Skip to content

Add Pulse: the fleet summarized, every number against its limit - #2274

Draft
G-r-ay wants to merge 1 commit into
feat/fleet-wallfrom
feat/pulse
Draft

G-r-ay wants to merge 1 commit into
feat/fleet-wallfrom
feat/pulse

Conversation

@G-r-ay

@G-r-ay G-r-ay commented Oct 14, 2026

Copy link
Copy Markdown
Collaborator

Stacked on #2273 (Fleet Wall); review that first. This PR's diff is Pulse only.

What

A Pulse console screen (nav: Overview → Pulse → Fleet Wall). It shows the fleet summarized, every number against its limit, in tabs that each go one level deeper:

  • Summary
    • Four gauges: handled without a person, closed today, median time to decide, and AI spend against the daily cap.
    • "What needs attention", built only from what the reads can see.
    • The day at a glance.
    • The fleet settings, read-only.
    • The main column and Adjustments scroll independently under a fixed head and tabs.
  • Throughput: the day by hour, and work by workflow.
  • Paused and stuck: what waits, bucketed by age, with the oldest one a click away; plus the intake queue against its high-water mark.
  • Decisions, Sources and memory, System: placeholders that say what they will show.

Day chart. Cases running against the cap is one panel; alerts in beside cases closed is another. They share the hour axis, each has its own labelled scale, and one crosshair tooltip spans both (it also works with arrow keys). Charts draw at their rendered width, so text keeps its size.

The API is a contract, not an implementation

GET /api/pulse/summary, /throughput and /work are defined by PulseSummary, PulseThroughput and PulseWork in core/monitor/models.py:

  • Each route validates its answer.
  • The console's types are generated from the models.
  • The open decision on which spend number is the source of truth is flagged in the model.

The reads (core/monitor/pulse.py) are not implemented:

  • Each route answers 501 and its tab shows its error state.
  • In demo mode they serve the seeded sample.

The docstrings list the sources and TODOs: hourly bucketing in SQL, and no counting dead runs as running.

Tests

  • Contract tests are extended to the three Pulse reads.
  • monitor.test.tsx: Summary gauges, levels, chart and attention links; the oldest waiting case opens.
  • The full vitest run has one timeout, "navigates primary screens and the More menu". It fails the same way on untouched main.

Origin

Taken from @krmayankb's krmayankb/vigil PR #7 (fleet/console-monitor), credited as co-author.

A console screen with tabs that go one level deeper each. Summary has
four gauges (handled without a person, closed today, time to decide, AI
spend against its cap), what needs attention, the day at a glance and
the fleet settings, read-only; its main column and the settings scroll
on their own. Throughput has the day by hour and work by workflow.
Paused and stuck shows what waits, by age. Decisions, Sources and memory,
and System are placeholders that say what they will show.

The day chart draws cases running against the cap and alerts in beside
cases closed as two panels on one hour axis, each on its own scale, with
one crosshair tooltip across both.

GET /api/pulse/summary, /throughput and /work validate their answers
against PulseSummary, PulseThroughput and PulseWork in
core/monitor/models.py. The reads are not implemented: each route
answers 501 and its tab shows its error state, except in demo mode,
where it serves the seeded sample in core/monitor/fixtures.py.

Taken from krmayankb/vigil PR #7 (fleet/console-monitor).

Co-authored-by: krmayankb <krmayankb@gmail.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Signed-off-by: gray <egboluche.g@gmail.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant