Pull context skill

Pull context from another session

Retrieval-on-demand: source sessions need no prep — their transcripts already live on disk at ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl. This skill finds the right one, distills it in a subagent (so the large transcript never enters the current session), and drops the digest into this conversation.

Arguments (free-form, all optional)

• A search term → pre-filter the list (e.g. /pull-context alerting).
• --all → search across every project, not just the current one.
• verbatim → pull the literal recent messages instead of a distilled summary.

Steps

1. List candidates. Run the lister, passing through any search term / --all:

   python3 ~/.claude/skills/pull-context/list_sessions.py [term] [--all]
   

   It prints rows of # | age | msgs | topic plus each session's transcript path. If the current project has no match it auto-falls back to all projects and prints a (... showing ALL projects) notice with [project] tags on each row — so you usually don't need to pass --all yourself.

2. Pick the source session.

   • If the user already named the topic clearly and exactly one row matches, use it.
   • Otherwise show the list and ask which number they want.
   • Never pick the current session. Identify it by matching the first user message in this conversation against the topic hints, and exclude it.

3. Distill (default) — do this in a subagent so the transcript stays out of the main context. Spawn an Explore (or general-purpose) agent with the chosen transcript path and this instruction:

   Read this Claude Code session transcript (JSONL, one event per line; user and assistant turns are under .message.content): <PATH>. Produce a tight digest, no preamble, in this exact structure:

   • Goal — what that session was trying to accomplish.
   • Current state — where it actually ended up (done / in progress / blocked).
   • Decisions & why — choices made and the reasoning, including rejected options.
   • Key pointers — files, paths, PRs, IDs, commands, dashboards referenced.
   • Open questions / next steps — what was unresolved or queued.

   Favor specifics (exact file paths, identifiers, values) over narration. Omit a section only if the transcript genuinely has nothing for it.

   For verbatim mode, skip the subagent and run the extractor to pull the last ~15 user/assistant turns (exact text, code, and tool commands preserved):

   python3 ~/.claude/skills/pull-context/extract_tail.py <PATH> [N]
   

   (Heavier and noisier than the digest — use only when the user needs literal wording or code from the other session.)

4. Inject the result into this conversation inside a clearly labeled block so it's obvious where it came from and that it's reference, not new instruction:

   --- Pulled context: "<topic>" (session <short-id>, <age>) ---
   <digest or verbatim slice>
   --- end pulled context ---
   

   Then give a one-line orientation ("That session left off at X; here's how it connects to what we're doing") and ask how they want to use it.

Notes

• The lister marks the current session as (THIS SESSION — skip) when CLAUDE_SESSION_ID is set; if it isn't, fall back to topic-matching (step 2).
• If the user wants context from several sessions, distill each in its own subagent and inject them as separate labeled blocks.
• Distilling keeps this session's context budget clean — prefer it unless the user explicitly needs exact wording or code (verbatim).