[chhoumann]

Why

t3-pstack should track Lauren Tan's pstack, so this moves all three components to cursor/plugins bf0c7cd (pstack #1236, pstack 0.15.37 through 0.15.42) as release 2.3.0. Lauren's text is adopted as written. Only the Cursor mechanics take a T3 form, and every model sheet a 2.x setup wrote keeps working.

What changed upstream

• why is one short file that answers from the commit and PR record, with no investigators, synthesizer, or references/.
• how is one file with one read-only subagent. poteto-mode no longer routes to it, and the playbooks trace code themselves. recall starts its own investigators.
• New defaults: hardest tasks on Fable 5.2 max, code playbooks on Opus 5.5 xhigh, swarm, reflect tooling, and verification on Opus 5.5 high fast, and each panel on Fable 5.2 high, Opus 5.5 high fast, and Gemini 4 Carbon high. Grok is no longer a default. Fable is its own family, and arena's judge comes from another vendor.
• setup-pstack's budget moves each role's own default effort, writes a # defaults: 2 line, and migrates old defaults.
• sharpen asks to prototype choices an experiment can settle, reflect swaps an example, and statechart reports SANY's line and column and draws arrows into a parent state through its border.

cursor-team-kit and thermos have no change between the pins, and their pins move with pstack's.

What the port adapted

• Defaults in models.json: claudeAgent/claude-opus-5-5 effort=xhigh (default), claudeAgent/claude-fable-5-2 effort=max (strongest), a new fast tier claudeAgent/claude-opus-5-5 effort=high fastMode=true, and the panel claudeAgent/claude-fable-5-2 effort=high, claudeAgent/claude-opus-5-5 effort=high fastMode=true, antigravity/gemini-4-carbon thinking_level=high. Gemini 4 Carbon is reachable through T3's Antigravity driver on this Mac, janus, and agents-fsn1, and a delegate_task to it returned OK.
• Fable is its own family in t3-sheet.mjs's family() and in the rule's text, so the default panel spans three families and two vendors.
• Retired how and why lines are harmless. Without a change, the checker printed "how explorer" is not a pstack role and three more lines for every 2.x sheet, and doctor would fail on all three machines. models.json now lists those four labels under retired, which the checker and doctor skip. setup-pstack drops them on its next run.
• The defaults marker is adopted in T3 form. The sheet block carries # defaults: 2, and the migration matches the defaults every 2.x setup wrote, since no T3 sheet ever held the Cursor slugs Lauren's version matches.
• Substitutions: 6 deleted with the text they anchored on, 7 re-anchored to Lauren's new sentences, 10 added, and the how explorer research-delegate rule now omits target, as Lauren's how names no model.

CHANGES.md has the full table and reasons.

Version

Minor, 2.3.0. setup-pstack gains the defaults migration, the defaults add a provider and a tier, and how and why lose their fan-outs. It is not major, because no user must edit their sheet or rerun the installer for things to keep working. The installer change only passes the retired list to doctor's sheet check. It writes nothing new to any config, and neither hooks file changes. Rerunning /setup-pstack is optional and takes the new defaults.

Rebase note

Open PRs #72 (2.2.9), #75, and #76 are not merged. Whichever lands first, this rebases onto it, and the version stays above 2.2.9.

Verification

• bun tools/generate.mjs --check current. GIT_CONFIG_GLOBAL=/dev/null bun test tests/ --timeout 30000 passes 564, skips 56, fails 0. The installer end-to-end test passes 18 of 18.
• Sync dry runs at bf0c7cd pass for pstack, cursor-team-kit, and thermos, with every fork's changed lines as before except setup-pstack's, which drop from 81 to 34.
• Skill script tests pass 373 of 373 with TLC and Temurin from mise. Without TLC, as CI runs them, 269 pass, 70 skip, and only the known launcher symlink test fails under /var. poteto-mode script tests pass 113 of 113, with typecheck and Prettier clean.
• Fresh T3 threads read poteto-mode and the changed skills from this tree in place of the installed 2.2.8. /why answered from the commit record and started no subagent. /how started one delegate_task research child with no target, and told the user once to rerun /setup-pstack because the sheet lacks # defaults: 2. setup-pstack's check on a copy of this Mac's sheet dropped the four retired lines, replaced no value, and the checker exited 0 on both the candidate and the unmodified copy. The real sheet was not touched. The threads, repo, and project are deleted.
• Review by codex/gpt-6-astra at reasoningEffort=xhigh found no critical, high, or medium issues. Its two low findings, a count in the notes and CONTEXT.md's family definition, are fixed.

---

Note

Medium Risk
Wide skill and model-default changes affect every delegated workflow and sheet validation, but backward compatibility for existing sheets and optional /setup-pstack limit mandatory user action.

Overview
Release 2.3.0 pins upstream Lauren Tan pstack at bf0c7cd (0.15.42) and rebases the T3 port on her latest skill shapes.

why and how shrink to single-file workflows. why no longer fans out investigators or uses references/; it answers from commits and PR discussion with explicit inference gaps. how spawns one delegate_task with role: "research" and no target (parent model). The how explorer, how explainer, why investigators, and why synthesizer roles move to retired in models.json so existing sheets still pass doctor until /setup-pstack drops those lines.

Default models and tiers change in models.json: hardest work on Fable 5.2 max, code playbooks on Opus 5.5 xhigh, a new fast tier for swarm/reflect tooling (Opus high + fastMode), and panels on Fable, Opus fast, Gemini 4 Carbon—Grok drops from defaults. t3-sheet.mjs treats claude-fable-* as its own family; arena cross-judge selection uses vendor rules (Fable + other Claude count as one vendor).

setup-pstack writes # defaults: 2, applies budget by adjusting each role’s default effort on a ladder, and migrates stale 2.x sheet values before re-setup. poteto-mode and playbooks stop auto-routing nontrivial work to how; they trace code or call slimmed skills instead. recall owns per-source investigators instead of delegating to why.

statechart reports SANY parse errors with line/column (not Java stacks) and fixes Mermaid edges into parent states via drawnEdges. Docs, manifests, and substitution/sync tooling update for the new upstream pin; many why reference files are removed.

^{Reviewed by Cursor Bugbot for commit 3ef8364. Bugbot is set up for automated code reviews on this repo. Configure here.}