macollins27/owner-rep

A Claude Code mod that watches a coding agent on behalf of the person who owns the work,
and reports back in plain words: what the agent has really done, what it has left out, and
what to tell it next.
It is built for the owner who directs agents but does not read code. The agent's own report
is the only account they get, and it is the one account that cannot be taken on trust. The
rep reads the same conversation, takes nothing on the agent's word, and says so briefly.
What it does
The rep draws one row above the prompt, in the same shape as Claude Code's built-in "You
should know" note: the result the owner wants, worked out from the owner's own messages, and
what the agent is doing right now, with a mark when the session needs a look (the agent has
stopped, is asking a question, or a plan is waiting for a yes). That status is plain words
built from the tool's name: a file's name for a read or an edit, the program's name for a
shell command, and never any text of the command itself.
Behind that row are three reads of increasing cost.
| Read | When | How | Cost |
|---|---|---|---|
| **Watch** | Every eighth model request of a turn, and at every stop | A fresh model instance over the agent's cached conversation, no tools | Mostly cache reads; never blocks the agent |
| **Briefing** | On a press (1 or click) | Same mechanism, longer brief | One model call; two when its answer is refused and asked again (below) |
| **Check** | On a second press, only when chosen | A separate headless claude -p run in a throwaway copy of the project, fenced, over the session record and the real files | Measured at about 210k fresh tokens plus 2.2M cache reads, about two minutes |
The **watch** stays silent unless it finds something with a cost to the owner: work dropped
or shrunk, a claim not shown, a test never run, the easy option taken in place of the right
one. Then one "Heads up" line appears, with the reason beneath it.
The **briefing** is a card: where the work really stands, what the agent's last message means,
what was left out, what should happen next, and a message to the agent, written and ready. One
press puts that message in the prompt box. Nothing is sent until the owner presses Enter. The
card is folded to fit the rows the band has, body first and buttons never, so the digit that
closes it or puts its message in the prompt always reaches a button on screen. Rows are
counted as the band wraps them, at words, because a count by cells says one row too few for
a long paragraph and the card then scrolls with its buttons out of reach.
The **check** reads outside the conversation. It opens the session's full record through
agentsview, a local tool that indexes every coding-agent session on the machine, opens the files the agent claims to
have changed, runs the project's own tests, and rewrites the card with each claim marked
**holds**, **false**, **not shown**, or **changed since** the briefing was written. When the
band is short, the claims that hold fold to a count first and the ones that failed stay
listed, because the check is paid for to find what does not hold.
How it decides
The rep measures against an objective: its own account, in its own words, of what the owner
wants on the job in hand. It is held in four parts: the result, what it is for, what "done"
has to show (a list of conditions), and the step the session is on now. It is never the
owner's text. The owner works by intent rather than exact wording, and a rep that only did
what the client literally said would not be a good one, so the objective includes what a
serious owner plainly wants without saying it. The owner's messages are weighed by kind:
reasons and complaints first, instructions next, and a "yes" to an agent's proposal last,
since that is the agent's idea with the owner's name on it.
The objective is written by a call of its own, never by the looks. Each message of the
owner's, and each note to the rep, starts one plain model call, not a fork of the
conversation, that is shown only the owner's numbered messages on the job (each kept whole
up to 6,000 characters, with the end of the agent's last words before it, so a short reply
can be understood), the objective as held, and the results of earlier jobs. The agent's
words reach the writer only as that line before each of the owner's messages, so they move