page-embedded chart blocks (Command Center 大屏) render axes and legend but no marks until a resize event — needs a controlled repro

[hotlong]

What was measured

Showcase → Analytics → Command Center (大屏) (page/showcase_command_center, a structured page whose chart blocks bind analytics datasets), objectui @ 67dadd6, ~800px console:

1. Load 1, screenshot at 5 s: KPIs populated (2 / 8 / 2 / 1 / 14 / 1.1M). 任务状态分布 shows y-axis 0-2 and the category labels Backlog / In Progress / To Do but no bars; 项目健康度 shows only the legend Green / Red / Yellow, no donut.
2. Load 2, SVG census at 10 s: the status chart's svg has axis texts and 1 rect, 5 paths; the health chart's svg has 3 paths and no text. Every POST /api/v1/analytics/dataset/query answered 200.
3. Then a viewport zoom reset (cmd+0, i.e. a resize event) — the same page immediately drew full bars and the donut.

The same datasets draw correctly on the Delivery Operations dashboard, so this is about the page-embedded chart block's first draw (a 0-size mount that never re-measures once data arrives is the usual shape), not the data.

Needs

A controlled repro without the automation harness: load the page, wait 30 s, do not resize, screenshot; then resize. If it reproduces, the chart block should re-measure on data arrival / ResizeObserver first callback, not only on window resize.

Found during the objectstack pin-bump dogfood (objectstack#14295); filed as "observed, not yet confirmed".

Generated by Claude Code

[hotlong]

Claim: PM loop round 4 (dogfood dispatch) — MEASUREMENT-FIRST card
Session: session_62849a16-0144-4728-8942-9f60bb3a73f1
Branch: claude/issue-7250-page-chart-first-draw
Worktree: objectui-issue-7250
Domain: domain:ui
File surface: measurement first (no edit until the repro exists); if confirmed, the page-embedded chart block renderer (the page:*/structured-page chart block under packages/app-shell or packages/plugin-dashboard) + tests, .changeset/ — NOT packages/plugin-charts/src/AdvancedChartImpl.tsx, which #7385 (in flight) and #7248 (serialized after it) own (stop on breach; explain in the report)
Container & model: M, model: opus, mode:subagent
Clause-②: no
Serial constraints cleared: #7385 (plugin-charts axis policy) is queued; this card must not touch AdvancedChartImpl.tsx — if the repro lands the defect there, stop and report; no open objectui PR touches the page chart block (REST pulls list 2026-09-03)

Maintainer direct-dispatch channel (2026-08-10 standing authorization); instruction in this session, verbatim: 「P1/P2 都开 issue，packages 500 单独修，服务先留着」 → 「使用 pm-dispatch 派发」. Dispatched by the dogfood session that filed the card, outside the lane seat; the seat holder keeps precedence on any later claim conflict.

Generated by Claude Code

[hotlong]

Claim: os-dev (measurement-first card) — starting the controlled repro now.
Session: session_62849a16-0144-4728-8942-9f60bb3a73f1
Branch: claude/issue-7250-page-chart-first-draw
Worktree: objectui-issue-7250
Base: b6e83be6a537eea8fdff45b19280591766140785

Protocol: cold-load the Command Center page at 800x900, no resize, sample every chart svg (rect / path / text counts + first path d length) at 5s / 15s / 30s, then dispatch one resize and re-sample; same protocol on the showcase_ops_dashboard control page. No edit until the measurement exists. packages/plugin-charts/src/AdvancedChartImpl.tsx is off limits here (#7385 / #7248 own it) — if the defect lands there I stop and report.

Generated by Claude Code

[hotlong]

Controlled repro: NOT REPRODUCED — the premise is falsified

Ran the protocol this card asked for (real Chromium via the repo's Playwright, no automation-harness clicking, viewport 800x900, no resize of any kind until the last step). Backend: showcase on :3911. Two independent cold loads on the target page, one on the control dashboard, one on the published console build, plus a first-draw profile and a zoom sweep.

Reading 1 & 2 — target page, page/showcase_command_center, two independent cold loads

Both loads produced byte-identical censuses. Per chart svg, rect / path / text / .recharts-bar-rectangle / .recharts-pie-sector / .recharts-area-area, plus first path d length:

t	任务状态分布	项目健康度	chart2	trend (632px)	chart4
5 s	r1 p5 t6 bars=5 d=213	r1 p3 t0 sectors=3 d=144	r1 p4 t7 bars=4	r1 p2 t12 areas=1	r1 p6 t7 bars=6
15 s	identical	identical	identical	identical	identical
30 s	identical	identical	identical	identical	identical

POST /api/v1/analytics/dataset/query: 10x 200, 0 failures. Screenshots at 5 s / 15 s / 30 s all show full bars and a full donut. Dispatching window.dispatchEvent(new Event('resize')) after 30 s changed nothing (every count identical). A real 1px viewport change (800 to 801) changed only geometry (d lengths shift, box 290 to 291) — mark counts unchanged.

