue scope tags: Feature: Doom Loop Detection — Pause on Repeated Identical Tool Calls (inspired by Kilocode) #512 (doom-loop handling), Feature: Tool-Call Loop Guard — SHA-256 Pattern Detection for Stuck Agent Loops (inspired by OpenFang) #481 (hash-based repeated-call signature detection)

Approach

• Build a call signature from tool name + normalized argument hash.
• Trigger protection after 3 identical consecutive calls.
• Exempt long-running process polling/wait/log actions.
• On trigger, inject an explicit recovery prompt to avoid repeating the same call.

Related

• Closes Feature: Doom Loop Detection — Pause on Repeated Identical Tool Calls (inspired by Kilocode) #512
• Addresses Feature: Tool-Call Loop Guard — SHA-256 Pattern Detection for Stuck Agent Loops (inspired by OpenFang) #481
• Split cleanup follow-up: refactor: split adapter and cli cleanup #939

Validation

• python -m pytest -o addopts='' tests/test_run_agent.py -q
• python -m pytest -o addopts='' tests/test_run_agent.py -k "doom_loop or resets_doom_loop_state_each_turn" -q

[teknium1]

I think that things like read() look like identical tool calls and aren't and there's too many edge cases of when it may need to repeat the same tool over and over to really implement this. I'm going to close for now. Feel free to reopen if a new formulation of this needs reconsidering