feat(desktop): surface parent task status in Side Conversations by hbw00111 · Pull Request #5232 · apache/maka

hbw00111 · 2026-09-12T13:23:26Z

Summary

Side Conversations now show when the parent task needs input or approval, is running, or most recently completed, failed, or was interrupted. The notice uses canonical Host execution and pending interaction kinds, with settled Turn records for sessions opened after a run ends. An idle session is not reported as completed.

All side panels share one parent observation. Terminal observation errors cancel the failed registration and resubscribe with exponential backoff from 100 ms to a 2-second cap, reset on a successful seed. Navigation and unmount cancel retries; callbacks from failed registrations are ignored. Navigation and reconnect invalidate stale reads, including lifecycle notifications delivered in one React batch. The action focuses the existing parent interaction or composer; responses still use the existing handlers.

Fixes #5230. Refs #4331 (parent-status item only).

Verification

Validated merge commit fc619651811a38cd80ac2e4694d00dabaf07fa2b against upstream 4410c3a2d19d20cfdc6b7815bdd2d72d33a7460f:

Production build, workspace typechecks, lint, formatting, Desktop/UI Knip, ASF headers, renderer architecture, E2E budget, locale checks and generated surface inventory passed.
Full Desktop Node/React suite: 2,709 passed. The coordinator separately rebuilt Desktop main and reran the parent-status, observation and Workbar controller suites successfully after restoring the merged tree from the upstream comparison.
Storybook build and render smoke: 368 stories / 397 theme renders passed, including the parent approval scenario and its focus assertion.
The previously failing side-chat-followups E2E passed both alone and in the full run. Upstream 99098aafb adds submission-readiness waits at the affected send sites; those changes are included unchanged by the merge. No assertions, timeouts or retry counts were relaxed for this PR.
Full Electron E2E: 37 passed, 1 failed. new-task-reload.spec.ts fails while archiving a Session with a live derived effect. The same failure reproduces on an upstream-only 4410c3a2d tree; it remains unfixed here. These macOS results do not establish that the hosted Linux check passes.
Conflict resolution preserves both parent-focus tests and upstream Browser activation coverage, adopts the upstream SessionUsageSummary rename, and regenerates the surface inventory. The PR still changes the same 27 files relative to upstream. Historical Base64 protocol declarations and the Biome hook/test match upstream.
Not exercised: packaged Windows/macOS behavior, prolonged real Host outages, or the complete Linux/container/release validation matrix. Hosted Linux CI passed for this head, including full Electron E2E, Storybook smoke and transcript geometry invariants.

Storybook, 1280×900, light/default. Before is the existing SideChat story from unmodified base 83aa12a29; after adds the parent sandbox approval scenario in the same Workbar frame. These are component scenarios, not live Electron screenshots.

Before	After

AI use

Generative tooling made a substantive contribution

Tool(s) and scope: Grok Build (grok-4.6, xhigh) implemented the feature, regression tests and Storybook scenario. OpenAI Codex reviewed the changes, independently reproduced a reconnect bug, verified the fixes and prepared this PR on behalf of @hbw00111. The feature commit includes both Generated-by trailers; OpenAI Codex also resolved the main conflicts, fixed terminal observer recovery, added regression coverage, and ran the revision checks listed above. Native Codex agents independently inspected the subscription contract and reviewed the fix; this is not independent human approval. Pi (Cline Pass DeepSeek V4.1 Flash, high) resolved the current upstream merge conflicts, ran local validation and the upstream E2E comparison; OpenAI Codex planned and reviewed this revision and independently reran the key regressions. The merge commit includes a Generated-by: Pi (DeepSeek V4.1 Flash) trailer.

Checklist

Tests cover the change and fail without it
Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

Yes — described under Summary above

hbw00111 · 2026-09-12T15:13:15Z

The latest upstream CI run failed in transcript-scroll-cost.spec.ts because boundaries.length was 0; paging to the first turn and the mounted-turn bound had already passed.

I reran the unchanged PR head (00fa56adffbfc5f0a75878311058a5befc07bf70) using the original CI workflow in my fork. The Desktop e2e step passed: https://github.com/hbw00111/maka/actions/runs/34700571003 . Other full-workflow steps are still running as of this note; this does not replace the upstream required check.

There is a plausible sampling race: the test disables displacement recording during wheel input and reenables it after the final animation frames. A range change completed in that interval is discarded. A controlled exercise of the existing probe reproduces zero recorded boundaries in that ordering. This supports a timing-sensitive test failure, though it does not prove the exact frame ordering in the failed CI run.

