Comparing changes

Keyman has no tree-sitter grammar anywhere, so it gets a standalone line-based extractor (the Liquid/Vue/Svelte/DFM pattern): - groups -> function nodes (callable units, span to next group) - user stores -> variable, &system stores -> constant (&NAME -> docstring) - use(g) -> calls edges (group->group, file->group for `begin`) - outs/any/notany/index/call/set/if/save/reset -> references edges, resolved intra-file (group->store, store->store) Wiring: .kmn in EXTENSION_MAP, 'keyman' in LANGUAGES, grammars.ts special-cases (no WASM), and a route to KeymanExtractor in extractFromSource. +8 extraction tests. Validated on real .kmn from keymanapp/keyboards across small/medium/large (2/251/1029 files): linear scaling, no node explosion, and 17776/17776 calls+references edges resolved intra-file (zero spurious cross-file edges). Co-Authored-By: Claude Opus 4.8 <[email protected]>

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Adds the connected-subgraph trio from the Keyman extractor roadmap so a keyboard's files stop being isolated islands: - .kmn: emit cross-file `references` from the &LAYOUTFILE / &VISUALKEYBOARD system stores to their named companion files (resolved against the .kmn's own directory so the resolver's file-path matcher links them). - .keyman-touch-layout (new extractor): a `component` node per layer (scoped by platform) + `references` layer->layer for every `nextlayer` switch, including subkeys/flicks/multitaps. Keys are not nodes (avoids graph blow-up). - .kps (new extractor): package manifest -> file node with the display name as docstring, plus cross-file `references` to each <Keyboard><ID>'s .kmn and any shipped keyboard-source companion. Build outputs, fonts, and docs are skipped. All three share the existing `keyman` language token, dispatched by extension in tree-sitter.ts. Validated end-to-end on khmer_angkor and baybayin: .kps -> .kmn -> touch-layout all resolve, with per-platform layer transitions. Tests cover detection, both new extractors, and the .kmn companion links.

Step 4 of the Keyman roadmap. Indexes on-screen (desktop) visual keyboard files: a file node (docstring = <kbdname>) plus a `component` node per layer, scoped by encoding and decoded modifier state (shift="RA" -> rightalt, "SRA" -> shift+rightalt). Desktop layers have no programmatic switches, so no inter-layer edges; keys stay data, not nodes (avoids graph blow-up). Indexing .kvks also lets the &VISUALKEYBOARD reference the .kmn extractor already emits resolve — validated end-to-end on khmer_angkor and baybayin (.kmn -> .kvks now resolves alongside .kmn -> .keyman-touch-layout). With this, the four core Keyman source formats (.kmn / .keyman-touch-layout / .kps / .kvks) are covered.

…eyboard_info Makes "which keyboards support language X" queryable uniformly across the whole repo, regardless of a keyboard's lineage: - .kps extractor now emits a `constant` node per supported language from <Keyboards><Keyboard><Languages><Language ID="...">DisplayName</Language>, with the tag as the node name and the display name in the (full-text-indexed) signature, plus a contains edge file -> language. This is the release/experimental source of language metadata. - New .keyboard_info extractor (keyman-keyboard-info-extractor.ts) does the same from the JSON `languages` map. This file is a build artifact for release keyboards (generated from the .kps) but is committed for legacy binary-only keyboards, which have no .kps — so indexing it gives them the same coverage. - Both build identical-shaped language nodes via a shared helper (keyman-shared.ts), so one search (e.g. "Tagalog") returns matches from both .kps (release) and .keyboard_info (legacy) at once. Validated end-to-end on a mixed sample (khmer_angkor/baybayin .kps + legacy european2/blackfoot .keyboard_info): language nodes resolve and FTS search by display name finds them across both lineages. All five Keyman source formats in the repo are now covered.

The .kps <RelatedPackages><RelatedPackage ID="..."> entries link one package to another (usually the legacy keyboard a newer one deprecates/replaces). The target lives in a different directory, but a package id equals its file's basename stem, so each related id is emitted as a cross-file `references` to both `<id>.kps` (release/experimental) and `<id>.keyboard_info` (legacy, which has no .kps). Whichever exists resolves by exact-name; external ids and the unused arm drop harmlessly. A self-referential related id is skipped. Most related targets (433 of ~611 in the repo) are legacy keyboards, so the .keyboard_info arm is what makes the bulk of these links resolve. Validated end-to-end: khmer_angkor.kps resolves to both basic_kbdkni.kps (release) and khmer10.keyboard_info (legacy), plus the reciprocal link back from basic_kbdkni. All 24 Keyman extraction tests pass.

…, touch-layout) + cross-file linking Brings in the roadmap work built in Claude Code web: - connect Keyman keyboards into a cross-file subgraph - visual-keyboard (.kvks) extractor - language coverage across .kps and .keyboard_info - link Keyman packages to their related packages Co-Authored-By: Claude Opus 4.8 <[email protected]>

…fo/touch-layout + cross-file linking) Co-Authored-By: Claude Opus 4.8 <[email protected]>

The README inherited upstream's install flow (curl install.sh / npm i -g @colbymchenry/codegraph), neither of which works for this fork — it isn't published to npm or as a GitHub Release. Rewrite Get Started, Quick Start, Manual Setup, and Library Usage to clone + build + npm link from the fork, fix the misleading npm/bundled badges, repoint the reinstall + Report Bug links, and add a fork-installs-from-source note. https://claude.ai/code/session_01YNaG1JupGkefkdP52Wo33e

docs: rewrite install instructions for from-source fork install

Navigation Menu

Comparing changes

Open a pull request

Uh oh!

Commits on Jun 5, 2026

This comparison is taking too long to generate.

Uh oh!