Phase 4 — Run Without You (the recurring loop)

A scheduled deployment fires a fresh session on a cron cadence — the agent runs without you. Each firing can carry its own outcome, nesting the bounded grade→iterate loop inside every recurring run.

See ../../references/loops-and-workflows.md.

Choose the trigger

Answer	Shape	Tool
"every morning / weekly / nightly"	recurring cron deployment	`deployment_builder.py` + `cron_validator.py`
"when X happens"	event-driven curl (documented, not scheduled)	`deployment_builder.py` (message only)
"only when I ask"	on-demand (no deployment)	none — just re-send a `user.message`

Workflow (recurring)

Validate the schedule.
```
python3 scripts/cron_validator.py --cron "0 9 * * *" --timezone Europe/Berlin
```
Invalid cron/timezone → exit 1. Read the DST note: wall-clock semantics mean spring-forward times are skipped and fall-back times fire twice — avoid 02:00–03:00 in DST zones if exactly-once matters.
Build the deployment payload.
```
python3 scripts/deployment_builder.py \
  --sheet ./my-agent/build-sheet.json --agent-id agent_123 --env-id env_456 \
  --nest-outcome --out ./my-agent/payloads/deployment.json
```
--nest-outcome includes the rubric so each firing self-grades. The tool prints the BYOK curl to create it and to test it once with the manual run endpoint before trusting the schedule.
Test before you trust. Fire one manual run, read the verdict, only then leave the cron in place. Pin the agent version in the deployment once it passes.

Finalize the roadmap.

python3 scripts/next_directions_writer.py \
  --sheet ./my-agent/build-sheet.json --loop-shape cron-loop --last-verdict satisfied --out-dir ./my-agent

Advance + hand to wrap-up. goal_state.py set --phase wrap-up, then invoke the wrap-up skill.

Hard rules

Test with a manual run first. Never commit a schedule you haven't fired once.
Safety rails on by default. always_ask MCP, limited networking where you can, read_only untrusted memory, max_iterations per firing, workspace spend limit. There is no spend cap inside CMA.
DST is wall-clock. Surface the note; pick safe times.
≤1,000 deployments/org.

Forcing-question library (recommend + cite)

"Cadence, event, or on-request?" Recommend: on-request v0 → cadence once graded. Cite: loops-and-workflows.md.
"Should each firing self-grade?" Recommend: yes — nest the outcome. Cite: loops-and-workflows.md (nesting rule).
"Which timezone, and is the time DST-safe?" Recommend: avoid 02:00–03:00 in DST zones. Cite: cma-primitives.md (wall-clock DST).
"Did you fire one manual run first?" Recommend: always. Cite: this SKILL.

Tools

scripts/deployment_builder.py — POST /v1/deployments payload (+ test-run curl).
scripts/cron_validator.py — 5-field cron + IANA tz + DST note.
scripts/next_directions_writer.py — write/refresh NEXT-DIRECTIONS.md.

run-without-you