Skip to content

ADR-0136: Synthesize before the identical-tool-call limit

  • Status: Accepted as an engineering decision within the owner's runtime repair request of 2026-09-27
  • Date: 2026-09-27
  • Related: ADR-0023, ADR-0115, ADR-0130; engineering plan Section 12.5
  • Detailed design: docs/plan/tool-system.md, docs/plan/model-gateway.md

Context

A repository research conversation ended with ToolLoopDetected after useful research. The model had fetched the same document repeatedly among other reads. Large results make the context builder replace older tool results with pointers; refetching then consumes the same run-wide identical-call counter. The breaker was enforcing its specified limit, but the run had no final-answer opportunity before another tool request failed it. Raising the limit would spend more on repetition without addressing that failure mode.

Decision

  1. After a completed batch leaves any identical-call counter at one below the configured threshold, the next model turn is synthesis-only. With the default threshold of five, four completed identical requests trigger this control. ADR-0130's evidence resets happen first: progressing browser observations do not trigger it. The checkpointed counter also governs resumed runs.
  2. The runtime adds a volatile platform instruction to give the best-supported answer, identify remaining gaps, and explain that research stopped because it was repeating. It must not claim unfinished work is complete. The control contains no tool arguments or retrieved text and does not enter the checkpoint or frozen prefix.
  3. ModelRequest.tool_choice is None by default, preserving provider defaults, or "none" for synthesis. OpenAI Responses and native Chat Completions send tool_choice: "none"; Anthropic sends tool_choice: {"type": "none"}. Definitions and prior call/result pairs remain intact for history replay. The XML fallback omits its tool-use instruction and schemas. The existing budget-synthesis control also uses this field.
  4. A tool request returned in repeated-tool synthesis fails with ToolLoopDetected before any dispatch, even if its arguments differ. A batch that reaches the hard threshold before it runs still fails under the existing breaker. There is no extra execution allowance and no automatic restart.
  5. All budget, deadline, cancellation, model-error and empty-turn rules continue to apply. Synthesis consumes a normal model call and its usage. Exhausting a hard budget can still prevent an answer; overlapping budget synthesis retains its existing failure classification. The denial and uncertain-effect rules are unchanged.

The provider mappings follow the OpenAI API tool-choice contract and Anthropic tool-choice contract.

Verification and limits

Regression coverage exercises the configured threshold, interleaved large web reads whose results are actually elided, batches, approval resumption, model retries, exhausted model-call budgets, and refusals before dispatch. The shared provider contract checks all three native transports and the XML fallback with both ordinary and synthesis requests. Existing changing-page, evidence-reset cap, denial and loop evaluations remain authoritative.

This repairs the abrupt loss of a concluding response, not the quality of a model's research strategy. The response may be partial. It does not add a tool for rereading an elided event, recover an already failed conversation, alter context budgets, add a gate or advance a milestone. No migration is required.