Add one evolving reasoning step. Use for iterative work, revisions, or branches.

Whole chain ready -> think_batch. Enforced multi-pass gate -> think_cycle. Code-audit checklist -> think_logic.

Returns progress, surfaced guidance, and one next action.

Submit an already-complete reasoning chain atomically. For exploration use think or think_cycle.

Rejects thoughts under 50 chars or over 60% similarity; warns when average confidence is below 4.

Verify and close a think/think_batch scope before a complex final answer.

For think_cycle use action=finalize. Rejects path gaps or unresolved blockers and returns the next correction.

Read-only search of an active/explicit scope or saved insights.

Use before repeating work or relying on prior reasoning.

Clear active scope. Irreversible.

TRIGGER:

• IF new_task_unrelated THEN reset

• IF all_paths_failed THEN reset

DO NOT RESET:

• IF user_says "tweak/fix/expand" THEN continue

• IF mid_execution THEN use branchFromThought instead

Run an enforced multi-pass reasoning loop for high-risk or ambiguous work.

Start once, then follow the single returned next step until the gate passes; finish with action=finalize. If start had constraints, include constraintCheck. For ordinary iterative work use think. Do not use think_done for a cycle.

Interop:

• backendMode=auto: mirror to think backend with fallback

• backendMode=think: strict think backend mode (no fallback)

• backendMode=independent: standalone cycle only

Return a read-only code-analysis checklist. It does not inspect code, store progress, or produce findings.

Use only when a model needs a methodology; use think or think_cycle to perform the analysis.