Skip to content

pstack 0.15.36 → 0.15.46: what Lauren changed #10

Open

Description

@github-actions

TL;DR

  • Model routing changed twice and settled on Opus 5.5. Code, prose and judgment default to claude-opus-5-5-xhigh-fast, and no role runs above xhigh. Fable 5.2 runs outside a panel only when you name it or after two failed Opus 5.5 xhigh runs. Panels seat Fable 5.2, Opus 5.5 and Gemini 4 Neon at gemini-4-neon-xhigh.
  • Older pstack-models.mdc rules keep their old lines until you rerun /setup-pstack, which migrates them. /poteto-mode reminds you once per session when the rule predates the # defaults: 3 marker.
  • Many skill texts got much shorter. /why, /how, /reflect and /architect are now short files with their reference files deleted, and the /poteto-mode playbooks trace code directly instead of running /how.
  • Every agent /poteto-mode starts now reads a new poteto-mode/worker.md first, so started agents see the operator's rules without a slash command loading them.
  • /statechart now has a path for native Windows, gives line and column for unreadable TLA+ specs, and draws parent states without arrows through their titles. The Windows work was not run on Windows.

Models and routing

  • Defaults moved in two steps. #1236 dropped Grok 4.7 as a default everywhere and routed to Fable 5.2, Opus 5.5 and Gemini 4 Carbon. #1249 then made Opus 5.5 first. Code, prose and judgment default to claude-opus-5-5-xhigh-fast, and Opus 5.5 always runs its -fast slug. No role runs above xhigh. The hardest tasks role is retired, and Fable 5.2 outside a panel runs only when the user names it or after two failed Opus 5.5 xhigh runs. #1236, #1249
  • The Gemini panel seat moved from Gemini 4 Carbon to Gemini 4 Neon (#1283), then from gemini-4-neon-high to gemini-4-neon-xhigh (#1284). This applies to arena runners, the arena cross-judge pool and interrogate reviewers. A Gemini seat runs at gemini-4-neon-xhigh unless its line names another level. Neon has no max or fast slug, so the seat runs at its top slug. #1283, #1284
  • /setup-pstack was reworked several times. It now writes # defaults: 3 and offers large, medium and small budgets on a ladder that tops out at xhigh. It retires the hardest tasks, architect runners and both reflect roles, along with the why and how roles. #1236, #1249
  • The unlimited budget is removed, not capped, because it could only mean max. An old unlimited rule becomes large. #1249
  • Migration of saved rules:
    • A saved gemini-4-neon-high panel entry, or any non-Neon gemini-4-* entry, becomes gemini-4-neon-xhigh. Other Neon levels a user picked stay.
    • The Gemini change keeps the # defaults: 3 marker, so /poteto-mode does not ask anyone to rerun /setup-pstack for it. Users with a saved rule move to Neon only when they rerun it.
    • #1283, #1284
  • poteto-agent pins claude-opus-5-5-xhigh-fast. #1249
  • The Claude Code adapter doc (docs/harness/claude-code.md) was updated in each of these four PRs to follow the new slugs. #1236, #1249, #1283, #1284

Changes to existing skills

  • /why is one file with a 48-word body that names no data source or tool. It answers from the record, meaning the commits and PR discussions behind the target and each later change, and looks at 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, synthesizer, output format and all 12 files under why/references/ are deleted. Answers are shorter, with no fixed headings. #1236
  • /recall no longer borrows the why investigators. It starts its own investigator for each source. #1236
  • /how is one file that starts one read-only subagent with a three-sentence brief and presents the output. The model routing, complexity gate, explorers, output format and both prompt templates are deleted. /teach still runs /how and /why. #1236
  • /reflect is one file with no subagents, and its reference files are deleted. #1249 Before that, its drift example changed from a model-name rename to a function rename (parseConfig to loadConfig), so the example does not age. #1236
  • /architect is a short brief that sends large or contested designs to /arena. Its reference files are deleted. The cost is that the interface-depth wording from the longer text is gone. #1249
  • /arena has shorter runner and judge steps. #1249
  • /sharpen now ends by having 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, every choice was left open. #1236
  • /unslop and /technical-writing are shorter. #1249
  • /blast-radius, /teach and /no-comments no longer point at parts of why and how that are gone. /no-comments now says to trace the symbol or run /why. #1236
  • Fable (claude-fable-*) counts as its own family, apart from other Claude models, so a panel with a Fable seat and an Opus seat stays diverse. #1236

Playbooks and principles

  • /poteto-mode has ten short triggers and shorter Subagents, reply and Playbooks sections. The Comments section is gone. The "nontrivial change, run how" trigger is gone, so a run no longer spends a step on /how before it designs or debugs. #1236, #1249
  • The new poteto-mode/worker.md is read first by every agent /poteto-mode starts. #1249
  • The Orchestrate playbook is removed. Program-scale work routes to /figure-it-out. #1249
  • The Investigation, Feature, Bug fix, Hillclimb, Perf issue and Refactoring playbooks say "trace" where they said to run how. #1236
  • The Opening a PR, Babysit and Multi-phase plan playbooks are much shorter. #1249
  • 14 principle-* skills are shortened. #1249
  • poteto-mode/scripts/check-plan.mjs and its test are deleted, since nothing runs them now. #1249
  • In the Benny automation skills, triage and reproduce trace the path themselves. how leaves the list of pstack skills the target repository must resolve, and they still use why. #1236

Agents

  • comment-sicko traces the symbol or runs /why, matching /no-comments. #1236
  • poteto-agent pins claude-opus-5-5-xhigh-fast. #1249

Fixes

  • /statechart ran /bin/sh for its group kill and janitor, so it could not run on native Windows. TLC there ends lines with \r\n, and the readers split on \n only. Marker matching failed, a broken rule produced "TLC exited 12 ... but its output names no rule that broke", and passing specs could report a wrong reason or leave raw @!@!@ markers in tlc.out. #1292
    • A shared splitLines helper now splits on \r?\n.
    • kill.ts and janitor.ts replace the shell. The kill is process.kill(-pid, "SIGKILL") on macOS and Linux and taskkill /t /f on Windows.
    • java and mise start with windowsHide, the worker gets TEMP and TMP, and Java gets UTF-8 stdout and stderr flags.
    • SKILL.md says to run node --disable-warning=ExperimentalWarning <skill dir>/scripts/statechart.ts on native Windows, because the launcher is a POSIX shell script.
    • On macOS and Linux the group kill, cleanup and output stay the same.
  • /statechart now reports the spec's line and column when SANY cannot read a TLA+ spec. Before, it showed the tail of SANY's XML run, a Java stack trace. check() reruns SANY in plain mode and reports from the first *** Errors, *** Abort messages or *** Parse Error line, capped at 40 lines. It falls back to the old tail if the rerun cannot start or has no error section. #1236
  • /statechart charts now enter a parent state through its border. Before, an arrow from outside to a state inside could cross the parent's title, which Pele 1.9.0 places after layout. The parent's start marker now carries the arrow on to the inner state, and a lone arrow keeps its label. #1236
    • It covers parents with one way in. A parent that two or more states lead into still draws direct arrows, and Pele can still route one through the title.
    • Each such parent now shows a start marker, but the chart states the same steps. Of the corpus charts, only Login changes.

Docs

  • The README and guide chapters 01 to 07, 09 and 10 follow the new routing, skill and playbook text. #1236, #1249, #1283
  • The docs check now fails if a Grok slug comes back into the README or guide. #1236

How she did it

  • She compared the two Gemini 4 Neon levels blind. Both caught every planted flaw. gemini-4-neon-xhigh found more real problems nobody planted and took about 50% longer per run. She moved the panel seat to xhigh because panels are where pstack pays for depth. #1284
  • For the Windows fix, the 34 new tests all fail on main and pass with the change.
    • With real TLC made to print \r\n, 8 of 33 corpus cases pass on main and 33 of 33 pass on the branch.
    • Everything ran on macOS. She states that nothing ran on Windows, so the Windows claims are inferred. #1292
  • For the SANY change, the three new tests fail without the run.ts change. The full statechart suite with the pinned tla2tools.jar was 50 of 50 on the corpus test and 373 of 373 after the parent-border change. #1236
  • She states that no subagent had run on the new Fable 5.2, Opus 5.5 and Gemini slugs from the #1236 branch. #1236
  • The test run in #1249 had one failure, the statechart symlink launcher test, which she says fails the same way on main on macOS. #1249
  • Searches backed the cleanups. #1249 searched pstack/ for max and found no routing above xhigh. #1283 searched pstack/ and harness/pstack/ for carbon and found 0 matches. #1236 searched for references to deleted why and how files and roles and found only the retired-role example in /setup-pstack.

Cursor diff 0f77cd5...0b46ec3 · 5 PRs · summarised by claude-sonnet-5-5

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions