fix(prompt): stop model from emitting slash commands as replies

[quiet-node]

Overview

Hardens the built-in system prompt so the local model never answers by emitting a slash command (e.g. /search "...") as its reply. Reported against gpt-oss 20B asking for a current net worth: the model replied with the literal text /search "Elon Musk net worth", nothing ran, and every follow-up re-emitted the same command in a loop.

Why

Slash commands in Thuki are user-input only: parseCommands runs solely inside handleSubmit (src/App.tsx), triggered by a user submit. A model-emitted /search ... never passes through it, so it is inert text rendered as a chat bubble. There is no agentic tool-call loop, so the model cannot self-invoke a command.

The prompt advertised /search and instructed the model to "tell the user to run /search". gpt-oss 20B, heavily tool-call-trained, misreads that as "invoke the tool by emitting it" (its reasoning trace literally plans to "reply with a slash command"). With no result coming back, each user nudge makes it re-reason and re-emit, producing the loop. Not a code bug: a prompt/model mismatch.

What changed

src-tauri/prompts/system_prompt.txt:

• Acting vs Asking: new rule stating slash commands are run only by the user, the model cannot run them, and it must never emit one or pretend to invoke one. When a command would help, it tells the user in its own words to run it, then answers with whatever it can already provide.
• Expertise and Limits: the /search guidance now asks the model to tell the user to run /search in its own words and vary the wording instead of repeating a fixed phrase.

Model-agnostic; targets the exact misread. Propagates to all non-customized installs via the existing refresh-on-load of the default prompt.

Testing

bun run test:backend passes. The change is prompt text embedded via include_str! with no Rust/TS logic added, so frontend coverage, lint, and typecheck are unaffected.

Note

Two optional follow-ups were considered and not included: a UI affordance that turns a model-emitted /search ... into a one-click run chip, and a full parse-execute-feedback tool-call loop. Both are larger and out of scope for this fix.