kandev/docs/plans/threads-layouts/plan.md

created, status, requirements, system_design, legacy_specs

created	status	requirements	system_design	legacy_specs
2026-09-11	implemented	REQ-UI-THREADS-DECK-004 REQ-UI-THREADS-DECK-005 REQ-UI-THREADS-SAVED-VIEWS-005	../../specs/ui/system-design/threads-conversation-deck.md ../../specs/ui/system-design/threads-saved-views.md

Implementation Plan: Threads Layouts

Overview

Add Columns and a two-row Grid, plus independent composer auto-hide, saved with each Threads view. Keep the existing Open task action. Phones keep one conversation at a time with swipe/picker navigation and the normal composer.

The user approved implementation on 2026-09-11. The initial five orders are done. The scoped polish requested on 2026-09-12 is also complete in Task 06. The feature merged in PR #3626 on 2026-09-13 at f718c50666eb7175e39087294afabc70ac39f3a7. Review remediation retains the existing implementation boundaries and is recorded below. The focused Grid clarification overflow follow-up records a later conformance repair; its completed results do not replace this package's historical evidence. Read the previews first, then the work order for the implementation boundary being changed.

Scope

In scope

• Persist layout and composer preferences through existing view drafts, Save/Save as/Discard, synchronization, and recovery.
• Render a stable two-row desktop/tablet grid with viewport-owned live chats.
• Reveal composers by pointer dwell or keyboard, preserving drafts, reading position, and access to required actions. Touch keeps the normal composer.
• Add Display controls, translated descriptions, and Maximum chats wording.
• Prove the flows with focused tests and update the existing user guide.

Out of scope

• New enlargement modes, arbitrary row counts, tile resizing, drag ordering, masonry, or more than one simultaneous phone transcript.
• New view tables/endpoints, browser storage for portable preferences, agent lifecycle changes, or changes to session-management/authorization rules.
• Automatically enabling Grid/auto-hide or changing the five-chat default.
• Native implementation delegation, publishing, commits, or runtime deployment.

Technical approach

Sources of truth

The UI system already owns the deck requirements and saved-view requirements. Extend those pairs rather than creating another feature specification.

The existing windowing package, saved-view package, phone package, and task-action package are implemented foundations. Their recorded results apply to their original scopes, not this package. Preserve their tests; add new cases without reopening or rewriting historical completion evidence. No sibling branch is required.

The existing ADRs already establish surface-owned views, portable user settings, and viewport activation. This feature applies those boundaries and does not need another ADR.

Persist presentation

Extend ThreadView/ThreadViewDraft in backend models and frontend types with layout and auto_hide_composer (frontend autoHideComposer). Defaults are Columns and false. Keep missing/unknown stored-field recovery independent; reject unsupported new writes before saving. Extend every clone, draft merge, and rollback snapshot. Keep queryFingerprint query-only.

Compose the board

Pass query.effectiveView presentation to ThreadsBoard. Preserve direct task children, keys, session choices, and the observer registry. Add the thread-layout.ts effective-layout helper. Grid fills down each pair, keeps the 360px minimum width, and uses two equal rows. A single task fills the height. A grid needs two 300px tiles plus the 12px gap in the board content box; below that, use Columns with a visible reason and no settings write. Extend useThreadSelectionRecovery to anchor layout changes.

Refresh viewport geometry when effective layout changes even if task IDs do not. Ignore stale observer callbacks. Both visible rows own live selected sessions, while the existing one-adjacent-task preload and single-phone-detail rules remain. Do not pin offscreen chats to retain a composer.

Disclose the existing composer

Add an optional Threads-only context scoped to the task and selected session. use-composer-disclosure.ts owns dwell/exit timers, explicit-open state, and keep-open conditions. Reuse existing editor state through a narrow interaction callback; audit native portaled model/mode/context/queue overlays. Do not duplicate editor, draft, submit, cancel, or recovery ownership.

