# Agent Biographer > Shadow and version-control another agent's conversation. Use when asked to track an agent, create agent biography, shadow an agent, document agent work, or make agent output more accessible. - Skill: `tankygranny05/agent-biographer` (Agent Skill) - Install (CLI): `npx skillmds@latest add tankygranny05/agent-biographer` - Raw SKILL.md: https://api.skillmd.com/api/skills/tankygranny05/agent-biographer/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: tankygranny05 (https://skillmd.com/u/tankygranny05) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/tankygranny05/agent-biographer --- # Agent Biographer **Purpose:** Shadow another agent's conversation, version-control it over time, and make dense agent output accessible for humans. --- ## Quick Start ```bash # 1. Find agent's CWD from SID find ~/.claude/projects -name "{sid}.jsonl" # 2. Dump conversation (--sid alone is sufficient) python ~/swe/telemetry_projects/agent_dump/launchers/dump_conversations.py \ --sid {suffix} # 3. Create archive and commit mkdir -p ~/KnowledgeBase/agent-archives/{name}-{suffix} # ... copy conversation.txt, create README, git init, commit ``` --- ## Step 1: Find Agent's CWD ### Method: Find JSONL File ```bash find ~/.claude/projects -name "{sid}.jsonl" ``` **Output format:** `~/.claude/projects/{encoded-cwd}/{sid}.jsonl` **Decode the CWD:** - Replace leading `-` with `/` - Replace all remaining `-` with `/` ### Examples (for inference) | JSONL Path | Encoded CWD | Decoded CWD | |------------|-------------|-------------| | `~/.claude/projects/-Users-sotola-agent-home/abc123.jsonl` | `-Users-sotola-agent-home` | `/Users/sotola/agent-home` | | `~/.claude/projects/-Users-sotola-swe-telemetry-projects/def456.jsonl` | `-Users-sotola-swe-telemetry-projects` | `/Users/sotola/swe/telemetry_projects` | ### One-Liner ```bash SID="f11289f0-b231-407d-b891-0fe64047ed5a" dirname "$(find ~/.claude/projects -name "${SID}.jsonl" 2>/dev/null)" | \ xargs basename | \ sed 's/^-/\//' | \ tr '-' '/' ``` ### If Method Fails If the JSONL file doesn't exist (old session), infer from context or ask the user for the CWD. --- ## Step 2: Dump Target Agent's Conversation ### Get SID Suffix The `--sid` flag uses **SUFFIX** (last 5 characters), NOT prefix! | Full SID | Suffix | |----------|--------| | `f11289f0-b231-407d-b891-0fe64047ed5a` | `ed5a` | | `019bfbb7-16c8-79e3-82e7-f6818860ca12` | `0ca12` | ### Dump Command ```bash # --sid alone is sufficient - no time constraints needed python ~/swe/telemetry_projects/agent_dump/launchers/dump_conversations.py \ --sid {suffix} \ --data-dir /tmp/agent-{suffix} ``` **Output:** `/tmp/agent-{suffix}/{agent-type}/{full-sid}-{pid}/conversation.txt` ### Find Active Session If multiple sessions exist, check round count: ```bash for dir in /tmp/agent-{suffix}/claude/*/; do rounds=$(head -5 "$dir/conversation.txt" | grep "Rounds:" | awk '{print $2}') echo "$dir: $rounds rounds" done ``` Most rounds = most recent/active session. --- ## Step 3: Create Archive ### Directory Structure ``` ~/KnowledgeBase/agent-archives/{descriptive-name}-{suffix}/ ├── README.md # Overview (you write this) ├── conversation.txt # Source of truth (copy from dump) ├── *.md # Any docs the agent generated (optional) └── .git/ # Version control ``` ### Initialize Archive ```bash # Create directory mkdir -p ~/KnowledgeBase/agent-archives/{name}-{suffix} # Copy conversation cp /tmp/agent-{suffix}/{type}/{full-sid}-{pid}/conversation.txt \ ~/KnowledgeBase/agent-archives/{name}-{suffix}/ # Copy any generated docs (if applicable) cp ~/agent-home/ai/generated_doc/*.md \ ~/KnowledgeBase/agent-archives/{name}-{suffix}/ 2>/dev/null # Initialize git cd ~/KnowledgeBase/agent-archives/{name}-{suffix} git init git add -A git commit -m "Initial commit: {agent description} Co-authored by [Claude: {your-agent-id}]" ``` ### README Template ```markdown # {Agent Name} - Agent Archive **Agent ID:** `{full-sid}` **Agent Type:** Claude/Codex **Date:** {date} **Rounds:** {count} ## Summary {2-3 sentences about what the agent did} ## Key Accomplishments 1. {accomplishment 1} 2. {accomplishment 2} ... ## Resume Command \`\`\`bash cd {cwd} && cclaude --resume {sid} \`\`\` ``` --- ## Step 4: The Update Loop (Version Control) When the user says **"update"**: ### 1. Fresh Dump ```bash rm -rf /tmp/agent-{suffix} python ~/swe/telemetry_projects/agent_dump/launchers/dump_conversations.py \ --sid {suffix} --data-dir /tmp/agent-{suffix} ``` ### 2. Copy and Diff ```bash cp /tmp/agent-{suffix}/{type}/{full-sid}-{pid}/conversation.txt \ ~/KnowledgeBase/agent-archives/{name}-{suffix}/ cd ~/KnowledgeBase/agent-archives/{name}-{suffix} git diff conversation.txt ``` ### 3. Analyze New Content - **DO NOT** re-read the entire conversation - **DO** read only the git diff output - Look for new