Claude History Ingest

Claude History Ingest — Conversation Mining

You are extracting knowledge from the user's past Claude Code conversations and distilling it into the Obsidian wiki. Conversations are rich but messy — your job is to find the signal and compile it.

This skill can be invoked directly or via the wiki-history-ingest router (/wiki-history-ingest claude).

Before You Start

Writing profile: Before drafting or rewriting natural-language Markdown, read and apply the Writing Profile Resolution section in llm-wiki/SKILL.md. Framework schema, provenance, safety, and operation-specific requirements take precedence. WRITING.md preferences apply only to newly drafted or rewritten natural-language Markdown; preserve source content and structured records.

1. Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → global config → prompt setup). This gives OBSIDIAN_VAULT_PATH and CLAUDE_HISTORY_PATH (defaults to ~/.claude)
2. Read .manifest.json at the vault root to check what's already been ingested
3. Read index.md at the vault root to know what the wiki already contains
4. Project Scoping — read WIKI_SKIP_PROJECTS from config (comma-separated substrings). Exclude any project directory whose name contains one of them from every step below (scan, delta, sampling, manifest writes). If the user names extra projects to skip this run, add them. Apply the exclusion once, uniformly — don't hand-write grep -v filters into individual commands, which drifts between the scan and manifest steps.

Ingest Modes

Append Mode (default)

Check .manifest.json for each source file (conversation JSONL, memory file). Only process:

• Files not in the manifest (new conversations, new memory files, new projects)
• Files whose modification time is newer than their ingested_at in the manifest

This is usually what you want — the user ran a few new sessions and wants to capture the delta.

Portable keys when comparing. Manifest keys follow the source key contract (v2) in llm-wiki/SKILL.md → .manifest.json. A session under $HOME is keyed ~-relative (~/.claude/projects/.../abc.jsonl), never by the expanded machine path; vault-relative and pseudo-keys are the other two forms. Before deciding a file is "new", resolve the stored key the same way the tool does (expand ~/env vars, resolve vault-relative against the vault root) — otherwise an already-tracked file looks new and gets re-ingested. The scripts/manifest.py helper does this for you:

# New/modified sources, honoring WIKI_SKIP_PROJECTS + --skip:
python3 "$OBSIDIAN_WIKI_REPO/scripts/manifest.py" delta "$OBSIDIAN_VAULT_PATH" \
  --scan "$CLAUDE_HISTORY_PATH/projects/*/memory/*.md"
# One-time repair if the manifest still holds legacy absolute keys:
python3 "$OBSIDIAN_WIKI_REPO/scripts/manifest.py" migrate "$OBSIDIAN_VAULT_PATH" --dry-run

The helper is optional — if it's unavailable, apply the same resolution inline before every manifest lookup and write.

Pre-extraction (recommended — run before ingest)

Raw JSONL files are 80-90% noise: tool_use blocks, thinking blocks, progress events, and file-history-snapshot entries dominate by byte count. The scripts/extract-jsonl.py helper strips all of that and writes compact signal-only JSON to ~/.claude/extracted/, achieving 50–200× file-size reduction (e.g. 12 MB JSONL → 64 KB extracted). This lets the skill read 5–10× more conversations per run within the same token budget.

Run it as a pre-step before invoking this skill:

# First run — extract everything (skip excluded projects)
python3 "$OBSIDIAN_WIKI_REPO/scripts/extract-jsonl.py" --skip tsg,autom8

# Incremental — only sessions modified in the last day
python3 "$OBSIDIAN_WIKI_REPO/scripts/extract-jsonl.py" \
    --since "$(date -v-1d +%Y-%m-%d)" --skip tsg,autom8

Extracted files live at ~/.claude/extracted/<project-dir>/<session-id>.json and contain:

{
  "session_id": "uuid",
  "project": "-Users-name-myapp",
  "cwd": "/Users/name/myapp",
  "start_ts": "...",
  "end_ts": "...",
  "n_turns": 18,
  "n_user_words": 620,
  "turns": [
    {"role": "user",      "text": "..."},
    {"role": "assistant", "text": "..."}
  ]
}

When Step 3 reads conversations, always prefer the extracted file over the raw JSONL. (See Step 3.)