Reading 3 — control page, dashboard/showcase_ops_dashboard

Same protocol, same outcome: bars=3 / bars=5 / sectors=4 / lines=1 all present at 5 s, unchanged at 15 s and 30 s, unchanged by the resize event. The target-vs-control asymmetry the card is built on does not exist.

Reading 4 — the PUBLISHED console build (:3911/_console), the artifact the dogfood actually drove

Identical to the dev build: bars=5, sectors=3, areas=1 at 5 s, unchanged through 30 s with no resize.

What the original census actually shows

The reported signature — status chart 1 rect, 5 paths, health chart 3 paths and no text — is exactly the signature of a fully drawn chart in this codebase, not of an axis-only frame. Dumped element by element:

• the status chart's 5 paths are all class="recharts-rectangle", parent recharts-layer recharts-inactive-bar, each 37x252 px, gradient-filled. Bars are drawn as path here, not rect.
• the health chart's 3 paths are all class="recharts-sector" — the three donut segments. A donut has no axes, hence no text; no text is its normal state, not a symptom.
• the lone rect in each is the unclassed chart background.

So step 2 of the original report was measuring a chart that had already drawn.

The real blank window — finite, self-resolving, ~1 s wide

A finer profile (same cold load, no resize) locates the transient the 5 s screenshot most likely caught:

t=500..2500ms surfaces=0 (charts not mounted yet)
t=3000ms surfaces=5 r=1 p=0 t=6 bars=5 sec=2 area=1 <-- axis text, ZERO paths
t=4000ms surfaces=5 r=1 p=5 t=6 bars=5 sec=3 area=1 <-- marks present
t=5000ms+ unchanged thereafter

t=3000ms is the genuine axis-only signature: p=0. It lasts about one second and clears on its own, with no resize. Under the dogfood's load (automation harness plus a full framework dev server) that window plausibly stretched past a 5 s screenshot — which explains observation 1 — while observation 2 at 10 s was already a drawn chart, and observation 3's zoom reset simply happened after the animation had finished. The "resize fixed it" causality is post hoc.

Zoom hypothesis — also negative

Because the original reset zoom with cmd+0, I re-ran cold loads at zoom 0.5 / 0.67 / 0.8. All three: bars=5, sectors=3, areas=1 at 5 s, unchanged at 15 s. Zoom does not produce a stuck blank.

Same code as the dogfood observed

The dogfood commit 67dadd602a3a is an ancestor of the measured HEAD b6e83be6a. The only commit touching packages/plugin-charts, packages/plugin-dashboard or the page renderer in that range is 4ca30d044 (#7355), whose diff is ObjectDataTable.tsx plus its test — a table, not a chart. AdvancedChartImpl.tsx is untouched in that range. So this is not a "already fixed since" result: the chart render path measured here is the same code the dogfood saw.

Verdict

No fix, no PR. Nothing in the page chart block re-measures only on window resize; the marks arrive on their own and stay. Recommend closing as not-reproducible. AdvancedChartImpl.tsx was not touched (#7385 / #7248 keep it).

If a slow-first-paint chart is worth pursuing as its own concern, it is a different card — an enter-animation window with no skeleton, visible as an axis-only frame for about a second — and it is cosmetic, not a stuck state.

Generated by Claude Code

[hotlong]

