Skip to content

fix(host-service): treat a check run that failed to start as a failure - #8963

Merged
Kitenite merged 1 commit into
ParadevAI:mainfrom
akshatcoder-hash:fix/startup-failure-check-status
Oct 30, 2026
Merged

Kitenite merged 1 commit into
ParadevAI:mainfrom
akshatcoder-hash:fix/startup-failure-check-status

Conversation

@akshatcoder-hash

@akshatcoder-hash akshatcoder-hash commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Fixes #8962

Problem

mapCheckRunStatus in pull-request-mappers.ts has no STARTUP_FAILURE case. A check run that GitHub reports as COMPLETED with that conclusion falls to the default branch and comes out as pending, so computeChecksStatus keeps the pull request pending. The workflow never started, so GitHub shows that check run as a failure, and the mobile client already maps this conclusion to failed.

Changes

  • Map STARTUP_FAILURE to failure in mapCheckRunStatus, next to TIMED_OUT and ACTION_REQUIRED.
  • Add two tests: the check status from parseCheckContexts, and the aggregate from parseCheckContexts plus computeChecksStatus.
  • Leave mapStatusContextState and STALE untouched.

Tests

  • pull-request-mappers.test.ts before the fix: 12 pass, 2 fail, both new tests. After the fix: 14 pass, 0 fail.
  • cd packages/host-service && bun test --pass-with-no-tests --parallel with bun 1.4.3: 2683 pass, 8 todo, 3 fail. The three failures are the terminal router integration tests that need a real pty daemon, and the same three fail on main.
  • A parallel run also times out at the 5 s limit on one or two unrelated tests now and then. Those tests pass in other runs and on main, so the extra timeout is run ordering, not this change. A run just now had only the three pty daemon failures.
  • biome check and biome format on the changed files report no issues. bun run typecheck is red on main because of unrelated @superset/trpc errors, and this change adds no new errors.

Checklist

  • PR title follows conventional commits (type(scope): subject)
  • bun run lint and bun run typecheck pass (CI fails on lint warnings too). Lint passes on the changed files. Typecheck fails on main in @superset/trpc and is unrelated to this change.
  • "Allow edits from maintainers" is checked on fork PRs
  • I have signed the CLA (the bot asks on your first PR)

mapCheckRunStatus had no STARTUP_FAILURE case, so a COMPLETED check run
whose workflow never started fell through to default and reported pending.
computeChecksStatus then kept the pull request pending forever.

Map it to failure next to TIMED_OUT and ACTION_REQUIRED, and cover both
the per-check status and the aggregate in tests.
@github-actions

github-actions Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

All contributors have signed the CLA ✍️ ✅
Posted by the CLA Assistant Lite bot.

@coderabbitai

coderabbitai Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

Review in Change Stack →

No actionable comments were generated in the recent review.

Recent review info
Run configuration
  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: 5aed3deb-7bbc-42fc-9e7b-d2a92a841f01
Commits

Reviewing files that changed from the base of the PR and between 3c59928 and 91bc4d1.

Files selected for processing: 2
  • packages/host-service/src/runtime/pull-requests/utils/pull-request-mappers/pull-request-mappers.test.ts
  • packages/host-service/src/runtime/pull-requests/utils/pull-request-mappers/pull-request-mappers.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 9 remain after this review.

Walkthrough

The check-run mapper now treats GitHub’s STARTUP_FAILURE conclusion as failure. Tests cover the parsed check status and aggregate pull request checks status.

Changes: Check-run failure mapping

Layer / File(s) Summary
Map startup failures and verify aggregate status
packages/host-service/src/runtime/pull-requests/utils/pull-request-mappers/pull-request-mappers.ts, packages/host-service/src/runtime/pull-requests/utils/pull-request-mappers/pull-request-mappers.test.ts
STARTUP_FAILURE maps to failure. Tests verify the parsed check status and aggregate status.

Priority: Normal

Estimated code review effort: 2 (Simple) | ~8 minutes

Change: Bug fix · Severity of issue fixed: Medium

Suggested reviewers: kitenite

Merge Risk: Minimal · up to 91bc4

No actionable merge-blocking risk is evident in this change.

