Renders the shapes a language server produces — locations, navigation targets, call hierarchies, outlines, diagnostics, source frames — as plain text meant to be read by an agent from a transcript.

What it is

Pure functions. Data in, string out. No colour, no terminal, no framework, no document layer. Every component takes the protocol's own values — an LSP Range, a severity number — and does the formatting itself, so a caller never formats anything before calling.

That constraint is the whole design. A component that took "12:7-12:18" instead of a Range would be asking its caller to do the library's job, and would leave every consumer with its own copy of the zero-based-to-one-based conversion. Here that conversion exists once, in range.ts.

Layers

source-offset.ts   position ↔ offset arithmetic. No strings at all.
range.ts           positions and ranges as text. One conversion point.
                   ↑ everything below builds on these

code-frame.ts      source excerpt with a gutter and span carets
locations.ts       located rows grouped by file (ripgrep --heading -n)
targets.ts         navigation targets (❯ file:line:col)
calls.ts           call hierarchy, one file, indented callables
symbols.ts         document outline, ctags kinds, depth nesting
workspace-symbols.ts   symbol search hits with containers
diagnostics.ts     compiler-style problems

divider.ts         section rule, centred or right-anchored text
summary.ts         right-aligned label column, singular/plural
time.ts            durations that stop being milliseconds past a second

Prior art

The layout primitives are adapted from Vitest's reporters — divider, padSummaryTitle, formatTime, the code frame's gutter and caret arithmetic, the ❯ frame, and the rule that a row never prints a zero state. Vitest solves the same problem this does: dense, scannable, colour-optional text about code, read quickly by someone who did not write it.

The transplanted source is kept at packages/core/src/markdoc/format/vitest-reference/ with an extraction log.

Status

Early. The components here are in use by Type Atlas, an MCP server for TypeScript code intelligence, which is where each one was shaped against real agent-facing output rather than designed in advance.

Not yet extracted from that project: the diff body (-/+ unified lines), state glyphs, count lines, stack frames, breadcrumbs, tree rows, and an indented-block writer.