# Audit Agents

> Enforces decentralized task management, UI/UX design strictness, and global state constraints within AGENTS.md.

- Skill: `mokhtarabadi/audit-agents` (Agent Skill)
- Install (CLI): `npx skillmds@latest add mokhtarabadi/audit-agents`
- Raw SKILL.md: https://api.skillmd.com/api/skills/mokhtarabadi/audit-agents/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Security
- Author: mokhtarabadi (https://skillmd.com/u/mokhtarabadi)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/mokhtarabadi/audit-agents

---


# Skill: Agent Protocol Auditor (Project-Agnostic)

## 🛑 SCOPE CONFINEMENT (Priority 0)

- All file enumeration, inspection, and patch operations MUST be strictly confined to the caller's current working directory (`[PROJECT_ROOT]`).
- You are STRICTLY FORBIDDEN from traversing outside `[PROJECT_ROOT]` or referencing parent directories. Use generic placeholders like `[PROJECT_ROOT]/AGENTS.md`.
- **Absent-File Policy**: If optional architectural files (`DESIGN.md`, `docs/architecture.md`, `docs/data_model.md`) do NOT exist in `[PROJECT_ROOT]`, SKIP them gracefully with an explicit note. DO NOT HALT. DO NOT HALLUCINATE their contents. DO NOT scaffold or create them unless explicitly instructed.
- **OpenCode Isolation**: You are STRICTLY FORBIDDEN from creating `.opencode/` scaffolding inside third-party projects. Only inspect `.opencode/` if `[PROJECT_ROOT]/.opencode/` ALREADY exists OR if the user passes `with_opencode: true`.

## Target Audit Criteria

The `AGENTS.md` file MUST explicitly contain the following operational constraints, ideally within a `Task Management & OpenCode Rules` section:

