vultr/openclaw
1
0
Fork 0
mirror of https://github.com/openclaw/openclaw.git synced 2026-05-06 07:30:43 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix(deepseek): backfill v4 assistant reasoning replay

Browse Source
This commit is contained in:
Peter Steinberger
2026-04-28 10:07:15 +01:00
parent 0876ff481b
commit 62997f7fce
9 changed files with 176 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
docs/providers/deepseek.md
View File

@@ -84,17 +84,17 @@ calls can continue.
## Thinking and tools
DeepSeek V4 thinking sessions have a stricter replay contract than most
OpenAI-compatible providers: when a thinking-enabled assistant message includes
tool calls, DeepSeek expects the prior assistant `reasoning_content` to be sent
back on the follow-up request. OpenClaw handles this inside the DeepSeek plugin,
so normal multi-turn tool use works with `deepseek/deepseek-v4-flash` and
`deepseek/deepseek-v4-pro`.
OpenAI-compatible providers: after a thinking-enabled turn uses tools, DeepSeek
expects replayed assistant messages from that turn to include
`reasoning_content` on follow-up requests. OpenClaw handles this inside the
DeepSeek plugin, so normal multi-turn tool use works with
`deepseek/deepseek-v4-flash` and `deepseek/deepseek-v4-pro`.
If you switch an existing session from another OpenAI-compatible provider to a
DeepSeek V4 model, older assistant tool-call turns may not have native
DeepSeek `reasoning_content`. OpenClaw fills that missing field for DeepSeek V4
thinking requests so the provider can accept the replayed tool-call history
without requiring `/new`.
DeepSeek `reasoning_content`. OpenClaw fills that missing field on replayed
assistant messages for DeepSeek V4 thinking requests so the provider can accept
the history without requiring `/new`.
When thinking is disabled in OpenClaw (including the UI **None** selection),
OpenClaw sends DeepSeek `thinking: { type: "disabled" }` and strips replayed

6
docs/providers/venice.md
View File

@@ -127,9 +127,9 @@ Use the table below to pick the right model for your use case.
If Venice exposes DeepSeek V4 models such as `venice/deepseek-v4-pro` or
`venice/deepseek-v4-flash`, OpenClaw fills the required DeepSeek V4
`reasoning_content` replay placeholder on assistant tool-call turns when the
proxy omits it. Venice rejects DeepSeek's native top-level `thinking` control,
so OpenClaw keeps that provider-specific replay fix separate from the native
`reasoning_content` replay placeholder on assistant messages when the proxy
omits it. Venice rejects DeepSeek's native top-level `thinking` control, so
OpenClaw keeps that provider-specific replay fix separate from the native
DeepSeek provider's thinking controls.
## Built-in catalog (41 total)