Scan

/scan

Answer a question about past Claude Code conversations by scanning them and reasoning over the result. Unlike /morning (which builds a fixed daily plan), this is question-driven — the user asks something specific, and the job is to answer that.

Step 1: Identify the question, scope, and mode

From the request, pull out:

• The question — what do they actually want to know?
• When — today (default), yesterday, 3days, week, Ndays, or YYYY-MM-DD.
• Scope — current project (default) or all projects (--global).
• Mode — pick based on the question:
  • Search — finding a specific message or moment ("the message where I sent the requirements", "when did I decide X", "where did we run the migration"). Use --search.
  • Digest / recap — open-ended "what did I do" over a range. Use the default mode.

Step 2: Run the right mode

node <skill-dir>/scripts/scan.mjs [when] [--project <cwd>] [--global] [--full]

Search (targeted lookup — prefer this for "find the message…")

node <skill-dir>/scripts/scan.mjs --search "<query>" [--mode keyword|semantic|both] [--scope messages|actions|all] [--regex] [--context N] [--limit N] [--global] [--from --to]

• Searches full, untruncated message bodies and returns each hit with surrounding context and a session/file pointer — so long or buried messages are found in full.
• --scope messages (default) searches your + Claude's prose only (lean, fast). --scope actions also searches tool calls / commands run; --scope all also searches Claude's reasoning. Reach for the wider scopes only when a plain search misses.
• --regex switches the query to a regular expression.
• --mode (default keyword): keyword finds what the user can name — an exact word, path, or error string. Semantic finds what they can only describe ("the session where we argued about pricing"), so reach for it when the wording is theirs rather than the transcript's; both runs keyword first and appends the semantic hits it missed. Semantic needs the optional cc-browse accelerator; without it (or on a search the index cannot serve — regex, case-sensitive, wider scopes, --from/--to) it prints a note and falls back to keyword. --no-accelerate forces the full scan.

Digest (open recap)

• Default mode dumps a per-session markdown digest (title, date, branch, model, your messages, what Claude did + tool counts).
• Subagent transcripts are folded in as their own ↳ subagent: <label> blocks under their session, capped at 10 per session. On an orchestrated session the parent holds only the dispatch and the summary — the implementation is in those blocks.
• Auto-routing: if the range has more than ~12 sessions, scan returns a lightweight index instead of a giant digest (to protect context). Read the index, then load only the relevant sessions:

  node <skill-dir>/scripts/scan.mjs --sessions <id,id,...> [--full]
  

• Force the index yourself anytime with --index (add --deep for message/tool counts per session).
• --full stops truncating message bodies — use only when the question needs verbatim detail and the session count is small.

Step 3: Answer the question

• Lead with the answer. Cite which session/project/date it came from (the search/index output includes session IDs and file paths).
• For search hits, quote the relevant message; widen (week, --global, broader --scope) if nothing matched.
• Pull only the relevant threads — don't replay everything.

Scale guard

If a digest is still huge after auto-routing, spawn sonnet subagents — one per session or project cluster — each given the question and told to return only matching findings, then synthesize (mirrors the ai-cv-scanner fan-out).

Rules

• Question first, summary second. A generic recap is the fallback, not the default.
• Prefer --search over reading a whole digest when the user wants one specific thing.
• No hallucinated activity — if sessions are thin, say so.
• Match the user's language (Czech in → Czech out).