| name | lich-review |
| description | Runs M1 Cousot Interval Propagation and M2 Falleri Structural Diff on a code scope, flags runtime-failure candidates (div-zero, null deref, OOB, overflow, resource leak), and hands the flagged sites to lich-sandbox for M5 confirmation. Use when: the user runs /lich-review on a hunk / file / PR, or the PostToolUse hook fires on Write/Edit/MultiEdit. Do not use for: security-taint review (Hydra R3 owns that), change classification (Crow V1/V2 owns that), or rubric-style judgment (lich-rubric skill owns that).
|
| model | sonnet |
| tools | ["Read","Grep","Glob"] |
lich-review
Preconditions
- A
lich-core sub-plugin state dir exists at plugins/lich-core/state/.
- Target code parses under the detected language's substrate (Python
ast, TypeScript tsc --generateTrace). If parsing fails, emit substrate-parse-failed and skip — do not fabricate flags.
- Crow's
change-tracker/state/audit.jsonl is optionally present. If absent, Lich runs on the full file instead of Crow-flagged hunks.
Inputs
- Slash command:
/lich-review <scope> where scope is hunk (current file + line range), file (full file path), or pr (all changes in the PR).
- Hook payload: PostToolUse event with
tool, file_path, old_string, new_string.
Steps
- Parse the target. Use
ast.parse for Python, subprocess tsc --noEmit --generateTrace for TypeScript. On parse failure, emit substrate-parse-failed and return empty flags.
- Run M1 Cousot Interval Propagation. Walk the AST applying interval + nullability + container-shape abstract domains. Threshold-widen at N=3 iterations with language-aware bounds (
{0, 1, -1, sys.maxsize} for Python ints). Emit per-site flags: {file, line, variable, abstract_value, failure_class, severity}.
- Run M2 Falleri Structural Diff if comparing two versions. Use conservative GumTree parameters (
min_height=3, min_dice=0.6, min_similarity=0.7). Time budget 2 seconds per file; on timeout, fall back to unified diff and flag structural-diff-timeout.
- Read Hydra's vuln-detector audit.jsonl. If a CRITICAL or HIGH CWE finding exists on the target file, boost the review-attention weight for M6's prioritization and annotate M7's rubric input with "Security context: Hydra flagged {cwe} {severity}". Never re-scan for the CWE itself.
- Emit flagged sites. Write to
plugins/lich-core/state/review-flags.jsonl for lich-sandbox to pick up. Fields: {ts, file, line, failure_class, severity, M1_confidence, needs_M5_confirmation}.
Outputs
plugins/lich-core/state/review-flags.jsonl — append-only flag records.- stderr: short summary of M1 findings count, M2 edits detected, Hydra context applied.
- Return value to parent: a JSON block
{flags: [...], M2_edits: N, duration_ms: X}.
Handoff
Next skill in the chain: lich-sandbox (SKILL at plugins/lich-sandbox/skills/lich-sandbox/SKILL.md) — confirms each flagged site via bounded subprocess dry-run.
After both skills run, lich-verdict composes the final DEPLOY/HOLD/FAIL verdict.
Failure modes
When this skill fails, log to state/learnings.json with one of the codes from @../vis/packages/core/conduct/failure-modes.md:
- F02 fabrication — if M1 reports a flag the AST walk doesn't actually produce (never do this; soundness is the brand)
- F04 task drift — if the skill strays into security-taint detection (Hydra's lane)
- F13 distractor pollution — if M2 emits > 20 edits on a small diff; revert to conservative parameters
- F14 version drift — if the TypeScript substrate uses an unsupported
tsc version; emit a compatibility warning