Could a maintainer rerun the failed upstream job at https://github.com/apache/maka/actions/runs/34697088741 ? GitHub rejected my rerun request because it requires repository admin rights. No assertions, retries, or code were changed for this rerun.

me2seeks

P2 — Restore the historical protocol declaration and remove unrelated hook changes

The Side Conversation parent-status implementation is sound. However, merge commit ed5a813 also changes work that is unrelated to #5230:

packages/runtime-host/protocol-compatible-changes/base64-length-allocation.json changes the existing declaration from epoch 147 to 149, although neither the declared protocol files nor the declared behavior are changed here. The upstream merge parent and current main both contain 147; the feature-branch parent contains 143. 149 is introduced only by this merge result.
scripts/biome-staged-check.mjs and its test add a rule that skips every staged .patch and .diff file. This is an independent pre-commit behavior change and is absent from both merge parents.

Please restore the declaration to 147 and remove the Biome hook/test change from this PR. If the hook behavior is desired, it needs its own issue, scope, and review.

中文说明

Side Conversation 显示主任务状态的实现本身没有问题。但合并提交 ed5a813 额外把既有 Base64 协议兼容性声明的 epoch 从 147 改成了 149：上游父提交和当前 main 都是 147，功能分支父提交是 143，只有这次合并结果出现 149。同时，它还引入了所有 .patch/.diff 文件跳过 staged Biome 检查的独立行为和测试。

请将 epoch 恢复为 147，并从本 PR 移除 Biome 钩子及测试改动；如需该钩子行为，请单独建有明确范围的 PR。

Project canonical pending interaction kinds and root execution into the companion panel, with settled-turn history fallback and lifecycle guards. Reuse the parent conversation controls and cover navigation, reconnect, and browser focus behavior. Refs apache#5230 Generated-by: Grok Build (grok-4.6, xhigh) Generated-by: OpenAI Codex

Generated-by: OpenAI Codex

Reuse the continuity fixture with idle session state and the matching host epoch so parent execution observation receives canonical interaction data. Generated-by: Grok Build (grok-4.6, xhigh)

Preserve parent conversation focus and the current Work Board callbacks and regression tests. Regenerate the surface inventory from the merged tree. Generated-by: OpenAI Codex

hqhq1025

Technical review: NO-GO (1 P2).

This change adds a parent-task execution projection and Side Conversation banner for waiting, running, terminal, and unavailable states. The data minimization and session scoping are sound, and the previous unrelated protocol/Biome diff is no longer present. One recovery path remains incomplete: a terminal observation error leaves the open Side Conversation permanently unsubscribed, so a recovered Host cannot restore the parent status without a session switch or reopening the panel.

Validation on caa2ba2580cc9caf8feb1f72733b2e7ea70426ea: build, full typecheck/lint/format, ASF headers, renderer architecture 112/112, E2E budget 38, focused tests 213/213, Storybook build and render smoke 361 stories / 390 theme renders, and hosted test passed. Desktop full suite completed with 2496 passes and 8 cancellations isolated to the unchanged MCP OAuth deadline suite, which reproduces when run alone. The head is mergeable with current main 4a42aaeab4a7435a84c7d512e81b4d803f9ee0aa.

The new tests cover normal reconnection through an existing callback, but not the production terminal-error contract where main removes the observer registration. Please add bounded resubscription and a regression requiring a second subscription after onSeedError. I did not validate packaged Windows/macOS behavior or a prolonged real Host outage.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

hqhq1025 · 2026-09-14T11:46:18Z

+      sessionId,
+      () => undefined,
+      undefined,
+      () => {


[P2] Resubscribe after an observation error

This callback only marks the projection unavailable, but an observation error is terminal for this registration. Main sends host_observation_error and immediately closes the observer state, which removes the observer ID; an initial sessions:observe rejection likewise leaves no active registration. Because this effect's dependencies do not change, a recovered Host can no longer deliver another seed, so every open Side Conversation keeps showing “Parent task status is unavailable” until the user switches Sessions or closes and reopens the side chat.

A compiled-hook probe where the first subscription called onSeedError and the second would seed a running Turn remained at subscriptionCount=1,status=unavailable after 250 ms. The current reconnect test injects a later projection through the same callback, so it does not cover this terminal path. Please cancel and retry the subscription with bounded backoff, as the App Shell observer does, and add a regression that requires a second subscribeEvents call after onSeedError.

Replace terminally failed observer registrations using capped exponential backoff, cancel retries on cleanup, and reject late callbacks. Cover synchronous failures, retry reset, session changes and history refresh with regression tests. Generated-by: OpenAI Codex

hqhq1025

Technical review: GO. I found no remaining P0-P3 issues on this head.

