Custom post-processing fails with DeepSeek because reasoning_effort:none is rejected

[sinco-z]

Problem

When using the Custom post-processing provider with DeepSeek's OpenAI-compatible endpoint:

• Base URL: https://api.deepseek.com
• Model: deepseek-v4-flash

post-processing fails with HTTP 400 because Handy sends:

"reasoning_effort": "none"

DeepSeek rejects this field/value:

reasoning_effort: unknown variant `none`, expected one of `high`, `low`, `medium`, `max`, `xhigh`

Context

PR #1221 added reasoning_effort: "none" for Custom providers to avoid thinking-mode latency in local models such as Ollama/Gemma/Qwen. That behavior is useful and should remain for those endpoints.

The issue is that not all OpenAI-compatible APIs accept the same reasoning-control parameters. DeepSeek uses a different thinking object for V4 thinking mode, and rejects reasoning_effort: "none".

DeepSeek docs: https://api-docs.deepseek.com/guides/thinking_mode

Expected behavior

Custom provider should still work with DeepSeek-compatible endpoints.

Ideally:

• Preserve reasoning_effort: "none" for local Custom endpoints where it works.
• Avoid sending reasoning_effort: "none" to DeepSeek.
• Optionally send thinking: { "type": "disabled" } for DeepSeek to avoid extra latency during transcript cleanup.

Actual behavior

Post-processing fails and Handy falls back to the original transcription.

Logs

LLM post-processing failed for provider 'custom': API request failed with status 400 Bad Request:
{"error":{"message":"Failed to deserialize the JSON body into the target type: reasoning_effort: unknown variant `none` ..."}}

Notes

This is related to #1221, but not a request to remove the Custom reasoning optimization globally. The local-model latency improvement should be preserved.

[cjpais]

Yeah we do need to fix this especially for Deepseek in particular

[Raymond-Ning]

Same issue encountered, waiting for a fix

[cjpais]

You should be able to use previous versions before this patch and should resolve for now