Module loop_guard 

Loop detection — catches runaway repeated tool calls. Loop detection for the inference loop.

Modeled after Gemini CLI’s approach: simple consecutive-identical-call detection + feedback injection instead of hard stops. No windowed fingerprinting, no name saturation heuristics, no tool-only suppression.

Design philosophy

Claude Code and Codex have zero loop detection — they trust the model. Gemini CLI has the only thoughtful approach: detect consecutive identical tool calls (same name + args), then inject a “take a step back” feedback message to nudge the model out of the loop. Hard-stop only on the 2nd detection (model ignored the feedback).

What we DON’T do (and why)

• No windowed fingerprint tracking — nobody else does this.
• No tool-name saturation — editing 12 files in a refactoring is normal.
• No tool-only response suppression — efficient models work silently.
• No per-turn tool call cap — frontier models emit 30+ parallel calls.
• No deduplication — if a model emits 66 identical calls, the user should see that and switch models, not have us silently paper over it.

What we DO

1. Consecutive identical calls — same (tool, args) called CONSECUTIVE_REPEAT_THRESHOLD times in a row → inject feedback.
2. Hard iteration cap — absolute ceiling on loop iterations. User can extend interactively.

Structs

LoopDetector

Tracks consecutive identical tool calls.

Enums

LoopAction

What to do when a loop is detected.

LoopContinuation

Options for continuing after hitting the hard cap.

Constants

MAX_ITERATIONS_DEFAULT

Default hard cap for the main inference loop.

MAX_SUB_AGENT_ITERATIONS

Hard cap for sub-agent loops.