drscrewdriver/dsh-llm-openai-completions

✅ Revived as 0.3.0 (2026-10-05). The transport takeover is back with real features pi-ai cannot offer: multimodal true-send (image/audio/video/PDF) with a modality hard gate, an llmOpenaiCompletions prompt-source seam for dsh-promptbook-class consumers, credentials-service keys, and full-rc dynamic compatibility across 0.1.0-rc.8 – 0.2.0-rc.2. The old settings.get-era architecture is gone — the judgment is enabled && providers.includes(route) (pure manual membership: the per-route checkbox is the truth), and the switch lives in this plugin's own settings section. Contract surfaces: AGENTS.md.

Positioning: revived as a short-circuit backend that dsh-thinking-levels can drive — speaking all three mainstream API formats per route api declaration (OpenAI Chat Completions · Anthropic Messages · OpenAI Responses), built primarily for Xiaomi MiMo multimodal: true audio/video/PDF send.

README languages / 说明文档

English · 中文 · 日本語 · 한국어 · Deutsch · Español · Français · Italiano · Русский

Install guides: EN · 中文 · 日本語 · 한국어 · DE · ES · FR · IT · RU — Changelog

OpenAI-completions-compatible transport takeover for custom gateways (vLLM / LM Studio / self-hosted OpenAI proxies): the plugin short-circuits the llm/stream waterfall for taken-over routes and serves them with behavior that never guesses — the data plane (baseURL / models / reasoningEfforts / compat / input) stays in llm-pi-ai, read live through settings.describe().

• system is always role: "system" — a custom gateway never receives developer, so the Unexpected message role 400 is gone even when thinking is enabled (pi-ai's detectCompat defaults supportsDeveloperRole: true for non-standard URLs, which breaks every custom gateway once a reasoningEfforts table marks the model as reasoning).
• Thinking is driven by the model's compat.thinkingFormat:
  • toggle models (reasoningEfforts table, no supportsReasoningEffort) → chat_template_kwargs.enable_thinking (+ preserve) — always, regardless of the configured format, because that is what bare vLLM Qwen3 templates honor
  • qwen → top-level enable_thinking: boolean (no reasoning_effort, no budget)
  • qwen-chat-template → chat_template_kwargs.enable_thinking (+ preserve)
  • effort-capable → reasoning_effort passthrough (Qwen3.8-style)
• The receive side splits Qwen3-style </think> content (vLLM renders thinking into content with no reasoning_content field) into a reasoning block, and reads reasoning_content / reasoning / reasoning_text.
• No thinking_budget is sent (deliberate: avoid truncation surprises).

🎉 0.3.0 — Multimodal true-send with a modality hard gate

The short-circuit consumes the waterfall before the host projects file blocks to text handles, so media arrives structural and is encoded for the wire — no host patches, no file-handle reconstruction:

• image → image_url data-URI parts (multi-image count and order preserved, text interleaved in source order)
• audio → input_audio parts (base64 + format)
• video → video_url data-URI parts (20 MB inline cap)
• PDF → OpenAI file parts
• Modality hard gate: every media block in the full history is checked against the model's llm-pi-ai input declaration before anything is read or sent; an undeclared modality is refused loudly (INVALID_REQUEST, naming the modality) instead of being silently degraded to text.
• One-shot calls (session-title) strip media blocks and never think.
• Bytes come from the attachment service (readImage / readFileStream); non-vision models still reject images loudly (UNSUPPORTED_CONTENT).

🎉 0.3.0 — The llmOpenaiCompletions prompt-source seam

ctx.provide("llmOpenaiCompletions", { listHostedModels, registerPromptSource }) — the same contract dsh-llm-mimo exposes as llmMimo, for dsh-promptbook-class consumers: replace the actually dispatched system prompt and tool descriptions per (provider, model), deterministically (cache-safe), text-only (never name/parameters), zero behavior with no source registered. Scope guarantee: consulted only on taken-over routes. Full contract: AGENTS.md §1.

