setup-pstack

Files (1)

• SKILL.md

SKILL.md (first lines)

# Setup pstack

Follow the [portable runtime contract](../pstack-pi/references/runtime.md) throughout this setup.

Write the optional pstack configuration at `$PSTACK_CONFIG` when that variable is set; otherwise use `.pstack/config.md` in the current project. This is an override layer, not a requirement. Never write to a vendor-specific home directory.

## Steps

Model discovery and model assignment are different capabilities. A model listed by
`pi --list-models` is available to the current Pi process, but that does not mean a
workflow can send work to it. Per-role assignment also requires a host task or
subagent facility that accepts a model choice. Native Pi intentionally does not
include subagents. Do not claim that a role was assigned when that facility is
absent.

### Pi delegation prerequisite

For native Pi, recommend the maintained `pi-subagents` package when the user wants
parallel workers or per-role models:

```bash
pi install npm:pi-subagents
```

Tell the user to restart Pi after installation, then run `/subagents-doctor`.
Continue setup only when the `subagent` tool and the expected agents are visible.
The package supplies the delegation facility; it does not choose pstack's model
policy by itself.

### OpenCode delegation prerequisite

OpenCode has its own native primary and subagent agents. Do not install
`pi-subagents` or write Pi-specific settings when the current host is OpenCode.
Use the agents and concrete `provider/model-id` choices reported by OpenCode, and
keep their model configuration in the host's `opencode.json` or
`opencode.jsonc`. If OpenCode does not expose a usable subagent, report that
limitation instead of substituting a Pi-specific mechanism.

### 1. Detect available choices

Detect both of these independently:

1. The model IDs the current host actually exposes.
2. Whether the current host exposes a task or subagent facility that can select a
   model for each child.

For Pi, `pi --list-models` is the model inventory and the `provider/model-id` form
is the concrete value to record. With `pi-subagents` installed, the `subagent`
tool and `/subagents-doctor` establish the delegation inventory. The live host
inventory is authoritative. Never copy a vendor slug from prose or guess that a
model is available.

If the host has a task facility but cannot enumerate models, use only a host role
mapping that the facility documents. Do not ask the user to invent a raw slug.
`inherit-parent` is valid only when the host can pass the current model to a child.

### 2. Load current state

Read the selected configuration path when it exists and treat its concrete values
as current choices. If it contains only portable role aliases from a host without
per-child delegation, treat those aliases as stale inactive state and start from
the detected models instead of carrying them into Pi's agent overrides.

### 3. Map and confirm

If both model inventory and per-child model selection are available, show every
workflow role with its current concrete `provider/model-id` choice. Mark an
explicit model absent from the live inventory as needing a replacement. Ask the
user to accept or change the choices, offering only detected model IDs and
`inherit-parent` when supported. Use the host's structured interaction tool when
available; otherwise ask one focused question in normal conversation.

If the host exposes models but no task or subagent facility, stop model mapping
with an explicit capability report. For native Pi, recommend
`pi install npm:pi-subagents`, a restart, and `/subagents-doctor`. For OpenCode,
explain that its native agents are not visible or configured and point the user
to the host's agent configuration. For any other host, name the missing
capability without proposing a vendor-specific substitute. Say that no pstack
role can receive a different model in this session and do not present the
portable defaults as an assignment or overwrite an existing role configuration
with inactive aliases.

For panel roles (`how critics`, `arena runners`, `arena cross-judge pool`, `architect runners`, and `interrogate reviewers`), one child runs per entry, so list length controls fan-out. Prefer diversity for judgment-sensitive panels. `swarm workers` is the default choice for every worker unless a race assigns a different choice per arm.

### 4. Validate

Every explicit model written must be present in the detected host inventory.
Role aliases pass only when the host task facility documents those aliases.
`inherit-parent` passes only when the host can pass the current model to a child. If
a selected explicit model is unavailable, stop and ask for a replacement. Never
write a configuration that requires another host.

### 5. Write the configuration

When the host supports per-child selection, create the parent directory when
needed and overwrite the selected file so re-runs remain idempotent. Use concrete
detected model IDs in this shape. Replace every placeholder with a model the
current host reported, and use `inherit-parent` only when the host supports it:

```md
# pstack role and model configuration
# Values are concrete provider/model-id choices confirmed by the host.
feature, refactoring: <implementer-model>
bug-fix: <reviewer-model>
perf-issue: <reviewer-model>
hillclimb: <implementer-model>
judgment and prose: <reviewer-model>
hardest tasks: inherit-parent
how explorer: <explorer-model>
how explainer: <synthesizer-model>
how critics: <reviewer-model>, <planner-model>, <designer-model>, inherit-parent
why investigators: <researcher-model>
why synthesizer: <synthesizer-model>
reflect tooling: <researcher-model>
reflect judgment: <reviewer-model>
reflect divergent: <designer-model>
reflect synthesizer: <synthesizer-model>
arena runners: <designer-model>, <planner-model>, <implementer-model>, inherit-parent
arena cross-judge pool: <reviewer-model>, <planner-model>, <designer-model>, inherit-parent

From shrimpwtf/oh-my-pstack, MIT.

More from shrimpwtf/oh-my-pstack