setup-pstack
Configure which models pstack's delegates use per task category and at what reasoning budget. Detects your authenticated models and writes the shared pi-herdr-agents task preferences, models and thinking, after you approve the change in the host's dialog. Use for /setup-pstack, "configure pstack models", "pstack budget", or changing pstack's model choices.
- License 许可
- MIT
- 最近更新
- 2026-11-13
- Stars
- ★ 3
- Forks
- 0
01 前言 Front Matter
- name
- setup-pstack
- description
- Configure which models pstack's delegates use per task category and at what reasoning budget. Detects your authenticated models and writes the shared pi-herdr-agents task preferences, models and thinking, after you approve the change in the host's dialog. Use for /setup-pstack, "configure pstack models", "pstack budget", or changing pstack's model choices.
03 SKILL.md 正文
Setup pstack
Write pi-herdr-agents' task preferences, models.tasks and models.tasksThinking, which set the models and the thinking level for each task category pstack's delegates launch with.
Pstack keeps no model configuration of its own. poteto-mode launches every ordinary delegate with a task:<category> model, and pi-herdr-agents resolves it from these shared preferences. A change affects every pi-herdr-agents workflow and role pack that uses the categories, not only pstack. The categories are coding, review, recon, qa, architecture and docs. An explicit model in the subagent call wins, whether it is an exact provider/model-id or a task:<category> selector that expands to these preferences. Only a call without model falls back to role, per-agent and default models, then the parent's model. Likewise an explicit thinking overrides a category's thinking.
Run this skill through /setup-pstack [request]. The command appends a setup brief that starts with "/setup-pstack started the setup flow." and makes the host writer subagents_write_task_models callable. Reached any other way, such as /skill:setup-pstack, there is no fresh command-generated brief, so the config revision, current choices and authenticated models are unknown: explain these steps, ask the user to run /setup-pstack, and do not call the writer even if it is loaded. Never edit the pi-herdr-agents config with the write, edit or bash tools. Paths here are relative to this skill directory.
Steps
1. Detect available models
The brief's "Authenticated models" list is the dependable source: every exact provider/model-id with configured authentication in this session, with the thinking levels it supports. Each model is a JSON string, as in "Current choices", and the ID is that string's value. An ID shown with escapes such as \n or \u200b holds control or invisible characters: point it out and do not offer it. Never write a model you have not confirmed is in that list, and never write a remembered or guessed ID. If the list is (none), ask the user to authenticate a provider and run /setup-pstack again. There are no model aliases: a category holds exact models only.
2. Load current state
The brief's "Current choices" are the saved choices: for each category, its ordered model list and its thinking level, where null means not set. There are no skill defaults, so an unset category needs a choice. A saved model that is not in the authenticated list cannot be written back. Drop it.
3. Budget, map, and confirm
(a) Ask for a budget. If the user's request in the brief already names one, use it. Otherwise ask with these four options and these exact labels, as a numbered list, and wait for the answer. When every category has the same saved thinking and it is one of the four targets, name that budget as the current one; otherwise say the saved thinking is mixed or not set.
unlimited — max reasoninglarge — xhigh reasoningmedium — high reasoningsmall — medium reasoning
(b) Apply it. Build the working table from the current choices. The budget changes thinking only, never models: unlimited, large, medium and small set every category's thinking to max, xhigh, high or medium. The host clamps a category's thinking to each candidate model at launch: the target when the model supports it, else the nearest supported level above it, else the highest one below. So unlimited runs a model that tops out at xhigh at xhigh, and a model without reasoning runs at off. Work out that effective level for every model from the levels the brief lists.
