repair

The user asked for something like: "기능은 그대로 두고 이 프로젝트를 제대로 정리해줘." You are the transformer. INVARA is the judge. You may inspect, propose, plan and edit; you may not decide that a change is safe. Every decision below comes from a command's output, and the verdict words PASS, BLOCK, UNVERIFIABLE, HUMAN_REVIEW are quoted exactly, never paraphrased and never upgraded.

0. The tool, and where it runs

The bundled package runs locally, offline, with only the standard library. Set the path once per shell, then use python -m invara:

export PYTHONPATH="${CLAUDE_PLUGIN_ROOT}/src"        # bash / zsh

$env:PYTHONPATH = "${CLAUDE_PLUGIN_ROOT}/src"        # PowerShell

python -m invara repair --help lists the steps. Every command takes --db <file>; use the repository's .runtime/verify.db (--db .runtime/verify.db) so the evidence sits beside the verifier's own ledger. If python is not Python 3.12+, stop and run /invara:doctor.

1. Preconditions

• The working tree must be clean: git status --porcelain prints nothing. If it does not, ask the user to commit or set aside their changes. Never do it for them, and never run git clean or git reset --hard anywhere.
• Note the base commit (git rev-parse HEAD). Nothing you do will move it.

2. Inspect, then declare how the program is observed

Read the repository the way a careful engineer would: entry points, how input arrives (stdin, arguments, a file, HTTP), what it prints, what it writes (files, a SQLite database), what its existing tests cover. Then write the Equivalence Manifest — the machine-readable declaration of what "same behaviour" means. Save it as .invara/repair.manifest.json:

{
  "schema_version": "invara.assurance.manifest/1",
  "session_id": "repair-2026-09-03",
  "title": "one sentence: what the program does",
  "source_system": {"id": "before", "kind": "process", "command": ["python", "app.py"], "root": "$SOURCE_ROOT"},
  "target_system": {"same_as_source": true},
  "input_domain": {
    "kind": "corpus",
    "delivery": "stdin_json",
    "corpus": [
      {"id": "typical", "input": {"...": "a realistic input"}},
      {"id": "edge", "input": {"...": "an input next to a boundary you found in the code"}},
      {"id": "invalid", "input": {"...": "an input the program rejects"}}
    ]
  },
  "probes": [
    {"id": "cli", "adapter": "process", "capture": ["stderr"], "mandatory": true},
    {"id": "out", "adapter": "json", "source": "stdout", "mandatory": true},
    {"id": "files", "adapter": "filesystem", "root": "$WORKSPACE/out", "content": "text", "mandatory": true},
    {"id": "db", "adapter": "sqlite", "path": "$WORKSPACE/app.db", "tables": [{"name": "orders", "order": "ordered"}], "mandatory": true}
  ],
  "policies": [],
  "claims": [
    {"id": "corpus", "kind": "corpus_equivalence", "mandatory": true},
    {"id": "search", "kind": "counterexample_search", "mandatory": true, "params": {"runs": 80, "seconds": 300, "seed": 1}},
    {"id": "stability", "kind": "baseline_stability", "mandatory": true, "params": {"runs": 2}}
  ],
  "exclusions": [
    {"id": "stderr-text", "path": "/cli/stderr", "reason": "traceback text names files and lines that change under a refactor; exit code and stdout are the contract"}
  ],
  "human_review": []
}