DeepSeek's official docs now publish an Oh My Pi compatibility contract with custom `models.yml`, reasoning mapping, and 400-error guardrails

Checked on June 19, 2026: DeepSeek's official Oh My Pi integration page says the built-in DeepSeek entries are not reliable enough on their own, requires a custom `~/.omp/agent/models.yml`, and documents the exact compat fields needed to avoid 400 errors in thinking-mode tool-call sessions.

英文原文

这篇动态暂时没有完整中文稿。为了避免中文 SEO 页面收录英文正文，详情页会标记为 noindex。

Accepted official-source monitoring note

Today's accepted item stays DeepSeek-first and uses a current official DeepSeek documentation page because the public X surface was still not safely crawlable in this run. DeepSeek's homepage continues to anchor the official @deepseek_ai account, but the direct X timeline did not expose readable current post content here, so the publish-safe choice is a docs-backed developer signal.

What we verified on June 19, 2026

DeepSeek now has an official Oh My Pi integration page in its agent-integrations docs.
That page explicitly says the built-in DeepSeek model entries are incomplete for reliable use and requires a custom ~/.omp/agent/models.yml.
DeepSeek publishes a full provider contract around https://api.deepseek.com, openai-completions, contextWindow: 1000000, maxTokens: 384000, and separate model definitions for DeepSeek V4 Pro and Flash.
The compatibility guidance is specific to DeepSeek's thinking-mode behavior, including supportsToolChoice: false, requiresReasoningContentForToolCalls: true, and requiresAssistantContentForToolCalls: true.
The same official page warns that missing those fields can trigger 400 errors in long tool-using reasoning sessions, and it maps Oh My Pi's xhigh effort level to DeepSeek's supported max value.

Why this is publishable

This is not a rumor, benchmark screenshot, or product-inventory claim. It is a current official DeepSeek integration contract that changes how developers should wire a real coding agent:

It gives a first-party workaround for incomplete built-in compatibility instead of asking users to guess at a custom schema.
It documents reasoning-mode and tool-call guardrails, which are stronger operational signals than a basic install command alone.
It is not a duplicate of the earlier GitHub Copilot, Copilot CLI, Hermes, Deep Code, OpenCode, Pi, WorkBuddy, nanobot, or Crush items already on the site.

Why this matters for DeepSeek-first SEO pages

Oh My Pi support pages can now target an official custom models.yml baseline instead of generic OpenAI-compatible advice.
DeepSeek setup content can distinguish ordinary endpoint wiring from thinking-mode tool-call compatibility, which is where many real 400 errors appear.
Routing pages can explain that DeepSeek's official docs sometimes override built-in agent defaults, which is exactly the kind of support-query intent these long-tail pages should capture.

Rejected candidates today

Official X timeline as the primary source: rejected for this run because the homepage anchor was visible but the direct x.com/deepseek_ai page still returned no readable timeline content here.
The homepage V4 Preview anchor: official, but older and already overlapped by existing V4 coverage.
The current Pi and WorkBuddy pages: still official, but already published yesterday and therefore duplicate-content risks.
The new Kilo Code page: official and accepted as a guide opportunity for this run, but weaker than Oh My Pi as the single news headline because it is a simpler /connect and /models workflow without the richer compatibility layer.
Status-page uptime alone: official, but weaker than a current docs page that changes real developer setup behavior.

Editorial takeaway

The safest official DeepSeek story today is an Oh My Pi integration documentation check: DeepSeek's own docs now say the built-in DeepSeek entries are not enough, publish a custom models.yml, and call out the exact compat fields needed to avoid thinking-mode tool-call failures. That is a stronger current developer signal than recycling an older X post that still could not be fully verified live.