fix(pstack): resolve rule conflicts and read the model rule the same way

[poteto]

Why

Three pstack rules contradicted other rules in the same skills, so an agent had to guess which rule wins.

• Babysit bans rebases and force-pushes and hands that work to the owner. An Autopilot-full owner runs babysit on its own PR, and it is also the owner that rebases on a trunk conflict, on a CI failure from trunk, and at merge prep. Autopilot-full did not say how the owner publishes those rebases.
• show-me-your-work calls the decision log append-only, but its end-of-run audit said to cut invented rows, drop padding, and fix wrong rows.
• opening-a-pr said a subagent that opens a PR does not babysit, while Autopilot-full and Autopilot-stack owners babysit their own PRs by design.

The skills that read the pstack-models.mdc rule also read it in different ways. arena, swarm, and interrogate used their line "when present". how, why, reflect, and architect said "your configured" model or runners without naming the rule or its line. Only interrogate said what to do when the Task tool rejects a slug.

Scope

• Autopilot rebases. Babysit step 4 names the Autopilot-full owner, and the root in Autopilot-stack, as the owner it defers to. Where babysit says to report a rebase, that owner rebases its own branch and publishes it with git push --force-with-lease. Autopilot-full step 2 adds the lease push after an ls-remote check and bans a force-push of a shared branch. A standalone babysit keeps the full ban.
• Append-only audit. The show-me-your-work audit never edits or removes a row. A wrong or invented row gets a new row that supersedes it. The audit checks only the current run's rows, and a run that adds to an existing log first logs a start row. The Rules list does not change.
• Owner babysit. opening-a-pr keeps the no-babysit rule for other PR-opening subagents. An Autopilot-full or Autopilot-stack owner starts the babysit loop after its code-ready report.
• Model rule. how, why, reflect, arena, swarm, and interrogate each name their role line. A missing rule or line means the default, and auto or inherit-parent leaves model unset. A rejected slug falls back to the default, and a rejected panel entry falls back to the default of its family. architect takes its runners from architect runners and follows arena's runner rules. poteto-mode names the lines that the code playbooks, the hardest changes, and prose and judgment read.
• Plan lint. multi-phase-plan runs the ten live lanes on the swarm workers model. check-plan.mjs accepts any model in the lane sentence and rejects the skeleton's unfilled <swarm workers model>.
• Setup. setup-pstack drops a line for a retired role, such as how critics, and lists it when it shows the roles.
• Version 0.15.5. No changelog file, as in docs(pstack): cut 19 more instructions Opus 5.5 does not need #419.

No default model changes.

Blast radius

Every pstack user gets the new text on update.

• An Autopilot-full owner may now force-push its own branch with a lease inside its babysit. Every other babysit keeps the ban on rebases and force-pushes.
• The decision-log audit no longer deletes rows, so an invented row stays in the log and a later row supersedes it. A run that writes to a log that already has rows opens each of its stretches with a start row.
• A user without the rule sees no model change. A user with the rule gets the same models for the lines it has. When the Task tool rejects a configured slug, the skill runs that role on its default and says so. In the arena, architect, and interrogate panels, a rejected entry runs on the default of its family.
• A plan copied from the multi-phase-plan skeleton fails check-plan.mjs until its lane model is filled in. Plans that name grok-4.7-xhigh-fast still pass.
• Not tested in a live agent run.

Verification

• node scripts/validate-plugins.mjs passes.
• In pstack/skills/poteto-mode/scripts, bun test orch watch-pr passes 52 of 52 and bun run typecheck passes. Neither covers check-plan.mjs, so the next check runs it directly.
• check-plan.mjs ran on the skeleton with every other placeholder filled. On main it passes only grok-4.7-xhigh-fast in the lane sentence. On this branch it also passes claude-opus-5-5-max, gpt-5.6-sol-max, and auto. Both reject a missing sentence, empty backticks, "Nine lanes", and a model without backticks. This branch also rejects the unfilled <swarm workers model>, so the raw skeleton fails with that one problem.
• git diff --check is clean, and no frontmatter changed.
• No README or guide page restates a changed rule.

---

Note

Low Risk
Documentation and agent-workflow skill text only; behavior changes are procedural (lease force-push for autopilot owners, append-only logs) with no runtime code or auth paths touched.

Overview
Resolves contradictory playbook and skill rules so agents no longer have to guess which instruction wins, and standardizes how every routed skill reads pstack-models.mdc (named role lines, missing-rule defaults, auto/inherit-parent omitting model, and family-based fallbacks when Task rejects a slug).

Autopilot vs babysit: babysit still forbids rebases/force-pushes for normal babysitters, but Autopilot-full/stack owners may rebase their own branch and publish with git push --force-with-lease after ls-remote. opening-a-pr exempts those owners from the “don’t babysit on open” rule so they start babysit after code-ready.

Decision logs: show-me-your-work keeps logs append-only; audits add start rows for multi-run logs and supersede bad rows instead of deleting or “cutting” them.

Models in plans: multi-phase-plan and check-plan.mjs tie ten live swarm lanes to the swarm workers line (regex accepts any filled model, rejects the unfilled skeleton placeholder).

Setup: setup-pstack drops retired rule lines (e.g. how critics) and surfaces them on re-run.

Version bumps to 0.15.5 in plugin.json. No default model slug changes.

^{Reviewed by Cursor Bugbot for commit be6bcd1. Bugbot is set up for automated code reviews on this repo. Configure here.}