Pre-merge checks: Passed 4 · Failed 1

Failed Failed checks: 1 warning

Check name Status Explanation Resolution
Docstring Coverage Warning Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Documentation-only edits and deletions within survi… Write docstrings for the functions missing them to satisfy the coverage threshold.
Passed Passed checks: 4 passed
Check name Status Explanation
Title check Passed The title "fix(host-service): treat a check run that failed to start as a failure" follows conventional commit format and states the primary change: a check run with STARTUP_FAILURE is now reported as…
Linked Issues check Passed Issue #8962 has one coding requirement: a completed check run with conclusion STARTUP_FAILURE must map to status failure, and the aggregate checks status must be failure. The reviewed head adds `case …
Out of Scope Changes check Passed The whole-PR diff against the merge base contains two files. In pull-request-mappers.ts, the change is one added case "STARTUP_FAILURE": line that returns "failure", which directly implements #8962.…
Description check Passed The description explains the problem, the code changes, test coverage, known unrelated failures, and checklist status. It uses Problem and Changes headings instead of the template's What & why heading…
Full details: Title check

Explanation

The title "fix(host-service): treat a check run that failed to start as a failure" follows conventional commit format and states the primary change: a check run with STARTUP_FAILURE is now reported as failure. The diff maps STARTUP_FAILURE to failure in mapCheckRunStatus and adds tests for it, which matches the title.

Full details: Linked Issues check

Explanation

Issue #8962 has one coding requirement: a completed check run with conclusion STARTUP_FAILURE must map to status failure, and the aggregate checks status must be failure. The reviewed head adds case "STARTUP_FAILURE": next to TIMED_OUT and ACTION_REQUIRED in mapCheckRunStatus (pull-request-mappers.ts), which returns "failure". Before this change, the value fell through to default and returned "pending". parseCheckContexts passes the check run's status and conclusion to mapCheckRunStatus, so the per-check status is now failure. computeChecksStatus returns "failure" when any check has status failure, so the aggregate is also failure. The PR adds two tests in pull-request-mappers.test.ts. One asserts parseCheckContexts returns { name: "build", status: "failure", url: null } for a STARTUP_FAILURE run. The other asserts that computeChecksStatus over that result returns "failure". These tests match the issue's expected results. The issue's reproduction steps (opening the PR in the review tab) are manual and not coding requirements. I read the code and test diff only. I did not execute the tests, so test results rely on the author's report.

Full details: Out of Scope Changes check

Explanation

The whole-PR diff against the merge base contains two files. In pull-request-mappers.ts, the change is one added case "STARTUP_FAILURE": line that returns "failure", which directly implements #8962. In pull-request-mappers.test.ts, the change adds a type import for GitHubCheckContextNode, adds a parseCheckContexts import, adds a checkRun fixture helper, and adds a describe("parseCheckContexts") block with two tests. All of these support the new tests for #8962. No other code changed. mapStatusContextState and the STALE handling are unchanged, as the PR description states and the diff confirms. No unrelated changes were found.

Full details: Description check

Explanation

The description explains the problem, the code changes, test coverage, known unrelated failures, and checklist status. It uses Problem and Changes headings instead of the template's What & why heading, but it provides the required information. The How I tested it content is complete.

Full details: Docstring Coverage

Explanation

Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Documentation-only edits and deletions within surviving functions are included; unchanged diff context and fully deleted functions are excluded. Analyzed 2 functions across 2 files.

Resolution

Write docstrings for the functions missing them to satisfy the coverage threshold.

Finishing Touches: 💡 1
🛠️ Fix failing CI checks: 💡
  • Commit to this branch
  • Create a new PR
🧪 Generate unit tests (beta)
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@greptile-apps

greptile-apps Bot commented Oct 27, 2026 •

Copy link
Copy Markdown
Contributor

RetriggerView in GreptileConfidence Score: 5/5

[Medium impact] This change appears safe to merge.

Summary

The host now treats a completed GitHub check with STARTUP_FAILURE as failed instead of pending.

  • Pull-request checks count failed workflow starts as failures.

Reviews (1) · Last reviewed commit: "fix(host-service): treat a check run tha..." · Reviewed by Greptile

@akshatcoder-hash

Copy link
Copy Markdown
Contributor Author

I have read the CLA Document and I hereby sign the CLA

@Kitenite Kitenite left a comment

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks! Verified the new tests fail without the fix and pass with it.

@Kitenite
Kitenite merged commit 9a87d57 into ParadevAI:main Oct 30, 2026
19 of 23 checks passed
Kitenite added a commit that referenced this pull request Oct 30, 2026
…#9161)

Host-service and the desktop pull request view each had their own copy of
the GitHub check run to status mapping and of the overall checks rollup.
The copies drifted: the desktop handled STARTUP_FAILURE, host-service did
not (#8962, fixed in #8963). Move both functions to
@superset/shared/pull-request-checks so a new GitHub conclusion is added
in one place.

The v2 workspace computeChecksRollup and the mobile effectiveCheckStatus
stay as they are: they count checks and treat skipped and neutral runs
differently, so sharing them would change what the UI shows.
Kitenite added a commit that referenced this pull request Nov 8, 2026
* chore: seed CLA signature file

* @tanxarx has signed the CLA in #8881

* @ozeitis has signed the CLA in #8888

* @shubhankiit has signed the CLA in #8894

* @chrisfentiman has signed the CLA in #8910

* @RohithReddy20 has signed the CLA in #8907

* @larcane97 has signed the CLA in #8922

* @moizghumann has signed the CLA in #8935

* @andyst-dev has signed the CLA in #7930

* @svenpeeters has signed the CLA in #8943

* @kenil-kanani has signed the CLA in #8940

* @akshatcoder-hash has signed the CLA in #8963

* @NilotpalK has signed the CLA in #8995

* @serhii-kucherenko has signed the CLA in #9000

* @tonydzi has signed the CLA in #9002

* @adaam2 has signed the CLA in #9038

* @tsavo-at-pieces has signed the CLA in #9059

* @ganeshmshetty has signed the CLA in #9130

* @aaronrosenthal has signed the CLA in #9135

* @alexnault has signed the CLA in #9138

* @kunalgoesbyken has signed the CLA in #9235

* @Joaco-gavernet has signed the CLA in #9236

* @theblondealex has signed the CLA in #8240

* @talarczykco has signed the CLA in #7274

* @throw99 has signed the CLA in #7906

* @lacherogwu has signed the CLA in #9310

* @KGKallasmaa has signed the CLA in #9316

* @Niltonsf has signed the CLA in #9331

* @Anmol1578 has signed the CLA in #9351

* @RyoKusnadi has signed the CLA in #9317

* @thedotmack has signed the CLA in #9414

* @abudker has signed the CLA in #9439

* @sanchitmonga22 has signed the CLA in #9443

* @Bruno-rose has signed the CLA in #9446

* @Ashad001 has signed the CLA in #9487

* @vivek-pothina has signed the CLA in #8359

* @colingreig has signed the CLA in #9489

* @anmol0b has signed the CLA in #9492

* @resteqs has signed the CLA in #9506

* @pgilmore has signed the CLA in #7878

* fix(desktop): name the DMG from the display name

The DMG file and its volume title now use DESKTOP_DISPLAY_NAME, so a
Paradev build ships Paradev-<version>-<arch>.dmg. Unset, names stay
Superset. The release keeps the Superset-<arch>.dmg alias beside the new
one, since the README, docs and download button link to it. The .app
bundle and the -mac.zip update files keep the Superset name.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
NicolaPanero added a commit to NicolaPanero/Paradev that referenced this pull request Nov 13, 2026
…d account reconnect (#34)

* feat(plugins): check GitHub setup in the environment skill (#9103)

Add a step 0 to the superset:environment skill. Before it creates an
environment, the agent checks that the organization's GitHub App is
installed and can see each repository (superset integrations github,
superset environments repos). It also offers a personal GitHub
connection, so pushes and pull requests act as the user and not as
the App.

Also tell the agent that boxes already have GitHub access for git and
gh, and renumber the steps.

* docs: clearer README introduction (#9101)

Adds a What is Paradev? section to the top of the README: what it is, who it is for, what it adds around CLI agents, platforms, the former name, and links to the comparison pages.

* feat(desktop): ask how people found Paradev during onboarding (#8534)

Adds one optional question to the Add your first project step: How did you hear about Paradev? Answers send acquisition_source_answered with a stable source key and is_ai_assistant, and set acquisition_source on the person. Skipping sends nothing. Specific AI assistant options (ChatGPT, Claude, Perplexity, Gemini, another AI assistant) because AI referrals often arrive without a referrer.

* docs(marketing): refresh Cursor, Codex app and Vibe Kanban comparison pages (#9106)

Refreshes cursor-alternative, superset-vs-cursor, codex-app-alternative, superset-vs-codex-app, vibe-kanban-alternative and superset-vs-vibe-kanban from primary sources. vibe-kanban-alternative becomes a migration page for the Vibe Kanban shutdown, with separate steps for continuing imported worktrees and starting new work. Unverifiable claims removed or softened.

* fix(chat-ui): match slash menu typography to the mention menu (#8984)

The / menu used 12px muted text while the @ menu uses 15px foreground
text. Use the mention menu's font sizes, colours, row padding and icon
size in CommandMenu.

Fixes SUPER-2962

Co-authored-by: Harshith Mullapudi <harshithmullapudi@gmail.com>

* fix(chat): match the live changes pill to the composer (#9114)

The pill used bg-muted/40 and border-border/60, which composite onto the
pane background instead of landing on the composer's bg-card and ring-border:
its fill rendered #1d1a1a against the composer's #201e1c. Its wrapper also
carried mx-4, the composer's inner text inset, so the pill lined up with the
placeholder rather than the box edge.

Verified over CDP: fill, ring colour and left edge now match the composer
exactly.

* feat(chat): a usage limit pauses the queue and offers to resume at the reset (#9109)

* feat(desktop): GitHub pull request and issue chips in sent messages

A GitHub pull request or issue pasted into a message rendered as the
same grey pill agent prose uses, with nothing behind it.

In a sent bubble it is now an 18px chip, violet for a pull request and
emerald for an issue, labelled "PR #N" or "Issue #N". Hovering it for
220 ms, or focusing it, opens a card: state, title, a plain-text
summary of the body, author, base and head branches, labels and
assignees, with a skeleton while it loads and a plain line when the
lookup fails. The click still follows the link settings. Agent prose
keeps its pill.

The card asks the host only once it opens. `github.getWorkItem` reads
one pull request or issue through the host's GitHub client into a
glance-sized shape, with the body stripped to plain text and cut at
280 characters, and remembers it for two minutes per repository, kind
and number; a failed lookup is not remembered, and a 404 answers
NOT_FOUND. The procedure is new, so released clients are unaffected.

The URL parser moved to a shared helper that the unfurl rule uses too.
Seven new strings, translated in every locale.

* feat(desktop): a shell command that only read, searched or listed says so

A `cat`, `sed -n`, `grep`, `rg`, `find -name` or `ls` the agent ran
through its shell rendered as the raw command in mono — or, for Claude,
as the agent's prose description in mono — and the run summary counted
every one of them as "ran N commands".

The row now reads like a native read or search: "Read package.json",
"Searched useQueue in src", "Listed components", with the target as a
chip and the command still in the body. A redirect is "Wrote notes.txt"
or, appending, "Edited". The classifier looks through the shell an
agent launched the command with (`/bin/zsh -lc '…'`), follows pipes and
chains to the stage that mattered, and leaves anything with
substitution, an unknown binary or a program of its own as a command.

Step counts and the run summary's icon use the same reading, so a run
of read-only greps is "4 searches" rather than "ran 4 commands". A
command row whose title is the command shows it unwrapped; a Claude
description is no longer set in mono.

Eleven new strings, translated in every locale.

* feat(chat): a usage limit pauses the queue and offers to resume at the reset

When the provider refused a turn for the account's usage, the refusal
was dropped on the floor: Claude's `rate_limit_event` was ignored by
the stream translator, the ACP agent's forwarded copy was never read,
the turn simply failed with two red lines, and anything queued was sent
straight into the same wall.

