PR evidence: Codex usage account bars

Usage account attribution screenshot evidence

Base: pingdotgg/t3code 0232932.
Fix: 94aef86.

reported-issue.png is the original user-supplied screenshot of the installed Nightly client.
before-1440.png and after-1440.png render the actual UsageLimitsPooled component and shared collectors from base and fix, using the same anonymized two-account fixture and production Tailwind CSS. The fixture reproduces the original 3 Session / 4 Weekly bars. The fix produces 2 account columns per row, with the missing Session value labeled Not reported.

This is focused component evidence, not a live authenticated end-to-end client or proxy verification. The renderer uses server-rendered React in headless Chrome; popover/button wrappers preserve children, classes and positions, while settings, icons, service callbacks, hover behavior and pace measurement are stubbed. No quota is fetched. The installed app was not replaced. A narrow web capture is included as supplemental evidence; it is not a native mobile runtime capture.

Anonymized fixture: first account has 21% Session and 7% Weekly remaining; it is reported by native Mac and agentbox logins without email, plus a CLIProxyAPI hub with the matching workspace ID and email. Second account has 100% Weekly remaining and no Session observation. Before, Weekly mean is 30% after rounding because the first account is counted three times. After, Weekly mean is 54% after rounding, and Session remains 21% over the one reported observation. Missing values do not contribute to aggregate calculations.

The public capture.mjs uses T3_USAGE_REPO to locate a prepared checkout. Run with node after installing locked dependencies. It writes disposable files beside the script and launches a temporary headless Chrome instance that it closes on completion. No server is started.

PR evidence: Working shelf exit grace + completion notifications (61c8efe41d)

Before/after UI evidence for pingdotgg#22458 (issue pingdotgg#24477). Before = upstream main d53873b, after = PR head 61c8efe. See README.md for method, timings and limits.

Evidence: thread-list snapshot index

PR evidence only (not a release): chart, harness and results.

Evidence: subagent questions go to the orchestrator first

PR evidence assets only; not a release.

Evidence: mobile fast resume after iOS socket kill

PR evidence only (not a release): screenshots, timing chart, harness and results.

PR evidence: Working shelf exit grace

Working shelf exit grace: UI evidence

Baseline: upstream 7e289ae1f77b9137580688aaccf00d48906984ad.
Patch: 005701b8f0882b020c3ac6ddd4a64c6fc208102a.

The actual web client ran in an isolated development instance on macOS with Chrome. The client atom registry received a synthetic thread shell: running, then completed for 500 ms, then running again, then completed with no continuation. No provider was invoked and no backend lifecycle behavior is claimed. An annotation explains each phase. Published videos are cropped to the sidebar; screenshots hide unrelated notification toasts.

• At 1.5 seconds, the synthetic partial step completes.
• At 2.0 seconds, its continuation starts.
• At 3.5 seconds, it completes for good.
• Upstream immediately moves the row into the Inbox during the gap, then back into Working.
• The patch retains the row in Working through the gap, and returns it after the final two-second grace period.

before-gap.png and after-gap.png show the same 500 ms gap. The videos include the complete transition sequence. capture.mjs is the browser capture script; it assumes an authenticated isolated dev instance, a synthetic project, and the installed workspace Playwright dependency. Authentication files are intentionally not included.

The T3 collaborative preview could load and inspect the app but timed out on screenshot and recording operations. These captures use local Chrome through Playwright after those repeated failures.

Validation: 332 focused tests across threadInbox, threadShell, Sidebar.logic and mobile threadListV2; scoped client-runtime/web/mobile typechecks; scoped lint and formatting. GPT-6.1-Sol approved the exact patch and independently passed all 19 tracker tests. Native Electron/mobile and real provider continuation were not exercised.

PR evidence: Codex app-server exit

Screenshots for the upstream PR fixing runs stuck after the shared Codex app-server exits. Not a release.

Terminal run-and-send: PR 20225 evidence

Before/after interaction screenshots and recording for pingdotgg#20225. Captured against the independently reviewed feature before the upstream integration merge. Public upload explicitly approved by the contributor.

Native subagent Stop PR evidence

Isolated reproduction using synthetic provider events and real Stop controls. Evidence only; not a software release.

Evidence: task_status and Working row report silent delegated turns

PR evidence assets only; not a release.