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

1. Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (inline @name override → walk up CWD for .env → ~/.llm-wiki/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.

Canonical paths when comparing. The manifest keys are absolute paths with ~ expanded (see llm-wiki/SKILL.md → .manifest.json). Before deciding a file is "new", expand its path the same way — otherwise a file already tracked as ~/.claude/... looks new when you scanned it as /Users/me/.claude/... (or vice-versa) and gets re-ingested. The scripts/manifest.py helper does this for you:

# New/modified sources, honoring WIKI_SKIP_PROJECTS + --skip, paths already canonical:
python3 "$LLM_WIKI_REPO/scripts/manifest.py" delta "$OBSIDIAN_VAULT_PATH" \
  --scan "$CLAUDE_HISTORY_PATH/projects/*/memory/*.md"
# One-time repair if the manifest already mixes ~ and absolute keys:
python3 "$LLM_WIKI_REPO/scripts/manifest.py" normalize "$OBSIDIAN_VAULT_PATH" --dry-run

The helper is optional — if it's unavailable, do the same expansion 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 "$LLM_WIKI_REPO/scripts/extract-jsonl.py" --skip tsg,autom8

# Incremental — only sessions modified in the last day
python3 "$LLM_WIKI_REPO/scripts/extract-jsonl.py" \
    --since "$(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

Step 3b: Parse Audit Logs (desktop sessions only)

For each audit.jsonl found under local-agent-mode-sessions/, read it line by line. Each line is a JSON record of one agent action:

{
  "type": "tool_call",
  "toolName": "Bash",
  "input": {"command": "npm test"},
  "output": "...",
  "timestamp": "2026-04-10T14:22:00Z",
  "sessionId": "..."
}

What to extract from audit logs:

• File access patterns — which files does the agent repeatedly Read or Edit? These are the high-value files in the project. Note them as project references.
• Shell commands — recurring Bash commands reveal the project's build/test/deploy workflow. Distill these into a how-to page in the procedure folder (e.g. "how this project is built and tested").
• Tool call sequences — if the agent always does Read → Edit → Bash in a particular order, that's a workflow pattern worth capturing.
• Error patterns — failed tool calls (non-zero exit codes, error outputs) reveal pain points, known rough edges, or recurring bugs.
• MCP tool calls — calls to MCP tools reveal which external services and APIs the project integrates with.

Skip from audit logs:

• Routine file reads with no pattern (e.g. reading config files once)
• Tool outputs that are just noise (long stack traces, verbose logs) — summarize the error class, not the full output
• Anything that looks like secrets, tokens, or credentials in command arguments or outputs

Cross-reference with the conversation transcript: The audit log tells you what happened; the conversation tells you why. When both are available for the same session, use them together — the audit log grounds the conversation in concrete actions.

Read the paired local_<uuid>.json session metadata before processing the audit log — it gives you cwd, startedAt, and title to contextualize the actions.

Step 4: Cluster by Topic

Don't create one wiki page per conversation. Instead:

• Group extracted knowledge by topic across conversations
• A single conversation about "debugging auth + setting up CI" → two separate topics
• Three conversations across different days about "React performance" → one merged topic
• The project directory name gives you a natural first-level grouping

Step 5: Distill into Wiki Pages

Each Claude project maps to a project in the vault. The project directory name from ~/.claude/projects/ encodes the original path — decode it to get a clean project name:

-Users/Documents/projects/my-Project   → myproject
-Users/Documents/projects/Another-app  → anotherapp

Read the vault's CATEGORIES table first (AGENTS.md, or llm-wiki layout --json). Pages go only in the folders it lists, and each kind of knowledge below goes to the folder holding its role — the paths shown are a standard vault's. A role no row claims falls to the catch-all row.

Project-specific vs. global knowledge

(A standard vault has no separate how-to folder, so procedure lands in concepts/; vaults that still have skills/ use it.)

