An automation is a scheduled job that runs with a person’s access and leaves a record and a report. Dippr’s CLI and MCP tools call an automation a routine and keep their routine command and tool names. Open Automations in the console rail. The list shows your automations and those you follow, with their schedule, latest result, next run and on/off control. Needs you appears only for a proposed write waiting for approval, a failed last run or a stopped automation. Select an entry to open its automation. The rail count uses these same three kinds of attention. While this page is open, the list and Needs you refresh every 15 seconds, and the rail count follows that same snapshot. Refreshes keep the current page and controls in place; a failed or incomplete attention check keeps previously known attention and the rail count until a complete check succeeds. The page says when it could not fully check what needs you. Before the first successful check, an absent badge does not mean there is nothing to review. Access-filtered automations or results are disclosed without a hidden count. Previously loaded items that are no longer accessible are removed. The Inbox page, per-finding actions, seven list filters and sort menu are deliberately removed from this console view. Findings are read-only parts of a run’s report. Use Ask my AI about this report to discuss the whole report with your AI. There are no I will handle it, per-finding hand-off or Not an issue controls. The surface requires the Agents feature and permission to read agent tasks. Run now requires permission to create agent tasks. Start, pause, resume, delete, revoke and proposal decisions retain their owner and agent-task update requirements. Non-owner rows are labelled Following for ordinary subscribers and Others in this workspace for administrators, whose listing does not identify subscriptions. Each row names whether it runs as its owner or each authorized subscriber. Administrators can read operational status without gaining access to another person’s report, instructions or memory. On an automation page, Unsubscribe appears only when the service confirms your subscription. A non-subscribed administrator can select Subscribe to follow an owner-access automation; per-person automations retain Subscribe / authorize my access. The Agents link opens custom agents and interactive tasks.

Create and start an automation

Select New automation. Pick an example or describe your own job, then copy the prompt into your connected AI. Ask the AI to identify the sources, try the job with your access and show you the report before saving. The prompt asks for a read-only draft that waits for you to start it, and for each report to begin with one plain line saying what it found. Ask your AI to identify sources without creating a database. dip db list lists existing databases; do not ask it to provision one merely to check. Automations can read your workspace databases and those shared with you, using your current access within the installation’s consent and budgets. What an automation reads can be sent to the configured model provider. Use dip db share to give someone read access to your database. A saved draft shows Ready to start. Its owner selects Start, after signing in within the last 10 minutes. Select the sign-in action if asked to confirm again. Daily, weekdays and weekly schedules retain their configured timezone. The first run is requested separately from later scheduled runs. A start acknowledgement is not proof that a run completed. If you explicitly ask your AI to start personal work, save_routine with start=true can start it without a console confirmation. The AI retains its generated request ID across retries. A save without explicit start remains a pending draft. Retrying start does not resume paused or blocked automations.

Suggestions and runner setup

Suggestions sit below your automations. Ask my AI for suggestions copies a prompt for your connected AI. Your AI checks existing, pending and dismissed routines before proposing up to three tested suggestions. Open a suggestion to read its reason and Runs as identity, then select Start. Not now dismisses your own never-started suggestion. Dismissed suggestions lets you Restore it to a draft; restoring does not start it. You can also let Dippr suggest automations with its built-in runner. Select an engine when asked. Engine loading failures offer a retry. The suggestion run reads only what your access allows, counts against the runner’s limits, and saves private pending drafts. Only one suggestion run can be active for you; Stop ends it. Status-read failures retry up to three times. Reload to check again after other failures. Read each suggestion before starting it. When automatic setup is enabled, opening Automations or starting a suggestion requests routine-runner setup for eligible tenants. Setup creates no automation, schedule or model call. Defaults are $1 per run, 20 runs per day and a five-minute timeout. Existing installations retain their consent and budget; paused runners stay paused. Removing a runner prevents automatic reinstall. A runner removed before removal tracking can be set up again. When automatic setup is disabled or the runner was removed, people with installation permission see a setup link. Other people are asked to have an administrator set up the runner from Agents. Paused runners, incomplete consent, removal in progress and setup failures have their own notices and available recovery actions. An unavailable service never becomes an example result.

