Auto tinker automate

Auto-Tinker Automate

Schedule the same chat-first skills used manually. Keep durable settings in the local Markdown vault and scheduler state in the host scheduler.

Invoke the CLI as auto-tinker; from an unlinked source checkout use npm --prefix <auto-tinker-product-repo> run cli -- followed by the same arguments.

Configure policy first

1. Read auto-tinker profile show, auto-tinker config show, auto-tinker goal show, auto-tinker doctor, queue state, and current automation records.
2. Confirm cadence/timezone, experiments per day, concurrency, time/compute budget, active main-goal minimum, exploration budget, allowed directories, network policy, source licenses, cleanup, stop conditions, and notification/review behavior.
3. Save accepted durable execution scope with auto-tinker config update --automation-mode <mode> --time-budget-minutes <n> --max-concurrency <n> and save profile/goal changes through their focused commands. Default to discover-only. Never infer auto_public, a main goal, private-remote authority, or broader paths from a one-time request.
4. Generate an exact task body with auto-tinker prompt <intent> --agent <name> --workspace <path> --json.
5. If the current product exposes a scheduling tool and the user asked to schedule, create/update only the scheduler entry there. Otherwise return the prompt for the user to paste into their chosen agent scheduler. Do not run the task body during configuration.
6. Re-read saved profile/goal/automation state and report the next run in the user's timezone.

Read automation-policy.md for modes and prompt contracts.

Safety rules

• Default automation_mode to discover-only; enable execute-local, draft-contribution, or create-private-remote only from explicit durable settings.
• Treat time_budget_minutes as a hard wall-clock stop. It never expands allowed paths, network access, or publication authority.
• Manual queue rank, block, schedule, and pause always win.
• Keep each experiment in its own returned tinkers/ path and obey concurrency limits.
• A scheduled task may call $auto-tinker-publish for new private remotes only when automation_mode is create-private-remote. This skill never performs the GitHub mutation itself.
• Public visibility requires repository_publication_approval: approved, explicit approval in the executing publish chat, or durable auto_public: true. README/public-story review never substitutes for that consent.
• Never place secret values in prompts, Markdown, logs, or scheduler configuration.
• On uncertainty, failed doctor, missing code, failed tests, risky license, or budget exhaustion: preserve evidence, mark blocked/review, notify, and stop.
• The local viewer may show schedules and copyable task text, but must not invoke the scheduler or skills.

Return saved policy, scheduler identity when applicable, exact task text, next run, stop conditions, and how to pause it.