dsh-inner-voice

dsh-inner-voice

Keep the agent's reasoning as a daily inner-voice log, rewrite it into one short line, and show it in a settings page.

A DeepSeek Harness plugin. Server half only reads session events; client half only reads a local HTTP route.

What it does

Part · Behavior A archive · Takes the reasoning of each assistant/message, keeps the first sentence (markdown markers stripped, clipped at maxChars), appends it to ~/.dsh/inner-voice/YYYY-MM-DD.md B rewrite · Sends the same reasoning to a small chat model and keeps one short line (≤20 chars). Pure procedural reasoning yields （无） and is dropped — no fake feelings C panel · Settings → Inner voice: entries by day, rewritten line plus grey raw excerpt, served by the read-only route /dsh-inner-voice-api

Log format:

- 22:12
  小鱼：又来了，这问题问第七遍了。
  raw: Now the user asks whether the plugin is loaded...

Install

dsh plugin --profile <your-profile> add dsh-inner-voice

Restart the app afterwards — server plugins have no HMR.

Config

Set these in your profile's cordis.patch.yml (override by id: dsh-inner-voice):

Key · Default · Notes enabled · true · Master switch minChars · 160 · Skip short reasoning maxChars · 400 · Raw excerpt cap label · inner voice · Prefix of the rewritten line rewrite · true · Turn B off to archive only model · deepseek-chat · Any OpenAI-compatible chat model apiKey · '' · Empty → read DEEPSEEK_API_KEY rewritePrompt · built-in · Override the instruction to fit another persona

Do not commit an API key into the package. Put it in your user-level profile patch or an environment variable.

Privacy

• Everything is written locally under ~/.dsh/inner-voice/.
• If rewrite is on and a key is present, the reasoning excerpt is sent to the configured chat endpoint (OpenAI-compatible) once per recorded turn. Nothing else leaves the machine.
• The panel only reads from the local route; it sends nothing.

Why rewrite at all

An agent's raw reasoning is verbose, often English, and packed with tool names (Installed (linked). Now check if profile patch...). Showing it verbatim is like opening the engine cover: accurate, unreadable.

The rewrite is a readability layer, not decoration. The silence rule guards the other side: when the reasoning carries no attitude, nothing is invented.

Known limits

• The rewrite costs one small chat call per recorded turn (roughly 100–200 tokens).
• The panel cannot restyle the chat transcript itself; it lives in Settings. Rendering reasoning inline would need a client renderer, which this plugin deliberately avoids.
• The log is plain text; no retention policy is enforced. Delete old files if you like.

License

MIT