Skip to content

README.md

Latest commit

 

History

History
64 lines (38 loc) · 5.05 KB

README.md

File metadata and controls

64 lines (38 loc) · 5.05 KB
Raw

@deepseek-ai/dsh-llm-engine

English | 中文

LLM-seam adapter for the local engine CLIs. Registers the claude-code and codex provider routes on ctx.llm, backed by the matching ctx.subagents backends: dsh-subagent-claude-code runs the official Claude Agent SDK, dsh-subagent-codex runs the official codex app-server. Selecting one of these routes as a session's provider — in the web Models picker, agent-default-model, or any model selection — routes every turn through the local CLI with its native OAuth state (claude.ai / ChatGPT login); no API key of any kind.

What the adapter does

stream() takes the latest DIRECT user prompt's text (synthetic context injections such as AGENTS.md and skill catalogs are excluded by source kind), starts the matching backend through ctx.subagents with a cwd-bearing stub parent, forwards the engine's live text deltas as text-delta chunks while the run settles, and closes with the assembled final answer and a terminal finish. The engines are agents, not model endpoints: they execute their own tools inside their own process, so the harness loop sees text-only turns and the session log records the answers like any other assistant message. Engine runs are never retried (maxRetries: 0): a retry would execute the CLI twice.

Long-lived sessions. With continuation: true (plugin config plus the matching backend rows), the adapter keys an engine session per harness session id and resumes it on every later turn: Claude Code via the SDK's resume (persistSession), Codex via thread/resume on a persistent thread. The engine then remembers earlier turns of the same harness session — one conversation, not fresh-per-turn.

Refusal surfaces (all error finishes, no engine process started):

  • No user text — EMPTY_PROMPT: the engines need a non-empty instruction.
  • Auxiliary purposes — UNSUPPORTED_PURPOSE: compaction and session-title calls would spawn a full agent run for a side task.
  • Non-completed results — ENGINE_FAILURE (or ABORTED for explicit cancellation), EMPTY_RESPONSE for a completed run with no text.

Each route advertises its real model catalog (claude-code: Claude Opus/Sonnet/Haiku family; codex: the installed app-server's GPT-5.3 Codex models) plus a native choice meaning "the CLI's own configured default — no override sent". Selecting a catalog model sends the id to the engine (SDK model / app-server turn/start model). Every catalog model advertises selectable reasoning efforts (off/low/medium/high, plus xhigh/max for Claude), mapped to the engines' native vocabularies (SDK effort/thinking, app-server effort); the harness materializes high when the caller omits one.

Composition

The @deepseek-ai/dsh-multi-provider bundle mounts this adapter together with the two backends; the runnable reference is examples/multi-provider. A profile needs dsh-subagent (in dsh-base) plus the two backend rows and this row:

- id: llm-engine
  name: '@deepseek-ai/dsh-llm-engine'

- id: subagent-codex
  name: '@deepseek-ai/dsh-subagent-codex'

- id: subagent-claude-code
  name: '@deepseek-ai/dsh-subagent-claude-code'

No keys: claude must be logged into claude.ai and/or codex into ChatGPT.

Model Experience

Prompt text

What the model sees

The engine receives the latest user message text only, as one task in a fresh run. It never sees the conversation history, the harness system prompt, or tool schemas; its own native system prompt, tools, and permissions apply inside its process.

Token effect

Data-dependent; each engine run pays for an independent context that never enters the harness's derived history beyond the recorded answer.

KV Cache effect

None; each run is a fresh engine process with its own provider state.

Permission decisions

The engine's native permission flow applies inside its process (Claude Code's approvals, Codex's approval_policy); the harness approval seam is not involved. See each backend README's permission posture.

Known Limitations and Deferred Work

  • Text-only turns — engine tool activity stays inside the engine process; the session log records no tool calls for engine turns (streamed text deltas arrive live, tool events do not).
  • Continuation is opt-in and native-stateful — continuation: true persists Claude session files and Codex threads under the native CLI config dirs; the one-shot default touches no native state.
  • Host cwd — engine runs happen in the host process's cwd, not the session's workspace.
  • Auxiliary calls refused — compaction and session-title generation fail for engine sessions; titles fall back to the fallback rules.
  • Catalog is static — model lists are curated from the installed SDK/catalog versions; a newer CLI may support models the list does not yet name (the native choice always tracks the CLI's own default).