Skip to content

feat(pstack): new default models, one-file why and how, statechart fixes, sharpen and reflect text - #1236

Merged
poteto merged 12 commits into
mainfrom
lauren/pstack-model-routing-46c0
Nov 15, 2026
Merged

poteto merged 12 commits into
mainfrom
lauren/pstack-model-routing-46c0

Conversation

@poteto

@poteto poteto commented Nov 15, 2026

Copy link
Copy Markdown
Collaborator

Six changes to pstack, each as its own commit followed by a version bump (0.15.37 to 0.15.42): text updates to /sharpen and /reflect, two /statechart fixes, one-file /why and /how, and new default models.

1. /sharpen asks for prototypes, and /reflect gets a new example

  • /sharpen: the last line of the skill now has the rewritten prompt ask the agent to prototype any choice an experiment can settle and to recommend an option from what it observed. Product and preference choices that no experiment can settle stay with you. Before, the skill left every choice open.
  • /reflect: the synthesizer's list of details that drift used a model-name rename as its example. It now uses a function rename (parseConfig to loadConfig), so the example does not age.
  • The README row for /sharpen and the guide's poteto-mode chapter describe the new last line.

Checks: node scripts/generate-pstack-harness-trees.mjs --check is up to date, node scripts/check-pstack-docs.mjs finds 0 problems, and node scripts/validate-plugins.mjs validates all plugins. Version is 0.15.37, and harness/pstack/ is regenerated.

2. /statechart reports the spec's line and column when SANY cannot read it

When /statechart could not read a TLA+ spec, the error showed the last lines of SANY's XML run, which is a Java stack trace. It did not say where the spec was wrong.

check() now runs SANY a second time in plain mode when the first run fails, and reports that run's error section. The message starts at SANY's *** Errors, *** Abort messages or *** Parse Error line, so it names the line and column in the spec. It is capped at 40 lines. If the second run cannot start or has no error section, the message falls back to the old tail.

Three new tests in test/corpus.test.ts write a broken spec and assert the message: an unknown name, a parse error, and twelve unknown names. Each asserts that no at tla2sany. stack frame is in the message.

Checks:

  • test/corpus.test.ts with TLA2TOOLS set to the pinned tla2tools.jar: 50 pass, 0 fail. The three new tests fail without the run.ts change.
  • CI's command, node --test pstack/skills/*/scripts/test/*.test.ts without TLA2TOOLS: 267 pass, 0 fail, 70 skipped. The new tests are among the skipped, like every test that needs TLC.
  • The same three script checks as in section 1 pass. Version is 0.15.38.

3. /statechart charts enter a parent state through its border

In a /statechart chart, an arrow from outside a parent state to a state inside it could run through the parent's title. Pele 1.9.0 places a parent's title after layout, in one of three slots, and arrows that go straight to the states inside can cover all three.

The Mermaid backend now draws those arrows so they cannot cross the title:

  • An arrow from outside a parent ends at the parent's border. The parent's start marker then goes on to the state inside (drawnEdges in backends/mermaid.ts).
  • A lone arrow into a parent keeps its label. When one state leads to several states inside, each label moves to the start marker's arrow to that state.

Scope:

  • Covers every parent with one way in, including a parent that holds a start state.
  • Leaves out a parent that two or more states lead into. One start marker cannot say which state led where, so those arrows stay direct, and Pele can still draw one through the title.
  • The report's arrow count stays the same, because each step's label is still on exactly one arrow.
  • Each chart that has a parent with one way in now shows a start marker inside that parent. The chart still states the same steps. Of the corpus charts, only Login changes.

Tests:

  • test/drawn.ts is a new shared reader of a chart's arrows. It replaces the copy in test/cli.test.ts and also reads start-marker arrows inside a parent.
  • test/chart.test.ts has two new tests: a parent with one way in, and a parent that two states lead into.
  • test/render.test.ts has a new test that renders a chart in Pele 1.9.0 and checks that no arrow enters a parent's title bar.
  • test/labels.ts builds its expected edges from drawnEdges, and five expected arrow lists in test/cli.test.ts change with the new drawing.

