Reading files

When an assistant asks to read a 2,000 line module, it almost never needs 2,000 lines. It needs to know what is in there: the imports, the classes, the signatures, the docstrings, the constants, the decorators. The bodies matter for the two or three functions the task is actually about, and those it can ask for.

So tokunseba keeps the skeleton and replaces each long body with one line saying how many lines went and how to get them back.

     1	import { readFile } from "node:fs/promises";
     2	import type { Config } from "./config";
     3
     4	export const DEFAULT_PORT = 7777;
     5
     6	/** Loads a config file and merges it with the defaults. */
     7	export async function loadConfig(path: string): Promise<Config> {
        // [tokunseba: 15 lines folded; whole file: tokunseba expand h_4b91c07e]
    23	}
    24
    25	export class Proxy {
    26	  private port: number;
    27
    28	  constructor(port: number) { this.port = port; }
    29
    30	  async start(): Promise<void> {
        // [tokunseba: 9 lines folded; whole file: tokunseba expand h_4b91c07e]
    40	  }
    41	}

The line numbers that survive are the file's own. The jump from 7 to 23 tells the model exactly which lines are missing, so asking for the right part is one targeted read rather than a search.

What is never foldedCopy Anchor Link

• Imports, and anything else at module level. That is the file's interface.
• The signature line, its decorators, and its docstring. The docstring is usually the most informative thing in a function and the smallest part of it.
• A body shorter than eight lines, where the marker is not smaller than the body.
• Anything at all, if the result would not be meaningfully smaller than the original.

LanguagesCopy Anchor Link

"Scanned" means a reader that tracks strings and comments, matches braces, and refuses to fold when the count does not balance. A real parser for each of these would be a dependency tree larger than the whole project, and a scanner that understands strings gets the same answer on anything that compiles.

What stops it firingCopy Anchor Link

Any of these leave the file exactly as it arrived:

• fewer than 60 lines
• no function body long enough to be worth a marker
• a file that does not parse, for Python
• unbalanced braces, for the scanned languages
• a file that is mostly signatures already, where the fold would remove under a quarter of it
• a minified bundle, which is one enormous line and is handled as junk instead
• no file path on the tool call, because the language cannot be known from content alone

Line-number guttersCopy Anchor Link

Most coding agents hand a file over with a cat -n style gutter. tokunseba computes the fold on the source underneath the gutter and emits the original lines, gutter and all, so what is kept is byte for byte what the tool produced.

Tab, arrow (→) and pipe (|) separators are all recognised, including the case where the separator on a blank line has been stripped by whitespace normalisation.

Getting the file backCopy Anchor Link

Every outline names a handle:

tokunseba expand h_4b91c07e

That returns the whole file, byte for byte, including the gutter if there was one. The model can do the same thing through the shell, or through the MCP expand tool if it has no shell.

Turning it offCopy Anchor Link

Outlining is part of tier 2:

tokunseba tier disable 2

That also turns off summarising and junk detection. There is deliberately no switch for outlining alone: the tier is the unit you consented to.