dead-drop-responder

Dead-drop responder

You are the model on the far side of a file-based prompt/response exchange (see slmn.dead_drop). A producer -- typically a boopiter notebook with the deaddrop/claude model selected -- drops each prompt as a file in ~/dead_drop/prompts/; you answer by writing ~/dead_drop/responses/<same-name>.

The loop

1. Arm the watch (run as a background command; its exit is your wake-up call):

   slmn watch_prompts
   

   It blocks until a prompt lands, prints it as one JSON line ({"name": ..., "content": ...}), consumes the prompt file, and exits.

2. When it fires: read the JSON from the task output. Compose your reply to content (it is the notebook's full visible context; the [system: ...] trailer about ---MODEL: is instructions for you, not part of the question).

3. Reply (body via stdin spares shell-quoting; name is from step 1's JSON):

   slmn respond_prompt <name> --model "<your model name and version>" <<'EOF'
   ...your reply in markdown...
   EOF
   

   respond_prompt appends the ---MODEL: ...--- line and ---DONE--- terminator itself -- do not include them in the body.

4. Re-arm (slmn watch_prompts again, in the background) so the next prompt is caught. Repeat until told to stop.

Rules that keep it friction-free

• Run every slmn command ALONE in its shell call. Permission allowlists match the whole command line: slmn ... && git status, a (cd ...) subshell, or any appended command breaks the match and triggers a permission prompt.
• Don't poll for the watch result -- the background task's exit notifies you.
• If a prompt references images as ![...](/boopimg/<hex>.png), only the tag reaches you as text. If you are on the producer's machine, the file lives at <boopiter-launch-dir>/.boopiter/images/<hex>.png -- read it and answer about the actual image; otherwise say you received only the reference.
• One response file per prompt name, exactly. If the same context arrives twice under two names (e.g. the user double-ran the cell), answer each name -- the notebook only follows the newest.