Skip to content

INSTALL.md

Latest commit

 

History

History
55 lines (44 loc) · 3.34 KB

INSTALL.md

File metadata and controls

55 lines (44 loc) · 3.34 KB
Raw

Installing the Fusha sarf/nahw engine

Fusha is the engine (Arabic morphology + syntax intelligence). The Qamus is the cart (the lexicon it pulls). This installs the engine as a skill for Claude or Codex, or you can use the repo directly.

Install for Claude

Claude Code reads skills from ~/.claude/skills/<name>/SKILL.md. The installer copies the full, self-contained sarf and nahw engine trees:

python scripts/install_claude_skills.py --dry-run     # preview (no writes)
python scripts/install_claude_skills.py               # install to ~/.claude/skills/fusha-sarf, fusha-nahw
python scripts/install_claude_skills.py --target /custom/skills/dir
python scripts/verify_skill_install.py --dir ~/.claude/skills/fusha-sarf   # verify an installed skill

After install, the skills appear as fusha-sarf and fusha-nahw (procedures, rules, evals, and each skill's own curriculum/drills included). The full zero-to-fluency tutoring runtime also uses the repo-level curriculum/ assessment/progress files; use the repo checkout or the claude.ai project pack when running a learner session.

Use with Codex

Codex reads skills from ~/.codex/skills/<name>/SKILL.md and instructions from AGENTS.md:

python scripts/install_codex_instructions.py --dry-run
python scripts/install_codex_instructions.py          # install + (re)write dist/codex/AGENTS.fusha.md

Then include dist/codex/AGENTS.fusha.md in your repo's AGENTS.md (copy the block between the <!-- BEGIN/END fusha-engine --> markers) so Codex consults the engine before any Arabic gloss/entry work.

Use the repo directly (no install)

Read sarf/SKILL.md and nahw/SKILL.md; open the procedure named by the gate; run the tools under tools/.

What you can do with it

  • Learn Fusha (ajami → fluency): curriculum/ (12-level roadmap, qamus/quran/hadith paths, drills, placement test, assessment rubric, progress templates) + curriculum/qamus-driven-fluency-engine.md.
  • Author Qamus entries from a corpus: tools/corpus_to_qamus_candidates.py (Nawawī40-proven, 0 live writes)
    • sarf/procedures/qamus-entry-authoring.md + nahw/procedures/qamus-entry-authoring.md.
  • Author hover glosses: sarf/procedures/hover-application.md, the four-gate pipeline, the token-addressed + suffix/pronoun layers (qamus/reports/token-addressed-hover-layer.md, suffix-pronoun-hover-report.md).
  • Generate learner drills from live Qamus/hover state: curriculum/drills/ + the proofing matrices.
  • Audit grammar safely: tools/run_grammar_evals.py + tools/grade_grammar_reasoning.py (the GrammarProblems gate: correct answer AND correct reasoning, else reject).
  • Repair entries from source photos: the tools/source_photo_* pipeline (rescue-first, not retake-first).

Invariants (always)

MCP-free skills (external sources are optional internal evidence via sources/README.md, never a dependency, never public); public gloss record stays {src:"qamus",kind:"authored",lang:"en"}; no raw source/media/secret committed; Qurʾān text unaltered; when uncertain, stay pending with a precise blocker.

Skill release state

The sarf/nahw skills are released at @2. Increments @2.1–@2.4 are candidate-only pending owner adjudication — do not treat them as installed/live until that ruling lands. See CONTRIBUTING.md for how to propose or track a candidate increment.