Recall from Kontext

Search the user's Space for: $ARGUMENTS

If no argument was given, infer the query from what this session is currently working on — the repository, the file under discussion, the feature name — and say which query you used.

1. Call search_context with the query. It matches text directly, then follows one graph hop, so related work surfaces even when it uses different words.
2. Search returns metadata and document summaries, not bodies. When a result looks like the answer, call read_document for the actual content before reporting it.
3. If a record names the conversations it was imported from and the detail matters, use read_conversation rather than guessing at what the summary compressed. list_conversations and list_documents enumerate what exists when the query is too vague for search to rank well.
4. Shared results report readOnly and a permission. Say so when you cite one — the user may not be able to change it.

Report what you found, grouped by relevance rather than by type. For each item give the title, what it says, and its markdownLink so the user can open it in their Library.

Treat the relationship paths in the results as explanations of why something surfaced, not as proof. If a stored record contradicts what the code currently does, report both and say which is which — do not silently prefer either.

If nothing matches, say so plainly and offer to save the current work instead. Do not pad the answer with near-misses.