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.
JavaScript agentic-codingai-agentsai-governanceautonomous-agentsclaude-codeclicode-reviewcodexcoding-agentsdeveloper-toolsgit-hooksllmlocal-firstmulti-agentsoftware-factoryspec-driven-developmentverificationvibe-coding
docs(skills): the run skill teaches pullboard flakes and its rerun setting
Corey Olson6422d005b021
diff --git a/skills/pullboard-run/SKILL.md b/skills/pullboard-run/SKILL.mdindex 5ff9fa9d..bb85f848 100644--- a/skills/pullboard-run/SKILL.md+++ b/skills/pullboard-run/SKILL.md@@ -87,7 +87,11 @@ When the same cause fails work twice, turn it into a draft row in the same sessiLand verified work in batches with `pullboard land`: preview with `pullboard land --dry-run`, then land with `pullboard land --max N`. Land takes accepted items in the order they were accepted. It sets aside one that carries an unverified commit (`STACKED_UNVERIFIED`) or a red it already proved (`LAND_KNOWN_RED`). Put every other reason an item waits on the board with `pullboard hold <id> "reason"`, so health and the view show it, and lift it with `pullboard hold <id> --off` when the reason is gone.-When a batch goes red, land runs each failing test file alone. A file that passes alone is a flake and holds the batch with `LAND_FLAKE` until the test is fixed or the person records an expiring waiver. A real red is split in halves until one item reproduces it alone; land returns only that item, with a fact naming the tests it fails, and lands the green rest. When green items fail only together, it stops with `LAND_INTERACTION` and lands nothing.+`pullboard flakes` lists retained failure observations, original and diagnostic receipt digests, and linked repair items. The ledger is always on: a later passing diagnostic never replaces the original red or permits submission, verification or landing.++`pullboard flakes rerun on`, run by the coordinator from the main checkout, opts this board into one exact-target rerun at the pinned candidate and a pinned-main comparison for supported native Node checks; `pullboard flakes rerun off` turns these pairs off again, and off is the default. An unsupported runner or selector stays unclassified. Full raw logs and profiles remain private under the local common Git directory; their portable digests remain on the board. A second confirmed flake sighting drafts ordinary strong repair work; it does not approve requirements or a waiver.++With paired diagnostics off, when a batch goes red, land runs each failing test file alone. A file that passes alone is a flake and holds the batch with `LAND_FLAKE` until the test is fixed or the person records an expiring waiver. A real red is split in halves until one item reproduces it alone; land returns only that item, with a fact naming the tests it fails, and lands the green rest. When green items fail only together, it stops with `LAND_INTERACTION` and lands nothing.A red batch reopens only the item a paired control proves. Before you reopen a culprit, run the same failing test or command twice: once on the control, which is main or the batch without the candidate, and once on the candidate. The control must pass and the candidate must reproduce the red. Keep both commands, both commits and both outputs as the evidence, and put them in the reopen note: `pullboard reopen <id> --note "<test file>: <test name>; merge commit <sha>"` names the red, and a verified item refuses any other note with `LANDING_FAILURE_REQUIRED`. When the pair does not prove it, keep the candidate held with `pullboard hold <id> "unproven culprit: <test>"` and leave it unreopened until a paired control does.