playbook/antigravity-awesome-skills/plugins/antigravity-awesome-skills-claude/skills/logic-fix-all/guide-phases-0-2-consent-scope-health.md

Logic-Lens — Logic Fix All — Phases 0-2 (Consent · Scope · Health)

---

Phase 0 — Pre-flight Notice & Consent Gate

0a. Estimate file count (git ls-files | wc -l in a git repo; find . -type f -not -path '*/.git/*' -not -path '*/node_modules/*' -not -path '*/target/*' -not -path '*/.venv/*' -not -path '*/build/*' -not -path '*/dist/*' -not -path '*/vendor/*' | wc -l otherwise). Display this notice verbatim with the estimate filled in — do not paraphrase:

⚠️  /logic-fix-all — Full Repository Logic Audit & Fix

Scope:    ENTIRE repository, not just recent commits or staged changes.
          Includes runtime-affecting files: source code, runtime
          config (.json/.yaml/.toml/.ini), constraint files
          (CLAUDE.md, .logic-lens.yaml, AGENTS.md, etc.), and
          behavioral documentation (README, ARCHITECTURE, ADRs).
          Auto-excludes .git, build artifacts, dependency caches,
          binary assets; respects .gitignore and .logic-lens.yaml
          `ignore:` patterns.
Estimated files to scan: ~N

Method:   Semi-formal execution tracing — Premises → Trace →
          Divergence → Remedy. This is a LOGIC review, not a
          syntax/style/lint pass.

Skills:   logic-health → logic-review → logic-locate → logic-explain
          → logic-diff, iterated until clean.

Token cost: HIGH. The pipeline uses ranked passes and scope caps, but
deep tracing still costs roughly 5k–15k tokens per reviewed file
(more for deeply interprocedural code, less for stateless utilities),
times ~1.3 for iteration rounds.
Your estimate: min(N, 100) reviewed files × ~10k tokens × 1.3 ≈
(compute and show here, e.g. "~1M tokens").

Git impact: The pipeline edits source files. It does NOT commit,
push, or amend. If you have uncommitted work, commit or stash first.

Iteration: Critical findings loop until resolved (no cap). Warnings
and Suggestions default to 3 rounds, configurable via
`.logic-lens.yaml` `fix_all.max_iterations:`.

Proceed with full autonomous run? [Y/n]

0b. Parse the user's reply. Maintain two phase-local counters (reset when Phase 0 exits): consecutive_pauses, consecutive_questions.

Signal sets (case-insensitive):

• Consent: Y, yes, ok, sure, proceed, go, continue, 继续, 好, 好的, 行, 可以
• Hard negation: no, n, abort, cancel, 取消, don't, 不要
• Soft-pause: wait, hold on, not yet, 一下, 先别, 等一下, 等我, let me

Decision (first match wins):

1. Hard negation → abort: "Aborted by user before scan — no files modified".
2. Consent + soft-pause → increment consecutive_pauses, acknowledge in one line, wait for next message. If consecutive_pauses would reach 3, re-show the full Phase 0 notice and reset to 0.
3. Consent (no negation) → proceed. Honor any scope/language instructions in one acknowledgment line.
4. Question → increment consecutive_questions, answer once, re-prompt. If consecutive_questions would reach 2, fall through to rule 5.
5. Unmatched (or falling through from rule 4) → re-show notice once; if next reply still unmatched, treat as abort.

0c. After consent, ask no further questions until Phase 8 cap escalation.

---

Phase 1 — Scope Enumeration

1a. Read .logic-lens.yaml (if present): load only ignore, focus, disable, custom_risks, severity:, trace.*, and fix_all.max_iterations. Apply ignore immediately.

1b. Detect project type from marker files and derive exclusions:

• package.json → exclude node_modules/, dist/, build/, .next/, .nuxt/, coverage/
• Cargo.toml → exclude target/
• go.mod → exclude vendor/ (unless project-owned — check modules.txt)
• pyproject.toml/requirements.txt/Pipfile → exclude .venv/, venv/, __pycache__/, *.egg-info/, .pytest_cache/, .mypy_cache/, .ruff_cache/
• Gemfile → exclude vendor/bundle/
• pom.xml → exclude target/
• build.gradle/build.gradle.kts → exclude build/, .gradle/
• build.sbt → exclude target/, project/target/
• mix.exs → exclude _build/, deps/
• composer.json → exclude vendor/
• *.csproj/*.sln → exclude bin/, obj/
• pubspec.yaml → exclude .dart_tool/, build/
• Always exclude: .git/, .DS_Store, lock files (*.lock, package-lock.json, yarn.lock, Pipfile.lock, poetry.lock, Cargo.lock, go.sum), log files, binaries (.png/.jpg/.gif/.pdf/.wasm/.zip/.tar/.gz/.woff*/.ttf)
• Respect .gitignore as a hint, not absolute — some ignored paths may still be relevant.

1c. Classify every non-excluded file into exactly one bucket:

• Source code: files whose extension matches a language the project uses (inferred from markers in 1b).
• Runtime config: .json, .yaml/.yml, .toml, .ini, .conf, *.config.js/ts — verify by grepping the codebase for the filename before classifying.
• Constraint files: CLAUDE.md at every level, .logic-lens.yaml, AGENTS.md, GEMINI.md, schema files (*.schema.json, openapi.yaml, *.proto, *.graphql).
• Behavioral docs: README.md, CONTRIBUTING.md, ARCHITECTURE.md, docs/**/*.md describing runtime behavior, .env.example. Skip changelogs, licenses, marketing copy, .editorconfig.

1d. Classify each file by risk tier:

• High: public API surfaces; files changed in the last 30 days (git log --since=30.days --name-only --pretty=format: | sort -u); core business logic without test coverage.
• Medium: utility modules, helpers, non-core configs, stable constraint files.
• Low: stable well-tested code, stable docs.

Newly added files are already High. Constraint/behavioral-doc files are Medium by default, upgraded to High if referenced by recently-changed code. When a file matches multiple criteria, assign the highest tier.

1e. Sort: High → Medium → Low; within each tier, descending line count.

1f. Scope caps:

• >20 files: Low-tier files reviewed at reduced depth (top 3 non-trivial functions only).
• >100 files: keep only top 100 by (tier desc, line-count desc); drop the rest. Note truncation in the Fix Report.

1g. State the final file list at the start of the Fix Report: file name + tier + role.

---

Phase 2 — Health Pass (logic-health)

2a. Apply ../logic-health/logic-health-guide.md methodology to the Phase 1 file list, including its module/function budgets. Output: per-module Logic Score, aggregated findings by L-code, systemic patterns.

2b. Record Phase 2 output for reference. Do NOT write remedies yet — health gives shape, not precision. Precise findings come from Phase 3.

2c. If the health pass reveals a systemic pattern (same L-code in 4+ modules), earmark the representative file for Phase 3 priority review. Phase 3 must produce the full Premises→Trace→Divergence triple before the pattern can enter Phase 6 as a fix candidate — a systemic observation without a trace cannot justify a remedy.