feat(agent-chat): the data box shows what it waits for by MarkusNeusinger · Pull Request #12221 · MarkusNeusinger/anyplot

MarkusNeusinger · 2026-10-20T22:52:45Z

Summary

Owner report 2026-10-21: dropping a Parquet file (or any other binary file read on the server), or clicking .parse(), greyed out the buttons and seemed to do nothing until the data suddenly appeared. The only sign of work was a quiet line under the buttons.
New BoxWorking:
- It is drawn over the data box while a server-side file read or a parse runs: # reading "x.parquet" on the server … · 4 s or # parsing your data … · 3 s.
- It has the existing pulsing Ellipsis, a seconds counter from the first second on, and a thin indeterminate bar along the box's foot.
- That is a plain indicator, not a spinner (docs/reference/style-guide.md §10.1).
- There is no bar motion under prefers-reduced-motion.
- The overlay is aria-hidden; screen readers keep hearing the existing live line. The drop zone gets aria-busy.
In DataPanel, the converting boolean became convertLine (the line text itself), so the overlay and the live line say the same thing.

Plan

N/A. Gemini 4 Argon reviewed it first; its verdict is in a comment.

Test plan

BoxWorking.test.tsx: the line shows at once, and the counter reads · 3 s after 3 s (fake timers).
DataPanel.test.tsx: the overlay shows during a conversion and during .parse(), and is gone after the file lands. aria-busy is set while parsing.
yarn lint, fm:check, type-check, test (1,163 passed) and build.
Not driven in a browser: the flow needs the agents backend. After deploy, check it on /debug/agent by dropping ~/anyplot-testdaten/line-timeseries/stromzaehler.parquet, in both themes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

A server-side file read (Parquet, Excel, ODS, Feather, NumPy) and a .parse() take seconds, but the box only greyed its buttons and changed a quiet line under them, so the page looked stuck. BoxWorking draws the step over the box with a pulsing ellipsis, a seconds counter and a thin bar (plain indicator, style guide 10.1; no motion under reduced motion). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

Copilot

🟡 Changes recommended

Elapsed time resets when the overlay remounts, and long filenames can be clipped on narrow screens.

2 open issues

What changed in this PR

Adds visible progress feedback to the agent-chat data box during server-side file reads and parsing.

Changes:

Adds an overlay with status text, elapsed seconds, and a reduced-motion-aware progress bar.
Shares conversion status text and marks the drop zone as busy.
Adds component tests and a changelog entry.

File	Description
changelog.d/agent-box-working.md	Documents the progress feedback.
app/src/sections/agent-chat/DataPanel.tsx	Integrates the overlay and busy state.
app/src/sections/agent-chat/DataPanel.test.tsx	Checks conversion and parsing feedback.
app/src/sections/agent-chat/BoxWorking.tsx	Implements the progress overlay and timer.
app/src/sections/agent-chat/BoxWorking.test.tsx	Tests status text and elapsed seconds.

_{Review effort: Balanced}

codecov · 2026-10-20T22:58:25Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

…icks Gemini 4 Argon review of #12221: - the overlay dims the text underneath, so the line stands out on the box's own surface in both themes; - a drag over the box hides the overlay instead of unmounting it, so the seconds counter keeps running; - only the conversion that set the line clears it (convertSeq), so a later read that converted nothing cannot leave the box waiting; - aria-busy sits on the textarea, not on its wrapper. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

MarkusNeusinger · 2026-10-20T22:58:31Z

First review: Gemini 4 Argon

Verdict: needs fixes. Applied in the latest commit:

Overlay surface: the pill shared the textarea's surface. The overlay now dims the text underneath (color-mix of --bg-elevated at 55 %), so the line stands out in both themes.
Stuck overlay: a later read that converted nothing could leave the overlay stuck, because the clear was guarded by readSeq. Only the conversion that set the line clears it now (convertSeq).
Drag over the box: a drag unmounted the overlay and reset the counter. It is now hidden instead, and the counter keeps running.
aria-busy: moved from the wrapper div to the textarea.

Not applied:

Disabling the textarea while it waits. Editing during a conversion is a designed path: the edit wins, and .replace() asks when the file lands (textRef, offerFile). That behaviour is unchanged by this PR, so I kept it.
Dropping the counter and the bar as over-designed. The owner asked for a visible loading animation. A sandbox read takes several seconds, so the counter shows that the page is alive. The bar has no motion under reduced motion.
Hiding the live line under the box. Screen readers need it, and the overlay is aria-hidden.

Fable was not involved: this is a UI-only change with no high-risk surface.

… wrap Copilot review of #12221: - the panel keeps when the conversion and the parse began (set where they start), so collapsing and reopening the card no longer restarts the overlay's seconds count; - the overlay's label wraps a long file or sheet name instead of being clipped on a narrow screen. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

Copilot

🔵 Needs a closer look

Browser verification of overlay layout and legibility on desktop and mobile in both themes remains outstanding.

