Repository navigation
Closed
Description
The biggest change is that Orchestrate no longer uses a bookkeeping CLI and store. The coordinator now runs the program with its own agent tools, and the orch scripts are deleted. /interrogate was also rewritten so that adversaries must prove each finding with a run or a repro. Both PRs port updates from Lauren's source skills into the public repo and keep internal references out.
Changes to existing skills
/poteto-mode: in its first message on a task, the agent now tells the user their goal, the problem, and the smallest change it expects. Subagent briefs and the PR's## Whyrestate that goal. "Explain the Number" now also applies before you measure. A new fallback chain applies when the Task tool rejects a model slug. #986/poteto-modeplaybooks: autonomous-run, autopilot-full, autopilot-stack, babysit, multi-phase-plan, opening-a-pr and orchestrate got the matching updates. #986- Autopilot-stack is now Autopilot-full without the merge, and a PR stacks only on a PR that it builds on.
- A subagent that opens a PR runs
/deslopand/no-comments. - An orchestrator copies each store file to local disk once.
- Orchestrate playbook (
orchestrate.md): rewritten. #1043- The coordinator decides which agent does each piece of work, checks what each one produces, and never writes code.
- It spawns workers with
Taskand keeps the program's state in the Project tracker (notes.md) and the Project store. - It watches PRs with PR and CI event subscriptions when the harness has them. It calls no scripts and runs no polling loops.
- A new Topology section sets these rules:
- one owner per PR;
- a fresh worker for each independent task;
- sub-coordinators only for areas with many parallel workers, at most three levels;
- forks for work that needs a worker's built state;
- a verifier from a different model family for each PR;
- one pilot unit before fan-out.
- A new Supervision section sets these rules:
- each kickoff names the evidence that proves the work is done;
- the coordinator reads the diff or the artifact, not only the summary;
- it sends back scope drift;
- it checks a PR's live state with
gh pr viewandgh pr checks(or the forge's equivalents) before it reports; - it retries stalled workers by cause;
- it asks the user only for real decisions or blockers.
- The Reply line is shorter: what landed, each PR's live state and verdict, what was dropped, and the decisions waiting on the user.
- Workers still default to
environment: "cloud".
/interrogate: rewritten. Adversaries must prove each finding with a run or a repro, and the lead keeps only the findings that a rerun proves. The four reference files are deleted. #986/ketchup-video: the video now has subtitles, with one short line per slide. The render step burns them in and checks the rendered file before it is embedded. #986/arena: #986- When the Task tool rejects a runner model, the skill tries a valid model of the same family. If none runs, it drops that runner and adds runners until there are two.
- The judge never comes from the parent's model family, and an
autoorinherit-parententry never judges. - When no model from another family runs, the agent skips the judge and picks the base alone.
/setup-pstack: the cross-judge pool line is now an order of preference, and Arena tries only its entries from a family other than the parent's. #986
Principles
principle-explain-the-number: a new rule says to write down the result that would change a decision before you run the measurement. #986
Scripts
poteto-mode/scripts/orch/(orch.ts,store.ts,orch.test.ts) is deleted, because the new Orchestrate playbook does not use it. #1043poteto-mode/scripts/package.json: thetestscript no longer runsorch.commanderandbun-typesstay becausewatch-prstill uses them. #1043store.ts:acquireLockdeclaredholderwithout an unused default value. This file was deleted in the next PR. #986
Docs and release
- The README, the guide, and
/poteto-helpdescribe the new interrogate, autopilot-stack and ketchup-video behavior. The guide also describes the new arena judge rule. #986 - Guide page 07 describes the new Orchestrate coordinator and no longer calls it "deliberately heavy machinery". #1043
- The generated Claude Code harness tree is regenerated in both PRs. pstack goes to 0.15.35 in #986 and 0.15.36 in #1043.
How she did it
- Both PRs port from the source skills and leave out internal-only model and tooling references. The public model defaults do not change. In #1043, internal tool names, commands and paths became public equivalents or generic wording, and runtime checks point to
control-uiandcontrol-clifromcursor-team-kit. #986, #1043 - Checks on #986:
validate-plugins,generate-pstack-harness-trees --checkandcheck-pstack-docs(58 skills, 23 playbooks, 0 problems) pass. poteto-mode script tests: 82 pass, 0 fail. statechart script tests: 267 pass, 67 skipped, 0 fail. #986 - Checks on #1043: the same three scripts pass. poteto-mode script tests: 61 pass, 0 fail, down from 82 because the 21 removed tests were in the deleted
orch.test.ts. The typecheck passes and statechart tests are unchanged. A scan of the added lines and changed files found no internal references, and a comment review found nothing to remove. #1043
Cursor diff 519d36a...0f77cd5 · 2 PRs · summarised by claude-sonnet-5-5
Metadata
Metadata
Assignees
Labels
No labels
Activity