Checks:

  • The skill's full suite with TLA2TOOLS set to the pinned tla2tools.jar: 373 pass, 0 fail, 0 skipped.
  • CI's command without TLA2TOOLS: 270 pass, 0 fail, 70 skipped.
  • The same three script checks as in section 1 pass. Version is 0.15.39.

4. /why is one short file that names no source

/why is now one file with a 48-word body. It names no data source and no tool, so it works the same in every project.

  • why/SKILL.md: the body tells the agent to answer from the record, not the code: the commits and PR discussions that introduced the target and each later change, and other sources if that trail is thin. It cites what the authors wrote, marks its own inferences, and says what the record does not answer. The investigators, the synthesizer, the model routing, the output format, and all 12 files under why/references/ are deleted.
  • recall: it borrowed the why investigators and source playbooks for its sweep of the shared record. It now starts its own investigator for each source. The source names stay generic.
  • blast-radius, teach: they no longer point at parts why no longer has (its step 2, its rules, its full sweep and narrowing advice).
  • poteto-mode: why stays a routed skill in the playbooks. It leaves the two lists of skills that set their own subagent type or model roles.
  • setup-pstack: the why investigators and why synthesizer roles are gone from the rule shape. A rule file that still has them loses those lines on the next /setup-pstack run, like any retired role.
  • poteto-help, the README and the guide's "Understand the code" chapter describe the new /why.

What a user sees: /why returns a shorter answer with no fixed headings, and it no longer searches each connected source in turn. It starts from source control and looks further only when that trail is thin. /recall still sweeps the shared record, with its own investigators.

Checks:

  • node scripts/generate-pstack-harness-trees.mjs --check: up to date. The generated why no longer carries the adapter line, because it uses no Cursor-only name.
  • node scripts/check-pstack-docs.mjs: 0 problems. node scripts/validate-plugins.mjs: all plugins validate.
  • A search of pstack/ finds no reference to a deleted file or to a why role, apart from the retired-role example in setup-pstack. Version is 0.15.40.

5. Nothing routes to /how, and /how is one file

/how stays as a short skill you invoke yourself. /poteto-mode and the skills it routes to no longer run it. They trace the code directly.

  • how/SKILL.md: one file. It starts one read-only subagent with a three-sentence brief and presents the output. The model routing, the complexity gate, the explorers, the output format, and both prompt templates under how/references/ are deleted.
  • poteto-mode: the "nontrivial change → how" trigger is gone, and how leaves the two lists of routed skills.
  • Playbooks: Investigation, Feature, Bug fix, Hillclimb, Perf issue and Refactoring say "trace" where they said to run how. The multi-phase plan template no longer names how in its reading list.
  • architect: Phase A and the scrap step trace the subsystems. The architect runners line and the interface-depth text are unchanged.
  • no-comments and its comment-sicko agent trace the symbol or run /why.
  • setup-pstack: the how explorer and how explainer roles are gone from the rule shape. A rule file that still has them loses those lines on the next /setup-pstack run, like any retired role.
  • Benny automation skills: the triage and reproduce skills trace the path themselves, and how leaves the list of pstack skills the target repository must resolve. They still use why.
  • Docs: the README and guide chapters 02, 03, 04 and 10 describe the new flow.

/teach still runs /how and /why, and the poteto-help catalog still lists /how.

What a user sees: /how returns one subagent's explanation with no fixed headings and no per-role models. A /poteto-mode run no longer spends a step on /how before it designs or debugs.

Checks:

  • The same three script checks as in section 1 pass.
  • A search of pstack/ finds no reference to a deleted file or to a how role, apart from the retired-role example in setup-pstack. Version is 0.15.41.

6. Defaults route to Fable 5.2, Opus 5.5 and Gemini 4 Carbon

pstack's default models change. The role labels and the four budgets stay, so an existing pstack-models.mdc rule keeps working. This changes which model every default user runs.

