[tellaho]

Demo

Human walkthrough with temporary, labeled simulated issues. The Developer simulation controls shown in this recording are excluded from this PR. Captions describe the actions; the original app footage plays at normal speed.

doctor-walkthrough.mp4

Fresh production-build captures below use controlled read failures and neutral fixture data. The failure image pauses a native app-generated shake frame while the next read remains pending.

One persistent toast	Individual retry: Retry all stays disabled

Retry all: active source and queued controls	Failed row shakes while the next source runs

Overview

Category: new-feature
User Impact: Buzz now shows background-read problems in one persistent notice, with a single place to review and retry them.

Problem: A shared outage can produce separate errors across the sidebar, workspaces and conversation, and their Retry actions do not all recover the same reads.

Solution: Buzz Doctor collects current read failures in one toast and a community-scoped dialog. Retry all runs affected reads in dependency order, disables the queue, shows the active attempt and shakes each failed row as it finishes; drafts, partial data and local write recovery stay intact.

File changes

docs/buzz-help.md
Explain the Doctor toast, issue dialog, sequential retries and failure feedback.

docs/design-system-adoption.md
Record Doctor as the shared home for background-read recovery.

docs/navigation-history.md
Document the community-scoped Doctor visit and its retained workspace.

src/app/App.tsx
Mount one Doctor for the active app session and restore focus to the current workspace.

src/app/navigation.test.tsx
Cover workspace retention, Back/Forward and explicit dialog dismissal.

src/app/navigation.ts
Treat Doctor as a global dialog while retaining the workspace presentation.

src/app/shell/PrimarySidebar.tsx
Remove the superseded sidebar recovery notice.

src/bundled/channels/LiveStatus.tsx
Suppress only live failures represented in Doctor; preserve diagnostics and local behavior.

src/bundled/channels/SoloWorkspaces.tsx
Suppress the workspace read error only when Doctor represents that source.

src/features/channel-navigation/ChannelInventory.tsx
Route represented catalog, sidebar, roster and visibility errors to Doctor without duplicate controls.

src/features/channel-navigation/SoloInventory.tsx
Suppress represented roster errors while keeping unrelated local feedback.

src/features/navigation/controller.ts
Preserve history return paths for global dialogs.

src/features/navigation/service.ts
Retain underlying page work while Doctor is open.

src/features/navigation/targets.test.ts
Validate Doctor targets, scope and distinct history keys.

src/features/navigation/targets.ts
Add the community-scoped Doctor target and shared dialog classification.

src/features/relay/BuzzDoctor.test.tsx
Exercise mounted toast/dialog recovery, disabled controls, history, failure replay and reduced motion.

src/features/relay/BuzzDoctor.tsx
Render the toast and issue dialog with shared controls, source-specific suppression and action results.

src/features/relay/DoctorRetryButton.tsx
Animate failed attempts using shared motion tokens, including disabled rows, and honor reduced motion.

src/features/relay/broker-live.test.ts
Verify successful retry control clears its own old error without hiding a newer stream error.

src/features/relay/broker-live.ts
Clear a recovered retry-control error at its owner.

src/features/relay/doctor-testing.ts
Provide isolated real-session read fixtures and explicit request gates for regression tests.

src/features/relay/doctor.test.ts
Cover dependency ordering, sequential admission, source-local results, session replacement and recovery.

src/features/relay/doctor.ts
Collect current read issues and sequence explicit retries through their existing owners without replaying writes.

src/shared/design-system/DESIGN.md
Align feedback guidance with Doctor and document explicit disabled appearance during loading.

src/shared/design-system/styles/components.css
Apply shared disabled colors when a loading control is explicitly disabled.

src/shared/design-system/ui/Button.tsx
Expose explicit disabled appearance separately from the internal loading lock.

src/shared/design-system/ui/Toast.tsx
Keep the toast body navigation action keyboard reachable.

src/shared/design-system/ui/registry.ts
Document combined loading and explicit-disabled behavior.

tests/browser/hidden-conversation-retry.spec.mjs
Exercise the production app and read owners through exhausted reads, history, sequential retries and native per-row failure animation.

tests/browser/history-loading.spec.mjs
Observe Doctor instead of the superseded inline recovery notice.

tests/browser/live-status.spec.mjs
Keep live-recovery keyboard and focus expectations aligned with the toast.

tests/browser/live.spec.mjs
Verify the shared notice, cooldown and healthy subscriptions with controlled recovery timing.

tests/browser/navigation-groups.spec.mjs
Observe grouped read recovery through Doctor.

tests/browser/presence.spec.mjs
Observe live presence recovery through the shared notice.

tests/browser/relay-recovery.spec.mjs
Verify automatic live recovery plus manual recovery of remaining finite reads.

tests/fixtures/design-system/ui/ButtonSpecimens.tsx
Show the real Button with loading and explicit disabled state together.

