# Tma1 Search

> Search past agent sessions on this project by keyword, and read back any session's conversation. Invoke when the answer lives in an earlier session — how something was fixed before, what was decided, what an agent tried last week — or when the user hands you a session id. Trigger phrases: "what did we decide about", "how did we fix this before", "search my sessions", "find that session", "read session <id>", "之前是怎么解决的", "上次讨论过", "找一下以前的 session", "/tma1-search".

- Skill: `tma1-ai/tma1-search` (Agent Skill)
- Install (CLI): `npx skillmds@latest add tma1-ai/tma1-search`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tma1-ai/tma1-search/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: tma1-ai (https://skillmd.com/u/tma1-ai)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/tma1-ai/tma1-search

---


# TMA1 Session Search

TMA1 keeps every agent session on this machine — yours and other agents' —
in a local database. This skill is the read path: find the session, then
read it.

## Two entry points

```
/tma1-search retry backoff            # keyword search, this project, last 7 days
/tma1-search flaky test --agent codex # only Codex's sessions
/tma1-search migration --days 30      # widen the window
/tma1-search pipeline --all-projects  # every project, not just this one
/tma1-search --session 3d8f1d0a       # read one session directly
```

`--session` takes a full id or a prefix of 6+ characters — the abbreviated
id in the `<tma1-context>` block works. Do **not** infer a session id from
a bare token: commit hashes look the same, and `oc:` / `cp:` ids don't look
like hex at all. Anything that isn't behind `--session` is search text.

## Workflow

1. **Search** — `mcp__tma1__search_sessions` with:
   - `query`: the keyword text (everything that isn't a flag)
   - `agent_source`: `--agent` value if given (`codex` / `openclaw` /
     `copilot` / `claude` / `self`); omit for every agent
   - `since_min`: `--days N` × 1440, default omitted (7 days)
   - `project`: `"*"` when `--all-projects` is given; otherwise omit
2. **Report** the matching sessions with their snippets, so the user can
   pick. One obvious hit and a clear question → go straight to step 3.
3. **Read** — `mcp__tma1__get_session_transcript` with the `session_id`
   from step 1. `has_more: true` means older messages exist; call again
   with the returned `next_offset` if the answer isn't in this page.
4. **Answer with quotes.** The value here is the earlier session's actual
   words, not your summary of them.

## Reading the search response

- `sessions[]` — `session_id`, `agent_source`, `last_activity_ago`,
  `hit_count`, and `snippets[]` (each a window around a match).
- `match_mode` — `term` is the indexed pass: whole words, case-sensitive.
  `substring` means that pass found nothing, so the search widened to a
  case-folded substring scan; results may include noise, say so.
- `scope_truncated: true` — this project has more sessions in the window
  than the search scanned. Narrow with `--agent` or `--days`.
- `note` — present only when nothing matched. Say so plainly; don't invent
  a plausible past session.

## Search behaviour worth knowing

The indexed pass matches whole words and is case-sensitive, so `Cwd` does
not find `peerCwdFilter` and neither does `peercwdfilter`. Search for the
identifier as it is written in the code. When that pass finds nothing the
search widens to a case-folded substring scan and tells you so with
`match_mode: substring` — that is the mode that rescues a mistyped case,
at the cost of noisier hits.

## Related

- Browsing by agent rather than keyword → `/tma1-peer`.
- Aggregates over the raw tables (cost by model, tool failure rates) →
  the `tma1` skill's `exec_query`.