Skip to content

Autonomous development for real products. Governed agents build from specs you approve, follow the rules you set, and nothing merges until another agent proves it. Your decisions, judgement and taste are the bar that work must clear. Declare your vision and let agents build it.

feat(merge): a replay bench predicts a board's flow [I11,R1,R2,H16,A1]

Corey Olson46587d7ec733
diff --cc src/templates.js
index 647c831b,3fc36b49..b3b56684
--- a/src/templates.js
+++ b/src/templates.js
@@@ -144,8 -144,10 +144,10 @@@ When a reject names a structured cause
6. A reject reopens the item. Fix it, commit, claim, submit again; the same head is refused. If you cannot get an item green, do not force it: \`pullboard escalate <id> --note "what you tried and how it failed"\` hands it one tier up.
7. Your lane is done when its items are verified, not when they are submitted. After your last submit, keep running \`pullboard next --wait 9\`: a rejected item comes back to your lane.
-The coordinator merges verified items into the main line; builders never merge. When a release ships, the coordinator runs \`pullboard milestone ship <name>\`: it records main's commit and moves the items that have not landed into the next milestone. \`pullboard health\` lists every stuck item: a lapsed claim, a review nobody took, accepted work that never landed, an old hold, an unanswered decision. Each line says why, since when, the next step and who takes it; run it at the start of a session and take the steps that name you.
+The coordinator merges verified items into the main line; builders never merge. When a release ships, the coordinator runs \`pullboard milestone ship <name>\`: it records main's commit and moves the items that have not landed into the next milestone. The queue orders itself by what each item unblocks and what it costs; the coordinator holds an item above that order with \`pullboard milestone pin <name> <ids> --reason "<why>"\`, below it with \`pullboard milestone sink <name> <ids> --reason "<why>"\`, and returns it with \`pullboard milestone lift <name> <ids>\`. \`pullboard health\` lists every stuck item: a lapsed claim, a review nobody took, accepted work that never landed, an old hold, an unanswered decision. Each line says why, since when, the next step and who takes it; run it at the start of a session and take the steps that name you.
+ **Replay flow.** Run \`pullboard flow replay --json\` to reproduce recorded landing order and stage quartiles from text-free lifecycle history. To compare a hypothetical order and board line settings, run \`pullboard flow replay --history history.json --policy policy.json --json\` with the exported version-1 history and a JSON policy containing \`order\` and \`settings\`. Recorded work durations stay fixed; alternate queue times are deterministic predictions, not measured speedups.
+
For recurring checks, \`pullboard sentinels\` lists this board's watches, their approvals and their latest results. Propose one with \`pullboard sentinel propose <name> --check "<command>" --every hourly --owner coordinator\`; a proposal does not run until the person approves its exact command, directory and schedule. Only the person runs \`pullboard sentinel approve <name>\` or \`pullboard sentinel decline <name> --reason "<why>"\` from their own terminal or Pullboard View. Agents never approve their own proposals; send the decision to the coordinator.
For spare test capacity, use \`pullboard runner add <ssh-host> --slots <n> --browser <n>\`. It registers SSH capacity on this machine; \`pullboard runner list\` shows each runner's registered slots and browsers, and \`pullboard resources\` shows their usage. Only committed source goes to a runner, and board moves stay here. To see whether this board is linked to the relay, and for which GitHub account, run \`pullboard relay status\`.