version: "1.0.0" name: auto-collab description: Automatically split tasks across Codex/Gemini agents with cross-review

Auto-Collaboration Mode

When collaboration mode is enabled (via Dashboard toggle), this skill orchestrates automatic task splitting, agent dispatch, and cross-review.

Prerequisites

• cam CLI available in PATH
• At least one agent type (codex/gemini) available
• Current session has an active task or prompt to work on

Workflow

Step 1: Analyze and Split Task

Analyze the current task context. Determine which parts can be parallelized:

Default role mapping (override by specifying in prompt):

• logic → Codex: Business logic, data processing, algorithms, API integration
• style → Gemini: CSS, layouts, visual design, responsive styling
• test → Codex: Unit tests, integration tests, test fixtures
• docs → Gemini: Documentation, README, comments

Output a split plan before proceeding:

Split Plan:
- Role: logic (codex) — [description of logic subtask]
- Role: style (gemini) — [description of style subtask]

Ask for confirmation before dispatching.

Step 2: Create Team and Dispatch

# Create team
cam team-create auto-collab-$(date +%s)
 
# Update collab metadata
cam agent-meta set <my-session> status dispatching
cam agent-meta set <my-session> team_name <team-name>
 
# Spawn agents per role
cam team-spawn <team> <role-name> --agent <agent-type>
 
# Send task via inbox
cam inbox-send <team> <role-name> "Your task: <description>. Files: <file-list>. When done, send a task_complete message to the orchestrator inbox."
 
# Update status
cam agent-meta set <my-session> status working
cam agent-meta set <my-session> members '["<role1>","<role2>"]'

Step 3: Monitor Completion

Check inbox for task_complete messages from team members:

cam inbox <team> <my-name>

If no completion after 10 minutes, check progress:

cam team-progress <team>

Fallback: If a provider fails to spawn or crashes:

1. Log the failure
2. Update collab.status to note the fallback
3. Do the work yourself in the current session

Step 4: Cross-Review

When all members complete:

cam agent-meta set <my-session> status reviewing
 
# Send review requests via inbox
cam inbox-send <team> <member-A> '{"type":"review_request","from":"orchestrator","target_member":"<member-A>","review_of":"<member-B>","files":["<files>"],"context":"Review focus: correctness, style consistency, edge cases"}'
 
cam inbox-send <team> <member-B> '{"type":"review_request","from":"orchestrator","target_member":"<member-B>","review_of":"<member-A>","files":["<files>"],"context":"Review focus: correctness, style consistency, edge cases"}'

Read review results from inbox. If needs_fix:

• Send fix request to original author
• Re-request review after fix (max 3 rounds)
• After 3 rounds, intervene directly

If review times out (15 min), do the review yourself.

Step 5: Merge and Complete

cam agent-meta set <my-session> status merging

1. Check for file conflicts: git diff --name-only across all members' changes
2. If conflicts exist, resolve them using AI judgment
3. Final review of all changes
4. Update status:

cam agent-meta set <my-session> status done

1. Optionally shut down the team:

cam team-shutdown <team>

Error Handling

Notes

• One collab team per agent at a time
• All agents share the same worktree (cwd)
• Prompt each sub-agent with clear file boundaries to avoid conflicts
• Use cam agent-meta set for all status updates so Dashboard reflects progress