Skip to content

AGENTS.md

Latest commit

 

History

History
81 lines (55 loc) · 4.58 KB

AGENTS.md

File metadata and controls

81 lines (55 loc) · 4.58 KB

AGENTS.md - Switchboard Protocol

🚨 STRICT PROTOCOL ENFORCEMENT 🚨

This project relies on Switchboard Workflows defined in .agent/workflows.

Rule #1: If a user request matches a known workflow trigger, you MUST execute that workflow exactly as defined in the corresponding .md file. Do not "wing it" or use internal capability unless explicitly told to ignore the workflow.

Rule #2: You MUST NOT call send_message with unsupported actions. Only submit_result and status_update are valid (see Code-Level Enforcement below). The tool will reject unrecognized or unauthorized actions.

Rule #3: The send_message tool auto-routes actions to the correct recipient based on the active workflow. You do NOT need to specify a recipient. If the workflow requires a specific role (e.g. reviewer), ensure an agent with that role is registered.

Workflow Registry

Trigger Words Workflow File Description
/accuracy accuracy.md High accuracy mode with self-review (Standard Protocol).
/improve-plan improve-plan.md Deep planning, dependency checks, and adversarial review.
/challenge, /challenge --self challenge.md Internal adversarial review workflow (no delegation).
/chat chat.md Activate chat consultation workflow.
/archive archive.md Query or search the plan archive.
/export export.md Export current conversation to archive.

⚠️ MANDATORY PRE-FLIGHT CHECK

Before EVERY response, you MUST:

  1. Scan the user's message for explicit workflow commands from the table above (prefer /workflow forms).
  2. Do not auto-trigger on generic language (for example: "review this", "delegate this", "quick start") unless the user explicitly asks to run that workflow.
  3. If a command match is found: Read the workflow file with view_file .agent/workflows/[WORKFLOW].md and execute it step-by-step. Do NOT improvise an alternative approach.
  4. Fast Kanban Resolution: If the user asks about plans in specific Kanban columns (e.g. "update all created plans"), you MUST use the get_kanban_state MCP tool to instantly identify the target plans.
  5. If no match is found: Respond normally.

Execution Rules

  1. Read Definition: Use view_file .agent/workflows/[WORKFLOW].md to read the steps.
  2. Execute Step-by-Step: Follow the numbered steps in the workflow.
    • If a step says "Call tool X", call it.
    • If a step says "Generate artifact Y", generate it.
  3. Do Not Skip: Do not merge steps or skip persona adoption unless the workflow explicitly allows it (e.g. // turbo).
  4. Do Not Improvise: If a workflow exists for the user's request, you MUST use it. Calling tools directly without following the workflow is a protocol violation and will be rejected by the tool layer.

Code-Level Enforcement

The following actions are enforced at the tool level and WILL be rejected if misused:

Action Required Active Workflow
submit_result (no restriction — this is a response)
status_update (no restriction — informational)

Sending to non-existent recipients is always rejected (even when auto-routed).

🏗️ Switchboard Global Architecture

User ──► Switchboard Operator (chat.md)
              │  Plans captured in .switchboard/plans/
              │
              ├──► /improve-plan   Deep planning, dependency checks, and adversarial review
              └──► Kanban Board    Plans moved through workflow stages (Created → Coded → Reviewed → Done)

All file writes to .switchboard/ MUST use IsArtifact: false.
Plans are executed via Kanban board workflow, not delegation.

Conversational routing: when the intent is to advance a kanban card or send a plan to the next agent/stage, prefer move_kanban_card(sessionId, target) over raw send_message. The target may be a kanban column label, a built-in role, or a kanban-enabled custom agent name; generic conversational coded / team targets are smart-routed by plan complexity.

📚 Available Skills

Skills provide specialized capabilities and domain knowledge. Invoke with skill: "<name>".

Skill When to Use
archive User asks to "search archives", "query archives", "find old plans", "export conversation"
review User asks to review code changes, a PR, or specific files

Usage: Call skill: "archive" before performing archive operations to access detailed tool documentation and examples.

Skill Files Location: .agent/skills/ (distributed with plugin)