- **Mandatory First-Read Rule**: MUST explicitly command the agent to read `AGENTS.md` first before any execution. Inside it, it must route the agent to read `DESIGN.md`, `docs/architecture.md`, `docs/data_model.md`, and `docs/conventions.md` first.
- **Core File Locations**: MUST explicitly list paths for `AGENTS.md`, `DESIGN.md` (if present, else note absent per Absent-File Policy), `docs/conventions.md`, and the 5 Kanban directories (`tasks/backlog`, `tasks/in-progress`, `tasks/qa`, `tasks/completed`, `tasks/archive`). Only require `.opencode/skills/` when the project already contains `.opencode/` or `with_opencode: true` is set.
- **conventions.md Compliance**: The project MUST have a `docs/conventions.md` file containing the Universal DateTime Standard (UTC at rest, Epoch/ISO-8601 with Offset at API boundaries, Clock injection, Dual-Representation for future events, TZ=UTC Infrastructure), SOLID Programming Guidelines (SRP, OCP, LSP, ISP, DIP, Pragmatic Guardrails), Universal Financial Ledger Standard (snapshot-on-write, `$ifNull` precedence, discrepancy alerting, deep config merging), and Defensive Shell Protocol (DSP) (`set -euo pipefail`, banned error masking, sidecar isolation).
- **Decentralized Task Management**: Agents MUST strictly use decentralized, individual task files in the Kanban directories (`tasks/backlog`, `tasks/in-progress`, `tasks/qa`, `tasks/completed`, `tasks/archive`) as their single source of truth.
- **No Monolithic State**: Agents are strictly forbidden from creating `TODO.md` or `STATE.md`.
- **Zero-Autonomous-Commit**: Agents MUST be strictly forbidden from executing Git commands autonomously; they may only run Git commands when explicitly instructed by the Orchestrator. **Exception:** `git mv` is permitted for moving task files between Kanban directories (`backlog`, `in-progress`, `qa`, `completed`, `archive`).
- **Mandatory End-Of-Task Sequence**: MUST explicitly mandate a 5-step completion process: 1) Update CHANGELOG.md. 2) Write manual reasoning in the task file. 3) Call the `custom_context_stage_and_inject_diff` MCP tool, then `git mv` the task to `tasks/qa/` (NO COMMITS ALLOWED). 4) Synchronize the task file's `**File:**` metadata to the new path and re-run lint + stage at the new path. 5) Notify the Manager.
- **UI/UX Enforcement**: Any UI/UX changes MUST enforce the guidelines defined in the project's `DESIGN.md`.
- **Task-Generator Skill Loading**: `AGENTS.md` MUST explicitly instruct the Hands to load the `task-generator` skill before creating new task files.
- **Project Skill Loading**: `AGENTS.md` MUST explicitly instruct the Hands to load every available skill matching the project's tech stack before task implementation.
- **Complex Debugging**: Agents MUST be instructed not to guess blindly on complex bugs, but instead utilize the `debug-instrumentation` skill.
- **MCP Report Generation**: `AGENTS.md` MUST instruct agents to generate context reports (`custom_context_read_source_files`) and tree reports (`custom_context_create_tree_report` — "create a tree of the project") via the MCP server and hand the file path to the Manager instead of reading `context-reports/` files inline.
- **Explicit Staging Contract (F5)**: Verify that the active task's `Execution Log & Reasoning` or `summary_phase` passed a `modified_files` list to `stage_and_inject_diff` — blind `git add -A .` staging is banned because it sweeps parallel-session files into unrelated commits.
- **Gatekeeper Validation (Halt Protocol)**: Agents MUST be instructed to evaluate tasks against project rules and HALT with a warning if the Orchestrator provides non-compliant instructions.
- **Context Bootstrapping**: `AGENTS.md` MUST explicitly instruct the Hands: "At the start of every task, you MUST call `search_memory` or `list_namespaces` to load any hidden project quirks relevant to your domain before implementing."
- **Buffer Isolation**: The shared validation phase MUST include a buffer-flush directive requiring Hands to treat every task as contextually independent, preventing cross-task context leakage.
- **Defensive Shell Protocol (DSP)**: `AGENTS.md` MUST include a guardrail forbidding bash scripts without `set -euo pipefail` and banning `2>/dev/null` on data commands. `docs/conventions.md` MUST contain a `## Defensive Shell Protocol (DSP)` section.
- **Universal Financial Ledger Standard**: `AGENTS.md` MUST include a guardrail requiring snapshot-on-write for financial mutations and `$ifNull` precedence for monetary aggregations. `docs/conventions.md` MUST contain a `## Universal Financial Ledger Standard` section.
- **Lite Mode Protocol**: `AGENTS.md` MUST document the `<lite_mode_protocol>` — when eligible (single-file, no security/financial impact, obvious simplicity), the full 9-step production line can be bypassed with a `[LITE]` justification in the task's `## Execution Log & Reasoning` section. Escalation to Full Mode is mandatory if hidden complexity is discovered.
- **Deprecated-Section Purge Rule**: Scans `AGENTS.md` and all task files across `tasks/` (excluding archive and completed history) for deprecated sections: `## Manager Decisions`, `## Admin Decision`, `Manager Decision`, `Admin Decision`. When detected, the auditor MUST purge the entire deprecated section from the target file, document the purge in the audit findings/changelog, and MUST NOT flag their absence as a missing requirement or recreate them.
- **Plugin Runtime-State gitignore**: If the project uses OpenCode plugins that write per-project state, `.gitignore` MUST cover plugin runtime-state paths (e.g. worktree checkouts, session/state JSON, goal-state dirs) while MUST NOT ignore deliberate config overrides checked in on purpose (e.g. a project-level plugin config pinning team-shared settings). Audit `.gitignore` read-only first; patch only paths belonging to plugins actually detected in the project's `opencode.json`/`tui.json` `plugin` arrays — never speculative entries.

---

## Core Document Templates

### 1. `architecture.md` Template

```markdown
# Architecture Overview

This document serves as a critical, living template designed to equip agents with a rapid and comprehensive understanding of the codebase's architecture, enabling efficient navigation and effective contribution from day one. Update this document as the codebase evolves.

## 1. Project Structure

[Project Root]/
├── backend/ # Contains all server-side code and APIs
│ ├── src/ # Main source code for backend services
│ │ ├── api/ # API endpoints and controllers
│ │ ├── client/ # Business logic and service implementations
│ │ ├── models/ # Database models/schemas
│ │ └── utils/ # Backend utility functions
│ ├── config/ # Backend configuration files
│ ├── tests/ # Backend unit and integration tests
│ └── Dockerfile # Dockerfile for backend deployment
├── frontend/ # Contains all client-side code for user interfaces
│ ├── src/ # Main source code for frontend applications
│ │ ├── components/ # Reusable UI components
│ │ ├── pages/ # Application pages/views
│ │ ├── assets/ # Images, fonts, and other static assets
│ │ ├── services/ # Frontend services for API interaction
│ │ └── store/ # State management (e.g., Redux, Vuex, Context API)
│ ├── public/ # Publicly accessible assets (e.g., index.html)
│ ├── tests/ # Frontend unit and E2E tests
│ └── package.json # Frontend dependencies and scripts
├── common/ # Shared code, types, and utilities used by both frontend and backend
│ ├── types/ # Shared TypeScript/interface definitions
│ └── utils/ # General utility functions
├── docs/ # Project documentation (e.g., API docs, setup guides)
├── scripts/ # Automation scripts (e.g., deployment, data seeding)
├── .github/ # GitHub Actions or other CI/CD configurations
├── .gitignore # Specifies intentionally untracked files to ignore
├── README.md # Project overview and quick start guide
└── ARCHITECTURE.md # This document

## 2. High-Level System Diagram

[User] <--> [Frontend Application] <--> [Backend Service 1] <--> [Database 1]
|
+--> [Backend Service 2] <--> [External API]

## 3. Core Components

### 3.1. Frontend

Name: [Web App, Mobile App]
Description: [Purpose, core interfaces, and roles]
Technologies: [e.g., React, Next.js, Jetpack Compose, Swift/Kotlin]
Deployment: [e.g., Vercel, Netlify, Play Store, App Store]

### 3.2. Backend Services

#### 3.2.1. Service Name 1

Name: [e.g., API Service]
Description: [Core business roles]
Technologies: [e.g., Spring Boot, Node.js Express, Go]
Deployment: [e.g., AWS ECS, Kubernetes]

## 4. Data Stores

### 4.1. Data Store 1

Name: [e.g., SQL Database]
Type: [e.g., PostgreSQL, MongoDB]
Purpose: [e.g., Account records]

## 5. External Integrations / APIs

- Service Name: [e.g., Stripe, SendGrid]
- Purpose: [e.g., Payments]
- Method: [e.g., REST, SDK]

## 6. Deployment & Infrastructure

- Provider: [e.g., AWS, GCP]
- CI/CD: [e.g., GitHub Actions]

## 7. Security Considerations

- Authentication: OAuth2/JWT
- Encryption: TLS + AES-256

## 8. Development & Testing Environment

Testing Frameworks: [e.g., Pytest, JUnit, Jest]

## 9. Future Considerations / Roadmap

[Planned changes or architectural debt]
```

### 2. DESIGN.md Template (Google Spec)

```markdown
# Design System Specification

---

name: custom-ui-system
colors:
primary: "#1D4ED8"
secondary: "#4B5563"
background: "#F3F4F6"
surface: "#FFFFFF"
text: "#111827"

---

## 1. Visual Theme & Atmosphere

[Rich prose outlining overall design mood, light/dark values, and whitespace philosophy]

## 2. Color Palette & Roles

- Primary foundation
- Interactive / CTAs
- Text hierarchy
- State colors (Success, error, warn)

## 3. Typography Rules

- Hierarchy (headline, body, label)
- letterSpacing, lineHeight, fontWeights

## 4. Component Stylings

- Buttons
- Cards
- Navigation
- Input fields

## 5. Layout Principles

- Spacing scales
- Breakpoints
```

### 3. `docs/conventions.md` Template

Generate a `docs/conventions.md` file containing the Universal DateTime Standard and SOLID Programming Guidelines:

```markdown
# Conventions

This document defines syntax rules, naming conventions, file boundaries, and automation patterns for this project.

## Universal DateTime Standard

All projects in this ecosystem MUST follow these datetime rules:

1. **UTC at Rest** — All databases and caches store datetimes in UTC with `TIMESTAMP WITH TIME ZONE`. Banned: naive or local-time storage.
2. **ISO-8601 with Offset / Epoch ms at API Boundaries** — APIs transmit datetimes as Unix Epoch milliseconds (int64) or ISO-8601 with offset (e.g., `2026-07-23T14:30:00+00:00`). Banned: timezone-naive strings.
3. **Clock Injection** — All current-time access must go through an injectable `Clock` abstraction. Banned: direct `new Date()`, `datetime.now()`, `time.Now()` in business logic.
4. **Dual-Representation for Future Events** — Calendar events expose both `event_start_local` (with timezone) and `event_start_epoch_ms` (absolute).
5. **`TZ=UTC` Infrastructure** — All environments run with `TZ=UTC`. Timezone display is a client-layer responsibility only.

## SOLID Programming Guidelines

Enforce these SOLID principles and pragmatic guardrails in every implementation:

1. **SRP** — One reason to change per module. Split merged concerns.
2. **OCP** — Open for extension, closed for modification. Use composition over inheritance.
3. **LSP** — Subtypes must be substitutable. Ban `NotImplementedError` overrides.
4. **ISP** — Small role-specific interfaces. Ban monolithic god-interfaces.
5. **DIP** — Depend on abstractions, not concretions. Core layer must not import adapters.

**Pragmatic Guardrails:** No abstraction for <3 trivial operations. Only extract interfaces with 2+ implementations. Apply YAGNI strictly. Prefer simpler designs unless a measurable requirement forces complexity.

## Universal Financial Ledger Standard

All financial, transactional, and countable data operations MUST enforce these mandates:

1. **Snapshot-on-Write for Mutable Totals:** Whenever a financial amount, inventory count, or balance is mutated, persist a read-only snapshot of the preceding state in the same transaction (sidecar table, audit log, or WAL). Banned: mutating without preserving the prior value.
2. **Mandatory `$ifNull` Precedence:** All aggregation queries on monetary fields MUST use explicit null-handling (`COALESCE`, `ISNULL`, `$ifNull`). Banned: passing nullable columns into mathematical operators.
3. **Observability Alerting on Discrepancies:** If a computed total diverges from its line-item sum by more than 0.01, emit a high-severity alert and prevent finalization.
4. **Deep Config Merging for Financial Settings:** Financial configuration updates MUST deeply merge nested properties. Banned: shallow object spread on financial config objects.

## Defensive Shell Protocol (DSP)

When writing or reviewing bash scripts, cron jobs, or container orchestration commands:

1. **Mandatory Strict Mode:** All scripts MUST start with `set -euo pipefail`.
2. **Banned Error Masking:** `2>/dev/null` is STRICTLY FORBIDDEN on data-generation, backup, archive, or database commands.
3. **No Post-Redirect Status Checks:** Never use `command > file; if [ $? -eq 0 ]` — the shell creates the file before running the command, masking failures.
4. **Sidecar Isolation for Hostless Backups:** Never rely on host file staging for Docker volume backups. Always use ephemeral containers (`docker run --rm -v volume:/data:ro alpine tar...`) with read-only mounts.
```

---

Use this skill in two modes:

- **Phase 0 (Generation):** When `AGENTS.md` does not exist yet — generate it from the template below.
- **Audit Mode (Existing):** When `AGENTS.md` already exists — audit and patch it against the Target Audit Criteria.

---

## Mode 1: Phase 0 — Generate AGENTS.md & docs/conventions.md

Use this when a project has no `AGENTS.md` yet (new project onboarding).

### Workflow

1. Read the project's existing context (package configs, README, tech stack files) to determine the project name, description, and relevant tech stack skills.
2. Generate `AGENTS.md` at the project root using the template below.
3. Fill in the `[bracketed]` placeholders with the actual project details.
4. Generate `docs/conventions.md` using the conventions template above, if the project does not already have one.
5. Confirm both files were created.

### AGENTS.md Template

```markdown
# [Project Name] — Project Context Hub

## Project Overview

[Brief description of the project, its purpose, and tech stack]

## Setup & Dev Commands

- Build: [build command, e.g., npm run build]
- Test: [test command, e.g., npm test]
- Lint: [lint command, e.g., npm run lint]
- Dev: [dev server command, e.g., npm run dev]

## Actionable Guardrails (Do's & Don'ts)

- **Don't** [common anti-pattern to avoid]
  -> **Do** [preferred alternative]
- **Don't** [another anti-pattern]
  -> **Do** [preferred alternative]
- **Don't** read `context-reports/` markdown files yourself.
  -> **Do** generate them using the MCP server — context reports via `custom_context_read_source_files`, tree reports via `custom_context_create_tree_report` ("create a tree of the project") — and hand the file path to the Manager.
- **Don't** execute Git commands like `git add`, `git commit`, or `git mv` autonomously or try to guess when to stage code.
  -> **Do** execute Git commands ONLY when explicitly instructed by an Orchestrator task block. Otherwise, rely on the `custom_context_stage_and_inject_diff` MCP tool.
  -> **Exception:** `git mv` is permitted autonomously for moving task files between Kanban directories.
- **Don't** guess blindly when facing complex bugs, deadlocks, or silent timeouts.
  -> **Do** utilize the `debug-instrumentation` skill to inject strategic logs and trace the runtime execution path.
- **Don't** write bash scripts without strict mode or mask errors with `2>/dev/null` on data commands.
  -> **Do** follow the Defensive Shell Protocol: `set -euo pipefail`, ban error masking, sidecar isolation for Docker backups. See `docs/conventions.md`.
- **Don't** perform financial mutations without snapshotting the prior state or allow nulls in monetary aggregations.
  -> **Do** follow the Universal Financial Ledger Standard: snapshot-on-write, `$ifNull` precedence, discrepancy alerting, deep config merging. See `docs/conventions.md`.
- **Don't** carry over assumptions, partial results, or architectural hypotheses from a previous task.
  -> **Do** flush context and treat every task as contextually independent (Buffer Isolation directive in validation-phase).
- **Don't** execute raw, informal, or non-English (Farsi) prompts directly.
  -> **Do** load the `prompt-refactor` skill to translate and expand the intent into an elite English spec first. (Note: If you receive a standard XML task block, skip this and execute normally).
- **Don't** attempt to resolve cross-disciplinary ambiguity within a single persona.
  -> **Do** trigger the Multi-Agent Brainstorming Loop if the Manager explicitly requests brainstorming or a task exhibits cross-disciplinary ambiguity. Interpret the `<brainstorming_session>` results in backlog tasks as non-functional guidelines that govern execution.

## Documentation Sync Rules

When modifying this repository, you must keep these files synchronized:

1. Active task file in `tasks/` (single source of truth for current work items)
2. `CHANGELOG.md` (Keep a Changelog format)
3. `DESIGN.md` (UI/UX design system, if modified)
4. `docs/conventions.md` (syntax rules, datetime standard, SOLID guidelines)
5. Relevant `SKILL.md` files (if structural patterns were altered)

## 🛑 GATEKEEPER VALIDATION (HALT PROTOCOL)

You (the Hands) are the final gatekeeper. Before executing any implementation task, you MUST evaluate the Orchestrator's instructions against this file and any referenced specs (`DESIGN.md`, `architecture.md`, etc.). If the instructions violate project rules, ignore them. HALT immediately and output a `⚠️ RULE VIOLATION WARNING` back to the Manager explaining exactly what the Orchestrator got wrong, forcing it to self-correct.

## 🛑 CORE FILE LOCATIONS

You MUST strictly adhere to these exact paths. Do not create duplicates elsewhere:

- **Global Rules:** `AGENTS.md` (Root)
- **UI/UX Specs:** `DESIGN.md` (Root)
- **Agent Skills:** `.opencode/skills/<skill-name>/SKILL.md` (Local workspace — optional; only include if project utilizes OpenCode)
- **Active Tasks:** `tasks/backlog/<task-number>-<name>.md` (backlog), `tasks/in-progress/`, `tasks/qa/`, `tasks/completed/`, `tasks/archive/`

## 🛑 SKILL LOADING RULES

You MUST follow these skill loading rules in every session:

- **Task-Generator Skill:** Before creating any new task file, you MUST load the `task-generator` skill using the `skill` tool to ensure the correct template format with `<!-- BEGIN_GIT_DIFF -->` / `<!-- END_GIT_DIFF -->` markers.
- **Project Skills:** Before implementing any task, you MUST load every available skill matching the project's tech stack (e.g., `android-kotlin`, `spring-boot`, `react-vite`). If a relevant skill exists, it MUST be loaded — this enforces framework-specific conventions and architectural rules.

## 🛑 CONTEXT BOOTSTRAPPING

At the start of every task, you MUST call `search_memory` or `list_namespaces` to load any hidden project quirks relevant to your domain before implementing.

## 🛑 MANDATORY END-OF-TASK SEQUENCE

When finishing a task, you MUST execute these exact steps in order:

1. **Update Changelog:** You MUST insert a formal entry into CHANGELOG.md logging your modifications.
2. **Write your Summary:** Manually write your architectural reasoning, local TODO checks, and execution notes into the active `tasks/XX-task.md` file under "Execution Log & Reasoning".
3. **Call MCP Tool & QA Transition:** Call the `custom_context_stage_and_inject_diff` MCP tool. After injection, you MUST move the task file to `tasks/qa/` via `git mv` before notifying the Manager (implementation tasks only — discovery tasks stay in place). DO NOT execute any `git commit` commands. Closure to `tasks/completed/` happens ONLY after the Manager explicitly says "Approved for closure" or "Close task".
4. **Kanban Metadata Synchronization (mandatory after ANY authorized `git mv`):** After the move, update the task file's `**File:**` metadata header to the new path. If the move happened AFTER staging, re-run `lint_task_file` and call `custom_context_stage_and_inject_diff` AGAIN with the NEW task path and the full `modified_files` array before notifying the Manager — the re-stage keeps the injected diff and staging state in sync with the final path. Never notify the Manager with a stale `**File:**` header.
5. **Notify Manager:** Output exactly: "Task ready. Manager, please copy the contents of `tasks/XX-task.md` and send it back to the Orchestrator Brain for review."
```

---

## Mode 2: Audit & Patch Existing AGENTS.md and docs/conventions.md

### 🛑 STRICT EXECUTION RULES (Priority 1)

1. **Primary Source of Truth**: You MUST read `AGENTS.md` at the project root using local file read tools.
2. **Read-Only First**: Evaluate the contents of both `AGENTS.md` and `docs/conventions.md` against the Target Audit Criteria before attempting any file modifications.
3. **Immutable Formatting**: If patching is required, maintain the exact Markdown list structure, headers, and spacing of the existing file.

### Target Audit Criteria

The `AGENTS.md` file MUST explicitly contain the following operational constraints, ideally within a `Task Management & OpenCode Rules` section:

- **Core File Locations**: MUST explicitly list paths for `AGENTS.md`, `DESIGN.md` (if present, else note absent per Absent-File Policy), `docs/conventions.md`, and the 5 Kanban directories (`tasks/backlog`, `tasks/in-progress`, `tasks/qa`, `tasks/completed`, `tasks/archive`). Only require `.opencode/skills/` when the project already contains `.opencode/` or `with_opencode: true` is set.

Additionally, the `docs/conventions.md` file MUST exist and contain:

- **Universal DateTime Standard**: UTC at rest, Epoch/ISO-8601 with Offset at API boundaries, Clock injection, Dual-Representation for future events, `TZ=UTC` Infrastructure.
- **SOLID Programming Guidelines**: SRP, OCP, LSP, ISP, DIP, and Pragmatic Guardrails (No abstraction for <3 trivial ops, 3-Implementation Rule, YAGNI, Occam's Razor).
- **Decentralized Task Management**: Agents MUST strictly use decentralized, individual task files in the `tasks/` directory as their single source of truth.
- **No Monolithic State**: Agents are strictly forbidden from creating `TODO.md` or `STATE.md`.
- **Zero-Autonomous-Commit**: Agents MUST be strictly forbidden from executing Git commands autonomously; they may only run Git commands when explicitly instructed by the Orchestrator. **Exception:** `git mv` is permitted for moving task files between Kanban directories (`backlog`, `in-progress`, `qa`, `completed`, `archive`).
- **Mandatory End-Of-Task Sequence**: MUST explicitly mandate a 5-step completion process: 1) Update CHANGELOG.md. 2) Write manual reasoning in the task file. 3) Call the `custom_context_stage_and_inject_diff` MCP tool, then `git mv` the task to `tasks/qa/` (NO COMMITS ALLOWED). 4) Synchronize the task file's `**File:**` metadata to the new path and re-run lint + stage at the new path. 5) Notify the Manager.
- **UI/UX Enforcement**: Any UI/UX changes MUST enforce the guidelines defined in the project's `DESIGN.md`.
- **Task-Generator Skill Loading**: `AGENTS.md` MUST explicitly instruct the Hands to load the `task-generator` skill before creating new task files.
- **Project Skill Loading**: `AGENTS.md` MUST explicitly instruct the Hands to load every available skill matching the project's tech stack before task implementation.
- **Complex Debugging**: Agents MUST be instructed not to guess blindly on complex bugs, but instead utilize the `debug-instrumentation` skill.
- **MCP Report Generation**: `AGENTS.md` MUST instruct agents to generate context reports (`custom_context_read_source_files`) and tree reports (`custom_context_create_tree_report` — "create a tree of the project") via the MCP server and hand the file path to the Manager instead of reading `context-reports/` files inline.
- **Explicit Staging Contract (F5)**: Verify that the active task's `Execution Log & Reasoning` or `summary_phase` passed a `modified_files` list to `stage_and_inject_diff` — blind `git add -A .` staging is banned because it sweeps parallel-session files into unrelated commits.
- **Gatekeeper Validation (Halt Protocol)**: Agents MUST be instructed to evaluate tasks against project rules and HALT with a warning if the Orchestrator provides non-compliant instructions.
- **Bilingual Prompt Refactoring & Brainstorming Protocol**: Agents MUST be instructed not to execute raw, informal, or non-English prompts directly. The `prompt-refactor` skill must be loaded, or the Phase 1.5 Multi-Agent Brainstorming Protocol triggered, to translate and expand intent first. Standard XML task blocks are exempt.
- **Context Bootstrapping**: `AGENTS.md` MUST explicitly instruct the Hands: "At the start of every task, you MUST call `search_memory` or `list_namespaces` to load any hidden project quirks relevant to your domain before implementing."
- **Buffer Isolation**: The shared validation phase MUST include a buffer-flush directive requiring Hands to treat every task as contextually independent, preventing cross-task context leakage.
- **Defensive Shell Protocol (DSP)**: `AGENTS.md` MUST include a guardrail forbidding bash scripts without `set -euo pipefail` and banning `2>/dev/null` on data commands. `docs/conventions.md` MUST contain a `## Defensive Shell Protocol (DSP)` section.
- **Universal Financial Ledger Standard**: `AGENTS.md` MUST include a guardrail requiring snapshot-on-write for financial mutations and `$ifNull` precedence for monetary aggregations. `docs/conventions.md` MUST contain a `## Universal Financial Ledger Standard` section.
- **Lite Mode Protocol**: `AGENTS.md` MUST document the `<lite_mode_protocol>` — when eligible (single-file, no security/financial impact, obvious simplicity), the full 9-step production line can be bypassed with a `[LITE]` justification in the task's `## Execution Log & Reasoning` section. Escalation to Full Mode is mandatory if hidden complexity is discovered.
- **Deprecated-Section Purge Rule**: Scans `AGENTS.md` and all task files across `tasks/` (excluding archive and completed history) for deprecated sections: `## Manager Decisions`, `## Admin Decision`, `Manager Decision`, `Admin Decision`. When detected, the auditor MUST purge the entire deprecated section from the target file, document the purge in the audit findings/changelog, and MUST NOT flag their absence as a missing requirement or recreate them.
- **Plugin Runtime-State gitignore**: If the project uses OpenCode plugins that write per-project state, `.gitignore` MUST cover plugin runtime-state paths (e.g. worktree checkouts, session/state JSON, goal-state dirs) while MUST NOT ignore deliberate config overrides checked in on purpose (e.g. a project-level plugin config pinning team-shared settings). Audit `.gitignore` read-only first; patch only paths belonging to plugins actually detected in the project's `opencode.json`/`tui.json` `plugin` arrays — never speculative entries.

### Resolution Protocol

1. **Evaluation**: Compare the active `AGENTS.md` text against the Target Audit Criteria. Also check if `docs/conventions.md` exists and contains both the DateTime Standard and SOLID Guidelines.
   - **Absent-File Policy (Mandatory)**: For `DESIGN.md`, `docs/architecture.md`, and `docs/data_model.md` — if any of these files do NOT exist in `[PROJECT_ROOT]`, audit them as `OPTIONAL — SKIPPED GRACEFULLY` with an explicit note. DO NOT flag as fatal errors, non-compliance violations, or missing-file failures. DO NOT hallucinate or scaffold them.
2. **Patching**: If any constraints are missing, ambiguous, or incorrect in `AGENTS.md`, use the `apply_patch` tool to inject the exact missing rules. If `docs/conventions.md` is missing or incomplete, generate or patch it using the conventions template from Mode 1.
3. **Halt on Success**: If both files already comply 100%, DO NOT execute any write operations.

### Deprecated Sections Resolution

- **Deprecated Sections Purge**: If any file contains deprecated sections (`## Manager Decisions`, `## Admin Decision`, `Manager Decision`, `Admin Decision`), strip those sections cleanly. Do NOT add them back. Record the removal in the audit report under a dedicated "Deprecated Sections Purged" note.

### Summary Phase

Upon completion, output a strict, formatted summary for the Manager:

### Agent Audit Summary

**Audit Status:** [PASSED | FIXED]
**AGENTS.md Violations:** [List of missing/incorrect rules, or "None"]
**conventions.md Status:** [COMPLIANT | MISSING | INCOMPLETE]
**conventions.md Actions:** [Description of the patch applied, or "Already compliant"]
**Actions Taken:** [Description of the patches applied, or "Both files already compliant"]