rounds, new tool calls, new accomplishments ### 4. Commit ```bash git add conversation.txt git commit -m "Round N: {summary of new content} Co-authored by [Claude: {your-agent-id}]" ``` --- ## Step 5: Make Content Accessible ### Your Job as Biographer The target agent's output is often **dense and terse** - good for agents, bad for humans. Your job: 1. **Translate** technical content into accessible explanations 2. **Don't dumb down** - maintain accuracy and depth 3. **Occasionally fact-check** - but don't overdo it 4. **Explain the "how"** - show tool call traces when interesting ### Example: Tool Call Trace When the agent does something interesting, show HOW: ```markdown ## How the Agent Found the Information ### Phase 1: Search (Calls 1-5) - Searched for "throughput" in docs → found references but no numbers - Read SKILL.md → design description only ### Phase 2: Experiment (Calls 6-10) - Built release binaries - Created test data samples - Ran actual benchmarks - **Result:** 2.4M lines/sec **Key insight:** Agent transitioned from researcher → experimenter ``` --- ## Step 6: Reasoning Content & Role-Based Judgment ### Extracting Reasoning Add `--include-reasoning` to include thinking/reasoning blocks: ```bash python ~/swe/telemetry_projects/agent_dump/launchers/dump_conversations.py \ --sid {suffix} \ --include-reasoning \ --data-dir /tmp/agent-{suffix} ``` **Default behavior:** Reasoning is excluded (it's exploratory, not final output). ### Judging Reasoning: Role Matters **CRITICAL:** Reasoning content is **exploratory thought**, not final judgment. How you judge it depends on the agent's **Role**. | Role Type | Overthinking Is... | |-----------|-------------------| | **Complex debugging / ecosystem cleanup** | GOOD | | **Simple execution task** | BAD | ### Example: Same Agent, Different Judgment **Agent:** 50+ tool calls, 29 minutes of reasoning about Docker containers **If Role = "Debug and fix the agent-box Docker image":** > EXCELLENT. Exhaustive exploration of edge cases (volume mounts, symlink detection, platform mismatches, port conflicts) led to correct fix on first try. For ecosystem debugging, you WANT the agent to think through every scenario. **If Role = "Run this Docker image for me":** > MASSIVE FAILURE. User wanted `docker run `. Agent spent 29 minutes overthinking irrelevant edge cases. Should have executed in 30 seconds. ### Default Stance No Role provided → Report reasoning content **without judgment**. Let the user decide if the depth was appropriate. --- ## Workflow Summary ``` ┌─────────────────────────────────────────────────────────────────┐ │ INITIAL SETUP │ ├─────────────────────────────────────────────────────────────────┤ │ 1. Get target SID from user (full SID or suffix) │ │ 2. Find CWD: find ~/.claude/projects -name "{sid}.jsonl" │ │ 3. Extract suffix (last 5 chars of SID) │ │ 4. Dump: --sid {suffix} (add --include-reasoning if needed) │ │ 5. Create archive, write README, git init, commit │ └─────────────────────────────────────────────────────────────────┘ │ ▼ ┌─────────────────────────────────────────────────────────────────┐ │ UPDATE LOOP │ ├─────────────────────────────────────────────────────────────────┤ │ On "update": │ │ 1. Fresh dump (rm old, re-dump with --sid {suffix}) │ │ 2. Copy conversation.txt to archive │ │ 3. git diff conversation.txt (NOT full re-read) │ │ 4. Analyze new rounds, make accessible │ │ 5. Commit with summary │ │ 6. Repeat │ └─────────────────────────────────────────────────────────────────┘ ``` --- ## User Preferences (Learned) | Preference | Do This | Don't Do This | |------------|---------|---------------| | **Check conversation first** | Read conversation.txt diff | Check docs first | | **Don't re-read everything** | Use git diff | Read from start on every update | | **Make accessible** | Translate dense output | Dumb it down | | **Commit each update** | git commit after every update | Batch commits | | **Show tool call traces** | Explain HOW agent found info | Just show results | | **Do own research** | Fact-check, add context | Blindly trust target agent | --- ## Common Mistakes to Avoid ### 1. Wrong --sid Format ```bash # WRONG - prefix --sid f11289f0 # CORRECT - suffix (last 5 chars) --sid ed5a ``` ### 2. Adding Unnecessary Time Constraints ```bash # UNNECESSARY - time constraints add nothing when --sid is specified --start-hour 720 --end-hour 0 --sid ed5a # CORRECT - --sid alone is sufficient and fast --sid ed5a ``` ### 3. Re-reading Full Conversation on Update ```bash # WRONG cat conversation.txt # Every time # CORRECT git diff conversation.txt # Only changes ``` --- ## Archive Naming Convention ``` ~/KnowledgeBase/agent-archives/{descriptive-name}-{suffix}/ ``` Examples: - `rust-db-expert-0ca12/` - `biography-skill-researcher-ac26/` - `docker-ecosystem-deployer-d8b8/`