@amartinr/pi-bifrost-reasoning-fix

Pi extension that restores DeepSeek reasoning on tool-calling turns when the provider routes through Bifrost.

Problem

Bifrost relays DeepSeek reasoning in non-standard fields instead of the OpenAI reasoning_content:

• streaming: delta.reasoning / delta.reasoning_details
• messages: reasoning / reasoning_details

DeepSeek requires every assistant message in a tool-calling context to carry reasoning_content. The requirement derives from history content — once an assistant has produced tool_calls — not from whether the current request still ships tools. pi often re-sends history after a tool call without tools in the payload; DeepSeek still drops reasoning and may return an HTTP 400. Bifrost does not re-expose reasoning_content, so the replayed history reaches Bifrost with reasoning / reasoning_details and DeepSeek stops reasoning on the next tool-calling turn.

Mechanism

The extension registers before_provider_request and rewrites the provider payload before it is sent:

1. Moves assistant reasoning / reasoning_details text into reasoning_content.
2. When the history contains a tool call (or the request carries tools), guarantees every assistant message carries a reasoning_content field (empty if none yet), so DeepSeek keeps reasoning.

The rewrite is deterministic and stable across requests: a given assistant message always serializes identically once tool-calling is in scope, and user/system/tool messages are never touched. A stable prefix is therefore preserved for the provider's prefix cache.

The rename step is content-driven — it fires only when an assistant message carries reasoning / reasoning_details. The forcing step is not content-driven: it fires for any tool-calling payload (request tools or tool-call history). Model gating is defined by the extension configuration (see Extension configuration).

Scope of application

The extension applies to the model IDs listed under models in the extension configuration, matched by exact ID or prefix.

• Missing config file or empty models list — the extension applies to every request. In this mode the forcing step also touches tool-calling payloads from other providers, which may reject unknown reasoning_content fields.
• Listed models — the extension applies only to those models. Listing the DeepSeek models routed through Bifrost is the recommended setup.

Streaming (outlet)

after_provider_response is a no-op by design. pi reads streaming reasoning via the OpenAI SDK, which parses delta.reasoning / delta.reasoning_details and tolerates SSE comments (: heartbeat) and data: [DONE]. The drop occurs only when history is replayed on the inbound (request) side, which this extension fixes. No stream rewriting is required on the pi path.

Provider configuration (models.json)

The extension performs the inbound rewrite, so models.json does not need workarounds for the reasoning / reasoning_details dialect. It requires a Bifrost provider whose models declare reasoning: true and a thinkingLevelMap aligned with DeepSeek's official effort levels.

{
  "providers": {
    "bifrost": {
      "baseUrl": "http://<bifrost-host>/v1",
      "api": "openai-completions",
      "apiKey": "<your-api-key>",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": true
      },
      "models": [
        {
          "id": "deepseek/deepseek-v4-flash",
          "name": "DeepSeek V4 Flash (Bifrost)",
          "reasoning": true,
          "thinkingLevelMap": {
            "minimal": null,
            "low": "low",
            "medium": null,
            "high": "high",
            "xhigh": null,
            "max": "max"
          },
          "input": ["text"],
          "contextWindow": 1000000,
          "maxTokens": 384000
        }
      ]
    }
  }
}

Notes:

• apiKey is your Bifrost credential. Keep it in the file as usual for pi; never commit it. The key shown here is a placeholder.
• supportsReasoningEffort: true lets pi forward the reasoning level from the toggle.
• thinkingLevelMap exposes DeepSeek's three official effort levels (low, high, max) using exactly those values, so only reasoning_effort values DeepSeek documents are ever sent. The intermediate pi levels (minimal, medium, xhigh) map to null (hidden). This avoids sending medium/xhigh — which DeepSeek silently re-maps to high and a strict gateway like Bifrost may reject.
• Gateway-specific headers (bifrost.headers) are deployment-specific; keep them out of public documentation.

Extension configuration (config.json)

Configuration is read from a JSON file at the extension's global config path:

~/.pi/agent/extensions/pi-bifrost-reasoning-fix/config.json

A missing file applies defaults. An invalid file falls back to defaults with a warning.

The models list determines which requests the fix rewrites. In the example below, only the two DeepSeek models are affected.

{
  "models": ["deepseek/deepseek-v4-flash", "deepseek/deepseek-v4-pro"],
  "forceReasoningContentOnTools": true,
  "logLevel": "off"
}

Key	Type	Default	Description
models	string[]	[]	Model IDs (or prefixes) to which the fix applies. An empty list applies the fix to every request.
forceReasoningContentOnTools	boolean	true	Ensures every assistant message carries reasoning_content when the history has a tool call (applies even without tools).
logLevel	string	"off"	off / error / info diagnostics to stderr.

Credentials

The extension does not send requests to Bifrost. It rewrites the provider payload in memory; pi performs the authenticated HTTP call using the apiKey and headers declared in models.json. The extension never sees or forwards credentials, and no secret belongs in its config file.

• Never store an API key in the extension config file; keep it in models.json (or the environment) as pi expects.
• If the extension ever calls Bifrost directly, use a configurable auth header (like GATEWAY_AUTH_HEADER/GATEWAY_AUTH_VALUE) rather than a hardcoded secret.

Installation

Install the package as a pi package:

pi install npm:@amartinr/pi-bifrost-reasoning-fix

Alternatively, add it to packages in ~/.pi/agent/settings.json and reload pi (/reload) so the extension is auto-discovered.

After installation, create the extension configuration and list the target models (see Extension configuration). Without a config file, the extension applies to every request.

Development

npm install
npm run build
npm run typecheck

Publishing

The package is scoped to @amartinr and published public. prepublishOnly runs the build so dist/ is always current.

References

• filters/bifrost_reasoning_content_fix — same normalization for the Open WebUI → Bifrost path.
• pipes/agent_loop_guard — robust SSE forwarding for Bifrost's non-standard stream.
• Upstream Bifrost issues:
  • maximhq/bifrost#5325 — reasoning emitted in non-standard fields.
  • maximhq/bifrost#974 — streaming delta.reasoning dropped for some providers.
  • maximhq/bifrost#6523 — opening role-only SSE chunks dropped.

License

MIT