Prompt assets

Prompt templates live alongside schema definitions so the assistant behaviour stays easy to review and update.

Layout

src/prompts/: text templates used by the planner, executor, intent classifier, and response helpers (for example, planner.md and executor.md).
src/lib/llm/templates.sh: helper functions that load prompt files, substitute runtime variables, and pass the final strings to llama.cpp.
src/schemas/: JSON schemas that constrain planner output and tool arguments.

Edit templates directly to adjust tone or required fields. Keep schema changes in sync to avoid model errors.
Store reusable snippets (such as safety disclaimers) in dedicated files and compose them within the main prompt templates.
Keep prompts minimal and version-controlled; avoid inlining large instructions in code so they remain discoverable for audits.
Oversized context (for example, verbose web_fetch results) is summarized automatically before llama.cpp is invoked so prompts stay within the configured token budget.
The planner prompt includes an intent context block plus a search_context block derived from a deterministic pre-plan web search. When llama.cpp is available, a rephraser model generates 1–3 queries; otherwise the raw user query is used. If search is disabled or fails, the block may be empty.
The search context formatter emits a default “No search results were captured for this query.” message when the pre-plan search yields no items.
Planner instructions require tool arguments to stay concise (single-line string fields, no code fences, and short summaries) to prevent oversized or unsafe payloads from reaching tool handlers.