Skip to content

README.md

Latest commit

 

History

History
48 lines (33 loc) · 4.07 KB

README.md

File metadata and controls

48 lines (33 loc) · 4.07 KB
Raw

multi-provider example

English | 中文

Runnable reference for the @deepseek-ai/dsh-multi-provider pack bundle: the agent spine with three key-based providers (Gemini, MiniMax, Kimi) activated on the generic pi-ai adapter, plus the Claude Code and Codex delegation backends, composed as one leaf. cordis.yml is the composition base; cordis.snapshot.yml is its replay twin for the keyless snapshot suites.

What this leaf proves

  • Key-based providers register and serve. The llm-pi-ai row activates google (Gemini, GOOGLE_API_KEY), minimax (MiniMax, MINIMAX_API_KEY), kimi-coding (Kimi, KIMI_CODING_API_KEY), and anthropic (Claude API, ANTHROPIC_API_KEY) with curated model catalogs. Keys resolve per request through the credential seam — no key is inlined. Selecting a key-based route without its key fails loud with MISSING_CREDENTIAL naming the key.
  • Delegation backends load with native auth. The claude-code (official Claude Agent SDK) and codex (official codex app-server --stdio) providers compose on the host and start no child process until a tool call. Their tool rows ship disabled: true (the shipped-profile posture); remove the flag in a copy of the row to expose subagent_claude_code / subagent_codex to every agent.
  • Engines are selectable providers. The dsh-llm-engine adapter registers claude-code and codex on the LLM seam (model native): sessions on those routes run every turn through the local CLI with its OAuth state — the same path the web Models picker uses.
  • End to end. The keyless suites boot the real composition through the Loader and pin the provider listing; the keyed suites run a live task per provider and a live delegation per backend.

Run it

# List the registered providers and model catalogs (keyless, deterministic):
pnpm run demo:multi-provider providers

# Run one task on a key-based provider (needs the provider's key):
pnpm run demo:multi-provider run --provider google --model gemini-2.5-flash "hello"
pnpm run demo:multi-provider run --provider anthropic --model claude-opus-5 "hello"

# Run the whole task through an engine provider — native OAuth, no key.
# The session's turns run through the local CLI via the engine LLM adapter:
pnpm run demo:multi-provider run --provider claude-code "hello"
pnpm run demo:multi-provider run --provider codex "hello"

Both commands boot this leaf's cordis.yml; the demo bin lives at packages/examples/multi-provider-demo/src/bin.ts.

Credentials

GOOGLE_API_KEY, MINIMAX_API_KEY, KIMI_CODING_API_KEY, and ANTHROPIC_API_KEY resolve per request through the credential seam: the process environment, then the managed credentials document ($DSH_HOME/.credentials.yaml, what the web Models page writes). Claude Code and Codex need no key — they authenticate with their native OAuth state (claude logged into claude.ai, codex logged into ChatGPT).

Tests

Suite Keys What it pins
tests/providers.spec.ts none Real Loader boot: the three routes with the curated catalogs, both backends registered, disabled tool rows absent from the tool surface; leaf providers dict stays in sync with the pack bundle patch
tests/listing.snapshot.ts none Demo bin providers stdout, byte-for-byte (re-record with pnpm run test:snapshot:refresh -- multi-provider)
tests/providers.e2e.ts per provider One live task per provider through the demo bin (Reply with exactly: PONG), self-skipping without its key
tests/delegation.e2e.ts none (native OAuth) One live delegation per backend through the real composition, self-skipping when the product CLI is missing or logged out

Known Limitations

  • Delegation is per-task, not whole-session: each run is one fresh query/thread (see the backend READMEs). Driving an entire session through Claude Code or Codex is a separate experimental surface.
  • The llm-pi-ai providers dict here must mirror packages/bundle/multi-provider/cordis.patch.yml; tests/providers.spec.ts enforces the match mechanically.