Reproduction steps

1. Open Messages with a draft, then induce failures in background reads. For an isolated reproduction, run bin/pnpm test:browser hidden-conversation-retry.spec.mjs --project chromium --project webkit --no-deps.
2. Check that one persistent toast offers Retry and View issues. Open Doctor: the underlying conversation and draft remain in place; Back/Forward restores the dialog and explicit Close ends its visit.
3. Retry one issue. Its row spins, all retry controls disable, and Retry all remains plain. Retry all instead spins the aggregate action and runs each affected source in sequence.
4. Let a read fail while the next source is held. Its Retry button shakes immediately while still disabled. After recovery, the toast clears and an open Doctor shows the healthy state. Reduced motion suppresses shakes.

Validation

• TypeScript and 241 tests across eight complete affected model/component files passed after rebasing onto main a2ab54fb5, including the newer partial-inventory reconnect behavior and Solo composer contract.
• All four cases in hidden-conversation-retry.spec.mjs passed in Chromium/WebKit (19.7s). One browser case was added and the existing history/exhaustion case was retained. It covers the production read-owner integration and native Web Animations that DOM tests cannot establish; request gates prove a row fails before the later source completes. Retry matrices remain in model tests.
• Those checks ran at b7d949ed3. History cleanup preserved the full source tree; the final rebase only adds main’s unrelated agent-prompt wording change (f293f0418). Range-diff confirms the feature patches are unchanged. Mandatory commit hooks and final-head push checks passed: TypeScript/related unit tests (206.58s) and design types/guards (5.97s). Hosted checks provide the remaining broad coverage.
• The rebase preserves main's removal of the Solo composer inventory notice and its partial-inventory reconnect recovery. The temporary Developer harness is absent.

Not yet tested

• Native desktop/webview behavior on this final branch.
• Recovery against a real live relay/network outage; browser runs use controlled relay policy and read responses.
• Final human acceptance of the harness-free build. The supplied walkthrough exercises the earlier temporary simulation.

This remains a draft; independent agent review is deferred.

[kalvinnchau]

🤖 Three actionable findings: the persistent recovery toast obstructs Send, live/connection retries do not track settlement, and the recovery test can mask an automatic-recovery regression.

[kalvinnchau]

🤖 [P2] Keep the persistent Doctor toast clear of Send

A background visibility-read refusal places this persistent toast over the healthy conversation composer. In the production frontend/broker harness, Chromium and WebKit at 1440×600 and 390×600 with 16px and 20px root text both resolve Send’s center to Open Buzz Doctor. A real mouse click opens Doctor and publishes no message; Send becomes pointer-accessible after recovery removes the toast. An unrelated finite-read failure can therefore obstruct messaging indefinitely. Reserve composer clearance or move persistent recovery outside its controls, and restore Send pointer-actionability assertions in live-status.spec.mjs and relay-recovery.spec.mjs; the replacement focus checks pass despite this obstruction.

[kalvinnchau]

🤖 [P3] Track live/connection retry settlement and selectable work

relay.retry() and session.live.retry() return void, so awaiting source.retry() does not wait for reconnection. The loop advances to finite reads, and the immediate finally check observes connecting/pending rather than the eventual failure, losing per-row failure feedback for these sources. Separately, when the only retained issue is busy live/connection, the toast Retry remains enabled but selection excludes it and returns without feedback. Either wait for bounded source settlement to provide the stated dependency ordering/results, or make the weaker contract explicit and disable Retry when no work is selectable. Add live/connection-specific retry coverage. Whether early finite reads themselves fail again was not runtime-verified.

[kalvinnchau]

🤖 [P3] Separate automatic recovery from manual Retry coverage

The notice.isVisible() branch races notice removal and permits manual Retry all to repair reads that previously had to recover automatically. An automatic finite-read recovery regression can therefore pass through the manual branch, while other runs never exercise Retry all. Use deterministic separate cases: assert automatic recovery clears all issues without clicking Retry, then force a persistent finite-read failure and assert manual Retry clears it.

[baxen]

🤖 Closing in favor of #2968, which takes the "consolidate even harder" direction: one quiet, buttonless reconnect toast after 30 s on every page, replacing the separate LiveStatus, timeline and agent-banner notices. Automatic recovery already retries on its own, so it drops the per-source Retry / Retry all / View issues controls.

The one piece of this PR we're keeping is the broker-live.ts fix that clears "Live retry control failed (…)" once a later control request succeeds. It's being folded into #2968.

Thanks @tellaho for pushing on this. The video and the whack-a-mole framing are what set the direction.