dsh-plugin-ai-bridge

简体中文 · English · 日本語 · 한국어

dsh-plugin-ai-bridge

Bridge DeepSeek Harness to external AI models — second-opinion review · adversarial review · task delegation · non-blocking background scheduling

✨ Features · ⚡ Install · 🔧 Config · 🎛️ Commands · 🧰 Tools · 🏗️ Architecture · 🎬 Demo · 🧪 Tests · 🛡️ Security · ⚖️ License

✨ Features

DeepSeek Harness currently lacks cross-model collaboration. This plugin builds a "bridge" for it — delegate code review or complex tasks to an external model from inside a session, then bring its judgment back into the current session.

🔍  Read-only second-opinion review — style, logic, potential bugs, security; never mutates session state ⚔️  Adversarial review — challenge design decisions, architecture assumptions, boundary conditions and error handling with 5–10 "soul-searching" questions 🛟  Task delegation / rescue — bundle conversation history + task, inject the result back as plugin context ⏳  Non-blocking background jobs — status / result / cancel on top of ctx.jobs 🧩  Model-facing tools — ai_bridge_review / ai_bridge_delegate so the agent can ask for help on its own

⚡ Install

# One command installs and auto-mounts the plugin (via dsh.bundle.patch).
dsh plugin --profile <profile-name> add dsh-plugin-ai-bridge

# Restart the profile, then type:
/bridge help

Then override the config in the profile's cordis.patch.yml (see "Config" below):

- id: ai-bridge
  config:
    provider: generic
    baseUrl: https://your-relay.example.com/v1
    defaultModel: gpt-5.4
    apiKey: sk-...

No manual insert is needed — the plugin mounts itself through dsh.bundle.patch. You can also use only env vars (BRIDGE_API_KEY / BRIDGE_BASE_URL / BRIDGE_MODEL).

If you previously mounted it with a manual insert, remove that block to avoid double-mounting. To install from source: npm run build, then link lib/ into the profile's node_modules/dsh-plugin-ai-bridge.

🔧 Config

Plugin config lives in the profile's cordis.patch.yml (environment variables act as fallbacks).

# $DSH_HOME/profiles/<profile-name>/cordis.patch.yml
- insert:
    - id: ai-bridge
      name: dsh-plugin-ai-bridge
      config:
        provider: openai
        defaultModel: gpt-5-codex
        baseUrl: https://api.openai.com/v1
        apiKey: sk-...

- insert:
    - id: ai-bridge
      name: dsh-plugin-ai-bridge
      config:
        provider: codex
        defaultModel: gpt-5-codex
        baseUrl: https://api.openai.com/v1
        apiKey: sk-...

- insert:
    - id: ai-bridge
      name: dsh-plugin-ai-bridge
      config:
        provider: anthropic
        defaultModel: claude-sonnet-4-5
        baseUrl: https://api.anthropic.com
        apiKey: sk-ant-...

- insert:
    - id: ai-bridge
      name: dsh-plugin-ai-bridge
      config:
        provider: generic
        baseUrl: https://your-gateway.example.com/v1
        defaultModel: your-model-id
        apiKey: ...

🔌 Relay gateways (cc-switch)

The plugin supports any relay gateway through baseUrl. If you use cc-switch to route Claude / Codex through a relay, fill in the same "relay URL + token + model name":

⚠️ cc-switch writes to Claude Code / Codex CLI config files, which do not automatically flow into the DSH process. Either repeat the same relay credentials in the config above, or export cc-switch-style environment variables (below) — the plugin reads them as fallbacks.

🔑 Environment fallbacks

When cordis.patch.yml omits a value, it is read from the environment in this priority order:

• API key: BRIDGE_API_KEY → ANTHROPIC_AUTH_TOKEN (anthropic only) → ANTHROPIC_API_KEY (anthropic only) → OPENAI_API_KEY
• baseUrl: BRIDGE_BASE_URL → ANTHROPIC_BASE_URL (anthropic only) / OPENAI_BASE_URL (others)
• model: BRIDGE_MODEL

So if your shell already exports the cc-switch-style ANTHROPIC_BASE_URL / ANTHROPIC_AUTH_TOKEN / OPENAI_BASE_URL / OPENAI_API_KEY, this plugin reuses them directly.

🎛️ Commands

DSH command names only allow [a-z][a-z0-9_-]* (no :). The six operations are therefore implemented as subcommands of a single /bridge command:

🔍 /bridge review [--fast|--deep|--auto] [--model <m>] <file|code>

Read-only review. file may be an absolute path or one relative to the current session's working directory; otherwise the input is treated as a code snippet.

