post-mortem

name	post-mortem
description	Find agent transcripts from a completed Nomic game
disable-model-invocation	true

Post-Mortem: Locate Game Transcripts

Find the clerk and player agent transcripts from a completed Nomic game, copy them to the project, and verify they look correct.

Step 1: Find transcripts

Agent transcripts are JSONL files stored in the Claude Code project directory:

~/.claude/projects/<project-path>/

where <project-path> is the working directory with / replaced by - and leading -. For example:

/mnt/nw/home/c.dumas/claude-playground/nomic-game-1
→ ~/.claude/projects/-mnt-nw-home-c-dumas-claude-playground-nomic-game-1/

List all .jsonl files there. These are agent transcripts — each is a separate agent session (clerk, players, or auxiliary sub-agents).

Step 2: Identify agents

For each JSONL file, read the first line (JSON) to determine the agent type:

• {"type": "agent-setting", "agentSetting": "clerk"} → Clerk
• {"type": "agent-setting", "agentSetting": "player"} → Player
• Files starting with {"type": "user"} with large line counts are likely the main orchestrator or player sessions
• Files starting with {"type": "progress"} are auxiliary sub-agents (check-bash, explore, etc.)

To identify which player is which, use the read_agent_transcript MCP tool (from transcript-reader) with limit=5 on each candidate — the initial prompt contains the player name and model.

Step 3: Copy transcripts

Copy the identified transcripts into a transcripts/ folder in the project with descriptive names:

transcripts/
├── clerk.jsonl
├── player-<name>-<model>.jsonl
├── player-<name>-<model>.jsonl
└── player-<name>-<model>.jsonl

Step 4: Verify transcripts

For each copied transcript, check:

1. Event count — use read_agent_transcript with defaults and note the total event count reported in the [events X-Y of Z] header.
2. Last 5 messages — use read_agent_transcript with offset=-5 to read the final events. Confirm they look like a reasonable end of game (e.g. game-over announcement, final scores, post-mortem discussion) rather than a mid-conversation cutoff.

Report a summary table to the user:

Role	Name (Model)	Events	Last message summary
Clerk	...	...	...
Player	...	...	...
...	...	...	...

Step 5: Generate readable transcripts

For each JSONL transcript, generate a human-readable .txt version using read_agent_transcript with output_file set to the same path but with a .txt extension. Use default settings (truncated tool outputs). Run all four calls in parallel.

transcripts/
├── clerk.jsonl
├── clerk.txt                        ← generated
├── player-<name>-<model>.jsonl
├── player-<name>-<model>.txt        ← generated
├── ...

Do NOT read the generated files — they will be large.

Step 6: Update README

This branch represents a finished game. Update the README to:

1. Note that this branch is a completed game (e.g. "This branch contains the completed Game N").
2. List ALL files in the project that are not already documented in the README's project structure — including transcripts, decrypted notes, post-mortem files, and any other game artifacts. Read the current README and do a file listing to find gaps.

Step 7: Publish to viewer

Publish this game to the web replay viewer at https://butanium.github.io/nomic-viewer/

The viewer repo is at /mnt/nw/home/c.dumas/claude-playground/nomic-games/nomic-viewer.

1. Parse and publish — run from the viewer repo:

   cd /mnt/nw/home/c.dumas/claude-playground/nomic-games/nomic-viewer
   python parse_transcripts.py <game-dir> svelte-app/public/data/<game-id>.json
   

   where <game-dir> is the current project root and <game-id> is e.g. game-8. This parses the transcripts AND auto-regenerates games.json.

2. Commit and push to the viewer repo:

   git add svelte-app/public/data/
   git commit -m "Add <game-id> to viewer"
   git push
   

   GitHub Actions will automatically build and deploy to Pages.