deepseek-harness/packages/llm/llm-pi-ai

@deepseek-ai/dsh-llm-pi-ai

DeepSeek adapter for the harness LLM seam backed by @earendil-works/pi-ai (the LLM library behind the pi agent).

Why a second adapter exists

@deepseek-ai/dsh-llm-deepseek already talks to the same endpoint. This package is its design-verification twin: same models, same wire protocol, completely different internals — a unified LLM library with its own event vocabulary versus hand-rolled fetch/SSE. Anything the harness StreamChunk protocol cannot express for BOTH implementations is a core-vocabulary bug. The differences it exercised on purpose:

• pi-ai hands tool-call arguments around as parsed objects; the harness keeps raw JSON strings. The adapter patches replay payloads back to the original raw strings before sending them, and re-stringifies parsed output tool calls at block-end.
• pi-ai reports failures as in-stream error events (it never throws mid-stream); these map to finish {kind:'error'|'aborted'} chunks — the protocol's other sanctioned error path besides throwing (which llm-deepseek uses).
• pi-ai folds reasoning tokens into usage.output; there is no separate reasoning count to map.
• pi-ai's options omit some DeepSeek/OpenAI-compatible details; the adapter uses its onPayload hook to preserve the harness contract (stop, scrubbing pi-ai's own per-tool strict default — the hand-rolled twin sends no such field — omitted reasoning effort, raw replayed tool arguments).

Config

Same shape as llm-deepseek (one-line swap in cordis.yml), with pi-ai's thinking-level vocabulary:

- id: llm
  name: '@deepseek-ai/dsh-llm-pi-ai'
  config:
    apiKey: !!js process.env.DEEPSEEK_API_KEY
    baseURL: !!js process.env.DEEPSEEK_BASE_URL
    models: [deepseek-v4-flash, deepseek-v4-pro]
    reasoning: high   # off | high | xhigh (xhigh → wire 'max')

App attribution

Every request carries the shared attribution header from dsh-llm's attributionHeaders(), passed through pi-ai's headers stream option (pi-ai merges caller headers last, so it always reaches the wire - the unit suite asserts arrival on the mock server, same as llm-deepseek). OpenRouter-specific app attribution headers are intentionally not sent by this adapter contract; they are deferred to a future explicit OpenRouter adapter or mode. See dsh-llm § App attribution.

Dependency weight

pi-ai declares the openai/anthropic/google/mistral/AWS SDKs as install-time dependencies. They are lazy-loaded — only the openai SDK actually loads for this adapter — but they do land in node_modules. Accepted for a package whose purpose is design verification.

Testing

Unit suites run against a local node:http mock SSE server (pi-ai's openai SDK happily talks to any base URL). Real-API coverage in tests/adapter.e2e.ts (pnpm run test:e2e, key-gated): V4 Flash + V4 Pro across all exposed reasoning levels (off/high/xhigh), the thinking+tools round trip, and a cross-adapter structural-equivalence check against llm-deepseek.

Model Experience

DeepSeek request through pi-ai

What the model sees: The selected model receives the same logical system prompt, history, tools, stop sequences, and raw replayed tool arguments as the hand-written adapter. This package adds no prompt prose and removes pi-ai's own per-tool strict default to preserve that contract.

Token effect: Provider tokenization governs exact input. Reasoning level changes generated and passback content; pi-ai reports reasoning inside output usage rather than as a separate count.

DeepSeek response

What the model sees: pi-ai events become harness reasoning, text, tool-call, usage, and finish chunks; parsed tool arguments are restored to raw JSON strings at the harness boundary.

Token effect: Generated content affects later inputs only after the loop records it; adapter conversion adds no model-visible text.

Known Limitations and Deferred Work

• tool_choice is not mapped — same MVP contract as llm-deepseek.
• In-history system-role messages fold into user-role wire messages — pi-ai exposes a single systemPrompt slot, diverging from the hand-rolled twin's role: 'system' passthrough.
• LlmError.status is never set — pi-ai reports failures as in-stream events with no HTTP status, so error codes are regex-classified from the error text.
• buildModel hardcodes descriptor metadata — contextWindow: 128000, maxTokens: 64000, zero cost, identically for every registered model name; not configurable.
• pi-ai's built-in retries are disabled (maxRetries: 0) — failures surface immediately; retry policy belongs to llm/stream listeners.