(c) Show the categories and confirm. Show every category with its models in order, each with its effective thinking, and the upstream roles it drives from the table below. Mark a category with no authenticated model as needing a choice, and list each model step 2 dropped. Ask whether to accept as-is or change specific categories, offering the authenticated models and the thinking levels. A category's list is ordered candidates: the first launches, and later ones are fallbacks when a launch fails. It is not a panel, and its length sets no fan-out count. Every category saved now stays in the proposal, changed or not. Removing one is an explicit user choice: do it only when the user asks to remove that category, and tell them the host's dialog will show it as removed. A saved category whose models step 2 dropped needs a new model or that explicit choice. Favor capability for coding, review and architecture, and efficiency for recon and qa.
| Upstream role | Category |
|---|---|
| feature, refactoring; bug-fix; perf-issue; hillclimb | coding |
| judgment and prose; hardest tasks | architecture |
| how explorer; why investigators | recon |
| how explainer; why synthesizer; reflect judgment, divergent, synthesizer | architecture |
| swarm workers | qa to verify or measure, recon to explore, coding to write |
| arena runners, arena cross-judge pool, architect runners, interrogate reviewers, reflect tooling | none: diversity seats get one exact model per family from the live catalog at launch, per ../poteto-mode/references/fan-out.md section 3 |
| (no upstream role) | review for reviewers and judges, qa for verifiers, docs for documentation workers |
4. Validate
Every model written must be in the brief's authenticated list, every category written needs at least one model, and every thinking level is one of off, minimal, low, medium, high, xhigh or max. If a chosen model is not available, stop and ask again.
5. Write the preferences
Call subagents_write_task_models once with:
tasks: every category "Current choices" lists with models, changed or not, and each category the user added, each with its complete ordered list. The writer replaces the whole task map, so a category left out is deleted. Leave one out only when the user chose to remove it in step 3c.tasksThinking: the thinking for every category intasks, unchanged ones included.tasksMeta: the brief's "tasksMeta for this write", exactly.basis:{"kind":"registry-only"}.expectedConfigRevision: the brief's value, exactly.
The writer replaces the task models and their thinking in one atomic write, so re-runs stay idempotent. The host shows the user the exact change and asks for approval. Nothing is saved unless the user chooses to save. If the call fails, say what the host said and do not call it again with the same arguments:
- Declined, cancelled or timed out: nothing was written. Ask whether they want to change something.
Stale task model config revision: the file changed after the brief was read. Ask the user to run/setup-pstackagain.- Any other error: repeat it. It says whether the file is unchanged, holds the approved preferences, or needs inspection.
6. Confirm
Tell the user, from the tool result's saved tasks and tasksThinking, what was saved, that it applies after /reload or in a new session, and that it is shared by every pi-herdr-agents workflow. Re-running /setup-pstack updates it.
7. Offer a verification skill (optional)
Check whether the project has a way to drive the real app for proof (a verify-* skill, or an existing harness). If not, offer once: "want a project-local verification skill, so agents can drive the app the way a user does and prove changes work? I can generate one with /skill:create-verification-skill." On yes, read ../create-verification-skill/SKILL.md and follow it. On no, move on without pushing.
8. Offer a formal verification skill (optional)
Check whether the project already has a formal verification skill (a project skill under .pi/skills/ that runs a model checker, a proof, or another verifier against the code). If not, offer once: "want a formal verification skill, so agents can prove a key property of your code holds in every case in scope, not only the cases tests reach? I can generate one with /skill:create-formal-verification-skill." On yes, read ../create-formal-verification-skill/SKILL.md and follow it. On no, move on without pushing.
04 同仓库更多 Skills
poteto-mode
Test fixture. A different skill that reuses the poteto-mode name, to check shadowing diagnostics.
why
Use for 'why does X work this way', 'why we picked Y', design rationale, regressions, postmortems, or data-backed thresholds. Discovers available MCPs and queries each evidence category (source control, issue tracker, long-form docs, real-time chat, infrastructure observability, error tracking, product analytics warehouse) in parallel, then returns a cited read on decisions and tradeoffs. Use how for runtime behavior.
unslop
Cut AI tells from any writing. Must always apply.
typescript-best-practices
TypeScript best practices. Use when reading or editing any .ts or .tsx file.
technical-writing
Layered technical-writing standard: Diátaxis structure, Google developer style sentences, STE instruction rules, Global English syntax. Use for /skill:technical-writing or when writing or reviewing docs, RFCs, readmes, PR descriptions, or commit messages.
swarm
Fan out N parallel workers, drain them, and return one report. Use for /swarm, 'swarm this', or parallel coverage, races, gauntlets, and exploration.