SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillssetup-pstack

setup-pstack

General & other

View repository

Updated 10-15 22:12

About this skill

Setup pstack for T3

Before resolving models or delegating, read T3 execution. It defines global configuration, capacity selection, and transport constraints.

Write pstack-models.md, a file that sets pstack's model selection per role. Default to ~/.agents/pstack-models.md; use the project root only for a requested project override. Read model selection for candidate-pool syntax, update behavior and Fleet persistence. Preserve existing choices. For this user's automatic setup, use open capacity and retain the skill's reasoning requirements; do not make them choose a fixed roster to start. This initial default does not replace existing pins or pools unless requested.

Steps

1. Detect available models

Call orchestrator_capabilities. It lists the providers and models you can pass to delegate_task in this session, custom models included, with each one's provider instance ID and any reasoning options. That is the only source. Use enabled providers that can run child tasks. If none qualify, report the limitation. Never add a fixed model or pool entry absent from that catalog. inherit-parent and capacity are policies, not model IDs. Read t3-capacity for current account availability, but persist no usage snapshots.

2. Load current state

The roles are the labels shown in step 5. Read the existing file and its optional # budget line. Otherwise start from the bundled defaults. Identify obsolete roles for review rather than dropping unrecognized user settings silently.

3. Budget, map, and confirm

Offer open automatic selection (capacity), a restricted pool (capacity[model-id-a | model-id-b]), an ordered pool that falls back by tier (capacity[model-id-a > model-id-b]), or a pinned account/model. A pool selects one alternative for each seat; a panel's top-level list still controls its seat count. Preserve the invoking skill's required count and diversity. Existing user choices are authorization to keep them; ask only for unresolved preferences.

(a) Reasoning preference. Keep the skill's requirements by omitting # budget unless the user chooses an override. When they ask to change reasoning, explain that this controls effort, not subscription spending. Show the current setting and the existing options:

  • unlimited — keep max
  • large — xhigh reasoning
  • medium — high reasoning
  • small — medium reasoning

(b) Apply it. On a fresh run, start from the bundled default-models.md in poteto-mode's references. Keep capacity values dynamic and preserve the panel seat counts. For a fixed-model request, resolve that choice from the live catalog. On a re-run, keep the user's existing choices.

When the user chose an override and the catalog exposes reasoning options, set the model's actual effort option: unlimited takes the highest reasoning value; large, medium, and small target xhigh, high, or medium. The ladder is max > xhigh > high > medium > low; if the target is unsupported, use the highest supported value at or below it, otherwise report the mismatch. Preserve catalog-specific option keys and distinguish effort from workflow modes. Apply the preference when resolving each capacity seat; keep policy entries dynamic.

(c) Show the resulting selection. Show roles grouped by identical policy, explicit pools and pins, reasoning overrides and unresolved choices. One delegate_task runs per panel seat, not per pool candidate. arena cross-judge pool retains Arena's own selection rules. swarm workers is the default for each worker unless its workflow assigns another choice. Ask before replacing an unavailable explicit choice; do not ask again to apply preferences already given by the user.

4. Validate

Validate every new pin and pool candidate against the live catalog. Unqualified pool IDs may resolve through several accounts. Preserve existing unavailable pins as unresolved until the user decides. Validate panel seat count and whether its candidates can meet the invoking skill's diversity rules. Report quota or host limitations without silently broadening a restricted pool. If a candidate cannot meet a requested effort, try another eligible candidate within the same policy; report a blocker when none qualifies. Never silently lower the invoking skill's required effort.

5. Write the file

For a Fleet global change, the chosen file is on the MBP source, even when the setup runs on PC. Use an available authorized connection to read and update that source, preserving its current contents. If unavailable, return a proposed edit and report persistence pending; do not write a competing local global file.

Write the chosen file with one line per role. Include # budget only for an explicit effort override. Use role values from the model-selection reference; record fixed-model effort as the catalog's option key and value when needed. Keep unrelated user settings intact. The role map is:

# pstack model configuration. One line per role. Delete a line to fall back to the skill default.
# `capacity` resolves model/account per wave via the execution reference.
# `inherit-parent` as a value: the role runs on the parent thread's model (omit the delegate_task model). Entries in a panel list still count toward its fan-out.
feature, refactoring: <providerInstanceId>/<model> (<effort>)
bug-fix: <providerInstanceId>/<model> (<effort>)
perf-issue: <providerInstanceId>/<model> (<effort>)
hillclimb: <providerInstanceId>/<model> (<effort>)
judgment and prose: <providerInstanceId>/<model> (<effort>)
hardest tasks: <providerInstanceId>/<model> (<effort>)
how explorer: <providerInstanceId>/<model> (<effort>)
how explainer: <providerInstanceId>/<model> (<effort>)
why investigators: <providerInstanceId>/<model> (<effort>)
why synthesizer: <providerInstanceId>/<model> (<effort>)
reflect tooling: <providerInstanceId>/<model> (<effort>)
reflect judgment, divergent, synthesizer: <providerInstanceId>/<model> (<effort>)
arena runners: <entry>, <entry>, <entry>
arena cross-judge pool: <entry>, <entry>, <entry>
swarm workers: <providerInstanceId>/<model> (<effort>)
architect runners: <entry>, <entry>, <entry>
interrogate reviewers: <entry>, <entry>, <entry>

6. Confirm

Report the path, scope and actual host delivery. The adapter reads it before each delegation wave; running children keep their existing selection. Open capacity considers newly exposed T3 models without a setup rerun. Explicit pools and pins change only when requested. For Fleet globals, update the MBP source and verify destination contents; a configured sync is not delivery proof.

7. Offer a verification skill (optional)

When setup is for an application project, check whether it has a way to drive the real app for proof (a verify-* skill or existing harness). If missing, offer /create-verification-skill once. Skip this step for global configuration without an application project.