The previous observation-recovery P2 is fixed. A terminal observation error now invalidates and releases the failed registration, retries with bounded exponential backoff, fences late callbacks by attempt, cancels pending retries on Session change or unmount, and rereads settled history after the replacement registration. The new regression coverage exercises synchronous failure, delay capping/reset, stale callbacks, cleanup, and stale history reads through the production hook contract.

Validation on 79792f04307d505d8eb4c03bf7653758151f86a8: build:test, full typecheck/lint/format, ASF headers, renderer architecture 112/112, E2E budget 38, focused production-path tests 176/176, UI 464/464, renderer and Storybook builds, Storybook render smoke 361 stories / 390 theme renders, and the full Electron E2E suite 38/38. The exact Side Chat follow-up scenario that failed once in hosted CI also passed 10/10 additional local runs. The Desktop suite completed with 2500 passes and 8 cancellations isolated to the unchanged MCP OAuth deadline tests, which reproduce when run alone.

The hosted test check is still red from that single unreproduced Side Chat E2E failure, so it should be rerun successfully before merge. The head merges cleanly with current main ea990cab7ffa768dc1a574b539f73a83f8faabb8. I did not validate native Windows/macOS behavior or a prolonged real Host outage.

Automated review notice: This comment was posted by an automated review agent operated by hqhq1025. It is not an independent human review and does not replace one.

Merge apache/maka main 4410c3a. Conflicts resolved: - workbar-controller.test.ts: keep the parent focus tests beside the upstream live browser view activation test. - session-workbar.stories.tsx: adopt the upstream SessionUsageSummary rename and keep the SessionExecutionProjection import. - docs/astryx-surface-file-inventory.md: regenerate the inventory. Generated-by: Pi (DeepSeek V4.1 Flash)

Astro-Han

Reviewed at head fc61965181. The observation architecture is right: one sessions:observe per WorkbarSurface riding the existing registry/observer seam (no parallel subscription), pendingInteractionKinds added to the existing projection, fan-out to all side panels, stale reads fenced by historyEpoch, retry lifecycle verified (failed registration is unsubscribed before backoff retry, 100ms→2s capped, reset on seed; callbacks from failed registrations ignored; session deletion does not cause a retry storm). Pending kinds → live rootTurn → settled listTurns precedence is correct, idle correctly renders nothing, and only interaction kinds cross the wire — no prompt content leaks.

One heads-up on the still-open review thread: the resubscribe-after-observation-error concern is addressed at this head — the hook cancels the failed registration and resubscribes with bounded backoff, with direct tests. The thread just isn't marked resolved.

Findings: one P2 on the WorkHub surface (inline), plus small items:

P3 — a failed listTurns read never retries (inline): latestTurnRead:'failed' maps to unavailable and nothing schedules a reread until a projection migration bumps historyEpoch, so a transient read failure can leave the warning banner up indefinitely on a settled parent.
P3 — "status unavailable" warning flashes on every mount (inline): execution === undefined (pre-seed) maps to unavailable, so every side-chat open briefly renders a failure banner that isn't one.
P3 — hasSideChat gates on the global panel topology, not visible tabs (workbar-surface.tsx:434-437): positionedTabs retains side-chat tabs belonging to other sessions, so the hook subscribes whenever any side-chat tab exists anywhere — an extra registration on the viewed session with no visible panel for it. Derive from visiblePanels (or the already-filtered quotes).
P3 — a second full observation: the hook's sessions:observe runs parallel to App Shell's existing one, and each mount/retry re-seeds the whole active transcript over IPC only to drop every event (() => undefined handler). Either port the App-Shell projection into the workbar or add an execution-only observe option.
P3 — accessibility: the notice has no role="status"/aria-live, so running→waiting→failed transitions aren't announced (chat-composer-region.tsx:312 has the precedent). Also permission kind maps to waiting_approval, but no respondable permission control exists on the desktop surface — the action then falls back to focusing the composer, which is fine, but the copy implies a button that isn't there.

Manual checks a reviewer cannot do from the diff: first-paint flash, the focus jump landing on the right control in both AppShell and WorkHub, banner wrap at the 320px panel floor, and reconnect → correct terminal status. The story SideChatParentWaitingApproval exists — please confirm light/dark captures ship in the body.

AI assistance: delegated reviewers audited the host observation path and the renderer lifecycle independently; I verified the load-bearing claims against the head source myself.

