search · diff
git:20260508.c6a2716 to git:20261001.84b0696
11 added, 15 removed. Audit B to A.
---
name: search
description: Hybrid search across Captain Memo's local memory + skills + observations. Use when the user types /captain-memo:search <query> to retrieve top hits without the model having to decide whether to call search_all on its own.
---
# Captain Memo — direct search
When the user invokes this skill, they want a direct hybrid search against Captain Memo's local index. Don't reason about it — just run it.
## What to do
1. Take the user's argument as the query string (everything after `/captain-memo:search `).
- 2. Run this Bash command and show the user the formatted result:
+ 2. Call the captain-memo MCP tool `search_all` with `query` set to that string and `top_k` set to `5`. Show the user the formatted result.
- ```bash
- curl -s -X POST http://127.0.0.1:39888/search/all \
- -H 'content-type: application/json' \
- -d "$(jq -nc --arg q "<USER_QUERY>" '{query: $q, top_k: 5}')" \
- | jq -r '.results[] | "[\(.score | (. * 100 | floor) / 100)] \(.channel) · \(.title)\n doc_id: \(.doc_id)\n \(.snippet[0:200])\n"'
- ```
+ The MCP server talks to the worker and sends the worker's secret for you: there is nothing to quote and no token to handle.
- Replace `<USER_QUERY>` with the actual query the user provided. Use `jq -nc --arg q ...` to safely quote the query (avoids shell injection on special characters).
+ The result is `{ results: [{ score, channel, title, doc_id, snippet, ... }] }`. Show `score` to two decimals and cut each `snippet` to about 200 characters.
## Format the output as
```
3 hits for "the user's query":
- [0.94] memory · feedback_no_clone_smy_in_cli_smoke
- doc_id: memory:feedback_no_clone_smy_in_cli_smoke:abc1234
- In the multi-tenant ERP at /home/kalin/projects/erp-platform/, do NOT use this pattern in CLI smoke tests…
+ [0.94] memory · feedback_no_mocked_db
+ doc_id: memory:feedback_no_mocked_db:abc1234
+ In this project, do NOT mock the database in integration tests…
- [0.87] memory · feedback_iron_backend_dynamic_promos
- doc_id: memory:feedback_iron_backend_dynamic_promos:def5678
+ [0.87] memory · feedback_use_real_fixtures
+ doc_id: memory:feedback_use_real_fixtures:def5678
…
- [0.82] observation · 2026-04-30 · "Fix off-by-one in pro-ration loop"
+ [0.82] observation · 2026-04-30 · "Fix off-by-one in pagination loop"
doc_id: observation:1700000000:ghi9012
…
```
Tip the user: they can run `/captain-memo:recall <doc_id>` to fetch the full content of any hit.
## On error
- - Worker returns 503 / not reachable: tell the user `captain-memo doctor` to diagnose
+ - The tool errors with 503 or the worker is not reachable: tell the user to run `captain-memo doctor` to diagnose
+ - The tool errors with 401: this session started before the worker began asking for its secret (0.52.0) and sends none. Tell the user to restart the AI session; `captain-memo doctor` lists any caller still sending none
- Empty results: say "no hits — try a more specific query, or check `captain-memo stats` to confirm the corpus is indexed"