@geohar/pi-plan

A plan sidebar for the pi coding agent. It renders your cribsheet plan and the live subagent fleet as a persistent TUI widget, wave-ordered by dependency so the work you can pick up now is up top.

What it does

pi-plan accumulates plan events off pi's in-process bus and renders them, alongside the live subagent fleet it reads directly:

• Wave ordering — each item's wave is the longest chain of unsatisfied deps above it, so wave 0 is everything that's free to start now; each wave is a set that can be worked in parallel. Blocked items show a dim ⋯N (N unsatisfied deps).
• Dep-aware — a dep is satisfied by kind: a note never blocks, a design decision blocks only while tainted, a plan item blocks until done (mirrors cribsheet semantics).
• Actionable first — plan items you can start are highlighted; design/note context is dimmed and trails; done items sink to the bottom.
• Live subagents — the running subagent fleet is read straight off the subagents:* bus and shown as a separate Agents group (toggle with /plan agents).

The crib plan is emitted by the companion source in @geohar/pi-cribsheet, but pi-plan is a generic consumer — any source can drive it by publishing the plan bus protocol below, with no coupling to cribsheet.

Plan bus protocol

pi-plan listens on two pi.events channels, split by operation. A source publishes a full snapshot to replace its slice, or part-by-part updates to patch it. State is kept per source namespace (ns), so multiple sources coexist in one view.

Channels

plan:snapshot — replace all items for a namespace:

{ "ns": "cribsheet", "seq": 7, "items": [ /* PlanItem[] */ ] }

plan:update — patch a namespace part-by-part:

{ "ns": "mytool", "seq": 8, "upsert": [ /* PlanItem[] */ ], "remove": ["id-1", "id-2"] }

PlanItem

Rich fields are optional — a simple source can emit just id + title:

{
    id: string        // stable, unique within the ns
    title: string     // display text
    status?: string   // "done" sinks to the bottom; "in-progress"/"active" is highlighted
    deps?: string[]   // ids (same ns) this item depends on; omit for a flat list
    kind?: string     // "plan" | "design" | "note"  (default "plan")
    tainted?: boolean // design-kind only: blocks its dependents while true
}

Semantics

• ns attributes the source; pi-plan accumulates a per-ns map. snapshot replaces the whole map for that ns; update upserts items by id and deletes the remove ids.
• seq (optional, monotonic per ns) drops out-of-order deliveries.
• Dep satisfaction is kind-aware: a note never blocks; a design blocks only while tainted; a plan blocks until its status is done.
• Wave = the longest chain of unsatisfied deps above an item; wave 0 is actionable now.
• Pick the channel that fits: a source that has the whole list emits plan:snapshot; a source that produces items piecemeal emits plan:update.

Minimal emit (any pi extension, in-process):

pi.events.emit("plan:snapshot", {
    ns: "mytool",
    items: [{ id: "a", title: "do a thing", status: "todo" }],
})

The same protocol is bridged to ACP by @geohar/pi-acp (it re-emits these events over RPC and maps them to an ACP plan), so a source that speaks it surfaces in both the pi TUI and ACP clients.

Install

pi install npm:@geohar/pi-plan

Run pi from a repo whose cribsheet project has plan items; the widget primes on session start and repaints when you edit the plan or the subagent fleet changes.

Control it with /plan

pi widgets are render-only, so the widget is driven by a slash command:

command	effect
/plan	toggle expanded ⇄ collapsed
/plan expand / /plan collapse / /plan hide / /plan show	set state
/plan agents	toggle the subagent group
/plan show agents / /plan hide agents	set the subagent group
/plan filter done	include/exclude done items
/plan filter context	include/exclude design/note context
/plan lines <n>	set the expanded row budget for the session (/plan lines reports it)

Collapsed is a one-line summary (▸ Plan N ready · M blocked · K agents /plan to expand); expanded is the wave-ordered tree followed by the Agents group.

Settings

pi-plan reads <PI_CODING_AGENT_DIR>/extensions/pi-plan.json (defaults are written on first run):

{
  "placement": "aboveEditor",   // aboveEditor | belowEditor
  "defaultState": "expanded",   // expanded | collapsed | hidden
  "showDone": false,            // include done items by default
  "showContext": false,         // include design/note context by default
  "showAgents": true,           // show the live subagent group by default
  "maxRows": 18                 // row budget before the expanded view caps ("… N more")
}

/plan overrides these for the session; the file sets the defaults.

Development

npm install
npm run build     # tsc → dist/
npm test          # tsc && node --test

The extension entry is ./src/index.ts (declared in package.json pi.extensions), loaded directly from source via jiti — no build step is needed for local loading. Point a local checkout at pi with a bare path in settings.json packages[] (there is no local: scheme): "/abs/path/to/pi-plan".

License

MIT © George Harker