Polly Pad Auditable Runtime

This runtime is headless by design: no GUI requirement.

Core requirement met:

• Auditable by owner
• Usable by agent AI
• Safe test spots
• Decommission + cousin takeover on shutdown/failure

Model

Polly Pad behaves like a small personal tablet for each agent:

• books/docs: task lane for reading and synthesis
• web: browser lane via HYDRA browser worker
• lab: sandbox lane for experiments
• memory: stable notes + compact snapshots

Lifecycle

1. Active pad executes tasks.
2. Failure/shutdown captured as exit_log.
3. Pad is decommissioned (not silently reused).
4. Last compact state is absorbed.
5. Cousin pad is spawned with inherited compact.
6. Takeover task is enqueued with handoff bundle.

Components

• Store: hydra/polly_pad.py
  • pad records
  • event log
  • compact snapshots
  • recovery ledger
• Watchdog: scripts/polly_pad_watchdog.py
  • scans failed tasks from Switchboard
  • decommissions source pad
  • spawns cousin
  • enqueues takeover task

Run

./scripts/run_polly_pad_watchdog.ps1

Optional:

python scripts/polly_pad_watchdog.py --switchboard-db artifacts/hydra/headless_ide/switchboard.db --pad-db artifacts/hydra/polly_pad/pads.db --scan-limit 100

Audit surface

Pad lifecycle DB:

• artifacts/hydra/polly_pad/pads.db

Switchboard DB:

• artifacts/hydra/headless_ide/switchboard.db

Key audit entities:

• pads
• pad_events
• recoveries
• tasks (switchboard)

Notes

• This is intentionally deterministic and machine-readable.
• For vehicle/fleet real-time systems, use turnstile policy to pivot/isolate rather than stall.