@jaco-tech/dsh-llm-claude-bridge

Claude Code subscription-backed model provider for DeepSeek Harness (ctx.llm).

Routes all LLM calls through your Claude Code subscription (the claude CLI / Claude Agent SDK) — allowing DSH to drive Opus/Sonnet/Haiku at no extra per-token cost — instead of using the pay-per-token Anthropic API.

This is a port of pi-claude-bridge — the extension that backs pi with a Claude Code subscription — adapted from pi's ExtensionAPI/pi-tui seam to DeepSeek Harness's Cordis ctx.llm/LlmAdapter seam.

Features

• Subscription Billing: Uses your Claude Code subscription (Pro or Max plan) via the claude CLI.
• Full Model Support: Opus 5, Opus 4.8, Opus 4.7, Opus 4.6 (1M), Sonnet 5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5, Fable 5.
• Extended Thinking: Full support for reasoning effort levels (off, low, medium, high, xhigh, max) with streamed reasoning deltas.
• DSH Tool Emission: Exposes DSH's tools to Claude Code via an in-process MCP server and emits standard DSH tool-call-delta chunks, so DSH executes tools within its own sandbox and approval policy. See Known limitations for the current state of the tool-result round trip.
• Multi-Turn Resumption: Automatically syncs and persists conversation history across turns using cc-session-io.
• Cordis Native: Plugs cleanly into ctx.llm via registerAdapter(["claude-bridge"], adapter) and registers a settings namespace for Web UI configuration.

Installation

Add to your DSH profile cordis.patch.yml (e.g. ~/.dsh/profiles/web/cordis.patch.yml):

- insert:
    - id: llm-claude-bridge
      name: '@jaco-tech/dsh-llm-claude-bridge'
      config:
        plan: 'max' # 'pro' or 'max'
        longContextExtraUsage: false
        strictMcpConfig: true
        autoMemoryEnabled: false
        pathToClaudeCodeExecutable: '$(which claude)' # or an explicit absolute path

Tool calling

Verified end-to-end by tests/test-tool-roundtrip.mjs:

1. Turn 1 — the model emits a tool-call chunk (get_weather, {"city":"Ghent"}). The adapter ends the turn at the tool-use boundary with finish {kind:'tool-calls'} (src/translate.ts), so the subprocess never resolves the MCP call against a placeholder.
2. DSH's loop executes the tool under its own sandbox/approval policy and appends a tool-result message.
3. Turn 2 — the result is replayed into the session history (src/serialize.ts), the prompt becomes (continue), and the model answers from the real result (verified: it referenced the injected 23°C result).

Known limitations:

• Attachments. Image blocks are passed through by attachment id and media type; binary payload wiring is untested.
• The MCP server still answers in-process tool calls with a placeholder under CLAUDE_BRIDGE_DEBUG — it is a debug aid, not a resolution path.

Billing

This routes requests through the claude CLI, so usage counts against your Claude Code subscription rather than pay-per-token API credits. Anthropic has changed its stance on SDK-vs-subscription usage before; check your plan's terms before relying on this.

Available Models

The model list is derived at runtime from pi-ai's built-in anthropic catalog (getBuiltinModels("anthropic")), projected through a small ordered allowlist — the same approach pi-claude-bridge uses. When pi-ai gains a new Claude model, it appears here automatically; only the Claude Code CLI-specific [1m] suffix / 1M-vs-200K context mapping is hardcoded, because that is CLI behavior the catalog does not encode.

The current set exposed under the claude-bridge provider:

• claude-bridge/claude-opus-5 (1M context)
• claude-bridge/claude-opus-4-8 (1M context)
• claude-bridge/claude-opus-4-7 (1M context)
• claude-bridge/claude-opus-4-6 (1M context on Max plan)
• claude-bridge/claude-sonnet-5 (1M context)
• claude-bridge/claude-sonnet-4-6 (200k context)
• claude-bridge/claude-sonnet-4-5 (200k context)
• claude-bridge/claude-haiku-4-5 (200k context)
• claude-bridge/claude-fable-5 (1M context)