Skip to content

V2: Support reasoning with tools across agents and orchestration #472

Description

@paullizer

Summary

Add broad, capability-aware reasoning-with-tools support at the model execution layer used by local agents and chat/workflow orchestration in React V2. This is not a YAMCS-specific change and should not require each action plugin to implement its own model reasoning logic.

This expands the existing Responses-for-thinking-models issue rather than creating a duplicate. The user-directed sequence is: first finish explicit None compatibility on existing tool paths (#1846), then enable deliberate reasoning plus tools across supported agent and orchestration execution paths.

User Value

Operators can use supported reasoning models with YAMCS, SQL/database, document search, MCP, and other authorized actions without being forced to disable reasoning, receiving unexplained provider errors, or relying on which internal helper happened to execute the request.

Existing V2 deployments already execute Sol/Luna/Terra actions successfully. Those successes must remain intact. A successful tool call alone does not establish that positive reasoning was enabled: some current paths correctly supply explicit None, and an agent uses its own configured model/effort rather than necessarily inheriting the composer's selection.

Proposed Behavior

  • Decide reasoning/tool compatibility from the selected provider, model, declared protocol capabilities, and effective request settings at the common execution boundary.
  • Keep existing supported Chat Completions paths working, including explicit None where required and accepted. Do not change working action plugins merely because their model transport changes.
  • Use a supported Responses path when the configured model and requested reasoning-plus-tools combination require it. Azure GPT-5.6's Chat Completions restriction cannot be bypassed by selecting a different dated API version.
  • Make selection and effective model/protocol/effort behavior clear in agent authoring and safe execution diagnostics. Do not silently downgrade an explicitly requested effort to None, silently switch models, or guess compatibility from a friendly display name.
  • Cover agents and the orchestration/workflow execution layer, including shared/group/global scopes and background/scheduled execution where applicable. Keep hosted Foundry agent behavior separate and unchanged unless an actual shared boundary needs integration.

Acceptance Criteria

  • Define a shared model/provider/protocol contract for reasoning with tools and record support using verified capabilities, including unsupported combinations and early actionable errors.
  • Provide the actual multi-turn reasoning-and-function-tool execution loop for supported OpenAI/Azure Responses endpoints, including function results and required reasoning items/context continuation across tool calls.
  • Integrate with local agent and applicable chat/orchestration/workflow execution without duplicating transport logic per YAMCS, database, document, or MCP action.
  • Preserve the selected deployment, actual agent-owned settings, requested/effective effort, tool schemas, tool authorization, provenance, source citations, token/output accounting, and safe user-visible diagnostics.
  • Preserve streaming, cancellation, bounded retries/timeouts, checkpoint/continuation behavior, and existing execution guards. A retry or disconnect must not replay completed or uncertain side-effecting tool work.
  • Preserve all working Chat Completions/None and GPT-5.1 paths. Do not force Responses onto providers that do not support it or claim every vendor/model has the same reasoning parameters.
  • Do not expose hidden chain-of-thought, secrets, raw provider bodies, unauthorized source names, or internal connection data in the UI or logs.
  • Add actual SDK/transport tests for reasoning plus tools over multiple turns, direct versus agent-owned model selection, streaming and nonstreaming results, orchestration/background parity, failures, retries, and nonregression of existing connectors.
  • Validate supported combinations in a controlled configured environment before claiming live model compatibility; clearly distinguish mocked request-shape coverage from real model acceptance.
  • Update agent/model connection/orchestration documentation and the application version when implementing.

Notes

Target: paullizer-react-v2-ui. Priority: P1. Assignee: Paul Lizer (@paullizer). Roadmap size: L.

First prerequisite: #1846 closes the narrower existing-tool-path explicit-None gap and should not be delayed by this larger capability.

Related: #1606 records GPT-6 Astra/Sol restrictions and explicitly identifies Responses migration as follow-up work; #978 tracks broader Microsoft model capability coverage. This issue owns the generic agent/orchestration reasoning-with-tools execution capability rather than a single action integration.

Primary areas: model_endpoint_clients.py, functions_model_capabilities.py, model endpoint/provider/runtime contracts, semantic_kernel_loader.py, agent_logging_chat_completion.py, and shared orchestration/workflow model invocation and result boundaries.

Official contract: https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning#tool-calling-with-reasoning-models

Activity

  1. linked a pull request that will close this issueControl center #567on Jan 8, 2026
  2. moved this from In progress to Pending Assignment in Simple Chat Roadmapon Feb 5, 2026
  3. changed the title [-]Add responses api support for thinking models[/-] [+]V2: Support reasoning with tools across agents and orchestration[/+] on Oct 27, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Labels

P1Priority 1: urgent, address before P2 workenhancementNew feature or request

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions