@jackice/hashline

A compact, line-anchored patch language and applier, packaged as a pi coding-agent extension.

Forked from @oh-my-pi/hashline (upstream sync: manual comparison; only stable APIs are forked). Differences from upstream: removed the @oh-my-pi/pi-natives dependency (inlined pure-TS xxHash32 and LCS line-level diff), and NodeFilesystem now uses node:fs — so this package loads under the Node runtime (pi's extension loader).

Hashline is a diff format designed for LLM-driven file edits. It binds every hunk to a file-content hash so stale anchors are rejected before they corrupt code, and it abstracts over the filesystem so the same patcher works on disk, in memory, over the network, or against any custom backend.

Use as a pi extension

Once installed into pi (packages or pi install npm:@jackice/hashline), the extension entry point index.ts is loaded automatically:

• rewrites native read output: [PATH#TAG] file header + global line numbers N:TEXT, for patch anchoring;
• registers the hashline tool: applies PUT/CUT/REM/MV patches anchored by line numbers + content-hash tags, rejecting stale tags or recovering via 3-way merge, with atomic preflight of all hunks before writing;
• the native edit tool is kept as a fallback.

Quick start

import {
	Filesystem,
	InMemoryFilesystem,
	InMemorySnapshotStore,
	Patcher,
	Patch,
} from "@oh-my-pi/hashline";

const fs = new InMemoryFilesystem();
const snapshots = new InMemorySnapshotStore();
const before = `const greeting = "hi";\nexport { greeting };\n`;
await fs.writeText("hello.ts", before);

const tag = snapshots.record("hello.ts", before);
const patcher = new Patcher({ fs, snapshots });
const patch = Patch.parse(String.raw`[hello.ts#${tag}]
PUT 1.=1:
+const greeting = "hello";`);
const result = await patcher.apply(patch);

console.log(result.sections[0].op); // "update"
console.log(await fs.readText("hello.ts"));

Format

See src/prompt.md for the user-facing description and src/grammar.lark for the formal grammar.

Each file section starts with [PATH#TAG]. The tag is a 4-hex content hash of the full normalized file text recorded by the SnapshotStore, and it is not meaningful outside that store. The patcher protects against stale anchors by resolving the tag, verifying the live file still matches the recorded content hash, and refusing or attempting session-aware recovery on mismatch.

Inside a section:

• PUT A.=B: — replace lines A through B (inclusive) with following +TEXT body rows.
• PUT A*: — replace the syntactic block beginning on line A.
• PUT <A: / PUT >A: — insert following body rows before/after line A (<1 = head, >$ = tail).
• PUT >A*: — insert following body rows after the resolved block's last line.
• PUT <A / PUT >A / PUT A.=B @name / PUT A* @name — paste a captured register at a gap, over a range, or over a resolved block (no : header or body rows; @name is optional only at gaps).
• CUT A.=B / CUT A* — delete concrete lines or a resolved block and capture them (anonymous, or @name when given).
• REM — delete the whole file named by the section header.
• MV DEST — move/rename the section file to DEST (optionally after line edits).
• +TEXT — literal body row (use + alone for a blank line).

Abstractions

Filesystem

Read and write text by path. The default implementations:

• InMemoryFilesystem — backed by a Map. Tests, sandboxes.
• NodeFilesystem — disk-backed via Bun.file/Bun.write. Default for CLIs.

Subclass Filesystem to wire hashline into any storage: VFS, S3, an LSP text-document protocol, a Git tree, anything.

SnapshotStore

Required. Hashline tags are full-file content hashes recorded per path, so Patcher must receive the store that observed them. Recovery replays edits against the cached pre-edit snapshot and 3-way-merges onto current content when the live file diverged.

Patcher

The orchestration class. Reads, normalizes line endings + BOM, applies edits, restores line endings, and writes via the configured Filesystem. Multi-section patches are preflighted up front so a partial batch never lands.