The session now carries `usageLimit` in its state: when it was hit,
which window closed and for which model, when the provider said the
window reopens, and whether the host has been asked to continue then.
Both Claude paths report it. The ACP adapter reads the agent's
`usage_update._meta` (`_claude/rateLimit`, `_claude/model`) and the
refused prompt; the stream translator remembers a refused
`rate_limit_event` and reports it with the turn that ended in error. A
refusal paid for by extra usage goes on and reports nothing; a turn
that survives one is not a limit. Any harness whose turn fails with the
limit named in its error gets the same, without a reset time.

The live session pauses the queue, holds new prompts in it, and takes
one of four answers: resume now, dismiss, resume at reset, or cancel
that. Resuming continues the cut-short work — the next queued prompt,
or a plain "Continue" when nothing is queued — and resume at reset does
the same from a host-side timer thirty seconds past the reset, so it
happens with the window closed. A plain queue resume clears the limit.

Above the composer, an amber notice says "Usage limit reached · 5-hour
limit · Sonnet 5.5 · the account", when it resets (ticking faster in
the last two minutes), Resume / Resume at reset / Resuming at reset, a
dismiss, a Switch account menu — the host's other logins for the agent
with their plan and tightest window; picking one makes it the default
and reopens the chat on it, the way a crash recovery does — and a link
to Settings → Usage. In the transcript, a failed turn is one card — an
alert icon, "Turn failed" or "Stopped by the usage limit", and the
agent's own words — rather than the same error twice in red, and an
ACP failure carries the agent's message without the RPC method.

`respondToUsageLimit` is a new procedure and `usageLimit` a new
optional field, so released clients are unaffected. Eighteen new
strings, translated in every locale.

* fix(chat): the usage limit's account button names the account

It said "Account" whatever login the chat was on. It now shows the
current login's email ("Account" until the host reports it), and the
detail line no longer repeats it.

* fix(chat): pause queued work before refusal and cancel dead-session resume

* fix(desktop): preserve queued prompts and identify the active chat account

* fix(desktop): read shell commands the way the shell does

Seven cases where a shell row said the wrong thing:

- `find … -delete` (or -exec, -ok, -fprint…) runs or removes something;
  it stays a command instead of reading as a search.
- `$(…)` and backticks inside double quotes still run; only single quotes
  stop them.
- A quoted redirect target (`> 'my notes.txt'`) is one path, with its
  quotes and escapes resolved.
- `cd /repo/src && cat index.ts` keeps the leading slash.
- Codex reports the command's working directory in `rawInput.cwd`; the
  row keeps the path as typed and the chip opens it there.
- Inside double quotes a backslash escapes only $ ` " \ and newline, so
  `grep "foo\.bar"` keeps its backslash.
- A redirect resolves against the directory current where it ran, not
  after a later `cd`.

* fix(desktop): a stderr redirect is not a shell command's output

`grep foo src 2> err.log` read as "Wrote err.log" and counted as an edit.
Only a stdout redirect (`>`, `1>`, `>>`) makes a command a write now.

German progressive shell labels read as commands ("Lese"); they are
status nouns now ("Lesen"). Indonesian "list" said "register"; Russian
"Searched" said "Found".

* fix(desktop): a redirect's target is not one of the command's arguments

After 2dd895c207, `grep foo src 2> err.log` was a search again but
searched err.log: the tokenizer read the redirect target as a path
argument. Redirections (`2> f`, `< f`, `2>&1`) are now removed from a
stage before its arguments are read; that also fixes the test added in
2dd895c207.

* fix(desktop): count a file that shell commands write once

A shell write carries no changed paths, so each call counted as its own
edited file: two writes to out.txt read "edited 2 files". The path the
command names, resolved against where it ran, now joins the set of
edited paths, and only a write with no path falls back to one per call.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(desktop): keep a shell command visible when its row reads as a read

A shell row shown as Read, Searched, Listed or Edited no longer showed
the command anywhere for agents that send no terminal block: the title
is the intent, and the body held only the description and the output.
The body now starts with the command as a "$ command" line, so a row
with no output can still be opened to see it. A terminal block already
prints the command, so the line is left out when there is one.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(chat): the usage limit owns a queue it paused, and counts down its resume

Three rough edges when a usage limit paused a chat with queued prompts:

- The queue bar said "Queue paused because you interrupted". It now says
  "Queue paused by the usage limit" when the limit paused it.
- The queue bar and the limit notice each had a Resume. While the limit
  holds the queue, only the notice offers it.
- After the reset, an armed resume waits the host's grace before it sends,
  and nothing said so. The host now reports when it will send
  (`usageLimit.resumesAtMs`, cleared on cancel), and the notice reads
  "Limit has reset · Resuming in 25s", ticking each second.

`resumesAtMs` is an optional field on a loose object, so older clients
ignore it.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(chat): a usage limit reads as a pause, in one panel above the composer

A usage limit is not an error, and the screen said it three times over:

- The transcript showed a red failure card with "Copy error". It now shows
  one muted line, "Stopped by the usage limit", like "Stopped" for an
  interruption; the notice above the composer carries the recovery.
- The limit notice and the queue each drew their own rounded tab, stacked
  with their corners showing. The notice is now the top row of the work
  drawer, with the queued prompts under a divider, so they read as one
  panel fused to the composer. With nothing queued, the notice is the
  whole panel. While the notice speaks for a queue the limit paused, the
  separate "Queue paused" bar is left out.

The notice moves from Composer to SessionView, which hands it to the
drawer; Composer only needs to know a limit holds the chat, to hide the
early "80% used" warning.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(chat): close the usage limit's review gaps

- A v2 turn that fails through a later `error` state runs the same usage
  limit check as a refused prompt (`failRefusedTurn`), so its reset time
  and the `usage_limit` kind are kept.
- LiveSession marks a failed turn `usage_limit` when the limit was reported
  for it or its error names the limit, so legacy Claude sessions and the
  text fallback show the quiet history line, not the failure card.
- `usage_update._meta: null` no longer fails the whole update.
- A throw from the agent when the timed resume sends fails the session
  instead of escaping the timer.
- A failed restart after an account switch stops holding the chat, and a
  failure with no stored session shows the recovery screen.
- `getQueue` adds each prompt's `fullContent`, hidden parts included, and
  an account switch restores that instead of the visible parts only.
- During the post-reset countdown the armed button stays and cancels.
- `queuePausedByLimit` keeps the pause reason after the notice is
  dismissed, so the bar no longer says "because you interrupted".
- Indonesian and Italian reset wording matches the usage page.

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(chat): keep the limit countdown out of the live region, and wrap its controls

- Only the notice's title is an `<output>` now. The whole notice was a
  polite live region, so the countdown, ticking every second, made screen
  readers announce the notice again and again.
- The notice row and its controls wrap. In a narrow pane the controls
  could not shrink, and the drawer's `overflow-hidden` clipped Resume and
  dismiss; they now step down to a second and third line instead.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

* fix(desktop): say when a diff is too large to expand or edit (#9100)

* fix(desktop): say when a diff is too large to expand or edit

A file past the Changes pane's 500k-character load cap failed silently:
expanding unchanged lines did nothing, and clicking a line left the file
in edit mode with Save/Cancel but no editor. After the first refused
load, the file header shows "Too large to expand" (clicking it opens the
file), later expand clicks skip the fetch, and line clicks no longer
start an edit.

* fix(desktop): mark the card that refused to load, and clear it when its diff changes

The too-large mark was keyed by path, so a staged and an unstaged copy of
the same file shared it. Key it by item, find the item by fileDiff identity,
drop the mark when the item's diff changes, and keep the text cursor off
marked lines.

* fix(desktop): skip a too-large mark when its diff was replaced during the load

* feat(desktop, cli): copy a workspace link from the sidebar and print it from the CLI (SUPER-2061) (#9112)

* feat(desktop, cli): copy a workspace link from the sidebar and print it from the CLI (SUPER-2061)

Integrations such as Linear and Slack accept only HTTP(S) links, so the
superset:// deep link cannot be pasted there. The public handoff page
(${NEXT_PUBLIC_WEB_URL}/workspaces/<id>) already exists; this makes the
link easy to get.

