Skip to content

Evaluate required tool choice and single-call enforcement in Cinference #33

Description

@janitor-manager

Cinference currently rejects tool_choice=required, a named/forced tool, and parallel_tool_calls=false when tools are callable. Prompt instructions alone cannot supply those guarantees. Track these separately from schema-constrained argument generation.

Define and evaluate enforceable required/named/allowed-tool selection and at-most-one-call behavior across Chat Completions, Responses and Messages. Handle text before/after calls, reasoning, truncation, cancellation, zero tools, malformed calls, streaming, replay and DFlash2 verification. Preserve current auto/none behavior and reject unsupported combinations explicitly.

Acceptance: implementation/feasibility decision with protocol conformance tests, failure semantics and performance evidence, or a documented unsupported boundary for dependent clients. Coordinate any shared constrained-decoding design with the strict-schema issue; do not silently rewrite required semantics to auto.

This issue tracks future engine work only. No fork, production change or GPU borrowing is authorized now. Related: #28/#29 and the harness compatibility issue. Source: https://github.com/satellitedown/cinference/blob/b74044fb0a319cd2a737cb7108012e6344b96dac/docs/serving.md

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions