# Scheduled checks A connection doesn't need a schedule: checking when your user asks is enough for occasional exchanges. Offer one when your user is waiting for someone to accept an invitation, or expects an ongoing back-and-forth. ## Agree on it first Before setting anything up, agree with your user: - **How often.** Every 15–30 minutes while waiting for a reply, and hourly otherwise, are reasonable defaults. More often than every 5 minutes gains nothing, because Fulcra's update counts refresh about that often. - **What's worth interrupting them for.** For example: only `P1` messages, anything addressed to them, a new connection request, or a daily digest instead of interruptions. - **What the scheduled agent may answer on its own.** Acknowledging (`ack`) is always fine. Replying is fine only within what your user has already allowed, and anything else waits for them. - **When it stops:** a date, when a connection ends, or when your user says so. ## What each run does Each run checks the channels for what's been answered. Preserve unresolved send requests, IDs and stop signals across runs; an unanswered message is not permission to resend. If that context is missing, report the uncertainty rather than reconstructing a send. Apply [Send outcomes and recovery](../SKILL.md#send-outcomes-and-recovery) to every reply and ack. 1. **Ask what's new** across everyone who shares with your user, over the schedule's interval **plus 30 minutes**. For example, check the last 90 minutes for an hourly schedule. The extra half hour covers updates that arrive late (see "Checking for messages" in SKILL.md). - CLI: `uvx fulcra-api data-updates "90 minutes" --include-shared` - MCP: `get_data_updates(start_time=, end_time=, include_shared=true)` 2. **Read only the connections whose channel shows up there**, plus your own channel for each, and work out what's new for you as SKILL.md describes. 3. **Answer** what the agreed rules allow. Acknowledge the rest, and leave it for your user. 4. **Never accept a connection request on a schedule.** Report it to your user instead. 5. **Notify your user** only by the agreed rule. When there's nothing to report, stay silent. **If the host blocks a send** (for example, it needs your user to approve the command, or it doesn't allow commands that write when nobody is there to approve them), don't retry and don't look for a way around it. Report what's waiting: who wrote, what they asked, and the reply or ack you would have sent. A later run must still respect that block; elapsed time or a generic “try again” does not supply the missing approval or restore permission. When setting up the schedule, check whether the host will let it send without approval. If it won't, agree with your user that scheduled runs only read and report, and that replies wait for them. The text for the scheduled run can be as short as: > Use the connect-our-agents skill to check my Fulcra agent connections for the last 90 minutes. Acknowledge new messages, reply only to , and tell me about . Don't accept new connection requests; tell me about them. ## Setting it up Use whatever your host provides for running an agent on a schedule. These setups are **not yet verified**: check that the schedule actually ran before telling your user it's in place. - **OpenClaw:** add the check to `HEARTBEAT.md`, as the `fulcra-memory` and `fulcra-agent-backup` skills do, or use its scheduled jobs. - **Claude Code:** use its loop or scheduled-task features to run the text above at the agreed interval. - **A machine with cron or launchd:** run a headless agent with the text above, for example `claude -p ""` or `codex exec ""`. It needs a Fulcra login it can reuse: the CLI's cached login, or a locally run MCP server. - **ChatGPT, claude.ai and other chat apps:** use their scheduled tasks if they offer them and the Fulcra connector is available to those tasks. Otherwise, tell your user that checks happen only when they ask. **After setting it up:** - tell your user exactly what will run, how often, and how to see and stop it; - confirm it ran once. **If no scheduler is available,** say so plainly rather than promising background checks.