babysitter:call

Orchestrate .a5c/runs/<runId>/ through iterative execution. Use the babysitter SDK CLI to drive the orchestration loop inside Codex CLI.

Dependencies

Ensure babysitter CLI is available:

npx -y @a5c-ai/[email protected] version --json

Workflow

1. Interview Phase (Interactive)

Gather user intent, requirements, goals, and scope. Research the repo structure, search the process library for relevant specializations and methodologies.

2. Process Creation

Create the process .js file with task definitions. Install @a5c-ai/babysitter-sdk in .a5c/package.json if not present.

3. Create Run

babysitter run:create \
  --process-id <id> \
  --entry <path>#<export> \
  --inputs <inputs-file> \
  --prompt "$PROMPT" \
  --harness codex \
  --session-id "$CODEX_SESSION_ID" \
  --plugin-root "$CODEX_PLUGIN_ROOT" \
  --json

4. Iteration Loop

babysitter run:iterate .a5c/runs/<runId> --json --iteration <n> --plugin-root "$CODEX_PLUGIN_ROOT"

For each pending task:

• agent — Build prompt from role/task/instructions, execute via Codex tools
• node — Execute Node.js script via shell
• shell — Run shell command directly
• breakpoint — Pause for user approval (interactive) or auto-resolve (yolo)
• sleep — Wait until timestamp
• skill — Load and follow SKILL.md instructions
• hook — Fire named lifecycle hook
• parallel — Execute sub-effects concurrently
• orchestrator_task — Delegate sub-orchestration

5. Post Results

Write result value to a separate file, then post via CLI:

babysitter task:post .a5c/runs/<runId> <effectId> \
  --status ok \
  --value tasks/<effectId>/output.json \
  --json

6. Check Completion

When run:iterate returns status: "completed" with a completionProof, the run is done. Return the proof wrapped in <promise>PROOF</promise> to signal completion.

Profile Integration

Read user profile for personalization:

babysitter profile:read --user --json
babysitter profile:read --project --json

Quick Reference