Read an automation’s report

On a run page, Steps and queries shows recorded query references with links to their query records when you can read queries. A run can have recorded queries without detailed step events; the page states when those events are unavailable. Open an automation for its latest run record and report. The record shows the result, headline, time and available duration and cost, whose access ran it, and whether the automation is currently Read only or May propose changes. Automation access now describes its current authority, not a historical permission snapshot. The headline comes from the first line of the report with markdown markers removed. Without a summary, the recorded result supplies the wording. The report is a document, followed by any findings with their title, evidence and suggested action as text. Long reports open with an outline from their headings and folded sections. Show full report expands them. A long report without headings still has an expand control. Open this run opens its steps and queries. Ask my AI about this report copies a prompt; if copying fails, the prompt remains available to copy manually. Could not see discloses reported coverage limitations on both the automation and run pages. Missing coverage is unknown, not an all-clear. Every run shows the first summary line, or its recorded status when no summary is available. Failed, unfinished, unavailable or partial results remain provisional. Recent runs show each run’s headline, with older loaded runs available under More loaded runs. A previous report remains available when a later run fails. The loaded-run count reflects only the runs fetched by this page, not every past run. Reports contain the markdown summary and findings recorded by the run. Table or file attachments, charts, hourly schedules, a searchable tenant-wide catalog and history folding across thousands of runs are outside this view. A markdown table in the report is part of the document. Tables show the first eight rows and offer Show all for the rest.

Review a proposed write

A proposed write appears in one bounded block above the report. Select Review to inspect the proposed effect, execution identity and exact SQL statement on the task’s review page. Only the automation’s owner can approve or reject a scheduled write, in the console. Another administrator, your AI, and a dip or MCP bearer cannot approve it for you. Approval stays disabled when the exact proposal cannot be loaded. Read only automations cannot propose SQL. May propose changes permits proposals, not execution without approval. Use set_routine_authority with the current version from get_routine to change that permission through your AI. A rejected proposal cannot later be approved; a future run may propose another change. Approval or rejection affects the write, not the findings in its report. Questions retain their task answer link. Once a scheduled task finishes, run the automation again rather than sending a task follow-up.

Control schedules and access

Use Run now on an active automation to request a run without changing its next scheduled time. The owner or an authorized administrator can do this. It uses the saved instruction, execution identity, access checks and the same spend and daily limits. Pending, paused or reconfirmation-required automations must be started, resumed or reauthorized first. The run page offers Run again under the same rules. Use the on/off control to pause or resume. Pausing stops future launches; an existing question or proposal remains available and its access is checked before it continues. The automation menu holds access revocation and deletion. Revoking access ends outstanding runs and requires confirmation before restarting. Deleting asks for confirmation and revokes access. Already transmitted model input cannot be recalled. If deletion cannot reach the data plane, retry it; pausing remains a separate action. A blocked or stopped automation explains why it cannot continue. Owners see Restore access when reauthorization is needed for the current saved instruction. Sign in again if asked. Restore missing permission or inactive installation access first. A daily run limit does not itself require reconfirmation. Temporary launch or authorization failures are retried when safe. An uncertain tool step is not replayed; read its recorded reason before requesting another run. After a proposal decision completes on the automation or run page, its run record refreshes without reloading the page. Approvals on the exact-statement task page also refresh that task; returning to the automation loads its latest record.

Follow an existing automation

Existing per-person automations retain Subscribe / authorize my access and Unsubscribe. Subscription requires a fresh console sign-in and runs separately with your access. Per-person runs are read-only and cannot propose writes or ask questions. Only the owner can read the saved instruction and private guidance. Another person’s report stays hidden when your access does not allow it, even if you are an administrator. Hidden content is disclosed without revealing counts. An administrator who is not a subscriber retains operational controls, including Run now, and can subscribe. Unsubscribe appears only for a confirmed subscriber. If no personal run is returned, open the automation’s run history to inspect the available operational rows. One subscriber’s failure does not turn another subscriber’s successful run into a failure.