Collapsed routine content occupies no height and is inert. Only the existing CI popover remains, when applicable; no Reply, Stop, or plugin row is added. Required actions/recovery force the normal surface open. Keyboard focus on the task tile reveals the editor. Manual Collapse preserves a draft and returns focus to the tile; Enter or a new pointer entry reopens it. Phone and coarse-pointer layouts keep the normal composer without changing preferences. Editor and plugin instances stay mounted. Opaque plugin activity does not hold the composer open, and collapse never cancels or revokes a plugin operation. The native transcript scroll owner retains bottom-follow or the current message/offset during footer resizing. A bounded expanded footer leaves an 80px transcript floor and keeps long input/action content reachable.

Expose controls

Keep Display exclusively in the existing view editor, using its draft and Save/Discard actions without a duplicate top-bar selector. Phone/tablet use the current drawer. Relabel the existing limit Maximum chats, retaining max_columns and five-task defaults. All copy is localized in five languages plus generated pseudo output.

ASCII UI preview

Spacing and labels are illustrative. The structural requirements are control order, stable task identity, two desktop rows, independent transcript scrolling, a CI-only collapsed footer, and one phone conversation. Use existing primitives/tokens; do not hardcode copy or pixel-perfect drawings.

UI-01: Columns, normal composers

Entry: Threads, existing/new view defaults. This preserves the current body.

[All threads v] [View settings]  [Listing views]
+------------------+------------------+------------------+
| A        [Open]  | B        [Open]  | C        [Open]  |
| status / agent   | status / agent   | status / agent   |
|                  |                  |                  |
| transcript       | transcript       | transcript       |
| scrolls          | scrolls          | scrolls          |
|                  |                  |                  |
| status controls  | status controls  | status controls  |
| [Reply editor]   | [Reply editor]   | [Reply editor]   |
| [Composer tools] | [Composer tools] | [Composer tools] |
+------------------+------------------+------------------+

Topbar and tile headers are fixed. The board scrolls horizontally when needed. Covers AC-UI-THREADS-DECK-004.1/.8 and AC-UI-THREADS-DECK-005.9. Rendered check: existing threads-view.spec.ts and UI-02 comparison.

UI-02: Grid, auto-hide enabled, no active draft

Entry: select Grid in Display and enable Auto-hide composer.

[Monitoring v] [View settings]      [Listing views]
+------------------+------------------+------------------+
| A        [Open]  | C        [Open]  | E        [Open]  |
| status / agent   | status / agent   | status / agent   |
| live transcript  | live transcript  | live transcript  |
| [CI]*            | [CI]*            | [CI]*            |
+------------------+------------------+------------------+
| B        [Open]  | D        [Open]  | F        [Open]  |
| status / agent   | status / agent   | status / agent   |
| live transcript  | live transcript  | live transcript  |
| [CI]*            | [CI]*            | [CI]*            |
+------------------+------------------+------------------+
                     horizontal board scroll ->

[CI]* appears only when applicable; otherwise no footer height is reserved. Stop is available in the revealed composer. Each transcript scrolls independently. Task order remains A, B, C, D, E, F. Five admitted tasks leave the last lower cell empty; one task fills the height. Both rows load visible selected chats. An empty/initially loading deck uses its existing message and keeps the topbar.

Covers AC-UI-THREADS-DECK-004.1 through .5/.8 and AC-UI-THREADS-SAVED-VIEWS-005.6. Rendered check: threads-layouts.spec.ts, including lower-row deep links, removal, bounds, and one versus two rows.

UI-03: Grid tile, hovered or actively composing

Entry: hover one tile for 150ms, or focus the task tile with the keyboard.

+-----------------------------------+
| B                         [Open]  |
| status / selected agent           |
| transcript (independent scroll)   |
|-----------------------------------|
| status / queue controls           |
| [Draft message                  ] |
| [model] [mode] [attach]     [Send] |
|                        [Collapse] |
+-----------------------------------+

Only this tile changes its internal allocation. Hover does not focus. Its routine footer uses one interruptible 200ms reveal/160ms collapse with a subtle fade/6px motion. CI stays outside that region. Reduced motion is instant. A portaled picker holds the composer open. A draft stays open during automatic hiding; explicit Collapse retains it and returns focus to the tile. A pending question/permission/recovery also expands this region without hover. Long footer content has bounded scrolling; required response actions stay reachable. The sibling tile keeps identical bounds.

Covers AC-UI-THREADS-DECK-005.1 through .7/.9/.10. Rendered check: threads-composer-disclosure.spec.ts and its mobile companion.

UI-04: Display settings and recovery

Entry: View settings.

+---------------------------------------------+
| View: Monitoring                            |
| Scope / filters / sort                      |
|                                             |
| Display                                     |
| Layout                  [Grid v]            |
| Two rows for monitoring more conversations. |
| Auto-hide composer                [On]      |
| Hover or focus a chat to reply.             |
| Touch screens keep the composer visible.   |
| Maximum chats               [5]             |
| Counts chats across both rows.              |
|                                             |
| [Discard] [Save as]                   [Save] |
+---------------------------------------------+

The existing editor owns one scroll region and Save/Discard. On short windows, the Layout helper shows Grid needs more height. Showing Columns.; the stored choice stays Grid. A failed sync uses the current error/retry surface and rolls back all view fields together. Empty selected scope still disables Save for its existing reason. Phone shows the same fields on the editor page inside its one inset drawer and explains that layout applies on larger screens.

Covers AC-UI-THREADS-SAVED-VIEWS-005.1 through .7 and AC-UI-THREADS-DECK-004.7. Rendered check: threads-display-settings.spec.ts and its mobile companion.

UI-05: Phone, visible composer and keyboard

Entry: Threads on a phone with auto-hide enabled.

READING                       COMPOSING
+-------------------------+   +-------------------------+
| Threads / Monitoring 2/6|   | Threads / Monitoring 2/6|
| Task B v         [Open] |   | Task B v         [Open] |
| status        [Agent v] |   | status        [Agent v] |
|                         |   | transcript scrolls      |
| one live conversation   |   |                         |
| transcript scrolls      |   | [Draft message        ] |
| [Message editor       ] |   | [tools]          [Send] |
| [tools]          [Send] |   |                         |
+-------------------------+   +-------------------------+
                              | software keyboard       |
                              +-------------------------+

Swipe or the task-title picker changes the single conversation. Display stays inside the existing view drawer. The normal composer stays visible even when auto-hide is saved; no Reply or Collapse control is needed on touch screens. Touch targets are at least 44px; fixed controls clear safe areas. The available app height responds to the keyboard. Test keyboard-like height reduction in emulation; a physical mobile-keyboard smoke check remains distinct from emulation evidence.

Covers AC-UI-THREADS-DECK-004.6, AC-UI-THREADS-DECK-005.3/.7/.8/.10, and AC-UI-THREADS-SAVED-VIEWS-005.5. Rendered check: both new mobile specs plus existing swipe/picker tests.

Tests

The table records the planned test contracts; each work order's Results records the actual targets, coverage, and counts. Paths without a prefix are relative to apps/web.

Acceptance criteria	Test file and case
AC-UI-THREADS-SAVED-VIEWS-005.1/.7	Backend internal/user/store/thread_views_test.go: TestThreadPresentationStoredDefaults; service thread_views_test.go: TestThreadPresentationValidation; DTO thread_views_test.go: null/omitted draft decoding.
AC-UI-THREADS-SAVED-VIEWS-005.1/.2/.7	lib/state/slices/ui/thread-view-wire.test.ts: “normalizes presentation independently”; backend handlers/store tests: view/draft JSON round trip, restart read, existing settings payload delivery.
AC-UI-THREADS-SAVED-VIEWS-005.2/.3	lib/state/slices/ui/thread-view-actions.test.ts: “retains presentation across every draft operation” and “rolls back rapid presentation edits”; lib/threads/thread-view-query.test.ts: “presentation changes preserve query fingerprint and admission”.
AC-UI-THREADS-SAVED-VIEWS-005.4/.5/.6	components/threads/threads-view-controls.test.tsx and threads-view-controls-recovery.test.tsx: shared draft entry, default/limit labels, short-height and phone descriptions, recovery.
AC-UI-THREADS-DECK-004.1/.2/.6/.7	New components/threads/thread-layout.test.ts: “derives rows for zero, one, odd and many tasks”, “preserves requested grid across phone and height fallbacks”.
AC-UI-THREADS-DECK-004.3/.4/.8	components/threads/threads-board.test.tsx, new use-thread-selection-recovery.test.tsx, and app/threads/threads-page-client.test.tsx: “retains a lower-row reader across layout and membership changes”; preserve existing Open task/session/deep-link assertions.
AC-UI-THREADS-DECK-004.5/.6	components/threads/thread-column-activation.test.tsx: “details both visible grid rows”, “ignores old-layout observer callbacks”, “keeps phone detail singular after a grid transition”; existing 30-shell budget.
AC-UI-THREADS-DECK-005.1/.2/.3/.4/.7/.10	New components/task/chat/use-composer-disclosure.test.ts: dwell/exit, focus/draft/portal/host-operation holds, manual-collapse override, touch fallback, session/timer cleanup using fake timers.
AC-UI-THREADS-DECK-005.4/.5/.9/.10	components/task/chat/chat-input-container.test.tsx, use-chat-input-container.test.ts, and new composer-disclosure.test.tsx: single mounted editor, interaction reporting, action/recovery force-open, cancellation pending, unchanged submission and session draft restoration.
AC-UI-THREADS-DECK-005.6/.8	components/task/chat/transcript-viewport-resize.test.ts, transcript-auto-scroll.test.ts, and clamped-scroll-restore.test.ts: footer resize retains follow/reading state; browser geometry proves rendered anchoring and bounded footer reachability.

