pi-jefferson

Pre-1.0. A statusline footer of configured rows for pi. Install with pi install npm:pi-jefferson.

A starship-style statusline for pi: instruments you place in a row's cells, a real layout, and per-instrument formatting.

The name is a nod to the lineage — Jefferson Airplane became Jefferson Starship became Starship.

Status

The statusline is a footer of one line per configured row, each spanning the full width. By default a cell takes the width its content needs and its neighbours sit against it. blank cells and cells with a $bar fill whatever the row leaves, and a row option restores evenly divided cells. The configuration table lists the available instruments, their variables and default formats. With no configuration you get the default single row of placeholder cells.

Install

pi install npm:pi-jefferson

Then /jefferson reports the footer this session installed and how many rows it renders.

Configuration

Create pi-jefferson.json in pi's agent directory, or .pi/pi-jefferson.json in a trusted project. The project document replaces the agent document whole; the two are not merged.

Each cell names its instrument; instruments configures each instrument once by name for every cell that uses it:

{
  "rows": [
    {
      "cells": [
        { "instrument": "model" },
        { "instrument": "context", "align": "right" }
      ]
    }
  ],
  "instruments": {
    "model": {
      "format": "🧠: $model from $provider",
      "style": { "fg": "accent", "bold": true }
    },
    "context": {
      "format": "$bar $percent"
    }
  }
}

A row sizes its cells by its distribution, either "content" or "equal". Without it a row is "content":

{ "distribution": "equal", "cells": [{ "instrument": "model" }, { "instrument": "path" }] }

With "content", a cell takes the columns its text needs, label included, so neighbours sit against it. Cells that fill (blank, and any cell whose format holds $bar) share the columns the row leaves, the earliest taking any odd one. When no cell fills, the last left-aligned cell with text takes them, else the first cell with text, else the first cell. A cell with no text, because its data is unknown, takes no columns. Two neighbours that both have text get one column between them, and that column gives way before any text does when the row is too narrow. No column goes between them when a filling cell sits there, or when the row's spare columns do. A cell with no spare columns has nothing for align to pad, so align changes nothing in it.

With "equal", every cell gets the same share of the width, and the earliest cells take one extra column each when the width does not divide. A value that is neither rejects the document, like a bad align: the default row renders and the warning names the value. distribution belongs on a row; beside a rows list it rejects the document. rowGap, below, is the one key that sits beside either shape.

Upgrading: earlier releases divided every row evenly. Set "distribution": "equal" on a row to keep that.

The footer leaves empty lines between rows. rowGap sets how many, at the top level of the document, beside rows or beside a single top-level row:

{ "rowGap": 0, "rows": [{ "cells": [{ "instrument": "model" }] }] }

It is a whole number from 0 to 3. Without it, one empty line sits between each pair of rows and none before the first. 0 renders the footer without that space. Any other value, such as -1, 1.5, "1" or 4, rejects the document: the default row renders and the warning names the value. A misspelt key such as rowgap is ignored, and the default space renders. The empty lines belong to the footer, not to a row; each row stays one line.

Upgrading: earlier releases drew no empty lines. Set "rowGap": 0 to keep that.

format is plain text with the instrument's already-formatted $variables, not Markdown. $ and | are its two special characters, and \ makes the character after it that character: \$, \| and \m render $, | and m, and a backslash with nothing after it stays itself. Control characters become spaces, then whitespace at the front and back of a format is removed; whitespace inside it is text. Omit format to use its default:

Every instrument also accepts label, a string placed in front of its format, whether the format is configured or the default. It is literal text: $ means nothing in it, and $label is not a variable (a format that names it rejects the document). Each control character in it becomes one space, then whitespace at its front and back is removed, and it takes the instrument's style. A label that is not a string rejects the document.

"thinking": { "label": "🤔" }

That renders 🤔 High. A label followed by a format always gets one space between them, so you never write it yourself. The space takes the label's style, counts toward the cell's width, and comes out of what $bar has left. A label that is empty after the trim, label: "" included, shows no label and no space. A format that is empty or only whitespace leaves the label alone, with no space after it.

