audit

Interactive, point-by-point walkthrough of a review report produced by any review skill (skill-adversary, critical-code-reviewer, or any other). Two modes: **orchestrator mode** (provide a target + optional `--reviewer` flag — detects deployment context, calibrates severity, launches the reviewer, then walks through its report) and **walkthrough-only mode** (processes an existing report from the conversation). Parses review findings and processes each one at a time: re-evaluates validity, proposes and applies fixes, checks impacted files for regressions, and waits for user approval before moving on. Adversarial cross-provider validation (L2) is always active on Blocking/Required findings when `OPENROUTER_API_KEY` is set. Usage: `/walkthrough [target] [--reviewer name] [--batch|--no-batch]` Orchestrator mode: provide a target (file, directory, or glob). The set of available reviewers is discovered at runtime by scanning installed Claude Code skills — no hardcoded list. If `--reviewer` is omitted, the orchestrator detects the target type, suggests an adapted reviewer from the scanned set, and asks the user to confirm or pick another — there is no silent default. Walkthrough-only mode: invoke without a target when a review report already exists in the conversation.

Full-coverage project review: detects project type and size, spawns specialist background agents with disjoint scopes, consolidates all findings into one deduplicated report sorted by severity, then offers interactive walkthrough. Intent: the user wants a comprehensive, multi-facet review of an entire project — not a single file, not a PR diff, not a non-code document. Trigger on: "full review", "review complète", "revue complète", "full audit", "audit complet", "lance la review complète", "/sweep", "review this project", "audit this repo", "go through everything", "check the whole project", "thorough review", "passage complet sur le projet", "tout checker", "état des lieux du projet", "diagnostic complet", or any request that asks for a multi-angle assessment of an entire codebase covering architecture, quality, tests, and/or docs. EXCLUSIONS OVERRIDE TRIGGERS — do NOT trigger on: single-file reviews (even if phrased as "full review of this file"), PR reviews or diffs, non-code document reviews (papers, resumes, CVs), CI/build commands ("lance tout" meaning "run tests + build"), or when the user explicitly asks for only /walkthrough.

Circularity-aware orchestrator for audit skills. Detects when an audit skill (skill-adversary, mcp-adversary, sweep, critical-code-reviewer) is about to review an artifact that shares its own codebase, prompts, or model family — then injects cross-model judging and transparency countermeasures via OpenRouter. EXPLICIT-INVOCATION ONLY: this skill never auto-triggers from implicit context (enforced by `disable-model-invocation: true`). Invoke directly via "/blindspot" or by name ("lance blindspot", "blindspot review"). Use when: the user explicitly invokes blindspot, asks for a "circular review" / "review circulaire" / "self-review check", flags that reviewer and target share the same origin ("review [skill] with [skill]", "review this skill with skill-adversary" on a Claude Code skill), or asks for circularity countermeasures alongside another audit skill ("check for circularity", "add a second opinion", "get an external judge"). Do NOT apply to: general code review, PR review, plain skill-adversary/mcp-adversary/sweep invocations without circularity concern, or LLM evaluation methodology discussions.

Adversarial critic for Claude Code skills. Reviews a skill's full directory (SKILL.md, agents, docs, templates) and produces a structured report of flaws: trigger edge cases (false positives and false negatives), instruction ambiguities, contradictions, cross-file coherence issues, and gaps. Use this skill whenever the user asks to critically review, audit, stress-test, attack, or find flaws in a skill (SKILL.md file). Also trigger on: "adversary review", "attack this skill", "find trigger edge cases", "test my skill description", "audit my skill", "what's wrong with this skill", "review my SKILL.md", or any request to find weaknesses in a skill's triggering or instructions. Also trigger on symptom-based requests: "my skill triggers on the wrong things", "skill fires incorrectly", "description is too broad/narrow", "skill triggers when it shouldn't". Do NOT trigger for: general code review (use critical-code-reviewer), reviewing non-skill files, creating/editing skills (use skill-creator), or any use of "trigger", "edge cases", "stress-test", "skill description" in non-skill contexts (state machines, CI/CD, job postings, APIs).

Adversarial critic for MCP servers. Reviews an MCP server's tool descriptions, parameter schemas, and implementation code to find flaws: inter-tool discrimination issues, discoverability gaps, schema anti-patterns, semantic drift between descriptions and behavior, error handling inconsistencies, and undocumented workflow dependencies. Use when the user asks to review, audit, stress-test, or find flaws in an MCP server. Also trigger on: "review my MCP tools", "audit my MCP server", "are my tool descriptions good", "find issues in my MCP server", "test my tool schemas", "what's wrong with my MCP server", "MCP tool quality", "review my tool descriptions". Do NOT trigger for: general code review (use critical-code-reviewer), security scanning (use mcp-scan), reviewing Claude Code skills (use skill-adversary), creating MCP servers, or non-MCP tool/API review.