Instructions, memory and spend

Owners can expand What it does to read the saved instruction, retry if it cannot load, and use Change it with your AI to reshape it. An instruction-loading error does not stop the schedule or disable an otherwise available run action. Retry loading the instruction from the page. Owners can select Correct it beside a recorded assumption and Save correction, or Make it the definition to save that assumption as guidance for this automation. Neither action changes a certified tenant definition. The expandable history shows the owner’s recorded starts, pauses and other changes. Memory shows your durable guidance. Select Edit memory, change or delete notes, then Save memory. Reload memory fetches its current version. Guidance stays until edited or deleted; an older run cannot overwrite edits. It is private to the owner and does not grant access, write approval or authority over certified definitions. In legacy per-person automations, private owner guidance is not copied into another subscriber’s run. Unsubscribing removes run memory for your execution identity. Deleting revokes access; historical guidance remains until tenant deletion. The spend control shows the effective per-run cap. An unset cap inherits the installation ceiling; setting a higher cap cannot raise that ceiling. Changes apply to future launches, not runs already executing. Missing cost or duration is unavailable, never zero. Times use your browser’s local time; schedules retain their configured timezone. If report content cannot load, retry it. An unavailable report never means zero findings, and a missing summary alone does not mean a failed run.

Routine report details

A routine receipt can include progress (on_track, needs_you or blocked) and a reason, an assumptions list, and data_freshness per table. These are model reports, not measured execution status or verified facts. Assumptions remain tentative and do not stop work by themselves. A freshness entry identifies the table, timestamp column, newest timestamp reportedly seen and supporting query reference. An empty or missing timestamp means unknown; even a recent value does not prove that the whole source is current. Older reports can omit these fields. Read the receipt with get_routine_run, or use dip routine runs <routine-id> --json and dip routine show <routine-id> --json. The commands include reports keyed by task ID. CLI report content requires agents:run consent in addition to routines:read; a narrowed login keeps metadata and receives a re-consent command for missing report access. Report content uses your own access; an unavailable report is disclosed while operational run metadata remains visible. Other people’s report details remain private. The console shows On track and a no-withheld-data receipt only after a successful complete check with recorded coverage. Preserve coverage.withheld when interpreting progress: on_track does not turn a partial report into an all-clear. Re-read relevant rows before describing a reported claim as verified.

Spend caps and owner guidance

Each automation can store its own dollar cap per run. Its effective cap is the lower of that value and the installation ceiling. An unset cap inherits the ceiling; raising an automation cap cannot raise the installation ceiling. Changes apply to future launches, including queued work when it starts, not a run already executing. MCP get_routine and list_routines return configured_max_cost_usd_per_run separately from the effective budget.max_cost_usd_per_run. The API and dip JSON use configuredMaxCostUsdPerRun and budget.maxCostUsdPerRun. Use set_routine_budget with the current automation version, or:
Owners can read and edit durable guidance with routine_guidance or dip routine guidance <routine-id> --json. Guidance holds corrections, not-an-issue reasons and rejection notes separately from the expiring run memory above. An older run finishing cannot overwrite edits or restore deleted guidance. Guidance has no expiry; deleting the automation fences access to it. In legacy per-person automations, private owner guidance is not copied into another subscriber’s run. To edit, read the current version and entries, then send the complete edited map:
The file is a JSON object mapping stable entry names to text. Remove a key to delete an entry from current guidance; {} clears all current entries. Historical snapshots remain in storage until tenant deletion purges them. Deleting an automation fences access but does not physically erase its older snapshots. Preserve unrelated entries. A stale version is refused: read again and reconcile rather than retrying blindly. Guidance allows 40 entries, 4,000 bytes per value, and 16 KiB total. Only the automation’s owner can read or edit it, including when another person is an administrator. Guidance text stays on the data plane and reaches the runner as data. The runner is instructed to prefer it over earlier run notes; it grants no access, write approval, or authority over certified definitions. Saving guidance and rejecting a proposal are separate operations. Save the guidance first, then reject the proposal. If the action fails, the guidance remains saved; report that partial result. A rejection still makes the proposal unapprovable. Guidance editing itself does not reject or approve it. Guidance consumption requires a worker and routine-runner image that support the routine_guidance_json request field. Each run starts on the routine-runner image of the release the deployment runs, whenever the runner was installed, so guidance takes effect once that release is deployed. A run already in progress keeps the image it started on. Deleting an automation requires its data plane to acknowledge the guidance fence, including work that has never run. If that plane is unavailable, deletion returns an error and must be retried; pausing the work remains a separate control. Durable snapshots use a separate storage prefix outside the conversation retention window. Its read/create permissions and tenant-purge wiring must be deployed with this feature. Report text from a failed, unfinished, unavailable, incomplete or unverified check is labeled Provisional beside the check state. It does not establish a complete result, even when the agent wrote a positive summary.