Message IDs

Every prompt Cole types into an agent gets an exact, permanent id: the record id of that turn in the session’s transcript. Cole’s tooling keeps the id attached to the words wherever they travel, into briefs, handoffs, filed issues and Drover’s truth file. That is what lets an agent quote Cole word for word instead of paraphrasing, prove that a person (not another agent) said something, and jump back to the original message to read what came before it.

Where the id comes from

A Claude Code transcript is a JSON Lines file under ~/.claude/projects/<project>/, one record per message, tool call or event. Each record carries a uuid (its id), the parentUuid it follows, the sessionId and a timestamp. A turn a person typed also says so:

{
  "type": "user",
  "uuid": "3f6b2c1e-8d4a-4f7e-9b1a-5c2d7e9f0a14",
  "parentUuid": "…",
  "sessionId": "…",
  "timestamp": "2026-10-27T18:04:11.502Z",
  "origin": { "kind": "human" },
  "promptSource": "typed",
  "turnOrigin": "human",
  "message": { "role": "user", "content": "…" }
}

Not every user record was typed by a user. Tool results are user records too, a message from another session arrives with an origin of peer, task notifications and scheduled wakeups have origins of their own, and text a harness injects is marked isMeta. The origin fields are what tell a person’s words apart from everything else. (Codex rollout files have their own turn records, and its subagent threads are marked as such.)

You can list your own typed prompts, with their ids, from any transcript:

jq -c 'select(.type == "user" and .origin.kind == "human")
       | {ts: .timestamp, id: .uuid, text: (.message.content | if type == "string" then .[0:60] else "[blocks]" end)}' \
  ~/.claude/projects/<project>/<session>.jsonl

Capturing every prompt

A UserPromptSubmit hook on Cole’s machines writes each submitted prompt to a local spool: the hook’s whole payload (session id, prompt, transcript path, working directory) plus the machine, the tmux pane and the entry point. It parses nothing and calls nothing; the median cost is about seven milliseconds. Each record is written as a hidden file and renamed into place, so a reader never sees half of one. Headless claude -p automation and the hooks’ own model calls are skipped, so only interactive sessions are captured.

The spool does not know who typed the prompt, because the hook payload does not say. A collector answers that later by joining each spooled prompt with its record in the transcript on the machine that owns it:

A turn marked as meta, from the system, from a scheduled task or wakeup, or with an origin of peer, task_notification, scheduled or system is not a person’s.
A prompt whose turn is not in the transcript yet waits up to half an hour; one that never appears is skipped as unverified.
Prompts that another tool wrote are filtered out by the marks those tools leave: handoff prompts start with [cc-handoff] and say they were written by the handoff tool, launch prompts that hand a session its brief have a fixed shape, and relays carry the sender’s tag.

What the ids are used for

Every request becomes an issue

Cole asked, in these exact words, for “a github issue for anything i ask an agent to do”. A scheduled job takes the verified prompts in batches, asks a cheap model which are requests worth tracking, and files or updates issues in Cole’s private repositories. The model may choose the issue, but never the words: every quote is replaced by the record’s own text, so an issue never contains the model’s paraphrase of Cole. A word-for-word repeat becomes a comment, titles and bodies go through a secret redactor and a scanner, and every @ gets a zero-width space so nobody is pinged by accident. A typical run costs under a cent.

Briefs and amendments

When Cole changes a task mid-flight, the coordinator appends the change to the worker’s brief as an amendment, never rewriting what was there:

### A3, 2026-01-15T14:02Z, turn <turn id>
Owner, verbatim: "<their exact words, nothing trimmed that changes the meaning>"
Reading: <what changes in scope or constraints, one line>
Supersedes: A1 (the line that said to skip the migration)

The turn id is the transcript record id of Cole’s message. With it, anyone can open the original and read the conversation around it. The same format settles decisions (“D2 decided at this time, in this turn, in these words”) and relays to workers who never hear Cole directly.

Handoffs

When a session hands its work to a fresh one (Handoff Chains), the brief’s “What happened” section quotes the latest unfinished request word for word, and the tool records the session’s first prompt. The prompt the successor starts with says, in so many words, that the handoff tool wrote it and no person typed it, so a successor never takes the summary for Cole’s instructions.

The truth file

Drover records every message it delivers as a message event with its own time-ordered id, the sender (actor), the target and how it was delivered. In the receiving Claude session the message arrives as a peer turn from drover (<actor>), so the transcript says who relayed it. When a message reaches a handoff successor instead of the session it named, the event records that too. See Agent Messaging.

This wiki

When Cole asked for this wiki, the coordinating session passed the request to its writers verbatim, each message with its timestamp and record id, and noted which sends were duplicates. The scope of every page here traces back to those records. See About This Wiki.

Why it matters

Provenance. “Cole said” and “an agent said Cole said” are different claims. The origin fields and the actor in the truth file keep them apart, and Cole’s rules say the owner’s own words beat any relay of them.
Fidelity. A verbatim quote with an id can be checked; a paraphrase cannot.
Context. An id is a pointer back to the moment. The words before and after a request often matter as much as the request.
Dedupe. The same sentence typed into two windows is one request, not two issues.

Doing it yourself

You do not need Cole’s whole pipeline to get most of the value:

Keep transcripts. Claude Code deletes old ones after cleanupPeriodDays; Cole sets it high enough that they are never deleted.
When you write a brief or a decision log, paste the person’s words verbatim with the UTC time and the record id.
When a tool writes a prompt for an agent, have it say so in the prompt, so the agent never mistakes it for yours.
If you want every prompt captured, a UserPromptSubmit hook that appends the payload to a file is a dozen lines. Print nothing from it: that hook’s standard output goes into the model’s context.