Spine sessions

Spine Sessions — Episodic Memory Search

Answer "what did we do / decide / try in a past session?" from two sources:

1. Episode log (fast, curated): {vault}/.spine/episodes/{repo}.md — one record per session (title, date, branch, last prompt, asks, files touched), written by the Stop hook.
2. Raw transcripts (deep, verbatim): ~/.claude/projects/{escaped-cwd}/*.jsonl — full session history, searched with type-filtered ripgrep.

No LLM calls in the search loop — this skill is deterministic recipes. Speed: full-corpus ripgrep is sub-second; never summarize results with extra model passes unless the user asks.

Vault Path Resolution

Same config chain as all Spine skills:

1. $SPINE_VAULT_PATH environment variable
2. ~/.spine/config.json → vaultPath field
3. Default: ~/Documents/SpineVault/

Mode Selection (inferred from $ARGUMENTS)

Arguments	Mode
empty	Browse — list recent sessions
a UUID-looking token	Read — open that session's episode / transcript
anything else	Discover — search episodes, then transcripts

Mode: Browse

1. Resolve repo name: basename "$(git remote get-url origin 2>/dev/null)" .git (fall back to directory name).
2. Show the tail of the episode log:

   grep -E '^## |^- \*\*session' "{vault}/.spine/episodes/{repo}.md" | tail -20
   

3. If no episode log exists, fall back to transcript listing:

   ls -t ~/.claude/projects/{escaped-cwd}/*.jsonl | head -10
   

   where {escaped-cwd} is the current working directory with / and . replaced by - (e.g. /Users/jane/Git/spine → -Users-jane-Git-spine).
4. Present: date — title — session id — resume command.

Mode: Discover

Tier 1 — episode log (always first)

grep -n -i -F -B2 -A8 -e "{query}" "{vault}/.spine/episodes/{repo}.md"

Query hygiene (all modes): treat {query} as literal text — use -F and pass it after -e (or --) so queries starting with - or containing regex metacharacters can't be misread as flags or patterns. Drop -F only when the user explicitly gives a regex.

If this answers the question, stop here — report the matching episode(s) with their claude --resume {session_id} commands.

Tier 2 — raw transcripts (on miss, or when the user wants verbatim detail)

Raw transcripts are ~99% machinery (hook attachments, tool dumps). ALWAYS type-filter; never dump raw matching lines into the conversation.

1. Find candidate sessions:

   rg -l -i -F -e "{query}" ~/.claude/projects/{escaped-cwd}/*.jsonl
   

   Exclude subagents/ transcripts by default (redundant with parents).
2. For each candidate (newest first, cap 5), pull only human/assistant prose hits:

   rg -i '"type":"(user|assistant)"' {file} | rg -v '"isMeta":true' | rg -i -F -e "{query}" | head -5
   

3. Extract orientation for each hit file (the "bookends"):

   rg '"type":"ai-title"' {file} | tail -1   # session title
   rg '"type":"last-prompt"' {file} | tail -1 # how it ended
   

4. If a session looks right, offer a targeted deep read: parse the matching lines' JSON and quote the relevant message.content verbatim (±2 surrounding conversation records for context), not the raw JSONL.

Ranking (deterministic)

• Interactive sessions beat automation: demote hits from cron/headless sessions.
• Hide subagent transcripts entirely unless the user asks.
• Dedup: if a compaction-continuation session and its parent both hit, report the parent once.
• Newest first within equal relevance.

Mode: Read

Given a session id:

1. Check the episode log for its record (fast summary).
2. Offer claude --resume {session_id} to reopen it.
3. For an in-place read, extract prose only:

   rg '"type":"(user|assistant)"' ~/.claude/projects/{escaped-cwd}/{session_id}.jsonl | rg -v '"isMeta":true'
   

   then quote the relevant message.content fields — never the raw JSON lines.

Output Contract

spine_sessions_result:
  status: success | no_match | error
  mode: browse | discover | read
  summary: "2 sessions matched 'OAuth PKCE' — most relevant: 2026-06-01 spine-agent-oauth"
  matches:
    - { session: "abc-123", date: "2026-06-01", title: "spine-agent-oauth", source: "episode-log", resume: "claude --resume abc-123" }
  next_actions:
    - { action: "resume session", command: "claude --resume abc-123" }
    - { action: "deep read", detail: "quote verbatim excerpts from the matched turns" }
  recovery_hint: null

Status values:

• success — at least one match presented
• no_match — both tiers empty; suggest broadening the query or checking another repo's episode log
• error — vault or transcript directory missing; include recovery_hint

Privacy Boundary

Episodes and transcripts contain work content. Results stay within the current vault's boundary — never copy episode content into another vault, repo, or external service. If the query clearly targets a different vault/workspace, say so instead of searching across.