os-dev-report

{
 "issue": 7250,
 "status": "done",
 "branch": "claude/issue-7250-page-chart-first-draw",
 "pr": null,
 "premise_still_valid": false,
 "summary": "Measurement-first card, outcome (b): the controlled repro does NOT reproduce, so no fix was written and no PR was opened. Two independent cold loads of page/showcase_command_center at 800x900 with no resize show every chart drawing its marks by t=5s and unchanged at 15s and 30s; a dispatched resize event after 30s changes nothing. The control dashboard behaves identically, so the target-vs-control asymmetry the card rests on does not exist, and the published console build at :3911/_console (the artifact the dogfood actually drove) behaves identically too. The issue's own census is the falsifier: 1 rect / 5 paths and 3 paths with no text is exactly the signature of a FULLY DRAWN chart here, since bars render as path.recharts-rectangle (5 of them, 37x252px each) and donut segments as path.recharts-sector (a donut has no axes, so no text is normal). A finer profile does locate a real but transient axis-only window at t=3000ms (r=1 p=0 t=6) that clears by t=4000ms on its own with no resize, which plausibly explains the original 5s screenshot under harness load; it is a ~1s enter-animation window, not a stuck state. Zoom 0.5/0.67/0.8 cold loads were also negative. The dogfood commit 67dadd602a3a is an ancestor of measured HEAD b6e83be6a and the only commit touching the chart packages in that range is 4ca30d044 (ObjectDataTable.tsx, a table), so this is a falsification of the premise, not an already-fixed-since result. packages/plugin-charts/src/AdvancedChartImpl.tsx was never touched. NOTE for PM: the dispatch brief summarized the 10s census as '1 rect / 3 path no text'; the issue body actually reads status chart 1 rect / 5 paths with axis texts, health chart 3 paths and no text. The body was complete, no sanitizer truncation.",
 "tests": "No code changed, so no gate run is owed (no changeset, no lint, no vitest, no type-check). Verification was the browser measurement, run through scripts/pm/os-verify-lock.sh, which printed: 'VERDICT command-exit 0 . UNLOCKED (declared) . no usable flock on this host, so the shared verify lock was NEVER taken and NOTHING was serialized . ran 34s'. Harness: repo Playwright 1.62.1 chromium headless, viewport 800x900, storageState from a real UI login, console dev server on :5191 with DEV_PROXY_TARGET=http://localhost:3911, plus the published console on :3911/_console. Readings, all with NO resize: (1) target page load 1, t=5s/15s/30s byte-identical: chart0 rect=1 path=5 text=6 bars=5 dLen=213, chart1 rect=1 path=3 text=0 sectors=3 dLen=144, chart2 rect=1 path=4 text=7 bars=4, chart3 rect=1 path=2 text=12 areas=1, chart4 rect=1 path=6 text=7 bars=6; analytics dataset/query 200=10 other=0. (2) target page load 2, independent cold load, identical to load 1 at all three sample points. (3) control dashboard/showcase_ops_dashboard, t=5s/15s/30s identical: bars=3, bars=5, sectors=4, lines=1; 200=18 other=0. (4) published console :3911/_console same page, t=5s/15s/30s identical to the dev build; 200=5 other=0. Post-sample resize: window.dispatchEvent(new Event('resize')) changed zero counts on both pages; a real 800->801 viewport change moved only geometry (box 290->291, dLen 144->160), mark counts unchanged. First-draw profile: t=500..2500ms surfaces=0, t=3000ms surfaces=5 with r=1 p=0 t=6 (axis text, zero paths), t=4000ms p=5/p=3 (marks present), stable after. Structural dump confirming the census reading: the 5 paths are all class=recharts-rectangle under recharts-layer recharts-inactive-bar, bbox 37x252 each, gradient fills; the 3 paths are all class=recharts-sector with solid fills. Zoom sweep 0.5/0.67/0.8: bars=5 sectors=3 areas=1 at t=5s and unchanged at t=15s in all three. Screenshots at every sample point are in the scratchpad (issue-7250/), not attached. Repro-only scripts lived in the scratchpad, never in the repo; the worktree diff against base b6e83be6a537eea8fdff45b19280591766140785 is empty.",
 "mcp_calls": "0 — all GitHub reads and writes went through the gh CLI; no MCP GitHub call was made",
 "open_questions": [],
 "out_of_scope_findings": []
}

Generated by Claude Code

[hotlong]

PM disposition — reviewer of record: PM session session_62849a16-0144-4728-8942-9f60bb3a73f1.

Measurement-first outcome (b): two independent cold loads at 800×900 with no resize show every chart on the Command Center page fully drawn by t=5 s and byte-identical at 15 s and 30 s (dev build and the published console at :3911 alike); a dispatched resize changes nothing; the control dashboard behaves identically. The card's own census was misread by the reporter: bars render as path.recharts-rectangle and donut segments as path.recharts-sector, so "1 rect / 5 paths" IS a drawn bar chart, and a donut has no axis text by construction. The only transient is a ~1 s axis-only enter-animation window around t≈3 s, which the harness-loaded 5 s screenshot caught. The dogfood pin 67dadd6 is an ancestor of the measured head and nothing touched the chart packages in between, so this is a falsified premise, not fixed-since.

Closing as not-reproduced; no PR. Lesson recorded on the PM side: timing/first-paint readings taken inside the Claude Code browser pane are not filing-grade evidence — measure in a clean Playwright context first (same failure as #7249).

Generated by Claude Code