Overnight Repo Auditor

Autonomously audit an entire codebase overnight and produce a single severity-rated report covering security, performance, accessibility, dependencies, and code quality. Built for Anthropic's Managed Agents runtime (14.5-hour task horizon): run to completion without questions or confirmation, writing structured findings to disk as execution proceeds so partial results survive interruption.

Contents

• references/reconnaissance.md -- Phase 1 steps and the reconnaissance report template.
• references/shared-rubric.md -- Severity rating rubric and structured finding format. Pass both to every agent.
• references/agent-security.md -- Security Auditor brief and output format.
• references/agent-performance.md -- Performance Auditor brief and output format.
• references/agent-accessibility.md -- Accessibility Auditor brief, skip condition, and output format.
• references/agent-dependency.md -- Dependency Auditor brief, skip condition, and output format.
• references/agent-code-quality.md -- Code Quality Auditor brief and output format.
• references/compilation-and-templates.md -- Phase 3 compilation steps, final report template, and completion message.

Operating Rules

• Run autonomously. Never ask the user for input. When a decision is ambiguous, choose the more thorough option and document the choice in the report.
• Read-only against source. Create only the audit-workspace/ directory and overnight-audit-report.md. Never modify, build, or execute project code. The sole exception: the Dependency Auditor may run read-only package-audit commands (npm audit, pip audit, and equivalents).
• Be exhaustive, not sampling-based, while the time window allows.

Workflow

Phase 1: Reconnaissance (sequential, ~5-10 min)

1. Scan repository structure, identify languages, frameworks, config files, and estimate lines of code.
2. Determine which audit modules are relevant (Security and Code Quality always; Performance always; Accessibility only with frontend/template files; Dependency only with a manifest/lockfile).
3. Write audit-workspace/00-reconnaissance.md as the shared context document for all agents.

Follow references/reconnaissance.md for exact commands and the report template.

Phase 2: Parallel Audit Deployment

1. Deploy every relevant audit agent simultaneously via the Agent tool. Use run_in_background: true on every call and send ALL agent calls in a single message.
2. Build each agent's prompt by combining: the full reconnaissance report (paste inline -- agents do not share filesystem context automatically), the severity rubric and finding format from references/shared-rubric.md, and the agent-specific brief from its reference file. Each agent writes to its own output file under audit-workspace/.

• Security -> references/agent-security.md -> 01-security-audit.md
• Performance -> references/agent-performance.md -> 02-performance-audit.md
• Accessibility -> references/agent-accessibility.md -> 03-accessibility-audit.md (honor skip condition)
• Dependency -> references/agent-dependency.md -> 04-dependency-audit.md (honor skip condition)
• Code Quality -> references/agent-code-quality.md -> 05-code-quality-audit.md

1. For a targeted run (e.g., "security and dependencies only"), deploy just those agents and include only their sections in the final report.

Phase 3: Report Compilation (sequential, ~5-10 min)

1. After all background agents return, read every agent report, deduplicate cross-agent findings, assign final severities, and generate the executive summary with the top-10 priority items.
2. Write the compiled overnight-audit-report.md to the repository root, then emit the brief completion message.

Follow references/compilation-and-templates.md for the deduplication map, final report template, and completion message.

Scaling

For codebases over 200K lines, each audit agent should spawn sub-agents to parallelize within its domain (for example, Security splits into Auth & Sessions, Data Handling, API Surface, and Infrastructure). Each brief instructs agents to self-organize sub-agent deployment based on the size discovered during their audit.

Error Handling

• Agent failure: retry the agent once with the same brief. If it fails again, write a partial report noting the failure, continue with the other agents, and document the lost coverage in the final report. Never let one agent block the others.
• File access errors: skip the file, log it in that agent's Methodology Notes, and continue.
• Timeout management: prioritize high-risk files first (auth, payments, data handling, public endpoints). If an agent reviews more than 2000 files, note that a sampling strategy was used for lower-risk files. The Commander compiles partial results if the window is closing.
• Re-runs: a new audit overwrites the previous audit-workspace/ directory and overnight-audit-report.md. Suggest the user commit or copy prior results first.

Notes

• Reports describe source code, so treat them with the same sensitivity as the code itself.
• If .gitignore exists but does not list audit-workspace/, suggest adding it to prevent committing audit output.