Enchanter AI Lich Lich Python

lich-python

Preconditions

• Target file extension is .py.
• config/ruff-rule-map.json exists (ships with the adapter; maps ruff rule IDs to M-engine severity + category).
• Optional: ruff is installed in the developer's environment for the richer invocation path. If absent, fall back to Python stdlib ast-based idiom checks using the rule-map's stdlib-feasible subset (~40 of ~120 rules).

Inputs

• Chained from lich-core: {file: "foo.py", ast_tree: <parsed>}

Steps

1. Detect ruff availability. Run ruff --version via subprocess. Record the version for version-drift detection.
2. Invoke ruff (if available). ruff check --output-format=json <file> — parse the JSON output.
3. Fall back (if ruff absent). Run the stdlib-feasible rule subset directly over the ast_tree from lich-core.
4. Map rule IDs to M-engine categories. For each finding, load config/ruff-rule-map.json and emit:
   • M1 runtime-failure candidate — if rule is in the "correctness" category (E501 exceptions, F401 unused imports that may cause NameError, etc.)
   • M7 idiom suggestion — if rule is in the "style/idiom" category (UP-series pyupgrade, SIM simplify)
   • skip — if rule is in the "Hydra-overlap" category (CWE-tagged security rules — these belong to Hydra R3)
5. Emit findings. Append to plugins/lich-core/state/review-flags.jsonl under the lich-core sub-plugin's collection point.

Outputs

• Appends to plugins/lich-core/state/review-flags.jsonl.
• Return value: {ruff_version: "...", rules_fired: N, skipped_hydra_overlap: N}.

Handoff

Findings flow back through lich-core → lich-sandbox (for M1-class flags) → lich-rubric → lich-verdict.

Failure modes

• F14 version drift — ruff's JSON output format changes; the adapter must pin a tested range and warn on mismatch.
• F04 task drift — re-implementing ruff rules in the adapter. Don't. The adapter maps; ruff does the work.
• F07 over-helpful substitution — running ruff with --fix and auto-rewriting code. Never. Lich is advisory.

Why Haiku tier

This skill is a thin mapping layer: invoke subprocess, parse JSON, lookup in rule-map, emit. No reasoning. Haiku is appropriate.

Source: enchanter-ai/lich — distributed by TomeVault.