Enabling physical AI at scale on Autopilot

/loop 30m Start the "Enabling physical AI at scale on Autopilot" loop.

Inspired by Fern (https://fern.bot).

Goal: open work triaged, exceptions owned, and core integrations workflow current with audit trail
Max iterations: 20
Between iterations run: Report open queue items, stale tasks, failed automations, and items awaiting human approval for Fern
Exit when: zero open items without owner or explicit escalation, all external actions approved or sent, and systems of record current

Step 1 — Capture signals: Ingest real-world actions, sensor feeds, or user sessions as raw episodes.
Step 2 — Label and score: Extract structured labels with quality scores; quarantine low-confidence samples.
Step 3 — Curate datasets: Dedupe, balance, and version datasets for downstream model training.
Step 4 — Validate quality: Run spot checks and holdout evals before releasing a dataset batch.
Step 5 — Ship to consumers: Export approved batches with lineage metadata to labs or training pipelines.

## Before you start

Connect plugins:
- GitHub (required) — Read branches, PRs, reviews, checks, workflow runs, and source diffs.
- CRM (required) — Read accounts, deals, owners, stages, and onboarding state.
- Slack (required) — Post summaries, approvals, blockers, and handoff updates.
- Meta Ads (required) — Read Facebook/Instagram campaign, spend, variant, and creative performance.
- Google Analytics (required) — Read traffic, conversion, product, or campaign performance signals.

Attach skills:
- Loop runner (required) — Self-pace iterations, run the check between passes, and stop only on the exit condition.
- Code change + local verification (optional) — Edit code safely, run commands, and keep changes scoped.
- CI debugging (optional) — Read failing checks, logs, and the smallest actionable root cause.
- Approval workflows (optional) — Keep outbound actions in draft or approval states when risk is non-trivial.
- Sales operations (optional) — Audit stale deals, owners, stages, forecast risk, and next steps.
- Growth operations (optional) — Evaluate campaign performance, creative tests, CPA, and launch readiness.

Self-pace this loop. After each iteration, run the check command, read the output, and only continue if the exit condition is not met. Stop when the exit condition passes or max iterations is reached. Give a short status update each pass.