Make your agent finish the review
I built Ponder, a free CLI and skill for Claude Code, Codex, and OpenClaw, because an agent reviewing a codebase alone skims the back half and stops early. Ponder writes every unit to disk first and blocks stop with exit 2 while anything is ungraded.
This site’s first pass
When I ran round one on this site, pass 1 listed 47 parts and blocked its own stop at 0.94 coverage. The Skeptic withdrew every finding on three parts as duplicates, which left those three ungraded again until pass 2 re-checked them.
No deleting files to fake 100%
When an agent gets tired, it will try to drop ungraded screens or claim a file never existed. Ponder checks the filesystem against every path a unit ever had, refuses the lie if the file is on disk, and keeps ungraded deletes in the count.
The only way past is grading it
Deleting an ungraded unit keeps it in the count, and a claim that it never existed is checked against every path it ever had. Neither round on this site deleted a unit: pass 2 regraded the last three parts, and coverage reached 1.00.
Four voices argue over the list
Every pass after the first sends four voices over the finding list. The Skeptic deletes false alarms, the Builder sharpens vague fixes, the Dreamer adds what pass 1 missed, and the Empath re-ranks by who gets hurt first.
Who argued about this site
These are real findings from round one on this site, with the move each voice made on them. Over five passes the Skeptic cut 63, the Builder sharpened 130, the Dreamer added 14 and the Empath re-ranked 48, all tallied from Ponder’s ledger.
Guess the churn before each pass
Before each pass, the agent predicts how much of the finding list will change, and Ponder measures the churn afterward. Watching the surprise gap shrink shows the agent now knows where the remaining bugs are.
Round one’s curve
Across round one’s five passes on this site, churn fell from 0.80 to 0.30, 0.07, and 0.02, while surprise dropped from 0.50 to 0.10, 0.08, and 0.03. By pass 5, the agent predicted 0.05 churn and measured 0.02.
It stops when two passes settle
An agent doesn’t get to stop because it wrote a long summary. Ponder lets the review finish when all four voices inspect the list for two passes in a row and churn stays under 0.10, with new findings under 0.05.
Round one went quiet
Round one settled by the rules: churn 0.07, then 0.02, and Ponder called it converged. Round two never did. The Builder kept re-anchoring line numbers in code I was still writing, so the agent stopped it with a written override.
One-line install and a fix plan
When the loop converges, Ponder hands back an ordered fix plan with the file and line range for every item, and it never touches your source code. I made it free and MIT licensed, with a one-line install for Claude Code, Codex, and OpenClaw.
The same files, in your project
Every number on this side came from plain files Ponder keeps in the project it reviews: the parts, an append-only ledger, each pass’s metrics and a report you can read and diff. It’s free and MIT licensed, and the install is one line.
Make your agent finish the review
I built Ponder, a free CLI and skill for Claude Code, Codex, and OpenClaw, because an agent reviewing a codebase alone skims the back half and stops early. Ponder writes every unit to disk first and blocks stop with exit 2 while anything is ungraded.