Desktop: the left sidebar row menu gets "Copy link" beside "Copy Workspace
ID" for every workspace type (local, remote host, cloud). It reuses
useCopyShareLink, so it copies the same URL and shows the same "Link copied"
toast as the Workspaces page and the cloud workspace screen. The copy lives
in the menu, as in V2WorkspaceContextMenu, so the test drives the real menu
and reads the clipboard. A local workspace's link opens only on a machine
that has that workspace; the desktop already shows its missing-workspace
state, so there is no new UI. The command palette and the v2 workspace view
have no "Copy Workspace ID" action, so they are unchanged.

CLI: `ws create`, `ws get` and `ws list` now return `url` (the superset://
deep link) and `webUrl` (the HTTPS handoff link), and `ws open` adds
`webUrl` beside its existing `url`, as the ticket asks. `webUrl` comes from
the CLI's SUPERSET_WEB_URL config, not a hardcoded host. Fields are only
added; no existing field is renamed or removed, and no host or cloud row
already had a `url` or `webUrl` key. `ws create` also prints the web link
on a second line of its human message, like `tasks create`. The `ws list`
table keeps its columns; the links are in --json. `superset schema` output
is byte-identical to before, since no option, argument or description
changed. The CLI reference documents the new output fields.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* feat(desktop): group the sidebar row's copy actions under a Copy submenu

The workspace row menu had grown to four flat copy items (Copy Path, Copy
Branch Name, Copy Workspace ID, Copy link), its largest group, in a menu of
12 to 16 items. The channel row menu in the same sidebar and the task menu
already put copy actions under a "Copy" submenu, so this does the same.

The submenu lists Copy link first, then Copy Branch Name, Copy Path (local
workspaces only) and Copy Workspace ID. Open in Finder stays on the top
level. Copy Path takes the clipboard icon the other Copy Path items use,
since the trigger now carries LuCopy. Every label already exists in the
catalogs, so no translation changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(cli): address review on workspace links (SUPER-2061)

- Tests compared webUrl with `${SUPERSET_WEB_URL}/workspaces/<id>`, so a
  base with a trailing slash failed them although workspaceLinks returns
  the right link. The command tests now compare with workspaceLinks(id),
  and open-url.test.ts checks the link format once by origin and path.
- The session and local-checkout create paths had no workspace id in their
  fixtures, so their links ended in /undefined unchecked. Each host create
  path now has an id and a link check.
- The CLI dev script maps SUPERSET_WEB_URL from NEXT_PUBLIC_WEB_URL when it
  is not set, as it already maps SUPERSET_API_URL. Both setup scripts write
  SUPERSET_WEB_URL; a .env copied from .env.local.example did not, and the
  dev CLI printed production links for local workspaces.
- The CLI reference says `ws open --print` prints the deep link and points
  to `ws get --field webUrl` for the HTTPS one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* feat(db): a Dev's task session points at the Paradev task it works on (#9098)

agent_sessions gains task_id (-> tasks, on delete set null) and a unique index allowing one queued or running writer session per task. A Dev's work moves onto Paradev tasks; the old task columns are dropped once no code reads them.

* feat(devs): a Dev's work is a Paradev task (#9116)

The work a Dev delegates lived in columns on agent_sessions: title,
instructions, status, dedupe key, result and pull request. Paradev tasks
already carry all of that, show up in task lists, and link to boxes, so a
Dev's work is now an ordinary task:

- start_task makes a Paradev task (through insertTask, which the task routes
  now share) or works on an existing one by taskId, and starts a task
  session on it. A reviewer works on the same task as its writer; a task has
  one active writer.
- A writer's box is linked to its task, and the task moves to the org's
  first started status when the writer starts (moveTaskToStarted, split out
  of task.start).
- The task's pull request is the one its box reported, else the one the
  writer named, and is kept in tasks.pr_url. A session's result is its last
  turn's reply rather than a copy.
- list_work, read_task, message_task and stop_task work on tasks and list the
  Dev's sessions on each; titles in turn credentials and listSessions come
  from the task.
- A task session whose box never came up is queued again for a new box,
  once, without telling the main session.