With a project folder (a row with the project role, projects/ by default): put project-specific pages under projects/<name>/<folder>/, and for each project with content, create or update the project overview page at projects/<name>/<name>.md — named after the project, not _project.md.

Without one: don't create projects/ (law 3). File project-specific pages in the global folders, add project: <name> to their frontmatter, and keep the project's summary in the manifest's projects section only. Obsidian's graph view uses the filename as the node label, so _project.md makes every project show up as _project in the graph. Naming it <name>.md gives each project a distinct, readable node name.

Important: Distill the knowledge, not the conversation. Don't write "In a conversation on March 15, the user asked about X." Write the knowledge itself, with the conversation as a source attribution.

Write a summary: frontmatter field on every new/updated page — 1–2 sentences, ≤200 chars, answering "what is this page about?" for a reader who hasn't opened it. wiki-query's cheap retrieval path reads this field to avoid opening page bodies.

Add confidence and lifecycle fields to every new page's frontmatter:

base_confidence: 0.42
lifecycle: draft
lifecycle_changed: <ISO date today>

On update, leave lifecycle and lifecycle_changed unchanged — only a human editor transitions lifecycle state.

Mark provenance per the convention in llm-wiki (Provenance Markers section):

• Memory files are mostly extracted — the user wrote them by hand and they're already distilled. Treat memory-derived claims as extracted unless you're stitching together claims from multiple memory files.
• Conversation distillation is mostly inferred. You're synthesizing a coherent claim from many turns of dialogue, often filling in implicit reasoning. Apply ^[inferred] liberally to synthesized patterns, generalizations across sessions, and "what the user really meant" interpretations.
• Use ^[ambiguous] when the user changed their mind across sessions or when assistant and user contradicted each other and the resolution is unclear.
• Write a provenance: frontmatter block on every new/updated page summarizing the rough mix.

Step 6: Update Manifest, Journal, and Special Files

Update .manifest.json

For each source file processed, add/update its entry with:

• ingested_at, size_bytes, modified_at
• source_type: one of "claude_conversation", "claude_memory", "claude_audit_log", "claude_desktop_session"
• project: the decoded project name
• pages_created and pages_updated lists

Also update the projects section of the manifest:

{
  "project-name": {
    "source_path": "~/.claude/projects/-Users-...",
    "vault_path": "projects/project-name",
    "last_ingested": "TIMESTAMP",
    "conversations_ingested": 5,
    "conversations_total": 8,
    "memory_files_ingested": 3,
    "desktop_sessions_ingested": 2,
    "audit_logs_ingested": 2
  }
}

In a vault without a project folder, set vault_path to null.

Create journal entry + update special files

Update index.md and log.md per the standard process:

- [TIMESTAMP] CLAUDE_HISTORY_INGEST projects=N conversations=M desktop_sessions=D audit_logs=A pages_updated=X pages_created=Y mode=append|full

hot.md — Read $OBSIDIAN_VAULT_PATH/hot.md (create from the template in wiki-ingest if missing). Update Recent Activity with a one-line summary — e.g. "Ingested 5 Claude conversations across 2 projects; surfaced patterns in API design and testing strategy." Keep the last 3 operations. Update Active Threads if any ongoing project is now better understood. Update the updated: field in the frontmatter to the current timestamp — this is easy to forget; the body edit and the frontmatter bump must both happen.

Privacy

• Distill and synthesize — don't copy raw conversation text verbatim
• Skip anything that looks like secrets, API keys, passwords, tokens
• If you encounter personal/sensitive content, ask the user before including it
• The user's conversations may reference other people — be thoughtful about what goes in the wiki

Reference

See references/claude-data-format.md for more details on the data structures.

QMD Refresh After Vault Writes

GUARD: If $QMD_WIKI_COLLECTION is unset — the default — skip this step. QMD is a search index, not the source of truth.

When it is set, read llm-wiki/references/qmd.md and follow the Refresh section after this skill has written vault markdown. If refresh fails, do not roll back the vault changes; report the QMD status separately.