E2E tests

Use the existing primary-session creation pattern from threads-view.spec.ts and traffic capture in e2e/helpers/ws-traffic.ts. Do not seed non-primary sessions and assume they enter Threads. Factor shared setup into tests/task/threads-presentation-helpers.ts only where reused. Capture/restore user settings after each scenario, including failures.

File / project	Flow and criteria
threads-layouts.spec.ts / chromium	Compare three columns with six visible grid tiles on a sufficiently wide/tall viewport; assert equal rows, minimum width, odd/single/empty cases, lower-row session/deep-link/removal, height fallback and restore, bounded selected subscriptions, and no reorder on a reply. AC-UI-THREADS-DECK-004.1 through .8; AC-UI-THREADS-SAVED-VIEWS-005.3/.6. Include a coarse-pointer tablet and 767/768px boundary cases.
threads-composer-disclosure.spec.ts / chromium	CI-only collapse; hover reveal/hide without focus or sends; keyboard tile reveal; nonempty/attachment drafts; portaled model/context/menu interaction; successful and failed send; required questions/permissions/recovery; revealed Stop; unchanged sibling bounds; history/bottom scroll with and without auto-follow; offscreen restore and normal task-page composer. AC-UI-THREADS-DECK-005.1 through .6/.9/.10.
mobile-threads-composer-disclosure.spec.ts / mobile-chrome	Normal composer remains visible with auto-hide saved; type/send, attachment/menu and required action; swipe/picker away and back; keep one detail stream, 44px hit targets, safe-area/short-height containment, and no document overflow. AC-UI-THREADS-DECK-004.6 and AC-UI-THREADS-DECK-005.3 through .10.
threads-display-settings.spec.ts / chromium	Change through UI, save/reload, switch views, duplicate/Save as/discard, observe a second client, reject a write/retry, keep ordering, and respect a five-chat cap in Grid. AC-UI-THREADS-SAVED-VIEWS-005.1 through .7.
mobile-threads-display-settings.spec.ts / mobile-chrome	Open the existing drawer, change/save both settings, keep one phone chat, restore Grid on desktop, contain long translated copy, and expose save recovery in the same drawer. AC-UI-THREADS-SAVED-VIEWS-005.2/.4/.5/.6.

Use explicit causal waits and finite-animation settling; the intentional 150/300ms disclosure timing can use categorized dwell for negative browser assertions and fake timers in unit tests. Seed a transcript longer than the viewport for anchoring proof. A DOM-visible message is not evidence that its screen position stayed fixed.

Do not add a new runtime instance, demo, or screenshots to this design turn. Implementation captures test screenshots for comparing UI-01 through UI-05. A physical-device keyboard check should be recorded if available; do not label a resized emulated viewport as a hardware keyboard test.

