Search sessions

Recover what a subject was actually said to be, out of the session transcripts on this machine. The transcript records the conversation; the repo records what got committed, and the two disagree often enough to be worth a skill. Report quotes and where they came from, never a paraphrase of what a decision probably was.

Transcripts sit one JSON object per line at ~/.claude/projects/<slug>/<session-id>.jsonl, with a subagent's own lines under <session-id>/subagents/*.jsonl. The slug is the working directory path with / replaced by -.

Search before deciding something that may already be settled.

Ask agent-record (github.com/cniska/agent-record) first, where installed: it indexes every session and the commits they led to. Run agent-record sync before the first query, so the record holds the latest sessions. If it fails with error.code lock_held, another sync is running: search anyway, say the record may lag one sync, and do not retry. agent-record query search "<words>" finds a message, agent-record query thread <session> --at <at> reads the exchange a hit sits in, with the hit's at copied whole, its Z included, since a time with no zone is refused not_a_time; agent-record query prior-art "<path fragment>" finds where a shape already exists on disk, and agent-record sql "<select>" reads anything else. The workflow below is for when it is missing or finds nothing.

Workflow

1. List every slug for the repo, not just this one. A worktree or a sibling checkout gets a slug of its own, so one project spreads over several — -Users-me-code-apps, -Users-me-code-apps--claude-worktrees-puzzles, and so on. Glob for the repo's name and take them all: searching only the current directory's slug is how a conversation held in another checkout goes missing.
2. Narrow by file first. A line carries whole tool results and file contents, so a transcript runs to megabytes. grep -l -F the subject's terms across the candidate slugs and parse only the files that hit. Honor a stated time window with find -mtime; with none given, start bounded, at the last few days and below the largest files (find -mtime -3 -size -20M), and widen only when that finds nothing. A scan that runs into the command timeout is stopped and narrowed, never waited on.
3. Read messages, skip the machinery. Keep events whose type is user or assistant, and take message.content — a string, or the text parts of a list. Drop tool calls and their results, skill preambles (Base directory for this skill:), briefs pasted in as an opening message, and image placeholders. Search both sides: a decision is often only stated in the reply.
4. Widen once you have a hit. Read the messages around it in that session. The exchange that settled a question is usually three or four messages, and the sentence that changed the answer is rarely the one that matched.
5. Collect what makes a quote actionable: session id, timestamp, who said it, and any commit, file, or path named in the same exchange. A message saying the thing was written down is the thread to the artifact.
6. Check the artifact against the current checkout. A decision committed in another checkout and never pushed, or written to a file on a branch you are not on, is invisible from where you stand. Verify on the current branch before reporting a decision as present or missing, and say which it is.