Interleaved sequential thinking tool that interleaves reasoning with tool execution to solve multi-step problems. When to use: plan, verify, hypothesize, branch, revise, integrate-tools, reflect, reason-about-evidence, decide, course-correct. Key features: - Revision support: revise any earlier step via isRevision + revisesStep - Branching: explore alternative hypotheses in parallel via branchFromStep + branchId - Automatic phase inference: thinking / tool_call / analysis cycle is auto-detected from input shape - Host-delegated execution: this server only registers tool calls; the MCP host actually invokes them and feeds results back via previousToolResult - Quality signals: each response carries convergence / evidence-coverage / hypothesis-coherence scores so the model can self-monitor - Anti-idle-loop guidance: nextHint flags dead-end patterns (2+ thinking steps without tool call, 3+ revisions of the same step, missing analysis after tool call, budget exhaustion) Workflow: thinking → tool_call → analysis → repeat until nextStepNeeded=false. Each call appends to an in-memory history; pass the same stepNumber sequence to keep state coherent. You should: 1. Start each step by stating a falsifiable hypothesis and how you will verify it before jumping to tools. 2. Use phase=thinking to reason, phase=tool_call to register a tool, phase=analysis to interpret a previousToolResult. 3. Pass the host's actual tool result back in previousToolResult on the next analysis call — do not synthesize results yourself. 4. Cite concrete values from toolResult (e.g., "result.count = 42") instead of generic "based on the result". 5. Use isRevision + revisesStep to correct earlier reasoning rather than silently ignoring it; after 3+ revisions of the same step, branch into an alternative hypothesis instead. 6. Use branchFromStep + branchId to explore parallel hypotheses; branches live alongside the main flow and can be merged in later analysis. 7. Keep toolCall.parameters specific and concrete — empty or vague queries waste the host's tool budget. 8. Treat nextHint as a system-level nudge; convergence / evidenceCoverage / hypothesisCoherence < 0.3 means you have not yet integrated enough evidence. 9. If a tool call fails, retry with adjusted parameters or switch tools — do not give up on nextStepNeeded after one failure. 10. Only set nextStepNeeded=false when truly satisfied with the integrated answer; do not stop after a single tool failure.