Role Old default New default
hardest tasks claude-opus-5-5-xhigh claude-fable-5-2-max
feature, refactoring, bug-fix, perf-issue, hillclimb grok-4.7-xhigh-fast claude-opus-5-5-xhigh
judgment and prose, reflect judgment, divergent, synthesizer claude-opus-5-5-xhigh unchanged
swarm workers, reflect tooling, verification, trivial mechanical edits grok-4.7-xhigh-fast claude-opus-5-5-high-fast
arena runners, arena cross-judge pool, architect runners, interrogate reviewers claude-opus-5-5-xhigh, grok-4.7-xhigh-fast claude-fable-5-2-high, claude-opus-5-5-high-fast, gemini-4-carbon-high

Grok 4.7 is no longer a default anywhere.

Rules that changed with the defaults:

  • Levels. Role defaults were all xhigh. They are now xhigh or high, and max for hardest tasks. Gemini has no xhigh, max or fast slug, so a Gemini subagent runs at high. Fable 5.2 has no fast slug, so an urgent Fable 5.2 subagent runs at standard speed.
  • Families. Fable (claude-fable-*) is its own family, apart from the other Claude models, so a panel with a Fable seat and an Opus seat stays diverse. A leading cursor- is dropped before families are compared. Any other slug's family is its text before the first hyphen.
  • Arena's cross-judge. The judge is never from the parent's vendor, and Fable and the other Claude models are one vendor. With the default pool, a Claude parent gets gemini-4-carbon-high, and arena skips the judge if that is refused. A parent from another vendor gets claude-fable-5-2-high first.
  • Budgets. Each role now has a target effort: its default's effort moved by the budget. large keeps it, unlimited raises it one step, medium lowers it one step, and small lowers it two. For a role whose default is xhigh, the result is the same as before. The change keeps large equal to the defaults now that the defaults are not all xhigh, and a re-run still gives the same file.
  • Trivial mechanical edits read the swarm workers line. pstack has no separate role for them, and this change adds none.

Existing rule files: /setup-pstack writes a new # defaults: 2 line. A rule without that line comes from an older setup, which wrote a line for every role, so some of its values are old defaults and not choices. On the next /setup-pstack run:

  • A grok-* value in a role that used to default to Grok 4.7 takes the role's new default.
  • A claude-opus-* value in hardest tasks takes the new default.
  • A panel list of exactly one claude-opus-* entry and one grok-* entry takes the new three-seat list. In any other panel list, each grok-* entry becomes claude-opus-5-5-high-fast.
  • Every other value stays, including a grok-* value in a role that never defaulted to it, and every inherit-parent or auto value.

Until that run, /poteto-mode keeps honoring the old lines and tells the user once per session to rerun /setup-pstack.

Where Grok is still named:

  • setup-pstack step 2 names the grok-* pattern once. It needs that to recognize an old default in an existing rule file.
  • "Grok Bot" in /make-bot-ui and the three places that list that skill is a product name, not a model choice. This change does not touch it.
  • scripts/check-pstack-docs.mjs and scripts/generate-pstack-harness-trees.mjs keep their existing guard patterns. With this change, the docs check fails if a Grok slug comes back into the README or the guide.

Files:

  • setup-pstack: rule shape, budget step, old-default step.
  • poteto-mode: the defaults sentence and "Match the model to the job".
  • arena, architect, swarm, reflect: default slugs, families and the cross-judge rule.
  • Playbooks feature, bug-fix, perf-issue, hillclimb, refactoring, multi-phase-plan: the "(default ...)" lines.
  • docs/harness/claude-code.md: the non-Claude slug row and the default mapping. In Claude Code, claude-opus-5-5-high-fast is opus high, claude-fable-5-2-max is fable max, and gemini-4-carbon-high is inherit-parent.
  • README and guide chapters 01 and 04.

