Redo

A run made at a lower effort or on a weaker model is a draft. Redo follows pstack's Session pickup, except that it redoes the work that ran below the current level and reuses everything else. Load pstack:poteto-mode first; its playbooks and the project's gates still own each step.

1. Find the draft. A plan path is the draft. For a page link, find the plan whose subject or plan file holds that Page: line (git grep -l 'Page: <link>' in the plans directory); for a PR, the plan its body links, or on another person's PR, the plan your own comment links. Read the plan, its subject file, its decision log and its run directory under <plans>/artifacts/. Never read the published page, because another account may own it.
2. Read the levels. node .agents/pstack/levels.mjs <log>... prints the current session's level and the configured seats, then marks each lead stretch and seat round of the draft's logs at or below, with the reason for each below. A stretch with no lead row, or a seat with no level, counts as below. A subagent role runs at its models-sheet line.
3. Write the ledger. Open a new iteration of the draft's subject with the plan-page skill. Its Brief names the draft and the level change, and its ## Redo table lists every unit of the draft with its producer, that producer's level and one verdict:
   • redo: judgment a producer made below the level the current setup gives its role. That covers the plan and its Defaults, design picks, each subagent or seat lane, each panel finding's disposition, the code each built: row names, and the Close and its claims. It also covers any unit the draft marked weak: a partial, gap, open or inconclusive row, a missing seat, a same-family round or a dismissed critical finding.
   • reuse: deterministic evidence on bytes the redo will not change, such as a test run, a proof log or a render, and any lane already at or above the current level. Cite it, and do not run it again.
   • recheck: reused evidence whose inputs a redo may change. It reruns after the redo. When no unit is redo, say so on the page. For a PR, go on to step 6; otherwise stop.
4. Redo the plan and design first. Work from the owner's asks and the draft's inputs before reading the draft's conclusions, then compare and log each pick kept, changed or dropped. A design step that runs architect or an arena takes the draft's design as one more candidate. An unbuilt draft plan becomes superseded; a built one stays and the new iteration continues it.
5. Rerun the lanes. Each redo lane runs again on its configured model and level, from its brief in the run directory. Re-judge every reused seat answer at the current level, and log a new panel row for each finding whose disposition changes.
6. Clean up the code. Run the project's Babysit playbook, or pstack's, over the draft's diff at the current level, without its level check. The fixes then take the project's usual gates and land by its Delivery rule. When the draft is a PR, push the fixes to that PR's head branch instead, as Babysit does.
7. Recheck and close. Rerun each recheck proof. The Close lists what was redone and what changed, what was reused and why, and any unit whose level stayed unknown.