tma1-ai/tma1/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 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 /

copilot / claude / self); omit for every agent

• since_min: --days N × 1440, default omitted (7 days)
• project: "*" when --all-projects is given; otherwise omit

1. Report the matching sessions with their snippets, so the user can

pick. One obvious hit and a clear question → go straight to step 3.

1. 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.

1. 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.