If extract-jsonl.py was not run first, fall back to raw JSONL — but note the coverage will be shallower because each raw file costs far more tokens to read.

Conversation Sampling Heuristic

A history path can hold hundreds of conversation JSONLs — do not try to read them all. Per project:

• If the project already has memory files (memory/*.md), ingest those first (they are pre-distilled signal), then also process conversations not yet in the manifest — new conversations should still be captured even for memory-rich projects.
• If the project has no memory files, read only the 3 most recent conversations (by mtime) to characterize it. Prefer pre-extracted files (see above) — they are cheap enough that you can read 5–10 in the same token budget as 1 raw JSONL.
• Always report what you sampled vs skipped (e.g. "agenttower: 7 memory files + 4 new conversations ingested, 14 unchanged conversations skipped"), so the coverage gap is visible rather than silent.

Full Mode

Process everything regardless of manifest. Use after a wiki-rebuild or if the user explicitly asks.

Claude Code Data Layout

Claude Code stores data in two locations. Scan both.

Source 1: ~/.claude/ (CLI sessions)

~/.claude/
├── projects/                          # Per-project directories
│   ├── -Users-name-project-a/         # Path-derived name (slashes → dashes)
│   │   ├── <session-uuid>.jsonl       # Conversation data (JSONL)
│   │   └── memory/                    # Structured memories
│   │       ├── MEMORY.md              # Memory index
│   │       ├── user_*.md              # User profile memories
│   │       ├── feedback_*.md          # Workflow feedback memories
│   │       └── project_*.md           # Project context memories
│   ├── -Users-name-project-b/
│   │   └── ...
├── sessions/                          # Session metadata (JSON)
│   └── <pid>.json                     # {pid, sessionId, cwd, startedAt, kind, entrypoint}
├── history.jsonl                      # Global session history
├── tasks/                             # Subagent task data
├── plans/                             # Saved plans
└── settings.json

Source 2: ~/Library/Application Support/Claude/local-agent-mode-sessions/ (Desktop app agent sessions)

Pre-check first. Many users are CLI-only and have no desktop sessions. Before walking the structure below, confirm it's non-empty:

DESKTOP_SESSIONS="$HOME/Library/Application Support/Claude/local-agent-mode-sessions"
[ -d "$DESKTOP_SESSIONS" ] && find "$DESKTOP_SESSIONS" -name "audit.jsonl" | head -1

If that prints nothing, skip this entire section (Source 2 + Step 3b) and don't narrate it.

The Claude desktop app stores local agent mode sessions here. The structure is deeply nested:

~/Library/Application Support/Claude/local-agent-mode-sessions/
└── <outer-uuid>/
    └── <inner-uuid>/
        ├── local_<session-uuid>.json          # Session metadata
        └── local_<session-uuid>/
            ├── audit.jsonl                    # Audit log — tool calls, file reads, commands run
            └── .claude/
                └── projects/
                    └── <path-encoded-name>/   # Same path-encoding as ~/.claude/projects/
                        └── <uuid>.jsonl       # Conversation transcript (same JSONL format as CLI)

How to find all local-agent-mode sessions:

# Find all session metadata files
find ~/Library/Application\ Support/Claude/local-agent-mode-sessions -name "local_*.json" -maxdepth 4

# Find all audit logs
find ~/Library/Application\ Support/Claude/local-agent-mode-sessions -name "audit.jsonl"

# Find all conversation transcripts
find ~/Library/Application\ Support/Claude/local-agent-mode-sessions -name "*.jsonl" -path "*/.claude/projects/*"

Session metadata (local_<uuid>.json) — JSON file with fields like sessionId, cwd, startedAt, model, title. Read this first to understand the session context before opening the transcript.

Audit log (audit.jsonl) — Each line is a JSON record of one agent action: tool calls (Read, Write, Bash, Edit), file accesses, shell commands executed, MCP calls. Useful for understanding what the agent actually did — often richer signal than the conversation text alone. Fields: type, toolName, input, output, timestamp, sessionId.

Conversation transcript (.claude/projects/.../<uuid>.jsonl) — Identical format to CLI conversation JSONL. Parse the same way as ~/.claude/projects/*/*.jsonl.

