reasoning-field-remap

An Oh My Pi (omp) / pi extension for OpenAI-compatible gateways that return chain-of-thought reasoning in non-standard response fields.

The problem

omp reads model reasoning from reasoning_content, reasoning, or reasoning_text — the fields used by DeepSeek, Kimi, and most OpenAI-compatible hosts. But some gateways ship reasoning in their own field instead (for example delta.thinking / message.thinking). The result: the model thinks, tokens are billed, but the thinking blocks are invisible in the TUI — or multi-turn tool use breaks on strict hosts that require reasoning replay.

Typical fix is a local rewrite proxy. This extension does it in-process, with no extra process: it re-registers your marked provider so the response body is rewritten on the fly, and omp's own streaming implementation (tool calls, usage, retries, auth) does all the real work.

Install

Copy index.ts into your agent extensions directory:

~/.omp/agent/extensions/reasoning-field-remap.ts   (user scope)
<project>/.omp/extensions/reasoning-field-remap.ts (project scope)

or install from npm (single package, no dependencies):

omp install npm:pi-reasoning-field-remap
# in pi:
pi install npm:pi-reasoning-field-remap

Restart the session. That's it — no dependencies, no build step.

Configure

Mark any provider in ~/.omp/agent/models.yml with a compat marker:

providers:
  my-gateway:
    baseUrl: https://api.example.com/v1
    api: openai-completions        # must be openai-completions
    apiKey: ...
    compat:
      wireReasoningSource: thinking          # wire field to remap
      # wireReasoningSourceStream: thinking  # delta field (defaults to source)
      # wireReasoningDropFields:             # fields to strip (defaults to
      #   - thinking_signature               #   [<source>_signature])
    models:
      - id: vendor/model-x
        reasoning: true
        input: [text, image]
        contextWindow: 1048576
        maxTokens: 32768

On startup the extension logs:

[reasoning-field-remap] provider "my-gateway": reasoning wire field "thinking" -> reasoning_content (1 model(s))

Unmarked providers are never touched.

Options

The remapped value always lands in reasoning_content (omp's canonical field). The source field is removed from the response so providers that echo it back never anchor on it.

How it works

1. At session start the extension reads models.yml via omp's own ModelsConfigFile loader and finds providers carrying wireReasoningSource.
2. Each marked provider is re-registered under the same name with a custom internal API whose streamSimple delegates to the built-in openai-completions implementation from @oh-my-pi/pi-ai.
3. Before delegation the model is rebuilt through @oh-my-pi/pi-catalog buildModel with api: "openai-completions", materializing the resolved compat record the builtin requires (custom API names get none otherwise).
4. A fetch wrapper rewrites the response stream: delta.<source> → delta.reasoning_content (line-buffered, safe for chunk-split JSON), message.<source> → message.reasoning_content, drop-fields removed.

Limitations

• Only api: openai-completions providers are supported (that's where the wire-field zoo lives).
• Config is read once at session start; edit models.yml → restart the session.
• The extension uses omp's host package rewrites (@oh-my-pi/pi-ai, @oh-my-pi/pi-coding-agent, @oh-my-pi/pi-catalog/build), so it must be loaded by omp/pi — not standalone.

License

MIT

Tests

bun test

The suite mocks omp's host packages and exercises the rewrite matrix: default marker, chunk-split JSON, and non-stream bodies.