sequential-thinking: readOnlyHint and idempotentHint annotations are inaccurate (server is stateful, non-idempotent)

Summary

The sequentialthinking tool annotations readOnlyHint: true and idempotentHint: true are inaccurate. The server is stateful and non-idempotent — it maintains per-session thoughtHistory and branches instance state that persists across calls and changes on every invocation.

These annotations were added per #3403's suggestion, but the suggested values were not verified against the server's actual stateful behavior.

Evidence (BETS cycle, 2026-08-30)

Ran a governed Benchmarking/Evaluation/Testing/Standards (BETS) cycle against @modelcontextprotocol/server-sequential-thinking (published npm version, via npx -y). 11 test cases across 4 lanes.

readOnlyHint: true is misleading

The SequentialThinkingServer class (confirmed in src/sequentialthinking/lib.ts on main) holds instance state:

• private thoughtHistory: ThoughtData[] = [] (line 23-24)
• private branches: Record<string, ThoughtData[]> = {} (line 25-26)

Each processThought call mutates this state:

• this.thoughtHistory.push(input) (line 101) — appends to history on every call
• this.branches[input.branchId].push(input) (line 111-112) — appends to branch when branchFromThought + branchId provided

Observed across 11 calls in one session: thoughtHistoryLength incremented linearly (1, 2, 3, ... 11). The server is not read-only — it accumulates session state with each invocation.

idempotentHint: true is false

A tool is idempotent if repeated calls with the same arguments produce the same result. Calling sequentialthinking twice with identical arguments produces different thoughtHistoryLength values (N, then N+1) and potentially different branches contents. The tool is non-idempotent by definition.

Suggested fix

annotations: {
  readOnlyHint: false,    // server maintains thoughtHistory and branches state
  destructiveHint: false, // accurate — no external side effects
  idempotentHint: false,  // repeated calls produce different history lengths
  openWorldHint: false,   // accurate — operates on internal reasoning state
},

Impact

Clients relying on readOnlyHint: true for safety assumptions (e.g., "safe to call in parallel", "no state side effects to roll back") will miss the session-state accumulation. Clients relying on idempotentHint: true for caching/retry (e.g., "safe to retry without checking result") will misbehave — a retried call produces a different history length and may create a duplicate branch entry.

Source confirmed on main

src/sequentialthinking/index.ts lines 201-209 (current main as of 2026-08-30):

annotations: {
  readOnlyHint: true,     // ← should be false
  destructiveHint: false,
  idempotentHint: true,   // ← should be false
  openWorldHint: false,
},

src/sequentialthinking/lib.ts lines 23-26, 101, 111-112 confirm the stateful behavior.

Context

• BETS receipt: agents/_state/bets/20260830T092705Z_sequential-thinking.json (HUMMBL internal, available on request)
• Related: #3403 (added these annotations without verification — the suggestion was accepted as-is)
• Discovered during a governed BETS cycle on the MCP fleet, not via automated scanner

Happy to open a PR if that's preferred.

Start in src/sequentialthinking/index.ts at the tool annotations, then read the state fields and mutations cited in src/sequentialthinking/lib.ts. Confirm the metadata matches the server's stateful behavior and that the read-only and idempotency claims are accurate when the change is complete.