Repository navigation
feat(task-panel): show the task as one screen with one state - #1017
Merged
Merged
Conversation
astyfx
requested review from
dodonmountain and
paul-seo73
and removed request for
a team
November 19, 2026 07:52
Upwind code scan | ✅ Passed
View full scan details in Upwind Console → Scan history (2 commits)
Last updated: |
Add resolveTaskUserState, which maps the facts surfaces already hold (the live turn, pending decisions, the agent run's badge, how the last turn ended and what it produced) to exactly five user states: working, waiting, blocked, finished (with its result) and stopped (with why). resolveTurnEnding reads one turn's end in a single order (receipt, retained outcome, live error, stop reason), so two surfaces can no longer disagree about the same turn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Task panel reads top to bottom: the state on one line (from resolveTaskUserState), what the user must do only while a question, approval or sign-off waits, the result (pull request, changed files, answer) or why it stopped, the completion conditions still open, and a collapsed Details with the existing activity list, steps, subagents and outputs. Picking an earlier turn moves the title and state to that turn; a failed last turn opens Details on its record. A turn's saved record decides how it ended, so "Completed" can no longer sit beside "Turn failed". Removed from the panel: the Handoff tab and its editor (stored records and their delivery to providers stay), the List/Timeline switch, the reviewed marks and unread filter in Outputs, the per-approver approval tallies, the run header, empty metric cells, the record's "left out to keep the record small" notes, and the run badge and stage track that repeated the state above the stage list. Openers now just open the panel; the layout no longer stores a tab. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…on hold A run paused for a usage limit with a known reset, or waiting for PR checks, read as Working with no live turn: describeAgentRunStatus says waiting, but the task state dropped it. The run's waiting flag and reason now reach resolveTaskUserState, which reports a sixth state, on hold: nobody has to act, and the panel says what it waits for and when it resumes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ate's request With an earlier turn's async question still open, a live turn asking for an approval read Waiting for approval, but Review request jumped to the old question: the state followed the live turn's prompt while the request was the newest question anywhere. The panel now picks the request by the state it shows (the live turn's own prompt first, else the open card of that kind), so the state, its words and the jump target belong to one request. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ages With no live turn the panel followed the newest turn in the messages, so a task whose restored messages stopped at an earlier successful turn read Finished with that turn's answer while the history held a newer failure. It now takes the newer of the two (by the history's order, or by start time when the history has not saved the messages' turn yet), and the answer and files come only from that turn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… saved record Picking an earlier turn the history has no record of read Nothing has run, even when its message said the user aborted it. A past turn now falls back to its own message, as the current turn does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…es from For an agent run, Result counted the run's changed files (its report stage's diff) while Review changes opened only the last turn's reported edits, so a run whose earlier turns made the changes showed a count the review did not match. Review changes now opens that stage when the count is the run's, and the turn otherwise; a run diff with no start commit falls back to the turn's own count and review. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…delta While a turn streamed, the panel subscribed to the task's messages and rebuilt its state on every text delta, rescanning the whole conversation for changed files and open requests even with Details closed. While it follows a live turn it now reads no messages: the open requests come from a selector whose string value only changes when a request does (each message is read once per identity, the list once per array), the live prompt from its own primitive selector, and the result is not computed for a turn that has not ended. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The state line said only Working. It now adds, on the same line, the running
step's own label from the turn's activity ("Run command · bun test"), muted
and cut with an ellipsis; nothing when no step runs. Only the state word is
a live region, so the changing step is not read out on every update.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…istence stub The Task panel preview installs a persistence bridge with only the two turn-history reads, so the result-review client saw a bridge without listResultReviews and reported its storage unavailable: the Outputs list never loaded and five results-clarity cases failed. The browser store's read and writes are now exported from the client, and the stub serves them, so the preview reads the results it seeds as the browser build does. The results-clarity cases also needed three test fixes: Details open by themselves on a failed history (a click could close them again), an output entry keeps an empty live region of its own, and the auto-read case now passes the newest finished turn that selectAutoReviewableResults requires. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sk panel's Details The agent-run-panel and results shots opened the removed tab=progress and tab=results views, and the results shot deleted the preview's persistence bridge to reach the outputs. Both now open the one-screen Task panel's Details and capture the run's steps and the Outputs section; the preview's bridge stub serves the results. Regenerated the English and Korean images. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e on hold The run-start failure and the empty agent performance screen still sent users to the removed Progress and Outputs tabs, and the architecture, provider and guide docs named the removed Progress, Activity Timeline and Subagents tabs. They now name the Task panel and its Details. The Task panel docs and both user guides describe the new Waiting (on hold) state, the step beside Working, which turn the panel follows, and that Review changes opens the files the count names. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…gle-view e2e The Task panel preview gains older=question (an earlier turn's async question open under a live approval), run=usage-wait (a run paused for a usage reset with no live turn) and restored=1 (messages stop before the newest saved failure). The single-view e2e now checks the activity words on the Working line, the on-hold state, that Review request jumps to the live approval, and the restored-history failure; the running case looks for the step text only inside Details, since the state line now names it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The badge surfaces test mapped a running run waiting for a usage reset or PR checks to Working on the Task panel, the bug the review found. It now expects the on-hold state for a waiting run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'s start time With no live turn the panel compared the restored reply's optional startedAt with the newest saved turn. A reply without one won, so an older successful turn read Finished with its answer while the newest saved turn had failed. The saved list is now the order. A reply whose turn the list does not hold is taken as newer only when this session's own snapshot is of that turn or the conversation places it after a listed turn's reply, the two signs that the list was read before the turn began. No timestamp decides. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
astyfx
force-pushed
the
feat/task-panel-single-view
branch
from
November 19, 2026 12:08
ec62355 to
b18da9e
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.

Summary
The Task panel showed five tabs, each its own dashboard; the owner stopped reading it. All four design reviews agreed on this change (synthesis #7 and #8).
resolveTaskUserState(src/lib/task-user-state.ts) maps a task\x27s facts to working / waiting for me / blocked / finished / stopped, plus on hold for work that waits for nobody (a usage-limit reset, with when it resumes). The panel uses only this function, which removes contradictions such as "Completed" next to "Turn failed".Review and verification
Written and reviewed by different providers: each branch was implemented by one model, reviewed independently by a model of the other provider, and fixed until the reviewer's blocking findings were resolved. Lead checks on the rebased head:
bun run typecheck,bun run lint,bun run check:i18n,bun run check:doc-paths,bun run check:max-lines-ratchet,bun run build:pages,git diff --check, and every changed unit test file run one per process: all exit 0. New regression tests fail on the previous code and pass now. Not verified: the installed Electron app, live providers, real macOS sleep, real GitHub calls.Part of the design-review follow-up (synthesis priorities 1–14).
🤖 Generated with Claude Code