Three wire faces (per-route api)

face	endpoint	auth	thinking	modalities
openai-completions (default)	{base}/chat/completions	Bearer	thinkingFormat matrix	image/audio/video/pdf
anthropic-messages	{base}/v1/messages	x-api-key + version header	off/one-shot → disabled	image only
openai-responses	{base}/responses	Bearer	reasoning.effort (honest fold)	image/audio/video/pdf

• Every face: max_tokens = request value → model row → route default → 8192; the leading system message is extracted into the face's system channel and stripped from the message list.
• Anthropic pairing discipline: tool_use/tool_result must pair (broken history = loud INVALID_REQUEST); assistant thinking is not replayed (no signature in the harness — strictly weaker than the chat face).
• Responses boundaries: stop dropped (no Responses spelling); reasoning effort folds honestly; probe calibration pending against real MiMo endpoints.

Takeover judgment (single source of truth)

takeover(route) = section.enabled === true
                  && section.providers.includes(route)

# ~/.dsh/profiles/web/cordis.patch.yml (or Settings → Plugins → dsh-llm-openai-completions)
llm-openai-completions:
  enabled: true
  providers:        # the takeover list — the single truth (see below)
    - local-35b

• The switch and the takeover list live in this plugin's own settings section (.volatile() — runtime-editable, host-generated form).
• dsh-thinking-levels drives both fields: its takeover switch mirrors {enabled} into this section, and its model panel's per-route takeover checkbox writes providers directly — the checkbox is the truth, so dispatch, listHostedModels, and the control layer's gating all read one set.
• Disabling the switch (or unmounting the plugin) falls through to pi-ai unchanged — no adapter registration to unwind.
• Your llm-pi-ai config stays as-is (one shared config; this adapter only reads it — full field table: docs/settings-spec.md, takeover contract: docs/takeover-spec.md):

llm-pi-ai:
  providers:
    local-35b:
      api: openai-completions
      baseURL: http://192.168.100.242:8200/v1
      models:
        - id: Qwen3.6-35B-A3B
          reasoningEfforts: { off: null, high: 'high' }
          compat:
            thinkingFormat: qwen        # enable_thinking only, no reasoning_effort

Preview

Live screenshot — powered together with dsh-thinking-levels: the custom model's "thinkable" level dropdown (Default / Off / On). It demonstrates the capability-detection → takeover chain:

Install

Host compatibility: every rc from 0.1.0-rc.8 through 0.2.0-rc.2, version-conditional in code (shape-detected settings faces, a ToolCallId/CallId shim, credentials → launch-environment → process.env key resolution). See the install guide for all routes:

# npm (recommended)
dsh plugin --profile web add dsh-llm-openai-completions -w
# local tgz
dsh plugin --profile web add /path/to/dsh-llm-openai-completions-0.3.0.tgz -w
dsh web

Known limitations

• PDF support is endpoint-dependent (MiMo chat rejects it outright; some relays accept it) — the input declaration is a capability claim; check your endpoint before declaring pdf.
• Non-media file attachments (e.g. .txt) have no cross-gateway wire shape; on taken-over routes they degrade to a text placeholder (content stays reachable through the file tools).
• No thinking_budget (deliberate).
• The thinking-level selector still comes from the model's reasoning metadata (llm-pi-ai + dsh-thinking-levels); this transport controls the wire behavior only.
• Restart dsh web after install; the takeover listener composes at boot.

Development

npm install
npm run typecheck   # tsc --noEmit
npm test            # vitest (config / identify / serialize / sse / translate / prompt-source / version-conditional)
npm run build       # tsc → lib/
node scripts/audit-host-faces.mjs   # 15-rc dependency-surface audit

Upstream proposal

This plugin doubles as a reference implementation for the llm-pi-ai / pi-ai compatibility gaps: custom URLs should default supportsDeveloperRole: false, the receive side should split </think> content, compat should expose supportsDeveloperRole for gateways that genuinely support it — and the host projection layer could grow the same capability-aware file-block gate this plugin applies per route.