Checks:

  • The same three script checks as in section 1 pass.
  • CI's test command, node --test pstack/skills/*/scripts/test/*.test.ts: 270 pass, 0 fail, 70 skipped. The model change touches no script.
  • A search of pstack/ and harness/pstack/ for Grok finds only the places listed above.
  • No subagent has run on the new slugs from this branch. Version is 0.15.42.

cursoragent and others added 12 commits November 14, 2026 01:48
- sharpen: the rewritten prompt now asks the agent to prototype any choice
  an experiment can settle and recommend an option from what it observed.
  Product and preference choices stay with the user. The README and guide
  describe the new last line.
- reflect: the synthesizer's example of a detail that drifts is now a
  function rename instead of a model-name rename.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
When SANY's XML run fails, check() reruns SANY in plain mode and reports
that run's error section, capped at 40 lines. The message used to be the
tail of the XML run, which is a Java stack trace. Three tests cover an
unknown name, a parse error, and twelve unknown names.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
…arts

An arrow from outside a parent state could run through the parent's title.
The Mermaid backend now ends such an arrow at the parent's border, and the
parent's start marker goes on to the state inside. A lone arrow keeps its
label. When one state leads to several states inside, each label moves to
the start marker's arrow. A parent that two or more states lead into keeps
its direct arrows.

Tests share one reader of a chart's arrows in test/drawn.ts. Two chart
tests and one render test cover the new drawing.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
why is now one file with a 48-word body. It tells the agent to answer from
the commits and PR discussions behind the code, to look further when that
trail is thin, to cite what the authors wrote, to mark inferences, and to
say what the record does not answer. The investigators, the synthesizer,
the model routing, the output format, and the files under references/ are
deleted.

- recall starts its own investigator for each source instead of borrowing
  why's investigators and source playbooks.
- blast-radius and teach stop pointing at parts why no longer has.
- poteto-mode keeps its route to why and drops it from the lists of skills
  that set their own subagent type or model roles.
- setup-pstack drops the two why roles.
- poteto-help, the README and the guide describe the new why.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
how is now one file. It starts one read-only subagent with a short brief
and presents the output. The model routing, complexity gate, explorers,
output format and both prompt templates are deleted.

poteto-mode no longer runs how. The trigger line is gone. The
Investigation, Feature, Bug fix, Hillclimb, Perf issue and Refactoring
playbooks, architect, no-comments and its comment-sicko agent tell the
agent to trace the code itself. setup-pstack drops the two how roles. The
Benny automation skills trace directly and no longer require how. teach
and the poteto-help catalog keep how. The README and guide describe the
new flow.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
The role labels and the four budgets stay. Only the defaults change.

- hardest tasks: claude-fable-5-2-max.
- feature, refactoring, bug-fix, perf-issue, hillclimb: claude-opus-5-5-xhigh.
- swarm workers, reflect tooling, verification and trivial mechanical
  edits: claude-opus-5-5-high-fast.
- arena runners, arena cross-judge pool, architect runners and interrogate
  reviewers: claude-fable-5-2-high, claude-opus-5-5-high-fast,
  gemini-4-carbon-high.

Fable is its own family, apart from the other Claude models, and the two
are one vendor for arena's cross-judge. Gemini tops out at high, and
Fable 5.2 and Gemini have no fast slug. Each budget now moves a role's
default effort by steps, so large still equals the defaults.

setup-pstack writes a defaults marker line. A rule without it is from an
older setup, and its old default values take the new defaults on the next
run. Other values stay as the user's choice.

Grok 4.7 is no longer a default. setup-pstack names its slug pattern once,
to recognize an old default in an existing rule file.

Co-authored-by: lauren <poteto@users.noreply.github.com>
Co-authored-by: lauren <poteto@users.noreply.github.com>
@poteto
poteto merged commit bf0c7cd into main Nov 15, 2026
1 check passed
@poteto
poteto deleted the lauren/pstack-model-routing-46c0 branch November 15, 2026 03:36
SmailG added a commit to SmailG/claude-cursor-plugins that referenced this pull request Nov 15, 2026
…ne-file why and how, statechart fixes, sharpen and reflect text (cursor#1236)

chore: sync with cursor/plugins — feat(pstack): new default models, one-file why and how, statechart fixes, sharpen and reflect text (cursor#1236)
shiro-el added a commit to shiro-el/pstack-shiro that referenced this pull request Nov 15, 2026
Takes upstream cursor#1236: why and how cut to one file each, poteto-mode no longer routes to how, new default models (Fable 5.2, Opus 5.5, Gemini 4 Carbon), sharpen asks for prototypes, and statechart fixes.

