Improvement Sweep
The five gaps Heimdall's sweep looks for in a project's record, how it proposes a card per gap on the review queue, why the same idea is never proposed twice, when it runs on its own, what accepting a card does and what it costs: nothing.
By the end of this page you will know what the improvement sweep checks, where its cards appear, how to run it now and when it runs on its own, what keeping or dismissing a card does, and why it costs nothing.
A planning record grows by accretion: a requirement written in a hurry has no measure, a story has criteria for the happy path and none for the refusal, a rule has no worked example, an entity has no owner, a risk sits on the log with no mitigation. None of these stops a sprint, and all of them cost a conversation later. Heimdall's sweep reads the record for exactly these gaps and leaves one card per record on the review queue, so the analyst fills them when there is a quiet moment rather than when a developer asks.
What the sweep checks
| Check | Flags | Skips |
|---|---|---|
| Measure | A non-functional requirement whose measure is empty | Functional requirements; deprecated requirements |
| Negative criterion | A story with no negative acceptance criterion, whether it has other criteria or none | Tasks, bugs and spikes; done or cancelled stories |
| Worked example | A business rule whose examples are empty | Deprecated rules |
| Owner | A data entity with no owner | Deprecated entities |
| Mitigation | A risk that is open, monitoring or mitigating with no mitigation | Assumptions, issues and dependencies; closed risks |
Each gap becomes an Improvement card titled with the record's reference, such as Add a measure to CRM-NFR-3, with a sentence on what to add and why, a one-line reason, and Open CRM-NFR-3, which opens the record's page with its flyout showing. The cards sit under Waiting for you on the Heimdall suggests rail of the project's overview, beside what the research left.
Running it
Check the record on the Heimdall suggests rail runs the sweep now and answers with a toast: how many gaps it found, or that nothing is missing. People who may start a brief import see it, on a project whose assists are on. The sweep also runs on its own:
- Weekly, on every live project whose assists are on, early on Monday.
- After a brief import is fully decided, since the records it just wrote are where the gaps usually are.
A sweep that finds nothing new still runs the retirement below, so the rail stays honest.
Why nothing is proposed twice
The card's title carries the record's reference, so the same gap on the same record has the same title every time. Before it proposes, the sweep reads every card the project already holds on the review queue and every idea people dismissed, and skips a gap whose title is among them: a card you kept stays kept, a card you dismissed is never raised again, a card still waiting is not doubled. A waiting card whose gap was filled since, because someone added the measure or the example, is taken down on the next sweep; a kept or dismissed card is left as a record of the decision.
Keeping and dismissing
The cards are suggestions like any other on the rail: Keep holds the note on the project for everyone to read, Dismiss drops it and remembers it. Neither writes to the record; the record is changed in its own editor, and the next sweep notices. Fill the gap first and the card leaves on its own; keep the card when the gap is meant to stay for now and you want it on the record; dismiss it when the check does not apply, such as a risk the client accepted as it stands.
What it costs
Nothing. The sweep reads the record and calls no model, so no usage is filed and the team's AI budget does not move. It is switched off with the project's assists, like every other suggestion.
The sweep is the deterministic half of Heimdall's standing suggestions. The judgement calls, whether a criterion can be verified or a rule contradicts another, stay with the quality review and the council, which read with a model and are metered.
What happens next
improvement.sweptgoes out with how many cards were proposed, how many were taken down and how many wait, so a dashboard can follow how complete a project's record is.- Keeping or dismissing a card emits
import.suggestion_acceptedorimport.suggestion_rejected, as on every suggestion. - Nobody is notified; the rail is where the cards wait.
From the API and MCP
Sweep a project's record
Run the sweep now and read what it proposed, retired and left pending.
sweep_improvements
An agent runs the sweep and reads the counts.
List a project's research suggestions
The cards themselves, beside the research notes.
improvement.* events
One event per sweep, with the counts.
Permissions
| Action | Permission |
|---|---|
| See the cards | import:view-any, as staff |
| Run the sweep now | import:create, with the project's assists on |
| Keep or dismiss a card | import:update |
Frequently asked
I fixed the record, but the card is still there. The card leaves on the next sweep: press Check the record, or wait for the weekly one. Keeping or dismissing it also closes it, but a dismissal is remembered and the same gap would not be raised again if the record regressed.
Can I turn one check off? Not per check. Dismiss the cards a check raises where it does not apply; the dismissals are remembered per record, so the check stays useful for the rest of the project.
Why did the sweep find nothing on a project with obvious gaps? It reads only the five shapes above. A vague criterion or a rule that contradicts another is not a structural gap; the quality review reads for those.
Related
- The project overview — the rail the cards wait on
- Research — the other cards on the rail, from the web
- Quality review — the metered reading of one record
- Brief import — the review queue, and the decisions the cards share
- Usage and budgets — what is metered and what is not
How is this guide?
Last updated on
Impact analysis
What Heimdall reads before it weighs a change request, what it is told to look for, how its answer becomes the draft and the numbers on the request, what a reading costs, how the council's change impact moment differs, and where it stops.
Usage & Budgets
How every model call is priced and filed, what a turn costs on each tier, the team's monthly AI budget with its warning line and its cap, who is told when a line is crossed, what is never metered, and where to read the statement.