TriRoute: Routing Tool-Call Failures Across Resampling, Stopping, and Editing

The structure of a failed tool call can reverse the value of its next recovery action: diffuse malformed outputs favor fresh samples, concentrated syntax-valid faults favor local edits, and low-separation states can make stopping preferable. TriRoute operationalizes this geometry with a frozen six-feature score that selects among Best-of-$5$, Keep, and a radius-three TriEdit search. To distinguish the value of the resulting policy from a stronger repair arm, routed and greedy policies share the generator, verifier, prompts, seeds, five-template edit library, and a hard ceiling of six verifier evaluations. With GPT-4o, TriRoute improves over always invoking that identical edit library by $0.73$–$0.76$ percentage points on APIBench-Tool, BFCL v3, and API-Bank, with paired intervals excluding zero, while using $0.23$–$0.25$ fewer verifier evaluations per recovery episode. Against schema-aware validator-feedback retry, it reaches $81.59$ on BFCL v3 and $75.23$ on API-Bank—gains of $1.05$ and $1.21$ points—and reduces recovery-episode calls from $5.62/5.58$ to $4.91/4.89$. Score strata and failure slices exhibit the predicted shift from resampling to editing; without generator-specific fitting, routed-minus-validator intervals exclude zero in all nine cells spanning GPT-4o, Llama-3.1-70B, and DeepSeek-32B, while same-library point estimates are positive in all nine and six intervals exclude zero. In verifier-rich structured generation, assigning an existing recovery repertoire from the observed failure state raises success while reducing verification demand, without enlarging the repair library.