[Juliusolsson05]

Fixes #984
Refs #944

Problem

The Performance Monitor (#951) could say Agent Code was heavy, but not which agent was responsible. Its charts were static pictures: no hover, no y-axis units, and each line scaled to its own peak, so a calm series and a stall drew the same shape. The process table showed raw session-ID prefixes nobody can find on screen. Two superseded debug surfaces (the old Performance panel and the system-perf badge/popover) and their backend endpoints were still in the tree, unreferenced.

Someone opening the monitor needs answers in order: is the app heavy right now, where did memory and CPU go over the last minutes, which agent is responsible (under the label the workspace already shows, e.g. A15), and is the UI keeping up.

What changed

Per-agent history (main). AgentUsageHistory keeps 15 minutes of per-session memory and CPU plus a per-owner composition, derived from the process pages the coordinator already caches. Attribution rules, unit-tested:

• a process owned by exactly one session counts toward that agent, children included;
• a process reachable from several sessions is counted once as shared, never once per owner;
• app / agents / terminals / shared / other always sum to the page total;
• a placeholder row for an undiscoverable root is a session, not a process;
• sessions absent from the newest sample drop out of the ranking, and live sessions rank memory-descending (pinned as an IPC contract).

It lives in main rather than the helper because the question needs the minutes before the monitor opened, and helper query replies are capped at 256 KiB — too small for a large fleet's history. Served read-only via performance:monitor-agents, so it never waits on the helper.

Shared charts (components/charts). TimeSeriesChart (real pixel axes with rounded ceilings; runs break at missing readings and time gaps instead of drawing through them; crosshair snaps to real samples; one shared hover across charts covering the same window; thresholds; incident markers; full keyboard access with a live screen-reader readout), BarChart (nested series) and Sparkline (fixed caller-supplied ceiling so heights compare across rows). Geometry rules are pure, unit-tested functions in chartMath.

Monitor Overview rebuilt. Totals strip (app memory with share of system RAM and 15-minute delta, app CPU, heaviest agent, running counts, UI responsiveness, main heap, incidents — toned with the incident engine's own thresholds), stacked memory/CPU charts with one crosshair, then Agents by resource use: every running session under its workspace label (A15) with title, provider and tab; memory and share of the whole app; signed growth; CPU; process count; sparkline. Sortable by memory, CPU or growth — a leaking agent is often not the largest. Go to wakes a detached agent, focuses it, and closes the monitor only when navigation succeeds.

Timeline and Processes. Timeline moved onto the same charts: memory/CPU/responsiveness lanes, 1h and 6h ranges, Earlier/Later panning (the 7-day retention is finally reachable), and incident markers that select their incident. Processes shows agent label and title instead of session-ID prefixes; labels resolve through the workspace's own pane-label resolution, tested to navigate where they claim.

Agent Analytics uses the same chart components: per-day bars with wall-clock nested inside agent-hours (the containment is real; the gap is the parallelism), hover/keyboard readout, a parallelism tile, per-project share of all agent-hours.

Dead code removed. The old Performance panel (features/performance) and system-perf badge/popover (features/system-perf) with their endpoints performance:pane-stats and performance:system-stats, plus — found by review — the fully orphaned chain behind pane-stats: the ProcessTelemetry class, PerformanceService.recordPaneProcessStats, and the PanePerformanceStats/PanePerformanceSnapshot types. The "performance" feature-reference page moved to the monitor directory that now owns its commands, keeping the historical page id.

Design decisions

• History in main, not the renderer: the trend must predate the panel being opened.
• Growth as a first-class sort: "largest" and "leaking" are different questions.
• Shared helpers counted once: charging a shared MCP server to every owner would overcount; its bytes are reported once under "shared".
• Gaps break lines: a sleep gap or a missing reading is an interval nobody observed; drawing through it invents data, and a null is never shown as zero.
• Thresholds reuse the incident engine's: a tile turning amber means the same thing as an incident being recorded; no second, drifting scale.
• Decimal GB labels on binary divisions: matches what Activity Monitor / Task Manager print; users cross-check our numbers against those tools.

Independent reviews

Three independent Agent Code orchestration reviews ran on this PR (data layer, renderer/charts, PR-level conventions). All three verdicts: approve-with-comments, no blockers. Every blocker/major/minor finding was fixed in response:

• the orphaned pane-telemetry chain was deleted (review C's major);
• "heaviest agent" no longer depends on undocumented sort order; the ordering is now a pinned contract;
• zero-share rows paint no bar; the growth column labels the measured window;
• tests added for memory-descending ordering, the 256-session eviction cap, coordinator usage wiring (atomic pages, children summed, stop clears), and the shared in-flight preload read.

Tests and verification

• NODE_ENV=test npx vitest run on: charts (geometry + renderer interaction), AgentUsageHistory (attribution, ordering, eviction, ranking freshness), MonitorCoordinator (incl. usage wiring), performance-monitor (overview ranking, labels, Go-to), agent-analytics, feature-reference coverage, preload performance API — 9 files, 39 tests, all passing.
• npx tsc -b clean (full monorepo build).
• Attribution, transport atomicity and lifecycle suites from feat(performance): expose live monitoring and managed process attribution #951/feat(performance): collect bounded baseline evidence for every app run #953 pass unchanged.
• Manual UI pass still recommended before merge (see limitations).

Limitations and follow-ups

• Drag-to-zoom on the timeline is out of scope; ranges + Earlier/Later panning only (noted in feat(performance): make the monitor's first page attribute memory and CPU to agents #984 as follow-up).
• Per-agent history starts at launch and is not persisted across restarts; it fills over ~15 minutes.
• RSS can double-count pages shared between processes — stated in the overview footnote.
• Worst-case IPC payload (256 sessions × 180 points ≈ 1.7 MiB, ~23 ms clone per open window per 5 s poll) was measured by review and judged acceptable for a diagnostics panel; a top-K-history or decimation mitigation (~10–16× smaller) is a recorded follow-up if real fleets approach the cap.
• useMonitor (pre-existing) stops polling if a snapshot invoke never settles; a watchdog re-arm is a recorded follow-up.
• Renderer smoke tests run in jsdom; the charts' pointer behavior is covered by the keyboard-path tests plus shared geometry tests, not pixel assertions.