Suggestions Legacy
The Suggestions panel lists a handful of legacy agent commands, ranked by how often they appear in the most recent command log, and runs one when you click it. It is a command shortcut list, not a recommendation engine.
Heads up
The ranking is computed over everyone’s logs, not yours. There is no user model, no similarity, no collaborative filtering, and nothing is learned from what you click. If you want suggestions about your own space, use the Composer: its quick-start pills and follow-up chips are generated from your selection and your conversation.
Where to find it
View > Agents > Suggestions. It is also part of the Rich layout preset.
What it shows
On open, the panel fetches predictions from the backend and renders:
- Reload, a button in the top right that refetches.
- The backend’s status line, normally Predictions based on command history.
- Predicted, up to three command strings, each a clickable monospace button.
- Recent, the single most recently logged command, also clickable.
- Total logs, the number of log rows the prediction was computed from.
If no commands could be ranked, you get No predictions available. If the request fails, you get the error and a Try Again link.
How the ranking works
The backend takes the 100 newest rows of the user log, across all users and all spaces, and counts each distinct agent_command. It then returns, in order:
- Up to three commands that most often followed the single most recent logged command.
- The most frequent commands overall, used to pad the list up to three.
That is the whole algorithm. Ranking is by raw count, and nothing in the request identifies you, your space, or your selection.
Running a suggestion
Click a command. The panel logs the click as a new user log entry, classifies the command by its prefix (plot, legend, panel, tool, views), runs it through the legacy command processor, then refetches the predictions.
Because the click is logged, running a suggestion makes that command slightly more likely to be suggested again, to everyone. That is the only feedback loop that exists.
Limits
- Not personalized. Global counts over the last 100 log rows, nothing more.
- Not live. The list refreshes on open, on Reload, and after you run a suggestion. It does not react to what you are doing on the map.
- Legacy commands only. The buttons contain raw command strings, and running one goes through the legacy command processor, not the Composer.
- Bounded by whatever is in the log. On a quiet deployment the same few commands appear every time.