Prerequisites

• I will write this issue in English
• I have searched existing issues
• I am using the latest version
• I have read the documentation

Bug Description

The agent (Sisyphus-Junior / gpt-5.2-codex) enters an infinite loop while attempting to fix Python type hint errors reported by Pyright/LSP.

Specifically, it struggles with missing type arguments errors and repeatedly cycles through the same failed strategies:

1. Switches from built-in dict/tuple to typing.Dict/typing.Tuple.
2. Checks lsp_diagnostics.
3. Claims the error persists (possibly hallucinating the persistence or misinterpreting the LSP output).
4. Switches back to built-in types or attempts to add suppress comments (# pyright: ignore).
5. Repeats this process dozens of times, consuming massive amounts of context (screenshot shows 140k tokens used).

Steps to Reproduce

1. Assign a task to Sisyphus-Junior to refactor or fix types in a Python file.
2. Trigger a scenario where strict Pyright rules complain about generic types (e.g., dict vs dict[str, Any]).
3. Observe the agent failing to satisfy the LSP and entering a "Code -> LSP -> Code" infinite loop without stopping or asking for user intervention.

Expected Behavior

The agent should detect that it is repeating the same actions (loop detection). After 2-3 failed attempts to fix the same line/error, it should:

1. Stop and report failure.
2. Or attempting a brute-force suppression effectively.
3. NOT continue burning tokens indefinitely.

Actual Behavior

Infinite loop of futile refactoring until manual termination or context exhaustion.

Environment

• Agent: Sisyphus-Junior
• Model: gpt-5.2-codex
• Task: Python Refactoring (Pyright)