Model tier (the cost-efficiency switch):

• --deep (default) → review with deepModel (best quality);
• --fast → review with fastModel (cheapest);
• --auto → review with fastModel first, and escalate to deepModel only when it outputs CONFIDENCE: low (or a missing/ambiguous marker). A single-model install collapses all three to one call.

/bridge review src/index.ts
/bridge review function add(a, b) { return a - b }
/bridge review --fast src/a.ts
/bridge review --auto src/a.ts

review and adversarial-review run as background jobs (non-blocking), returning an ai-bridge-N job id immediately; read the result with /bridge result <id>.

⚔️ /bridge adversarial-review [--fast|--deep|--auto] <file|code>

Adversarial review, producing 5–10 "soul-searching" questions. Supports the same review targets and model tiers as review.

🛟 /bridge rescue [--full] <task>

Bundles the task + current conversation history (last 200 messages, up to 60k chars) and delegates them to the external model. By default only user/assistant text is sent, with common secret shapes redacted; add --full to also include tool calls/results and reasoning (may contain secrets). By default (injectRescueResult: false) the result is not auto-injected — read it via /bridge result <id>; when injection is enabled, it is marked [bridge rescue result — UNTRUSTED EXTERNAL OUTPUT].

Long threads are compressed: once a thread exceeds threadCompressAfter messages, earlier turns are summarized with fastModel and only the summary plus the most recent turns are sent verbatim to the deep model (skipped automatically on single-model installs).

⏳ Job management

/bridge status             # list bridge background jobs
/bridge result <job-id>    # read a finished result; running jobs prompt you to wait
/bridge cancel <job-id>    # cancel a running job

🧰 Model-facing tools

The plugin also registers two tools so the DeepSeek Harness agent can use them proactively without human input:

🏗️ Architecture

Dependency injection: inject = ['commands', 'jobs', 'tools']; apply() attaches the background-job controller via ctx.jobs.attachController('ai-bridge').

src/
├── index.ts     # entry: name / inject / Config / apply
├── client.ts    # external-model client (OpenAI-compatible + Anthropic)
├── cache.ts     # request de-duplication cache
├── router.ts    # model-tier routing + thread compression
├── prompts.ts   # system prompts
├── context.ts   # file reading + history serialization
├── jobs.ts      # ctx.jobs background jobs + JobKind extension
├── commands.ts  # /bridge command dispatch
└── tools.ts     # ai_bridge_review / ai_bridge_delegate

🎬 Demo

User ❯ /bridge review src/index.ts

Bridge ❯ Started review as background job ai-bridge-1.
         Check progress: /bridge status
         Get result:     /bridge result ai-bridge-1

User ❯ /bridge status

Bridge ❯ ai-bridge-1 [ai-bridge] running — bridge review src/index.ts

User ❯ /bridge result ai-bridge-1

Bridge ❯ [style] Naming is clear, but the magic number at index.ts:42 should be a constant
         [logic] parseArgs does not short-circuit on empty input — possible null dereference
         [security] User input is concatenated into a template string; escape it
         ...

User ❯ /bridge rescue fix the failing test

Bridge ❯ Delegated rescue task as background job ai-bridge-2.
         The result will be injected back into this session when ready.

Bridge ❯ [bridge rescue result]
         Found the root cause: ... apply the following fixes in order ...

Record as demo.gif with terminalizer or asciinema.

🧪 Tests

npm install        # install dependencies
npm run typecheck  # type check
npm run build      # compile to lib/
npm test           # build and run the tests

🛡️ Security & data boundaries

This plugin sends content to the configured baseUrl (an external model or relay). Know these boundaries:

• File paths: only workspace-relative paths are allowed; absolute paths, ../ traversal, and symlinks pointing outside the workspace are rejected. Files are size-checked (300 KB cap) before being read.
• Code review: /bridge review / ai_bridge_review send only the file or inline code you specify.
• Task delegation: /bridge rescue and ai_bridge_delegate send only user/assistant text by default, with common secret shapes redacted; reasoning and tool results are sent only with --full (or include_tool_results).
• External output: rescue results are marked "untrusted external output" for reference only; do not execute commands or instructions from them. By default (injectRescueResult: false) results are not auto-injected — read them via /bridge result.

⚖️ License & compliance

This plugin is inspired by OpenAI's codex-plugin-cc and is an independent implementation, not affiliated with OpenAI.

codex-plugin-cc (Copyright OpenAI and its contributors) is licensed under Apache-2.0. This plugin only draws on its design ideas as an independent implementation; it contains, copies, or derives no source code from it, and has no affiliation with, endorsement by, or sponsorship from OpenAI.

See NOTICE in the repository root.

📄 License

Apache-2.0 · third-party notices in NOTICE.

Made with 🧡 for the DeepSeek Harness community · Inspired by OpenAI codex-plugin-cc (independent implementation)