No open issues 🎉

2 resolved since last review

_{Review effort: Balanced}

…ude (#12226) ## Summary - Google's Antigravity FAQ ("Wrapping `agy` inside third-party coding agents or applications … is not supported under consumer plans") and a staff reply on 2026-10-16 ([forum topic 189905](https://discuss.ai.google.dev/t/clarification-on-consumer-plan-policy-for-invoking-official-agy-via-local-subprocess/189905): "No under consumer plans", even for the unmodified binary as a local subprocess) rule out the read-only `agy -p` second opinions #12223 introduced. Claude stopped running agy on 2026-10-21. - Second opinions and first PR reviews go through the new `zweitmeinung` project skill: `.claude/skills/zweitmeinung/gemini_review.py` calls Gemini through Vertex AI in project `anyplot`, location `global` (the site's agents stay on `eu`), with read-only tools over tracked repository files, prefetched data, and the public API; no shell, no credentials. The owner's split (2026-10-21/22): **Gemini 3.8 Flash on Flex PayGo** (half price) for routine first reviews, on probation against Copilot's findings; **Gemini 4 Argon**, thinking high, for plans and second opinions. - Cost work after a comparison run on three merged PRs (#12203, #12221, #12222): every tool turn re-sends the conversation, so `--diff-base` inlines the diff and changed files and `read_files` reads several files per call. Two loop bugs found there are fixed and pinned by tests: a text part beside the last function responses made Vertex answer 400, and a last turn without tools lost the implicit cache while Flash kept calling tools. The budget is now enforced through the tool responses with one tool configuration for all turns. Measured: Flash $0.07–0.28 per review, Argon $0.57–1.16. - Flex gets 5 minutes per run (and per request); after that, or after a timeout or a 429/5xx, the run goes online (`flex_fallback`, `per_turn_seconds`). The SDK retries timeouts too: the first Flash review of this PR waited over an hour on 15-minute attempts, and with a 5-minute cap per request the next one still ran for over 25 minutes. - Hardening after Copilot's reviews: a symlink is visible only when its target is itself a visible entry (a tracked `review.txt -> .env` exposed the untracked `.env` before), secret-name patterns apply to every path component of `--data` files, lock files stay out of the inlined diff and the diff is capped at 120,000 characters, a backtracking search pattern is stopped after 10 seconds, a line longer than the tool limit is cut instead of looping, and an unencoded URL is refused instead of crashing the run. - `gemini-task`: Gemini pushes its `gemini/*` branch and opens a **draft PR** labelled `gemini`. CI runs on drafts, Copilot does not (`review_draft_pull_requests: false` in the ruleset), and Claude marks the draft ready after its own review. The owner may let Gemini run tests and builds unattended in his own agy session; only he starts agy. - `CLAUDE.md` and `.claude/guardrails.md` describe the lanes and costs; the guardrail test anchors pin the obligations (Flash/Argon split, draft PRs only, Claude never runs agy). ## Plan N/A. The owner chose the Vertex route on 2026-10-21 after reading a research report on the policy, then the Flash/Argon split, the draft-PR hand-over, and moving the skill into the repository on 2026-10-22. ## Test plan - [x] `uv run pytest tests/unit tests/integration -n 8` (new `tests/unit/test_gemini_review.py`: confinement, http_get refusals, inlining caps, and the tool loop against a fake Vertex client) - [x] `uv run ruff check . && uv run ruff format --check .` - [x] `uv run python -m tools.changelog check --base origin/main` - [x] Flash smoke run with `--max-turns 1` on Vertex: forced last turn answers (exit 0), `traffic_type` `ON_DEMAND_FLEX` - [x] Gemini 3.8 Flash first review of this PR through the repository script: 3 of 5 findings applied (owner-only PR comments for Gemini, `git diff -z`, `*api_key*`); 2 skipped (callers already resolve `check_attachment`'s root, and the SDK returns `finish_reason` as an enum). Flex gave no answer on the first turn, so the run went online: 9 turns, $0.86. - [x] Copilot: three rounds on the rework, 9 findings, all applied 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Copilot AI balanced review requested due to automatic review settings October 20, 2026 22:52

chore: reference PR #12221 in the changelog fragment

9e97095

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

Copilot started reviewing on behalf of MarkusNeusinger October 20, 2026 22:53

View session

Copilot AI reviewed Oct 20, 2026

View reviewed changes

Comment thread app/src/sections/agent-chat/BoxWorking.tsx Outdated

Comment thread app/src/sections/agent-chat/BoxWorking.tsx

MarkusNeusinger requested a balanced review from Copilot October 20, 2026 23:12

Copilot started reviewing on behalf of MarkusNeusinger October 20, 2026 23:12

View session

Copilot AI reviewed Oct 20, 2026

View reviewed changes

MarkusNeusinger merged commit c87940f into main Oct 21, 2026
16 checks passed

MarkusNeusinger deleted the feat/agent-convert-loading branch October 21, 2026 06:08

MarkusNeusinger mentioned this pull request Oct 22, 2026

docs(agents): second opinions through the Gemini API, no agy from Claude #12226

Merged

6 tasks done

Navigation Menu

Repository navigation

feat(agent-chat): the data box shows what it waits for - #12221

feat(agent-chat): the data box shows what it waits for#12221
MarkusNeusinger merged 4 commits into
mainfrom
feat/agent-convert-loading

MarkusNeusinger commented Oct 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Oct 20, 2026

Uh oh!

MarkusNeusinger commented Oct 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

MarkusNeusinger commented Oct 20, 2026

Summary

Plan

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Uh oh!

codecov Bot commented Oct 20, 2026

Codecov Report

Uh oh!

MarkusNeusinger commented Oct 20, 2026

First review: Gemini 4 Argon

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🔵 Needs a closer look

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants