Repository navigation
chore(p3stack): sync with cursor/plugins at 424f2f6 - #26
Merged
Merged
Conversation
* feat(pstack): port recent skill updates - poteto-mode: restate the user's goal in the first message, briefs, PRs, and final reports. Background steers and resumes. Chain fallbacks when the Task tool rejects a slug. Apply Explain the Number before measuring. - Autopilot playbooks: block a PR only for a problem a run or a wrong doc line shows, push each finished unit, give wording-only changes no verifier, and make Autopilot-stack Autopilot-full without the merge. - interrogate: adversaries prove each finding with a run, and the lead keeps only findings a rerun confirms. Drop the reviewer and rubric references. - arena: drop or replace a rejected runner seat, and walk the cross-judge pool by family before falling back to the parent model. - ketchup-video: subtitle each slide and measure holds from the file. - principle-explain-the-number: write the decision before you run. - orchestrate: copy store files to local disk once. - Update the README and guide where these changes made them wrong. * chore(pstack): bump to 0.15.35 Co-authored-by: lauren <[email protected]> * fix(pstack): rebase stacked PRs onto their parent branch Co-authored-by: lauren <[email protected]> * pstack: never pick a cross-judge from the parent's family Arena's judge now comes only from a family other than the parent's. An auto or inherit-parent entry never judges. When no entry from another family runs, the agent skips the judge and picks the base alone. Setup describes the cross-judge pool the same way. Co-authored-by: lauren <[email protected]> * pstack: describe the arena judge fallback in the design guide The guide now says the judge comes from a different model family, and that the coordinator skips the judge when no model from another family can run. Co-authored-by: lauren <[email protected]> * pstack: let arena Phase D and Outputs handle a skipped judge Phase D compares with the cross-judge only when Phase C ran one, and the synthesis note says when Phase C skipped the judge. Co-authored-by: lauren <[email protected]> --------- Co-authored-by: Cursor Agent <[email protected]>
…1043) * feat(pstack): make Orchestrate agentic and delete the orch scripts The Orchestrate playbook now runs the program with the coordinator's own tools: Task for workers, the Project tracker and store for state, and PR and CI event subscriptions where the harness has them. It no longer calls the orch CLI, so the orch scripts and their test entry are removed. The overnight guide's description of the coordinator matches the new playbook. * chore(pstack): bump to 0.15.36 --------- Co-authored-by: Cursor Agent <[email protected]>
# Conflicts: # p3stack/.cursor-plugin/plugin.json # p3stack/docs/guide/10-recipes-and-pitfalls.md # p3stack/plugin.json # p3stack/skills/arena/SKILL.md # p3stack/skills/interrogate/SKILL.md # p3stack/skills/poteto-mode/SKILL.md # p3stack/skills/poteto-mode/playbooks/autonomous-run.md # p3stack/skills/poteto-mode/playbooks/autopilot-full.md # p3stack/skills/poteto-mode/playbooks/autopilot-stack.md # p3stack/skills/poteto-mode/playbooks/multi-phase-plan.md # p3stack/skills/poteto-mode/playbooks/orchestrate.md # p3stack/skills/setup-pstack/SKILL.md
Upstream 424f2f6 deleted p3stack/skills/poteto-mode/scripts/orch, so the bun test filter no longer has an orch suite to run. Co-Authored-By: Claude Opus 5.5 <[email protected]>
- interrogate: hand adversaries a patch of uncommitted changes, since a delegate_task worktree starts from the last commit. Adversaries write repros, so they spawn as role test, not the read-only review role. - arena: say the two-runner minimum overrides the one-seat-short panel rule. - poteto-mode: a rejected target falls back to inherit only when the seat is not a panel seat and the calling skill doesn't drop or skip it. Co-Authored-By: Claude Opus 5.5 <[email protected]>
…fallback Co-Authored-By: Claude Opus 5.5 <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Upstream pstack shipped 2 commits since the last sync. They make
/interrogatekeep only findings a run proves, and they replace the Orchestrate playbook and itsorchCLI with a short agentic playbook. This PR brings both into p3stack and ports the conflicting parts to T3 Code.What changed
inheritseat never judges. Dropped seats are refilled so at least two runners remain, which overrides poteto-mode's rule to run a panel one seat short.delegate_taskin its own worktree withrole: "test", and it applies a patch of the uncommitted changes, because a worktree starts from the last commit. Every follow-up round is still a fresh task.delegate_task, the run store,watch_pull_request,max parallel writers, and a verifier from another lab.delegate_tasktarget. That rule never falls back toinheritfor a panel seat.orchfrom the bun test command in CI and in sync step 6, because upstream deleted that suite.Needs a human
/interrogateworks. Reviewers who sorted findings into Act on, Consider, Noted, and Dismissed are replaced by adversaries who must prove each finding with a run, and the reviewer prompt, rubric, and lead-judgment references are deleted. Adversaries now write repro files, so I spawn them withrole: "test"and give each a patch of the uncommitted changes. A reviewer should confirm the patch handoff.Lines not from upstream
p3stack/docs/guide/04-design.md: "different lab" in place of "different model family", to match p3stack's lab terms.p3stack/skills/arena/SKILL.md: the runner top-up and the cross-judge lab rules, in T3 seat and role terms.p3stack/skills/interrogate/SKILL.md:delegate_taskspawn, the patch handoff for uncommitted changes, and p3stack's follow-up-round rule.p3stack/skills/poteto-mode/SKILL.md: upstream's rejected-slug fallback chain, as a rule for a rejecteddelegate_tasktarget.p3stack/skills/poteto-mode/playbooks/autonomous-run.md: upstream's "that support the goal", merged into p3stack's line withoutAskQuestion.p3stack/skills/poteto-mode/playbooks/autopilot-full.md: upstream's rebase, push, and wording-only verifier changes, merged into p3stack's T3 lines.p3stack/skills/poteto-mode/playbooks/autopilot-stack.md: "heartbeat tick" in place of "loop tick".p3stack/skills/poteto-mode/playbooks/multi-phase-plan.md: upstream's playbook picker, verifier, and blocking-rule changes, merged into p3stack's lines with<p3stack>paths.p3stack/skills/poteto-mode/playbooks/orchestrate.md: T3 rewrite of the Task, Project store, cloud VM, fork, and model family lines.p3stack/skills/setup-pstack/SKILL.md: upstream's cross-judge pool rule, in seat and lab terms.Blast Radius
Users of
/interrogate,/arena, Orchestrate, and Autopilot-stack get the new workflows. TheorchCLI is gone, so any in-flight Orchestrate program that called it must switch to the new playbook.Verification
scripts/sync-upstream.shfinished with checks passing.scripts/check.shpasses with 0 files left to port.scripts/check-docs.mjsreports 0 problems./interrogateran with Claude Opus 5.5, GPT-6.1-Sol, and Muse Spark 1.3. The four port findings from Claude and the one from Muse are fixed in this branch. GPT found nothing.🤖 Generated with Claude Code