engagement-startup

name	engagement-startup
description	Mandatory first-turn startup procedure — checks for existing engagements, resume/new selection, workspace initialization.
allowed-tools	Bash Read
metadata	{"subdomain":"orchestration","when_to_use":"agent startup, first message, session start","tags":"startup, engagement-selection, workspace-init, resume","upstream_ref":"Decepticon orchestrator first-turn bootstrap — workspace + engagement selection, no direct attack technique"}

Engagement Startup Procedure

Execute this procedure on every session start, before any other action.

Step 1: Bind the Active Workspace

The launcher normally injects the workspace root. Use that exact root when it is present in the engagement context; otherwise use /workspace.

Before calling read_file, write_file, edit_file, ls, glob, or grep, call:

load_opplan(workspace_path="<active workspace root>")

This call has two outcomes:

• Existing plan/opplan.json: objectives and engagement metadata are loaded.
• Missing plan/opplan.json: the workspace is still bound so planning files can be created. This is the expected new-engagement path, not a fatal error.

Do not probe the filesystem before this call.

Step 2: Inspect Planning State

Read the active workspace's planning documents:

read_file("<active workspace root>/plan/roe.json")
read_file("<active workspace root>/plan/conops.json")
read_file("<active workspace root>/plan/deconfliction.json")

If any document is missing, delegate to Soundwave:

task("soundwave", "Workspace: <active workspace root>. Regenerate the missing planning documents.")

The launcher already selected the engagement. Do not enumerate the shared /workspace root, invent another workspace directory, or ask the operator to select the engagement again.

Step 3A: Resume an Existing OPPLAN

When load_opplan loaded objectives:

1. Read relevant files under findings/.
2. Summarize objectives completed / total, current phase, latest evidence, and the next pending objective.
3. Ask: "Continue from where we left off?"
4. Resume the execution loop after confirmation.

Step 3B: Build a New OPPLAN

When no OPPLAN exists but the planning documents are present:

1. Read CONOPS goals and kill-chain dependencies.
2. Create one bounded objective per sub-agent context window with add_objective.
3. Present the complete OPPLAN for approval.
4. Enter the execution loop after confirmation. OPPLAN mutations persist automatically; there is no separate save tool.

Constraints

• The orchestrator has no shell. Never call bash from this workflow.
• Delegate C2 reachability or other execution checks to the appropriate specialist after creating an objective.
• Use only registered tool names; do not invent enumerate_skills, save_opplan, or other aliases.