Work orders

• Task 01: Persist presentation preferences
• Task 02: Render the conversation grid
• Task 03: Disclose the existing composer
• Task 04: Expose Display settings
• Task 05: Document Threads presentation
• Task 06: Polish Threads presentation

Execute 01 -> 02 -> 03 -> 04 -> 05 sequentially in the primary session after the user requests implementation. Work orders share state types, tile composition, locales, and browser fixtures. No wave authorizes delegation. Each work order owns RED/GREEN evidence and its exact verification commands; new source/test paths are explicitly identified as planned.

Verification results

Design validation (2026-09-11):

• python3 scripts/lint-spec-files.test.py: 36 tests passed.
• python3 scripts/lint-spec-files.py --all: all specification files passed.
• git diff --check -- docs/specs docs/decisions docs/plans/threads-layouts: passed. The status inventory includes this manifest and all five new orders.
• Read-only package audit: five pending orders, 42 valid criterion references, local links/anchors, dependency/design paths, and 38 verification commands checked. Command targets resolve to existing files or explicitly planned additions; pnpm scripts and working directories exist.

The design turn changed only requirements, designs, and plans. Implementation began after the user's later approval; current product evidence is recorded in each work order below. No commit or publication is requested.

Implementation completed on 2026-09-11. Every work order passed its checks. Task 01 passed user-settings Go packages, race-enabled persistence conformance, SQL guard, 160 unique targeted frontend tests, typecheck, targeted lint, and whitespace checks. Detailed RED/GREEN and commands are in its Results section. Task 02 passed 62 focused unit tests, typecheck, targeted lint, localization, and its final rebuilt browser suites: 20 desktop and 10 phone/tablet tests with retries disabled. Its Results include the native boot correction, responsive reader/deep-link regressions, and rendered checks. Task 03 passed its scoped composer/scroll tests and 6 desktop/2 mobile browser cases. Task 04 passed 102 focused unit tests, typecheck, scoped lint, localization, and final rebuilt integration: 23 desktop and 19 mobile browser tests, with retries disabled. These integrated suites overlap earlier cases; their counts are not added to earlier totals. Task 05 updated the public how-to and scoped ownership notes; public-doc validators (46 pages), 36 spec linter tests, complete specification lint, and diff checks passed.

The paired requirements/designs are active/current and this plan is implemented. Work-order Results hold the command and RED/GREEN details. No physical-device keyboard verification was available. No commit, push, publication, deployment, or delegated work occurred.

Implementation checkpoint

The user resolved the plugin question: the whole composer hides, leaving only the CI popover. This supersedes the earlier compact Reply/Stop proposal. Plugin instances and their existing capability stay mounted while hidden; opaque plugin activity alone has no keep-open guarantee. No plugin activity API or external repository change is needed. See Task 03.

To retain keyboard and touch access without another persistent control, the task tile receives keyboard focus to reveal, and phone/coarse-pointer layouts keep the normal composer. The latter is an explicitly communicated implementation assumption, not a separate user answer. Composer disclosure passed its scoped unit, type, lint, localization, and rebuilt desktop/phone browser checks. Display controls passed 102 focused unit tests, typecheck, targeted lint, localization, and the final integrated browser runs: 23 desktop and 19 mobile tests, with retries disabled. Final public documentation and its validators are complete. The package is ready for user review of the local changes.

Polish verification (2026-09-12)

Task 06 passed 137 scoped unit tests, typecheck, scoped ESLint, localization, and final rebuilt browser runs: 10 desktop and 4 mobile tests, with retries disabled and strict WebSocket accounting. Screenshots were inspected for the desktop grid and phone configurator; the phone remains the same single-chat composition. Public-doc validation (46 pages), its test, 36 spec-linter tests, complete spec lint, and whitespace checks passed. Detailed RED/GREEN evidence and test-harness corrections are in Task 06 Results.

The existing isolated playground on :48431 now serves the same verified web build. Only its frontend assets changed: no backend restart, reseed, view/draft reset, or operation against :9998. Old hashed assets remain available for already-open tabs; refresh loads the new interface. No commit or publication was requested.

PR review remediation (2026-09-12)

PR #3626 review fixes preserve the existing feature boundary:

• Saved-view selection cannot discard an unresolved draft. Desktop and touch lists explain that Save or Discard in View settings is required first.
• Replacement viewport observers measure current geometry before their first delivery, retaining visible chats and retiring offscreen subscriptions.
• The Grid height explanation stays in Display settings so it cannot consume the allocation being measured. Wheel interaction records the reader anchor.
• Deep-link scrolling honors reduced motion. Chinese hidden-chat wording and mobile locator/overflow assertions now describe the actual chat unit/target.
• Base-branch integration preserves native in-drawer deletion confirmation.
• Recovery skips zero-offset scroll assignments, which otherwise cancel an in-flight phone deep-link scroll during the first size measurement.

Regression evidence: six initial draft/observer unit failures, two board unit failures, two touch draft browser failures, and the near-threshold Grid browser failure reproduced before their fixes. Final focused unit coverage is 123 passing tests across these ten files:

cd apps/web
pnpm exec vitest run components/threads/threads-board.test.tsx components/threads/thread-column-activation.test.tsx components/threads/use-thread-selection-recovery.test.tsx components/threads/threads-view-controls.test.tsx components/threads/threads-view-controls-recovery.test.tsx components/threads/threads-view-editor-actions.test.tsx components/threads/threads-view-editor-utils.test.ts lib/state/slices/ui/thread-view-actions.test.ts lib/threads/thread-view-query.test.ts app/threads/threads-page-client.test.tsx
pnpm run typecheck
pnpm run i18n:check
pnpm run i18n:ratchet

Typecheck, localization, and scoped ESLint passed. Browser verification uses one worker, strict WebSocket accounting, and retries disabled:

cd apps/web
pnpm e2e:run --project chromium tests/task/threads-display-settings.spec.ts tests/task/threads-layouts.spec.ts tests/task/threads-composer-disclosure.spec.ts tests/task/threads-view.spec.ts -- --retries=0
pnpm e2e:run --no-build --project mobile-chrome tests/task/mobile-threads-display-settings.spec.ts tests/task/mobile-threads-composer-disclosure.spec.ts tests/task/mobile-threads-view.spec.ts tests/task/mobile-threads-swipe.spec.ts tests/task/mobile-threads-task-actions.spec.ts -- --retries=0

The original mobile sequence reproduced an initial deep-link failure twice. Three instrumented reproductions confirmed a zero-offset recovery assignment interrupting the browser scroll. A focused regression test failed before the guard and passed afterward. All three rebuilt browser reproductions passed (56.3 seconds), then the original 21-test mobile sequence passed in 3.8 minutes with diagnostics removed. The desktop command passed 24 tests before that final no-op-scroll correction; the focused recovery tests and complete mobile rerun cover the correction. Fresh post-commit desktop/capture and remote CI evidence will be recorded in the external task plan, without another code change.

Documentation validation from the repository root passed:

node --test scripts/validate-public-docs.test.mjs
node scripts/validate-public-docs.mjs
python3 scripts/lint-spec-files.test.py
python3 scripts/lint-spec-files.py --all
python3 scripts/lint-harness-files.test.py
python3 .github/scripts/lint-harness-files.py --all

This validated 46 public pages, 36 specification-linter tests, and 19 harness linter tests. Public guidance and the saved-view/deck contracts reflect these fixes. Review/CI completion is tracked externally for the pushed head, not inferred from local results. Physical-device checks remain unavailable.

PR CI remediation (2026-09-12)

The merge of main 140ef39c preserves immediate archive removal, raw focus request identity, fallback activation, and Grid reflow. Combined board tests retain both behavior sets; shared setup is extracted to satisfy the file-size limit without dropping assertions. The 145-test board/view/archive suite and 32-scenario desktop integration run passed before the composer correction.

