Wiki Agent

Wiki Agent — Targeted Cross-Agent History Search + Ingest

You are doing a query-driven targeted ingest from one specific AI agent's raw conversation history. The user is typically working in a different agent right now and wants to pull in context from another agent's past sessions.

This is not bulk ingest. You find sessions about a specific topic, extract the relevant blobs, distill them into the wiki, and return a synthesized answer the user can act on immediately.

Command Routing

Parse the invocation to determine the target agent and optional query:

If no query is given, default to recent sessions mode: ingest the last 5 unprocessed sessions from that agent and return a summary of what was found. This is equivalent to a focused wiki-history-ingest for that agent only.

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.
2. Read $OBSIDIAN_VAULT_PATH/.manifest.json → know what's already ingested.
3. Read $OBSIDIAN_VAULT_PATH/hot.md if it exists → warm context on recent wiki activity.

Step 1: Locate the Agent's History Root

If the history root doesn't exist, stop and tell the user: "No <agent> history found at <path>. Have you run <agent> on this machine? You can set a custom path with <CONFIG_VAR> in .env."

Step 2: Build Session Inventory

Use the cheapest index source for each agent — don't open session files until you know which ones are relevant.

Claude

Primary index:   ~/.claude/projects/  (directories = projects, files = sessions)
Session files:   ~/.claude/projects/*/*.jsonl
Desktop index:   find ~/Library/Application Support/Claude/local-agent-mode-sessions -name "local_*.json"
Signal fields:   sessionId, cwd, startedAt, title (in local_*.json)

Build a list of sessions: {path, project_dir, modified_at, already_ingested}.

Codex

Primary index:   ~/.codex/session_index.jsonl
Session files:   ~/.codex/sessions/**/rollout-*.jsonl
Signal fields:   thread_id, name/title, updated_at (in session_index.jsonl)

Read session_index.jsonl as the inventory. Each line: {thread_id, name, updated_at}. Map thread IDs to rollout files by matching directory names.

Hermes

Primary index:   ~/.hermes/memories/*.md  (fast to scan)
Session files:   ~/.hermes/sessions/**/*.jsonl
Signal fields:   file names, memory titles, first 3 lines of each memory

Scan memory filenames first (they're often titled by topic). Fall back to session listing.

OpenClaw

Primary index:   ~/.openclaw/workspace/memory/MEMORY.md  (structured long-term memory)
Daily notes:     ~/.openclaw/workspace/memory/YYYY-MM-DD.md
Session index:   ~/.openclaw/agents/*/sessions/sessions.json
Session files:   ~/.openclaw/agents/*/sessions/*.jsonl

Read MEMORY.md sections first — it's the pre-compiled summary of everything. Daily notes give recency signal.

Copilot

Primary index:   session filenames / directory listing
Session files:   varies by client (VS Code: ~/.copilot/sessions/*.jsonl or similar)
Signal fields:   session timestamps, file names

Pi

Primary index:   ~/.pi/agent/sessions/--<cwd>--/ directories
Session files:   ~/.pi/agent/sessions/--<cwd>--/<timestamp>_<uuid>.jsonl
Signal fields:   cwd (decoded from dir name), session_info.name, timestamp in filename

Scan session directories first. Decode --<cwd>-- to get the working directory. Read the first line (session header) and any session_info entries for the session name. No separate index file — the filesystem is the index.

Step 3: Score Sessions Against the Query

If a query was given, score each session in the inventory without opening full session files:

1. Name/title match — does the session name or thread title contain the query terms? Score: +3

2. CWD/project match — does the working directory suggest the right project? Score: +2

3. Recency — apply exponential time decay with a 90-day half-life, as a multiplier on the match score rather than a bonus added to it:

   base  = name_match(3) + cwd_match(2)
   score = base * (0.35 + 0.65 * 0.5 ** (age_days / 90))

   The 0.35 floor is deliberate: an old session that matches the query exactly must still outrank a recent one that barely matches, or the skill can never answer "how did I first solve this?". This is the same decay session-brain uses, so the two skills rank consistently.

4. Already ingested — if this session was previously ingested and the wiki page already covers the query (check hot.md + index.md), flag as "covered" but still show in results

Select the top 3–5 sessions by score. If no query was given, select the 5 most recent unprocessed sessions.

Step 4: Extract the Relevant Blob

Open each selected session file and extract only the content relevant to the query. Do not read the full session if it's large — use targeted extraction.

Per-Agent Extraction Strategy

Claude (JSONL conversation):

• Each line: {role, content, timestamp, ...}
• Search with: rg -i "<query terms>" <session.jsonl> to find the relevant lines
• Extract: the surrounding conversation window (10 lines before + 20 lines after each hit)
• Special signal: tool calls (Read/Write/Bash/Edit) reveal what was actually done — extract these even without keyword matches if they're in the relevant window

Codex (rollout JSONL):

• Each line: {type: "session_meta|turn_context|event_msg|response_item", ...}
• Filter to type: "event_msg" (user turns) and type: "response_item" (model output)
• Search with: rg -i "<query terms>" <rollout.jsonl>
• Extract: matching turns + their parent context (the turn_context preceding the match)
• Skip: session_meta events (operational metadata, not knowledge)

Hermes (memory files + session JSONL):

• For memory files: read the full file (they're short — typically <500 words each)
• For session JSONL: rg -i "<query terms>" + surrounding window
• Memory files with title matches → read fully; others → grep only

OpenClaw (MEMORY.md + daily notes + session JSONL):

• MEMORY.md: grep for section headers containing query terms → extract that section
• Daily notes: grep most recent 30 days for query terms → extract matching paragraphs
• Session JSONL: same grep-window approach as Claude
• Prefer MEMORY.md/daily notes over session JSONL (they're pre-synthesized)

Copilot (session JSONL):

• Same grep-window approach as Claude
• Look for checkpoint files if available (pre-summarized)

Pi (structured JSONL with tree layout):

• Each line is a tree entry: {type, id, parentId, timestamp, message?, ...}
• Build the active branch: map entries by id, find leaf (last entry with no children), walk parentId to root
• Search with: rg -i "<query terms>" <session.jsonl> to find matching entries
• Extract: the matching entries + their ancestors on the active branch (follow parent chain)
• Special signal: toolCall blocks inside assistant messages reveal what was actually done — extract these even without keyword matches if they're in the relevant window
• Prefer compaction and branch_summary entries when available — they're pre-synthesized summaries
• Skip thinking content blocks (noise) and model_change / thinking_level_change entries