中文

Head fc61965181。观测架构是对的：每个 WorkbarSurface 一条 sessions:observe，走既有 registry/observer 缝（没有并行订阅），pendingInteractionKinds 加在既有投影上，所有 side panel 共享一份；stale read 有 historyEpoch 防；重试生命周期核验过（失败注册先退订再退避重订阅，100ms→2s 封顶，seed 后重置，失败注册的回调被忽略，会话删除不会引发重试风暴）。优先级待处理 kind → 活跃 rootTurn → settled listTurns 正确；idle 不渲染；wire 上只有 kind 没有内容泄漏。

那个还开着的评审 thread（观测错误后重订阅）在当前 head 其实已经修了——hook 会取消失败注册并按界退避重订阅，有直接测试，只是没人点 resolve。

发现：一个 P2 在行内（WorkHub 表面上"前往主对话"按钮静默失效），小项：P3 listTurns 读失败不重试导致警告常驻；P3 挂载瞬间闪 "unavailable" 横幅（未 seed 时应渲染 null 而非警告）；P3 hasSideChat 按全局面板拓扑判定会多挂一条观察；P3 第二条完整观察每次挂载把整段活跃 transcript 过 IPC 推来又丢弃；P3 notice 缺 role="status"，且 permission kind 在桌面端没有可应答控件。

Astro-Han · 2026-09-16T09:50:29Z

 }

+const PARENT_INTERACTION_FOCUSABLE =
+  '.maka-composer-interaction-slot button, .maka-composer-interaction-slot [href], .maka-composer-interaction-slot input, .maka-composer-interaction-slot textarea, .maka-composer-interaction-slot [contenteditable="true"]';


P2 — "Go to parent conversation" silently does nothing on the WorkHub surface, exactly when it matters. Both selectors require the target inside .maka-composer-interaction-slot or a visible .maka-composer, but WorkHub renders FormInteractionPrompt/UserQuestionPrompt directly inside .workHubComposerSurface (workhub-root.tsx:311-314) and hides its .maka-composer while a question/form is active (:315 hidden) — closest('[hidden]') then excludes it, so there is no focus target and the click is a silent no-op. The notice says the parent is waiting for input and the button does nothing. Smallest fix: search from the notice's own .maka-detail-with-artifacts ancestor instead of a global .mainColumn, and give prompt containers a semantic contract attribute (e.g. data-maka-contract="session-interaction") both hosts share, rather than depending on composer internals.

Astro-Han · 2026-09-16T09:50:29Z

+  readonly latestTurnRead: ParentTaskLatestTurnRead;
+}): ParentTaskStatusKind {
+  const execution = input.execution;
+  if (!execution?.available) return 'unavailable';


P3 — pre-seed undefined is not the same as unavailable. Before the first projection arrives, execution is undefined and this maps to unavailable, rendering a warning banner for one IPC round-trip on every mount/session switch — a failure claim that isn't one. Return null while execution === undefined; real losses still arrive as available: false.

Astro-Han · 2026-09-16T09:50:29Z

+  }
+
+  const latest = input.latestTurnRead;
+  if (latest.status === 'pending' || latest.status === 'failed') return 'unavailable';


P3 — a failed listTurns read never retries. latestTurnRead:'failed' maps to unavailable, and shouldInvalidateHistoryRead only refires on availability loss or live-root transitions — a settled/idle parent may produce neither, so one transient persistence_failed leaves the warning up indefinitely. Treat available && !rootTurn && latestTurnRead.status === 'failed' as a refresh trigger in applyExecution (one condition + epoch bump; frames on a settled session are rare, so it stays bounded).

me2seeks

Reviewed exact head fc619651811a38cd80ac2e4694d00dabaf07fa2b. The unrelated protocol-epoch and staged-Biome changes from my earlier review are gone, and the observation/resubscription architecture is sound. One user-facing path remains blocking: on the WorkHub surface, "Open parent conversation" cannot focus the question/form that caused the parent to wait. I also recorded two non-blocking recovery/display gaps inline.

me2seeks · 2026-09-16T10:57:23Z

 }

+const PARENT_INTERACTION_FOCUSABLE =
+  '.maka-composer-interaction-slot button, .maka-composer-interaction-slot [href], .maka-composer-interaction-slot input, .maka-composer-interaction-slot textarea, .maka-composer-interaction-slot [contenteditable="true"]';


P2 — "Open parent conversation" is a silent no-op on the WorkHub surface. This search only recognizes controls below .maka-composer-interaction-slot, then falls back to a visible .maka-composer. WorkHub renders FormInteractionPrompt and UserQuestionPrompt directly under .workHubComposerSurface, while .workHubComposerContent (and therefore its composer) is hidden whenever either prompt is active. In the exact state where this notice says the parent is waiting for input, neither selector can return a usable target, so clicking the action does nothing. The current focus tests construct the App Shell wrapper and do not exercise WorkHub's real DOM.