Key data sources ranked by value (both locations combined):

1. Memory files (~/.claude/projects/*/memory/*.md) — Pre-distilled, already wiki-friendly. Gold.
2. Conversation JSONL (both ~/.claude/projects/*/*.jsonl and desktop app transcripts) — Full conversation transcripts. Rich but noisy.
3. Audit logs (audit.jsonl in desktop sessions) — Tool-call level record of what was done. Useful for extracting concrete actions, file patterns, and command patterns even when the conversation is sparse.
4. Session metadata (sessions/*.json and local_*.json) — Tells you which project, when, and what CWD.

Step 1: Survey and Compute Delta

Scan both data locations and compare against .manifest.json:

# --- Source 1: CLI sessions (~/.claude) ---
# Find all projects
Glob: ~/.claude/projects/*/

# Find memory files (highest value)
Glob: ~/.claude/projects/*/memory/*.md

# Find conversation JSONL files
Glob: ~/.claude/projects/*/*.jsonl

# --- Source 2: Desktop app local-agent-mode sessions ---
DESKTOP_SESSIONS="$HOME/Library/Application Support/Claude/local-agent-mode-sessions"

# Session metadata
find "$DESKTOP_SESSIONS" -name "local_*.json" -maxdepth 4

# Audit logs
find "$DESKTOP_SESSIONS" -name "audit.jsonl"

# Conversation transcripts
find "$DESKTOP_SESSIONS" -name "*.jsonl" -path "*/.claude/projects/*"

Build a unified inventory and classify each file:

• New — not in manifest → needs ingesting
• Modified — in manifest but file is newer → needs re-ingesting
• Unchanged — in manifest and not modified → skip in append mode

Report to the user: "Found X CLI projects, Y desktop sessions. Memory files: A. Conversations: B. Audit logs: C. Delta: D new, E modified."

Step 2: Ingest Memory Files First

Memory files are already structured with YAML frontmatter:

---
name: memory-name
description: one-line description
type: user|feedback|project|reference
---

Memory content here.

For each memory file:

• Read it and parse the frontmatter
• user type → feeds into an entity page about the user, or concept pages about their domain
• feedback type → feeds into skills pages (workflow patterns, what works, what doesn't)
• project type → feeds into entity pages for the project
• reference type → feeds into reference pages pointing to external resources

The MEMORY.md index file in each project is a quick summary — read it first to decide which individual memory files are worth reading in full.

Step 3: Parse Conversation JSONL

Always check for a pre-extracted file first (see Pre-extraction section above). For each conversation ~/.claude/projects/<proj>/<uuid>.jsonl, look for its counterpart at ~/.claude/extracted/<proj>/<uuid>.json. If found, read that instead — it is already filtered to user + assistant text turns and costs 50–200× fewer tokens than the raw JSONL.

# Resolution order for each session:
1. ~/.claude/extracted/<project>/<session-id>.json   ← prefer (compact, signal-only)
2. ~/.claude/projects/<project>/<session-id>.jsonl   ← fallback (raw, noisy)

Reading a pre-extracted file: it already contains only the turns you need. Iterate turns[].{role, text} directly. The top-level fields (cwd, start_ts, n_user_words, etc.) give you project context without any further parsing.

Reading raw JSONL (fallback): Each line is a JSON object:

{
  "type": "user|assistant|progress|file-history-snapshot",
  "message": {
    "role": "user|assistant",
    "content": "text string"
  },
  "uuid": "...",
  "timestamp": "2026-03-15T10:30:00.000Z",
  "sessionId": "...",
  "cwd": "/path/to/project",
  "version": "2.1.59"
}

For assistant messages, content may be an array of content blocks:

{
  "content": [
    {"type": "thinking", "text": "..."},
    {"type": "text", "text": "The actual response..."},
    {"type": "tool_use", "name": "Read", "input": {...}}
  ]
}

• Filter to type: "user" and type: "assistant" entries only
• For assistant entries, extract text blocks (skip thinking and tool_use — those are noise)
• The cwd field tells you which project this conversation belongs to
• Skip type: "progress" — internal agent progress updates
• Skip type: "file-history-snapshot" — file state tracking
• Skip subagent conversations (under subagents/ subdirectories) — unless the user asks