| name | search |
| description | LogosDB — concise semantic search over indexed memory (optional time window). |
/search
User invoked with: $ARGUMENTS
Instructions
Search LogosDB for semantically similar content and respond concisely (file paths + scores only — never echo or paraphrase the full chunk text in the slash-command reply).
Parse the arguments:
- Everything before any flag is the search query.
--namespace=<name> or -n <name> sets the collection (default: code).
--top-k=<n> or -k <n> sets the number of results (default: 5, hard cap at 8 for slash use — keep output small).
- Optional ISO 8601 timestamp window (inclusive), same as MCP
ts_from / ts_to:
--from-ts=<iso> or --ts-from=<iso> → pass as ts_from
--to-ts=<iso> or --ts-to=<iso> → pass as ts_to
- Optional
--candidate-k=<n> → pass as candidate_k when using a timestamp window (default: 10 × top_k).
Call the logosdb_search MCP tool with the query, namespace, top_k, and (when bounds are set) ts_from, ts_to, and candidate_k if provided.
Output format (strict)
Respond with only the following block. Do not wrap it in fences, do not include the raw tool JSON, and do not paste chunk text.
Searching '{namespace}' (top_k={top_k}{ts_note}): {N} matches
1. {file_path} (score: {score})
2. {file_path} (score: {score})
...
{file_path} is extracted from the [file:...] prefix in the result text.
{ts_note} is , ts=[{ts_from}..{ts_to}] when either bound is set, else empty.
- If no results:
No matches found in '{namespace}' namespace. — single line, nothing else.
Quiet mode for conversational use
When the agent invokes logosdb_search outside of /search (i.e. as part of answering a normal question), it should still:
- Keep
top_k 3–5.
- Cite source files briefly (e.g.
src/foo.ts) in the final prose.
- Not quote the full chunk text returned by the tool.
Related
skills/semantic-memory/SKILL.md