me2seeks · 2026-09-16T10:57:23Z

+  readonly latestTurnRead: ParentTaskLatestTurnRead;
+}): ParentTaskStatusKind {
+  const execution = input.execution;
+  if (!execution?.available) return 'unavailable';


P3 — The normal pre-seed state is rendered as an outage. execution is undefined before the first observation seed arrives, but this maps it to unavailable, so every mount/session switch can briefly show "Parent task status is unavailable" even though no read has failed. A missing pre-seed projection should render no status; actual observation loss already arrives as available: false.

me2seeks · 2026-09-16T10:57:23Z

+  }
+
+  const latest = input.latestTurnRead;
+  if (latest.status === 'pending' || latest.status === 'failed') return 'unavailable';


P3 — A transient settled-history read failure can leave the unavailable banner indefinitely. After listTurns fails, latestTurnRead: 'failed' maps to unavailable, but the invalidation logic only retries on availability loss or live-root transitions. An idle/settled parent may produce neither, so there is no automatic reread until some unrelated projection change.

github-actions Bot added the effort/XL Under 2500 readable lines label Sep 12, 2026

me2seeks requested changes Sep 13, 2026

View reviewed changes

hbw00111 force-pushed the feat/side-chat-parent-status branch from ed5a813 to 35bde23 Compare September 13, 2026 07:31

hbw00111 and others added 5 commits September 13, 2026 15:34

chore(desktop): register parent status notice in surface inventory

14d92f4

Generated-by: OpenAI Codex

test(desktop): provide complete runtime host candidate snapshots

7aeff56

Reuse the continuity fixture with idle session state and the matching host epoch so parent execution observation receives canonical interaction data. Generated-by: Grok Build (grok-4.6, xhigh)

chore(desktop): refresh surface inventory

55c3b3d

test(desktop): complete workhub execution story fixture

fa16fcf

hbw00111 force-pushed the feat/side-chat-parent-status branch from 35bde23 to fa16fcf Compare September 13, 2026 07:37

chore(desktop): merge main into parent status change

caa2ba2

Preserve parent conversation focus and the current Work Board callbacks and regression tests. Regenerate the surface inventory from the merged tree. Generated-by: OpenAI Codex

hqhq1025 reviewed Sep 14, 2026

View reviewed changes

Astro-Han reviewed Sep 16, 2026

View reviewed changes

me2seeks requested changes Sep 16, 2026

View reviewed changes

Navigation Menu

Uh oh!

feat(desktop): surface parent task status in Side Conversations - #5232

feat(desktop): surface parent task status in Side Conversations#5232
hbw00111 wants to merge 8 commits into
apache:mainfrom
hbw00111:feat/side-chat-parent-status

hbw00111 commented Sep 12, 2026 •

edited

Loading

Uh oh!

hbw00111 commented Sep 12, 2026

Uh oh!

me2seeks left a comment

Uh oh!

hqhq1025 left a comment

Uh oh!

hqhq1025 Sep 14, 2026

Uh oh!

hqhq1025 left a comment

Uh oh!

Astro-Han left a comment

Uh oh!

Astro-Han Sep 16, 2026

Uh oh!

Astro-Han Sep 16, 2026

Uh oh!

Astro-Han Sep 16, 2026

Uh oh!

me2seeks left a comment

Uh oh!

me2seeks Sep 16, 2026

Uh oh!

me2seeks Sep 16, 2026

Uh oh!

me2seeks Sep 16, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

hbw00111 commented Sep 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Verification

AI use

Checklist

Uh oh!

hbw00111 commented Sep 12, 2026

Uh oh!

me2seeks left a comment

Choose a reason for hiding this comment

P2 — Restore the historical protocol declaration and remove unrelated hook changes

Uh oh!

hqhq1025 left a comment

Choose a reason for hiding this comment

Uh oh!

hqhq1025 Sep 14, 2026

Choose a reason for hiding this comment

Uh oh!

hqhq1025 left a comment

Choose a reason for hiding this comment

Uh oh!

Astro-Han left a comment

Choose a reason for hiding this comment

Uh oh!

Astro-Han Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

Astro-Han Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

Astro-Han Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

me2seeks left a comment

Choose a reason for hiding this comment

Uh oh!

me2seeks Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

me2seeks Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

me2seeks Sep 16, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

hbw00111 commented Sep 12, 2026 •

edited

Loading