Deepseek Bridge — Development skill for Claude Code

deepseek-bridge

Use DeepSeek as the implementer inside Claude Code, while Claude (Opus, Sonnet, whatever you run) stays the orchestrator on your own subscription.

It adds one tool, `mcp__deepseek-bridge__implement`. Claude plans and reviews; when it has a well-scoped piece of implementation work it hands it to a DeepSeek agent that runs through the [opencode](https://opencode.ai) CLI (`opencode run -m deepseek/deepseek-flash`) and edits your working directory with its own tools. It is not a model entry in `/model`: Claude Code has one provider per session, so a tool is how a second model joins.

What it gives you

• Full context. The first call briefs DeepSeek with the task, your global and project CLAUDE.md, the git state and the conversation so far. Later calls in the same Claude session continue the same opencode session and send only what happened since.
• Long, agentic runs. Tested with 4 to 24 minute calls; the default limit is 60 minutes, configurable. Progress shows in the status line and in a log that updates as it goes. caffeinate keeps a Mac awake for the length of a run.
• A goal loop that does not take the model's word for it. Pass acceptance_cmd (a command that exits 0 when the work is done, such as python3 -m pytest -q). The bridge runs it itself after each run and sends DeepSeek back with the failing output until it passes. With only acceptance (criteria in words) it sends the model back until it says DONE: or BLOCKED:, and the result says that this was self-reported.
• Output that survives a lost pipe. opencode's output goes to a file the bridge reads as it grows. If nothing arrives, or the run ends with nothing received, the report, actions and cost are rebuilt from opencode's own database, so a finished run is never reported as stalled. A silent run is looked up in the database before it is killed.
• Resumable. A run that hit the time limit, stalled or was recovered keeps its session; the next call continue