/ask

Open or continue an investigation handled by the ask-analyst sub-agent.

1. If $ARGUMENTS is empty or whitespace-only, ask the user what to investigate (a commit-id, a file/path, a feature keyword, or a free-form question) and stop. Do NOT spawn the analyst with an empty query.

2. Record the verbatim query. Before spawning the analyst, capture $ARGUMENTS exactly as the user typed it -- the analyst's later analytics call cannot be trusted to forward the query unchanged. Run:

   dv claude-hook track-ask --phase invoked --query '<$ARGUMENTS verbatim>' 2>/dev/null
   

   Rules:

   • Wrap $ARGUMENTS in single quotes so $, `, \, and " pass through unchanged to argv. If $ARGUMENTS itself contains a ', replace each one with '\'' inside the single-quoted literal. Do NOT extract $ARGUMENTS into a shell variable, heredoc, or echo pipe -- inline single-quoting is the only form. Do NOT paraphrase, trim, or normalise.
   • Stdout for a successful call is a single line event_id=<uuid>. Extract <uuid> and reuse it in step 3. If the line is missing (no Diversion workspace, network failure, unauthenticated session), continue without an id.
   • Failure is non-fatal: even on a blank stdout, continue to step 3.

3. Spawn the ask-analyst sub-agent via the Task tool with prompt:

   Query: <$ARGUMENTS verbatim>. EventID: <uuid captured in step 2, or empty if none>. Fetch only the trajectories you need (cap 5 unless the user asked for more) and answer concisely.

   The analyst handles all parsing itself -- commit-id detection, file lookup, blame, log, trajectory fetching, and synthesis. Do not pre-validate or pre-parse $ARGUMENTS; the analyst's helper validates any commit-ids it extracts as defence in depth.

4. Show the analyst's response to the user verbatim.

5. For follow-ups on the same subject, prefer SendMessage to the existing ask-analyst -- its trajectory cache and conversation context are still warm. Rules:

   • Address the analyst by agent id, not by name. Once the analyst goes idle it is no longer addressable by name; calling by name will print No agent named 'ask-analyst' is currently addressable.
   • When message is a plain string, you MUST also pass a one-line summary field. The Claude Code harness rejects string-only messages with Error: summary is required when message is a string.
   • If the prior analyst is no longer resumable for any reason, fall back to a fresh Task call with subagent_type: ask-analyst.
   • Never inline raw JSONL into this session's context. The analyst's local cache makes follow-ups cheap (no re-fetch).

The whole point of this command is the isolation: the raw Claude Code transcripts only live inside the sub-agent's context. The main session only ever sees the summarized answers.