The shared composer previously remounted when the first queue row appeared: the collapse action changed a keyed input child into an unkeyed child array. Moving that action outside QueueAffordance preserves input identity, focus, submitted-draft clearing, and normal task/Quick Chat/mobile hosts. The new identity regression failed before this one-line placement correction; all 111 composer unit tests passed afterward:

cd apps/web
pnpm exec vitest run components/task/chat/chat-input-area.test.tsx components/task/chat/chat-input-area.test.ts components/task/chat/queued-ghost-list.test.tsx components/task/chat/use-chat-input-state.test.ts components/task/chat/use-chat-input-container.test.ts components/task/chat/composer-disclosure.test.tsx components/task/chat/use-composer-disclosure.test.ts

A fresh managed build passed 15 desktop scenarios covering every failed queue, startup/resume, and plan-comment assertion plus all five setup-script cases. The 29-scenario mobile run passed all failed phone cases, both mobile flaky scenarios, and Threads display/disclosure/navigation/archive flows. Both runs used one worker, strict WebSocket accounting, and no retries. Existing describe-level retry overrides were temporarily set to zero for these runs, then restored without changing their committed policy.

Two test-only timing corrections address additional CI evidence: the mobile branch-picker test waits for the previous popover's exit before matching the next list, and mention recency clears through editor transactions before retyping a dismissed trigger. The latter matches the desktop correction landed in #3594. Both passed three repeats under a two-CPU limit (six tests, 42.9 seconds), with retries disabled:

cd apps/web
taskset -c 0,1 pnpm e2e:run --no-build --project mobile-chrome tests/chat/mobile-mention-recency.spec.ts tests/settings/mobile-repository-branch-policies.spec.ts -- --repeat-each=3 --retries=0

The setup-streaming failure did not reproduce in three additional two-CPU repeats immediately after its preceding CI stream-budget scenario (six tests, 57.2 seconds). No setup production or test change was made:

cd apps/web
taskset -c 0,1 pnpm e2e:run --no-build --project chromium tests/session/session-stream-budget.spec.ts tests/session/setup-script-progress.spec.ts:32 -- --repeat-each=3 --retries=0

Claude's two clarification suggestions are reflected in comments: the footer reserve is an 80px transcript floor below a separate header, and strict JSON decoding rejects invalid explicit zero defaults while service validation owns unknown layout strings. No validation boundary or product behavior changes. Typecheck, scoped zero-warning lint, formatting, i18n ratchet, and specification lint passed. Exact-head post-commit checks, additional ordering probes, media, later-base synthetic integration, and remote CI/review outcomes are recorded in the external task plan; local passes do not establish remote CI success.

Process-stop CI test correction (2026-09-12)

The backend API shard exposed a pre-existing test timing assumption: TestHandleStopProcess_RetiresRunningProcess checked for removal immediately after Stop returned. The runner's force-kill path acknowledges signalling before its wait goroutine retires the process. Existing runner tests already wait for that observable completion boundary.

The test now uses its existing awaitProcessRetired helper to wait for HTTP 404 before asserting an empty list. Production lifecycle behavior, timeout values, public contracts, and UI are unchanged; no public-doc or specification change is needed. The related orphan-workspace cleanup PR does not change this API.

A temporary signal-ignoring fixture reproduced the exact CI assertions in all three race-enabled attempts. With the retirement wait it passed 20 repetitions under GOMAXPROCS=2 in 42.470 seconds. The temporary fixture was then removed. The final original fixture passed 20 repetitions in 2.014 seconds, and the complete API package passed three race-enabled runs in 64.467 seconds:

cd apps/backend
GOTOOLCHAIN=go1.26.0 GOMAXPROCS=2 go test -race ./internal/agentctl/server/api -run '^TestHandleStopProcess_RetiresRunningProcess$' -count=20
GOTOOLCHAIN=go1.26.0 GOMAXPROCS=2 go test -race ./internal/agentctl/server/api -count=3

Post-commit and exact-head remote results remain tracked in the external task plan. A locally corrected test does not clear the failed remote backend gate.

Kubernetes fixture readiness correction (2026-09-12)

The container-backed CI shard failed before in-cluster task seeding: its first settings boot request returned HTTP 503 on all three attempts. The fixture's pod probe and forwarded HTTP wait still used /health, even though platform startup deliberately reports liveness before application routes are ready. The normal backend fixture already waits on /ready.

