Skip to content

Requirements & costs

This is a static Astro/Starlight site. Reading it, searching it and building it require no model credentials and no backend service. Search uses a build-time Pagefind index served with the site. Fonts are bundled locally; the site does not load remote font services.

Installing a Pi extension is different: it runs code inside your agent process. A shell task can run whatever command you give it, and a planner or subagent can consume your configured model access.

The configured SDK test target is Pi 1.2.0, covering loader selection, local session lifecycle and memory TUI input—not a live PTY compatibility matrix. Classifier-dependent components generally require Pi 0.99 or later; next-prompt requires Pi 1.0 or later. Private terminal hooks can change independently of those minimums.

Component or feature Requirement and boundary
Logical todos No external executable, classifier, credentials or required sibling. pi-todo stores snapshots in session tool results; task text reaches the model and session history. UI needs an interactive session; rpiv-i18n is optional and rpiv-config is not required.
Background jobs A shell environment with bash; commands may need their own tools or credentials. Process groups are used for stopping work.
Subagents A runnable pi command and usable model/provider configuration. Child sessions make model calls.
Subagent worktrees Git and a suitable repository, only when worktree: true is requested. Subagents use git worktree add directly; no external worktree helper is required.
Plan mode Credentials or subscription access for the selected planner and helper models. The /planners workflow runs two planner sessions, each with separate usage; main-session plan mode does not start them. Optional tools bring their own requirements.
Diff review and rewind A Git repository. Snapshots include non-ignored untracked files and are not a backup service.
Herdr integration An interactive Pi in Herdr. Herdr’s own Pi integration owns working/blocked/idle state. PR metadata also needs an authenticated gh command.
Herdr worktree children Git plus the active Herdr socket’s worktree create/open operations. No external worktree CLI; normal child tabs share the existing checkout.
Notebook observation pi-marimo (a separate package): A running marimo server with the registry/session/SSE interfaces the extension uses. Token-protected servers need MARIMO_TOKEN. Editing is a separate marimo-pair workflow.
Status footer Optional provider quota sources; its row icons are Nerd Font glyphs that need terminal font support. Basic session rows do not require a Claude Code account.
Clipboard picker Pi’s clipboard support in your terminal environment. Optional System 1 classifier ranking.
CC render @shikijs/cli is an optional dependency of the component but a regular dependency of the root suite; it supplies diff syntax highlighting; hover, click and wheel need fullscreen mode and SGR mouse reporting. Patches private Pi rendering members.
Markdown extras grok-mermaid (a suite dependency) for diagrams; without it diagram fences pass through unchanged.
Terminal refinements Relevant TUI hooks and terminal behavior; clear-screen needs fullscreen mode, scrollbar-width only acts inside Herdr.

The documentation does not assert a Windows compatibility matrix for the whole collection. Several components use POSIX shell/process behavior. Check each component’s implementation before adopting it in a different environment.

Classifier features are optional, not free

Section titled “Classifier features are optional, not free”

System 1 configures an ordered route for the suite’s classifier features. The global default is Jev only; /system1 can select either native classifier alone or an ordered Jev and OpenAI Decisions chain. On an operational failure it can notify you and try the next configured provider; it does not shop outside that chain or bypass a safety refusal. Each attempted provider can receive the data needed for the judgment and can bill for requests.

Plugin What the classifier sees Without it
Plan and approval effort Recent task and conversation excerpts Keeps the current effort
Cache guard Tool output or conversation items being trimmed/compacted Warnings and clock still work; normal Pi compaction remains available
Tool gate Proposed calls, recent user intent and rules Blocks calls that require a classifier verdict; fixed checks and explicitly allowed/read-only paths retain their ordinary behavior
Subagents Delegation task and candidate choices Uses agent configuration and parent/default choices
Copy picker Candidate excerpts and recent messages Recency ordering
Next prompt Suggested next step and answer Can show the suggestion without vetting
Plan mode Planning task, plan and conversation excerpts for effort; available tools for preselection You can select tools and effort manually
Sidebar Proposed guarded action Refuses an unconfirmed call

The order is global to the Pi agent directory, not a per-project or per-feature backend switch. For most plugins, "jev": { "enabled": false } in the plugin’s own configuration still disables that feature’s classifier use; plan mode uses "jevToolSelection": false. Jev mode retains existing provider/model settings and discovery; Decisions overrides those targets. See the reference pages for filenames and merge precedence.

The default TypeSafe provider uses Pi’s credentials, such as TYPESAFE_API_KEY or /login typesafe. OpenAI Decisions instead uses the isolated system1-openai/gpt-6-luna classifier via the real OpenAI Decisions API, authenticated only by /login system1-openai or SYSTEM1_OPENAI_API_KEY. It does not use the ordinary OpenAI chat credential. Do not put credentials in these docs, a public repository, or a front-end environment variable.

Before selecting Decisions, claim the eligible ChatGPT Pro API credits into an API organization with active included credits, no purchased balance, no auto-recharge and no paid or postpaid continuation. Check those conditions in OpenAI billing; Pi cannot verify entitlement, credit balance or which funds pay for a request, and does not enforce a credits-only spending cap. Do not include Decisions in the order if the prerequisites cannot be established. A configured fallback may use a different provider after an operational failure; /system1 status shows the order and recent route.

The /system1 classifier route is independent of /waterfall, the subscription chat-model router. API credits, subscription usage windows, throttles and catalog-estimated cost are different quantities; none proves the state of the others. Footer cost is a reported token-price estimate, not an invoice or a subscription charge. Cache guard uses published model prices where available and an idle-time heuristic where a cache TTL is unknown. Classifier latency and savings depend on the workload and provider; benchmark anecdotes are not performance guarantees.

Quota lookup can make authenticated HTTP requests without making a model call. The status footer’s Claude account reading may belong to a different account from Pi’s model session. Manage context and cost explains the distinctions.