Skip to content
Talon

July updates: reasoning controls and safer provider sessions

Talon’s latest provider updates make long-running agents easier to tune without sacrificing session safety. Personas can choose how much reasoning work a compatible model should do, stateful Responses API sessions no longer leak across model changes, and context rotation now distinguishes the final model step from cumulative usage across a tool-heavy turn.

Choose reasoning effort per persona

Section titled “Choose reasoning effort per persona”

Personas can now set a reasoningEffort alongside their provider and model:

personas:
  - name: deep-researcher
    provider: codex-cli
    model: gpt-5.4
    reasoningEffort: high

Supported values are none, minimal, low, medium, high, and xhigh. Talon passes the value to Codex CLI and to OpenAI-compatible providers running in Responses API mode. Use the provider’s base model name rather than encoding effort in a model-name suffix.

Reasoning effort is also part of a resumable session’s identity. If a persona changes from medium to high, Talon starts a fresh provider session and restores assembled conversation context instead of resuming a chain created with different reasoning settings.

See the persona reasoning effort commit.

Resume Responses sessions without crossing models

Section titled “Resume Responses sessions without crossing models”

OpenAI-compatible providers now separate the API shape from session behavior:

agentRunner:
  providers:
    omlx-local:
      type: openai-compatible
      options:
        baseUrl: http://mac.local:8000/v1
        defaultModel: qwen3.5-9b-optiq-4bit
        providerId: omlx-local
        apiMode: responses
        sessionMode: previous_response_id

apiMode: responses selects /v1/responses. sessionMode: previous_response_id additionally enables stateful response chaining for endpoints that support it, allowing them to reuse their response state and prefix or KV cache between turns.

Stored response IDs are scoped by thread, provider, model, and configured reasoning effort. Changing the model or effort starts a fresh provider session, so Talon never resumes a stale response chain under incompatible settings. The deprecated omlxResponses option remains available as a compatibility alias, but new configurations should use apiMode and sessionMode explicitly.

Only enable previous_response_id for foreground endpoints that implement stateful Responses API chaining. Leave it off for ordinary chat-completions endpoints and providers that do not support cross-run response IDs.

Read the merged session-scoping pull request.

Context rotation now follows the final model step

Section titled “Context rotation now follows the final model step”

Multi-step turns can accumulate a lot of tokens while an agent calls tools. Using that cumulative total to decide when to rotate context can make a healthy session rotate too early, even when the final model step is small.

Talon now keeps cumulative usage for accounting and Langfuse, while using the final model step for context-rotation decisions when the provider exposes it. Codex CLI supplies this through its token_count.last_token_usage events. Tool-heavy work can continue without triggering a rotation solely because earlier steps in the same turn were expensive.

Pass vendor options to local models

Section titled “Pass vendor options to local models”

The same update adds providerOptions for vendor-specific request-body controls. Configure the fields directly under the provider; Talon wraps them under the configured providerId before calling the OpenAI-compatible runtime:

agentRunner:
  providers:
    ollama-mac:
      type: openai-compatible
      options:
        baseUrl: http://mac.local:11434/v1
        defaultModel: qwen3-coder:30b
        providerId: ollama-mac
        providerOptions:
          chat_template_kwargs:
            enable_thinking: false

This makes provider-native controls available without hard-coding them into Talon. For example, a Qwen deployment can disable its own thinking mode while Talon continues to use the standard OpenAI-compatible provider path.

See the context-rotation and provider-options commit.

Together, these changes make provider sessions more predictable: the right reasoning budget, the right resumable chain, and rotation decisions based on the model step that actually determines remaining context pressure.