Both fixture checks now use /ready. The E2E regression asserts application readiness immediately when the in-cluster fixture returns, before seeding. Focused unit tests exercise the rendered pod probe and the actual HTTP polling path against a controlled liveness-before-readiness transition. Both failed before the endpoint correction; all 21 Kubernetes fixture unit tests passed afterward in 10.11 seconds. Typecheck, formatting, and the E2E wait-policy lint also passed:

cd apps/web
pnpm exec vitest run e2e/scripts/kubernetes-readiness.test.ts e2e/scripts/kubernetes-fixture-policy.test.ts e2e/scripts/kubernetes-kubeconfig.test.ts e2e/scripts/kubernetes-pins.test.ts e2e/scripts/kubernetes-helpers.test.ts
pnpm run typecheck
pnpm exec prettier --check e2e/fixtures/kubernetes-tools.ts e2e/scripts/kubernetes-readiness.test.ts e2e/tests/kubernetes/kubernetes-executor.spec.ts
pnpm exec eslint --config eslint.e2e-sleeps.config.mjs e2e/fixtures/kubernetes-tools.ts e2e/tests/kubernetes/kubernetes-executor.spec.ts

The existing backend contract tests passed three race-enabled repetitions in 1.634 seconds:

cd apps/backend
GOTOOLCHAIN=go1.26.0 GOMAXPROCS=2 go test -tags fts5 -race ./internal/backendapp -run '^(TestBootstrapDelayedSuccessKeepsLivenessUntilReadiness|TestHealthHandlerBodyIncludesVersionRegardlessOfReadiness|TestReadyHandlerBodyShapesByReadiness)$' -count=3

Two isolated, single-worker local Kind attempts with retries disabled stopped before the test body: image loading exceeded its existing 180-second bound, and Docker cleanup reported that it could not receive the container exit event. The second attempt reused the fresh build and cached CI image. These setup failures do not count as behavioral regression evidence or passing E2E verification. Only the two runs' owned leftover clusters and incomplete image exports were removed; the shared Docker daemon, main instance, and user demo were not restarted. Real-cluster verification remains an exact-head CI gate, tracked in the external task plan alongside post-commit and review results.

Production liveness semantics, timeout values, and UI are unchanged. Public documentation already describes the correct boundary, so no public-doc change is needed.

Base reconciliation after CI passed (2026-09-12)

The base advanced during the final PR check and introduced equivalent fixes for the process-stop test and the forwarded Kubernetes readiness wait. The process test now matches the base exactly: its existing helper establishes HTTP 404 before the list assertion. The Kubernetes merge retains the base's readiness explanation, both /ready probes, and the focused polling and pod regressions. No production lifecycle or timeout contract changed.

Focused integration validation passed 367 tests across 34 web unit files, including Threads, composer disclosure, Kubernetes fixtures, the incoming mention Escape handling, and compact-kanban sizing helpers. Web typecheck and 20 race-enabled repetitions of the process-stop regression passed. Harness, specification, documentation catalog, and public-doc validation passed. The new documentation-coverage validator accepts all six work orders and their nine linked artifacts without an override.

Managed browser checks passed seven desktop composer scenarios and four phone disclosure/mention-recency scenarios. Both used one worker, strict WebSocket accounting, and zero retries after fresh backend, frontend, and fixture-plugin builds:

cd apps/web
GOTOOLCHAIN=go1.26.0 GOMAXPROCS=2 pnpm e2e:run --host --shards 1 --project chromium tests/task/threads-composer-disclosure.spec.ts -- --retries=0
GOTOOLCHAIN=go1.26.0 GOMAXPROCS=2 pnpm e2e:run --host --no-build --shards 1 --project mobile-chrome tests/task/mobile-threads-composer-disclosure.spec.ts tests/chat/mobile-mention-recency.spec.ts -- --retries=0

The normal merge-commit receipt and new exact-head CI/review results remain tracked in the external task plan. Prior green CI does not establish completion for the merged branch.

Risks

• A full composer can dominate a half-height tile. The grid-height fallback, bounded footer, scroll anchoring, and existing Open task action address this.
• Portaled controls, pending actions, uploads, and delayed sends need explicit interaction ownership; hover CSS alone is insufficient.
• Extra visible chats increase legitimate live rendering. Offscreen cleanup and both-row subscription assertions must remain intact.
• Old view/draft constructors can silently drop new fields. Test every clone, update, recovery path, and query fingerprint.
• Phone keyboard behavior varies across browsers. Emulation proves geometry and touch paths; record any physical-device coverage separately.