fix(cron): fall back to a local schedule when Cloud cannot reach this computer

[cpacker]

Summary

• Bare letta cron add on a desktop-local or unregistered runtime now creates a local schedule with a warning, instead of erroring. After fix(cron): preserve active listener execution by default #3665, those runtimes had no working no-flags path: Cloud scheduling cannot deliver into a computer that is not in the environments registry, so the local scheduler is the only placement that keeps the work executing where the conversation lives.
• Recurring schedules that hit this fallback get a louder caution — "every Monday" pinned to a session that must stay open is usually not what the user meant — with the durable alternatives (--runner cloud, --computer <deviceId>) spelled out.
• Rewrites the scheduling-tasks skill around intention rather than the flag matrix: omit --runner/--computer and the CLI keeps the schedule on the computer that created it, so scheduled work never conflicts with the active conversation from a second computer. Adds fast-followup vs recurring-job guidance and UTC-conversion instructions for cloud cron expressions (agents were previously shown examples that would silently schedule 9am UTC for a user asking for 9am local).
• Folds in the verified CLI-reference corrections from docs(skills): clarify cron clock and execution targets #3675: --once, id-or-name for get/delete (+ remove alias), runs history, --every compiling to a UTC cron expression on cloud schedules, numeric-only cron fields, and no in-place editing. Cameron is co-authored for those.

Stacked on #3683 (system prompts defer cron mechanics to the skill), so this PR touches only code and the skill — the skill is the single place placement behavior is documented.

Resulting default (no flags, cloud agent)

Runtime	Placement
Registered external listener	Cloud schedule targeting that listener (#3665)
Managed cloud sandbox	untargeted Cloud schedule → fires in the sandbox
Desktop-local / unregistered computer	local schedule + warning (was: error)
Local-backend agent / self-hosted	local schedule (unchanged)

Explicit flags stay authoritative and unchanged: --computer picks a registered device, --runner cloud picks the sandbox, --runner local pins this computer.

Implementation

resolveInferredTargetDevice now returns device | cloud-sandbox | local-fallback (the error arm is gone); handleAdd routes local-fallback to the local runner and threads the reason into the JSON output's warning field, merged with the existing no-scheduler-running warning when both apply. User-facing messages use plain wording ("not connected to your Letta account"), and guidance text avoids runtime taxonomy — the CLI reports its placement in output, and the skill teaches reading that instead of predicting it.

Follow-up (separate track): registering desktop-local listeners in the Cloud environments registry would let those runtimes get durable Cloud schedules targeting the desktop, at which point this fallback narrows naturally. Until then, local is the honest placement.

Supersedes #3675.

Test plan

• bun test src/cli/subcommands/cron.test.ts src/cli/subcommands/cron-runner.test.ts src/tools/shell-env.test.ts — includes new HTTP-boundary tests: unregistered runtime creates a local task (no schedule POST) with fallback + recurring warnings; one-shot fallback omits the recurring caution
• bun run check (12/12)

👾 Generated with Letta Code