Every instrument except blank has a default label, so model reads Model: anthropic/claude-opus-4 and thinking reads Thinking: High. A configured label replaces the default, and label: "" removes it. When an instrument's data is unknown the cell renders nothing, label included. With known data and "format": "", the cell is the label alone.

Upgrading: a default label now appears in front of a format you configured, too, on model, path, context, tokens and cost as well as thinking. If that format has no label, the cell starts with the label and one space, and "format": "" renders the label alone. Set label: "" to get the previous text back. model also changed its default format to $provider/$model, so to restore anthropic: claude-opus-4, set both label: "" and format: "$provider: $model". A space you wrote at the front or back of a label or a format is now removed, and a label's one space comes from the gap.

A |name after a variable, or after the text in front of it, modifies that text. Modifiers chain, each applying in the order the format writes them. On plain text a modifier covers the run of non-whitespace characters directly before it, so a multi-word phrase takes a modifier on each word:

🧠 $model via|dim $provider|capitalize

Here via renders dim and the provider's value is capitalized.

$path renders the directory's own name, pi-jefferson. Three forms give you the whole directory or the part of it under your home directory, and a form comes directly after the variable, before any modifier:

A form is not a string modifier: it picks which text arrives. A form applies before the modifiers after it, so $path\|relative\|capitalize capitalizes the relative text and $path\|bold renders the basename bold. Write a form anywhere else and the document is rejected, as it is for an unknown modifier. A form may follow only the variable that declares it, once, and before anything else. $model\|base is one of those rejections, because model declares no form.

A | begins a modifier only when a name follows it directly. A name is a run of letters, digits and underscores, so 10|20 and $model|_x start a modifier and an unknown one rejects the document. Followed by anything else (a space, the end of the format, $, | or other punctuation) the pipe is the character itself, so $model | $provider renders claude-opus-4 | anthropic. Where a literal pipe has to sit beside a name, write \|.

That escape changed meaning in an earlier release: a format written \| used to render \|, because a backslash before anything but $ was a literal backslash, and it renders | now. This release generalizes the escape itself: a backslash yields whatever character follows it, so a\b renders ab, and $idle\m renders the minutes with an m straight after them.

$bar is available on context: it takes the width the rest of the format leaves in the cell, and shrinks to nothing when none is left. A cell with a $bar fills the row's spare columns, sharing them with any blank cell; drop the $bar from its format and the cell takes only its text's width. Use at most one per format, and no modifier after it. Its filled portion changes to the theme's warning color above 70% and error above 90%; without an estimate from the host, it draws an unfilled, dim track.

style uses pi's ThemeStyle: fg accepts foreground theme color tokens, bg background tokens, or either slot can use a color pi's parser accepts. The six boolean attributes are bold, dim, italic, underline, inverse and strikethrough, and each is also a modifier inside a format. A modifier only adds: $model|bold renders bold even where the instrument's style sets bold: false, and nothing in a format switches an attribute off. fg and bg are not modifiers; colors belong to style.

thinking's $level is the one segment style does not fully control. It takes the foreground pi gives the thinking rules on the input box for the current level: the theme's thinkingOff through thinkingMax token, read from the theme current at render time. That replaces a configured fg on $level only. A configured bg and the boolean attributes stay, and the label and any other text keep the instrument's fg.

The notice instrument

notice shows the line pi would have drawn in the transcript: a cache miss of at least 20,000 re-billed tokens or ten cents, the usage a compaction or a branch summary billed, or the thinking blocks Anthropic dropped during provider recovery. One notice shows at a time. A newer notice replaces the one showing and counts its own seconds from its own arrival.

It shows for seconds, a whole number from 1 to 60, and ten when the document sets none, then disappears:

"notice": { "seconds": 20 }

Placing the cell is the switch. Jefferson never reads or writes pi's settings. pi's own showCacheMissNotices stays off at its default and draws nothing, and turning it on as well shows that event twice: once in the footer and once in the transcript.