Conflicts kept the fork's model-to-job table and hc routing, with upstream's structure on top:
- poteto-mode: the table drops the how explorer, how explainer, why investigators, and why synthesizer roles, and the routed-skill list drops how and why. The family rule now says Fable and the other Claude models are one Claude family, the arena skill's vendor.
- Playbooks keep the fork's spec-driven text and no default slugs, and trace the subsystem directly instead of running how.
- setup-pstack takes upstream's text and defaults and keeps the debate panel and debate judge pool roles.
- The Claude Code adapter takes upstream's default slug mapping and keeps the hc setup line.
caelaxie added a commit to pstack-ports/droid-pstack that referenced this pull request Nov 16, 2026
Ports cursor/plugins#1236 (bf0c7cd) onto the Droid seats.

Co-authored-by: Kaitian Xie <caelaxie@users.noreply.github.com>
caelaxie added a commit to pstack-ports/agy-pstack that referenced this pull request Nov 16, 2026
<!-- CURSOR_AGENT_PR_BODY_BEGIN -->
## summary

Ports upstream
[cursor/plugins#1236](cursor/plugins#1236)
(`pstack/` only) from `0.15.36` at
[`c742dd61706fbacc1b95f1a20b7b988cf096ce6c`](cursor/plugins@c742dd6)
to `0.15.42` at
[`bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7`](cursor/plugins@bf0c7cd).

The monorepo range also contains commits outside `pstack/`. Those are
not in this port. `harness/pstack/claude-code/` is not ported. This repo
does not ship `docs/harness/`, so the Claude Code adapter edit in #1236
has no file here.

### what landed

- `why` is one short file. Its references, investigators, synthesizer,
and model roles are gone. `recall` starts its own investigator per
source. `blast-radius` and `teach` no longer point at parts `why` no
longer has.
- `how` is one file and one read-only subagent. Poteto-mode no longer
runs `how` on the nontrivial-change trigger. Investigation, Feature, Bug
fix, Hillclimb, Perf issue, Refactoring, architect, no-comments, and
comment-sicko trace the code themselves. Benny traces directly and no
longer requires `how`. `teach` and `poteto-help` still catalog `how`.
- `sharpen` asks the rewritten prompt to prototype any choice an
experiment can settle. `reflect`'s drifting-detail example is a function
rename.
- Statechart: when SANY XML fails, `check()` reruns SANY in plain mode
and reports that run's error section, capped at 40 lines. Mermaid arrows
from outside a parent state end at the parent's border. The start marker
still continues to the inner state.
- Version `0.15.42`.

### antigravity-only adaptations

- Version bump is only in `.antigravity-plugin/plugin.json`. This port
has no root `plugin.json` and no `.cursor-plugin` manifest.
- `how` spawns `invoke_subagent` with `TypeName: "research"`
(read-only). `recall`'s per-source investigators use `TypeName: "self"`,
because `research` has no MCP tools. Existing path, tool, and transcript
adaptations stay.
- `invoke_subagent` still takes `pro`, `flash`, `flash_lite`, or
`inherit`, so the new slugs map onto those tiers. Code delegates
(feature, refactoring, bug-fix, perf-issue, hillclimb), prose, and
judgment are `pro`. The hardest-tasks role stays `pro`: Antigravity has
no tier above `pro`, so `claude-fable-5-2-max` and
`claude-opus-5-5-xhigh` are the same tier. Swarm workers and reflect
tooling stay `flash`. Review and design panels stay `pro, flash`. There
is no third tier for the upstream fable / opus / gemini panel.
- Budgets move each role from its own default on `flash_lite` < `flash`
< `pro`. `large` keeps the default. `unlimited` raises one step, capped
at `pro`. `medium` lowers one step. `small` lowers two. `inherit`,
`inherit-parent`, and `auto` stay. `setup-pstack` writes `# defaults:
2`. A rule without that line is an older setup. It names `grok-*` once
as an old slug. Old `flash` defaults on the code-delegate roles take
`pro`. Other chosen values stay. The `how` and `why` roles are dropped.
- `arena`, `swarm`, and `reflect` already stated those tiers, so their
default lines did not need new wording. The reflect synthesizer example
did change.

## harness proof

Head `210afbe6c9a7b1732946a5c4e882d4ecf6bf66be`. `agy` 1.5.9. Commands
ran on this commit. Statechart sources are unchanged from
`d045f325ce86b8eaf68ca40a38e84395e448e1cb`, where the suite was 340
tests, 270 pass, 0 fail, 70 skipped (TLC corpus runs skip without
`TLA2TOOLS`).

```
$ agy --version
1.5.9

$ agy plugin validate --strict .
Validating pstack (.antigravity-plugin/plugin.json)
  skills 61 · agents 2 · rules 0 · hooks 0 · mcpServers 0 · sidecars 0
✔ Validation passed

$ agy plugin install https://github.com/pstack-ports/agy-pstack/tree/cursor/port-pstack-0-15-42-00a0
Cloning plugin from https://github.com/pstack-ports/agy-pstack.git...
  [ok]    pstack
          ✔ skills      : 61 installed
          ✔ agents      : 2 installed (comment-sicko, poteto-agent)

$ agy plugin list
Installed plugins:
- pstack (enabled, 0.15.42)
    Source:     antigravity
    Components: 61 skills, 2 agents

$ git -C "$AGY_HOME/.gemini/config/plugins/pstack" rev-parse HEAD
210afbe
```

Loaded skills include `how`, `why`, `recall`, `setup-pstack`, `sharpen`,
`reflect`, `poteto-mode`, `poteto-help`, `statechart`, `teach`,
`blast-radius`, `architect`, `arena`, `no-comments`, and `swarm`.
Benny's three skills are in the installed tree under
`automations/benny/skills/` with valid frontmatter. They are not
registered slash skills. Strict validate of the installed copy passed.
The repo CI structural check printed `Validated 61 skills, 2 agents, and
23 playbooks.`

```
$ bash -n skills/poteto-mode/scripts/worktree-audit.sh
$ bash -n skills/show-me-your-work/scripts/log.sh

$ bun run typecheck   # skills/poteto-mode/scripts, bun 1.4.4
$ bun run test        # 62 pass, 0 fail
```

<!-- CURSOR_AGENT_PR_BODY_END -->

<div><a
href="https://cursor.com/agents/bc-a0f24530-c1c1-55bf-889a-619c03d400a0?cursor_ref=pr_footer&cursor_cta=open_in_web"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-web-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-web-light.png"><img
alt="Open in Web" width="114" height="28"
src="https://cursor.com/assets/images/open-in-web-dark.png"></picture></a>&nbsp;<a
href="https://cursor.com/background-agent?bcId=bc-a0f24530-c1c1-55bf-889a-619c03d400a0&cursor_ref=pr_footer&cursor_cta=open_in_cursor"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://cursor.com/assets/images/open-in-cursor-dark.png"><source
media="(prefers-color-scheme: light)"
srcset="https://cursor.com/assets/images/open-in-cursor-light.png"><img
alt="Open in Cursor" width="131" height="28"
src="https://cursor.com/assets/images/open-in-cursor-dark.png"></picture></a>&nbsp;</div>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Kaitian Xie <caelaxie@users.noreply.github.com>
kaneyxx added a commit to kaneyxx/omc-pstack-fy that referenced this pull request Nov 16, 2026
- pstack moves from v0.15.36 (1ba91fc) to v0.15.42 (bf0c7cd), all
  from cursor/plugins#1236.
- New defaults. Code roles, judgment, and reflect judgment run on
  `opus xhigh`. `hardest tasks` runs on `fable xhigh`, the owner's
  choice over upstream's `max`. `swarm workers` and verification run on
  `opus high`, because the Agent tool has no fast mode. The four panels
  default to `fable high, opus high, sonnet high`. An Agent spawn runs
  only Claude models, so sonnet takes the Gemini seat and every default
  panel keeps three distinct models.
- `reflect tooling` defaults to `sonnet high`, the owner's choice over
  upstream's fast Opus. The Judgment and Divergent lenses run on
  `opus xhigh`, so an Opus Tooling lens would run all three of reflect's
  lenses on one model. A targeted edit in `80-personal.json` changes the
  lens table, and the setup-pstack rule shape carries the same default.
  An older `sonnet` value for the role still migrates to the new
  default.
- Budgets follow upstream's per-role target effort. Each role's default
  effort moves by the budget, so `large` keeps every default.
- setup-pstack writes `# defaults: 2` and migrates an older rule. A
  `sonnet` value in a role that used to default to sonnet, `opus` in
  `hardest tasks`, and a panel of exactly one `opus` and one `sonnet`
  entry take the new defaults. A `sonnet` entry in any other panel list
  stays, because sonnet is still a default panel seat and changing it
  could shrink the panel. poteto-mode honors an old rule and tells the
  user once per session to rerun setup-pstack.
- arena's judge still never runs on the parent's model and tries other
  vendors first. It now says that Fable and the other Claude models are
  one vendor, so a Claude judge is a same-vendor judge.
- `how` is one read-only subagent and `why` is one short file. The four
  `why` rules, both old `how` rules, and the `gpt-4` synthesizer rule
  target removed text and are deleted. The `how explorer`, `how
  explainer`, `why investigators`, and `why synthesizer` roles are gone.
  New rules translate `how`'s `readonly: true` and recall's agent-mode
  investigators to `general-purpose` agents.
- Upstream reworded three passages that rules matched, so those rules
  match the new text. The comment-sicko rule keeps `why` and drops
  `how`. The poteto-help rule drops `/why` from the list of skills that
  spawn subagents and says only some use per-role models. The
  multi-phase-plan rule drops `how` from the docs a plan names. Upstream
  also reworded the README passage on old pins, so the README note in
  `80-personal.json` anchors on the new last sentence.
- A global substitution maps `claude-opus-5-5-<effort>-fast` to
  `opus <effort>`. The `grok-4.7` substitution is gone, because upstream
  no longer names Grok as a model.
- `docs/claude-code-mapping.md` gains the new Models rows, the budget
  rule, the old-default migration, and the retired roles. A separate
  row records the `reflect tooling` default and its reason. The row for
  three-seat panels says `inherit` took no seat because it repeats a
  seat's model under an opus or fable parent. The README's `pstack-ask`
  example keeps a three-model panel.
- The README's Personal choices section, and the two Chinese READMEs,
  disclose that `hardest tasks` defaults to `fable xhigh` and
  `reflect tooling` to `sonnet high`.
elricli added a commit to elricli/skills that referenced this pull request Nov 16, 2026
Simplify how and why, update their callers, and remove retired prompts.
Port statechart diagnostics and parent-state entry rendering fixes,
alongside sharpen and reflect updates.

Retain runtime model discovery and disable-model-invocation: true.

Upstream: cursor/plugins@bf0c7cd
Source: cursor/plugins#1236
Validation: node scripts/check.mjs; 373 statechart tests passed with
TMPDIR=/private/tmp, TLA2TOOLS and JAVA set.
PlagueFPS added a commit to PlagueFPS/skills that referenced this pull request Nov 16, 2026
Ports cursor/plugins#1236 onto the T3-adapted pstack skills:
- how and why collapse to one-pass skills; retire their role lines
- setup-pstack moves role defaults and per-role effort to upstream's new scheme
- statechart enters parent states through their border; SANY errors cite line and column
- sharpen asks for prototypes; reflect example swapped
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants