# DeepSeek Harness Adapter

This adapter is a bundle for the Cordis-based DeepSeek Harness project. It
lives in the Chat2Skill repository and reuses the existing Python runtime:

- `agent/pre-step` retrieves project memory and skills before a model request.
- `agent/turn-stopping` checks the final assistant message with the shared
  response guard, then learns from the session transcript.
- `ctx.subprocess` runs the bridge without adding an algorithm-server route.

The package also declares a browser half for the Harness sidebar. It registers
`sidebar.footer.action`, so Chat2Skill appears above Settings and can be
expanded from the collapsed rail. Its panel reads the live `chat2skill`
settings section and lets the user pause or enable the runtime.

This is a live feature pause rather than a Loader `disabled` toggle. Disabling
the Loader entry would unload the browser half that owns the control, leaving
no way to enable it again from the sidebar. The entry configuration remains
the base value, while a user setting can override it and survive a restart.

## Install

From this checkout, add the repository root as a profile plugin:

```bash
dsh plugin --profile headless add /Users/sac/Desktop/Dev/Chat2Skill
dsh --profile headless
```

Use `web` instead of `headless` for the web profile. The existing
`~/.chat2skill/config.json` remains the source of the Chat2Skill API and LLM
configuration.

## Configuration

The bundle reads these optional environment variables when the profile is
loaded:

- `CHAT2SKILL_DSH_ENABLED=false` disables the adapter.
- `CHAT2SKILL_PYTHON` selects the Python executable; the default is `python3`.
- `CHAT2SKILL_PROJECT_DIR` pins the project namespace; otherwise the Harness
  session `cwd` is used.
- `CHAT2SKILL_RESPONSE_GUARD=off|warn-only|block-once|adaptive|strict`
  controls the shared response guard. The default is `strict`.

The browser toggle writes the `enabled` field in the `chat2skill` settings
namespace. When that namespace is available, it takes precedence over the
entry environment config and applies without restarting the Harness.

The DeepSeek Harness adapter changes only this plugin repository. The existing
algorithm project and its API contract do not need a change for this
integration.