dsh-cli-bridge

🟢  ━━  🔵  ━━  🟣

the three pillars, bridged

╭───────────────────────────────────────────────────────╮
│                                                       │
│               dsh ⇄ claude code · codex               │
│                                                       │
│    Run Claude Code & Codex inside DeepSeek Harness    │
│               — watch every step, live                │
│                                                       │
│            live · autonomous · zero setup             │
│                                                       │
╰───────────────────────────────────────────────────────╯

dsh-cli-bridge

The built-in delegates hand you the final result. This one hands you the whole run.

What it does

One tool call from DeepSeek Harness. A real agent doing real work — on screen, in real time, under your control.

dsh-cli-bridge delegates work to Claude Code and Codex, then keeps you and DeepSeek Harness in the loop for the entire run:

📺 Watch it live	🧠 Give DSH the wheel	🛠️ Nothing to set up
Every tool call, every file changed, every line of reasoning streams into DSH as it happens — not a spinner, not a final paste.	DeepSeek Harness can steer, answer, and review the delegate's work on its own, round after round — or hand you the question when it should not decide.	The plugin finds, installs, and updates Claude Code and Codex itself. You install one plugin; it manages the rest.

[!IMPORTANT] Why this exists — the built-in claude and codex delegates call the CLI like a tool and return only the final message. The whole run happens in the dark, and the transcript costs you tokens twice. dsh-cli-bridge puts the run on screen, keeps the transcript out of your model's context, and gives the harness the controls.

The difference

A built-in delegate call	✨ dsh-cli-bridge
❌ hands back only the final result	✅ the whole run, streamed live into DSH
❌ DeepSeek just reads the answer	✅ DeepSeek steers, answers, and reviews the delegate
❌ you install and update the CLIs yourself	✅ the plugin installs and updates them
❌ one login, configured by hand	✅ many isolated accounts, side by side
❌ tasks share one workspace	✅ a git worktree per task, merged back automatically
❌ Claude Code talks to Anthropic only	✅ Claude Code + Codex, and any Anthropic-compatible endpoint

And the rest

The headline is the live run. The rest is what makes it safe to let an agent loose on a repository:

• 👥 Multi-account, isolated — every account is a directory; each login lives in its own home and never touches another.
• 🌐 Any endpoint — Claude Code can reach DeepSeek, OpenRouter, or any Anthropic-compatible provider, by base URL + token.
• 🌿 Worktrees & automatic merge — parallel tasks each get their own branch; finished work merges back --no-ff, one at a time, conflicts kept for you.
• 🧭 A deterministic hand-back — the delegate states when it needs a decision, and the protocol is fixed rather than guessed from prose.
• 🔒 Inherited permissions — the harness's own Read Only / Workspace Write / Full Access mode is the whole policy; the plugin invents none.
• ⚙️ Model & effort per call — each delegation names its own model and effort, or inherits sensible defaults.

Install

dsh plugin --profile web add dsh-cli-bridge

Support

dsh-cli-bridge is free to use and funded by the people it saves money. If it keeps delegate spend off your model bill, sponsor it.

• Use it freely — personally or commercially, with your own modifications.
• Sponsor to redistribute — publishing, forking, or offering it as a service is the one paid gate (a sponsorship or a commercial licence).
• Contribute — the project stays source-available while its author can still license it commercially. See LICENSE and CONTRIBUTING.

Licence

Source-available — © 2026 hviana. Free to use, funded by sponsors, redistribution gated. The inlined open-source components keep their own licences; see THIRD-PARTY-NOTICES.md.

Source-available  ·  © 2026 hviana  ·  sponsor  ·  third-party notices
built for DeepSeek Harness