Transcript Grep
Find where a pattern appears in a session transcript and show the matches in context.
Input
The user provides: $ARGUMENTS
Interpreted as a session reference plus a search pattern, e.g.
<session-id> "rate limit" or latest TypeError. Parsing rules:
- The session reference is the first token if it looks like an id, or the words
latest/last(most recently updated session). - The remainder is the search pattern (string or regex, quoted if it contains spaces).
- If no session is given, default to the most recent session.
Data Sources
| Endpoint | Returns |
|---|---|
GET /api/sessions?limit=N | session list to resolve latest/last and to confirm the id exists |
GET /api/sessions/:id/transcript | the ordered transcript messages (role, content, tool calls/results, timestamps) for the session |
Report Sections
1. Resolve the session
If latest/last (or no id), call GET /api/sessions?limit=1. Otherwise verify
the id with GET /api/sessions?limit=1000 (or GET /api/sessions/:id). Report
the resolved id, status, and model before searching.
2. Fetch and search
Call GET /api/sessions/:id/transcript. Walk the messages in order and match the
pattern against message text, tool_name, and tool input/output content.
Case-insensitive by default; treat the pattern as a regex if it contains regex
metacharacters, otherwise as a literal substring.
3. Matches with context
For each match show:
[#N HH:MM:SS role(:tool_name)] … preceding line of context … > matching line with the **pattern** emphasized … following line of context …
Number matches sequentially. Include ±1–2 messages (or lines) of context so the match is interpretable. If a tool call matches, show the tool_name and a trimmed view of its arguments/result.
4. Summary
Report: total matches, how many distinct messages matched, the roles involved (user / assistant / tool), and the timestamp span of the matches. If there are zero matches, say so plainly and suggest a looser pattern.
Output
- Lead with the match count and session header, then the contextual snippets.
- Keep snippets trimmed — truncate long tool payloads with
…rather than dumping them. - Cite only transcript content returned by the API — never fabricate messages.
- If the dashboard is unreachable, tell the user to start it with
npm startfrom the repo root.