Why there is no cache-warm notice. pi draws one when a cache warm finishes, and no event tells an extension that a warm finished. Showing it late, when the next turn found it, was the alternative, and that would report an old event. The warm is left out.

The words are yours. Each kind's text is a format, and the values in it are variables:

The defaults read as pi's English, and no translations ship. If English is not your language, write a kind's text in formats and the notice reads in your words:

"notice": { "formats": { "miss": "Miss de cache: $tokens jetons refacturés$cost" } }

The values are bare quantities; the words around them belong to the kind. $tokens is pi's count, 125k for 125,300 and 20k for 20,000. $cost is pi's whole suffix, the leading space and the parentheses included, and nothing below a cent, which is why the default texts hold no parentheses of their own. $idle is the minutes, 6, which is why missIdle writes $idle\m for pi's 6m, and $blocks is the count, which is why the thinking notice is two kinds: one for a single block, one for any other count. A kind that carries no value for a variable renders nothing there: a thinking notice has no cost, and a compaction has no block count.

A kind's format is plain: it names that kind's values and takes no modifier and no form. Styling belongs to the cell, so style still sets colours and attributes over the whole notice, and a |name inside a kind's text rejects the document. The cell's own format names $notice, the text composed for the kind; $tokens belongs inside a kind's text and is no variable of the cell.

If the extension cannot read a document's style or format, it rejects the whole document, renders the default row, and warns with a message naming the offender. The offenders include a modifier it does not implement, a form on an instrument or in a position that has none, a modifier on $bar, a modifier with nothing before it, a seconds value that is not a whole number from 1 to 60, and a formats entry it cannot use: an unknown kind, a value that is not a string, or an undeclared $name.

Why another statusline

pi-powerline-footer is excellent and is the reason this exists. It renders its row as a joined string with a fixed set of presets, which means segment spacing is three columns by construction, a segment cannot carry another segment's value, and the left / right / secondary groups are one flat list rather than rows or columns.

This package is for the case where you want a layout — rows, alignment, per-gap spacing, a stated truncation priority — and per-instrument formatting, with the pi session's own data behind each instrument.

Development

Trunk-based: main is the only long-lived branch, and every merge to it is a release.

pnpm install
pnpm typecheck
pnpm test

Tests live in tests/, never beside a source file. The pi.extensions manifest in package.json names the one entry, extensions/jefferson.ts, and pi loads only what the manifest lists, so the files under extensions/jefferson/ are never entries. A test file there would still ship in the package, because files includes extensions.

@earendil-works/pi-coding-agent and @earendil-works/pi-tui are peerDependencies, not dependencies, because pi supplies them at runtime. A physical copy bypasses pi's module mapping and duplicates classes and registries; pi reports a warning when it sees a manifest configured that way.

lodash-es is the package's one runtime dependencies entry. Its upperFirst, toUpper and toLower perform every string mutation a modifier makes. pi installs a package's dependencies when it installs the package, so it arrives with this one rather than shipping in the tarball; why this build and not the CommonJS one is ADR 0004.

A workstation cannot publish

publishConfig.provenance is true, and provenance is signed from a CI OIDC provider. A workstation has neither that provider nor a trusted-publishing credential, so the identical pnpm publish fails before it reaches the registry. Publishing is CI's job, and that is enforced by the mechanism rather than by a note in a file.

Releasing

Merging to main releases. There is no separate release step and nothing to remember to do.

The bump comes from a label on the pull request: release:major, release:minor, or release:patch. No label means a patch. It defaults low on purpose — an unlabelled change should never ship a major version by accident.

The base comes from the git tags, not from package.json. Main accepts no direct pushes, so the workflow cannot commit an updated version back to it, and package.json would drift from what actually shipped. Every release writes a v… tag, and those tags are the record. Specifically, the base is the highest release tag reachable from the commit before the one being released, which makes a re-run compute the same version it did the first time.

Authentication is npm trusted publishing over OIDC. The workflow holds no token and the package accepts publishes from this repository's release.yml, so there is nothing to rotate and nothing to leak. Provenance is attached automatically.

License

MIT