The old columns stay until a follow-up drops them.

Verified on a local stack against a production copy: a DM became task
SUPER-2998, moved to In Progress, its writer's box linked to it, opened
ParadevAI/acme-demo#7 and recorded it on the task; a reviewer session on the
same task approved it and the Dev reported back.

* refactor(db): drop the task columns agent_sessions no longer uses (#9117)

A Dev's work lives on Paradev tasks since #9116: the title and pull request are the task's, the result is the session's last turn reply, and one active writer per task replaces the dedupe key. Drops title, dedupe_key, result and pull_request_url, and the dedupe index.

* feat(desktop): Linear-style Workspaces list (views, filter chips, Display menu, virtualised rows) (#9115)

* feat(desktop): rework the Workspaces list into a Linear-style page

Views row on top (All, Needs me, Working, Cloud, plus saved views with a
Save view / Reset flow), filter chips with one "+ Filter" menu, and a
Display menu for layout, grouping (status, project, host, creator) and
ordering. Default is a flat list ordered by last agent message; grouping
and ordering persist per profile alongside the saved views (capped at 20,
sanitised on rehydration, UUID ids).

Rows: title, creator avatar, place mark (cloud / other device / offline
device), PR pill, project, then a trailing cell that shows the live status
glyph or the compact age — the old cloud tab's rule and icon set
(WaveSpinner, pulse, info mark). The list is virtualised with
@tanstack/react-virtual; the active section header stays pinned.

Deep links gain tab, status, group and order; an unknown tab lands on All
and keeps its filters. WindowControlsInset no longer reserves a right
margin when the OS draws the window controls on the left.

* fix(desktop): apply view filters on tab deep links; harden view ids

A ?tab= link now starts from the view's own filters before URL overrides,
so ?tab=working shows the Working view instead of everything. View ids are
own-property checked, so ?tab=constructor lands on All instead of
throwing on Reset. The no-cloud fallback switches the tab without dropping
the filters the link carried. Grouping controls hide in board layout,
whose columns are always the status buckets.

* fix(desktop): complete the views tablist semantics; fix four translations

Tabs carry ids, aria-controls and a roving tabIndex; the list or board is
the tabpanel labelled by the active tab; delete and New view sit beside
the tablist instead of inside it. German, Dutch and Polish "Archived
shown" read as full phrases; Japanese "Needs me" no longer reuses the
"Needs attention" label.

* fix(desktop): archived rows keep their own last section; Clear lands on All

Under no grouping, project, host and creator grouping, archived
tombstones now sit in a trailing Archived section instead of sorting among
live rows, as the store's contract says. An archived row is never shown
as live, whatever stale agent status it carries. Clear resets to the All
view rather than leaving a built-in tab selected with no filters. Deleting
a saved view asks for confirmation.

* docs(desktop): describe the Display menu's archived toggle

---------

Co-authored-by: Harshith Mullapudi <17528887+harshithmullapudi@users.noreply.github.com>

* chore(desktop): remove the "No agent" toggle from New Workspace (#9121)

The agent is picked in the composer's model picker since #8971, so the
separate opt-out pill below the composer is no longer needed. A stored
"none" choice now falls back to the first available agent; a host with
no agents still launches without one.

* fix(desktop): open the file when an edited-file row has no diff (#9122)

Rows in chat's "Edited N files" card open the file's diff on click. A file
outside the worktree (e.g. /tmp) has no diff to review, so the row was
disabled and only the small arrow opened the file. The row now falls back
to opening the file, and the stats column is part of the clickable row.

Fixes SUPER-3024

Co-authored-by: Harshith Mullapudi <17528887+harshithmullapudi@users.noreply.github.com>

* fix(marketing, shared): point the YC job board and careers links at the paradev slug (#9119)

YC renamed the company, so /companies/superset now redirects and
/companies/paradev is canonical. The job board embed and every product
surface that links to the formal listing still used the old slug.

Both slugs serve the same board today, so nothing was broken; this stops
us relying on YC's redirect.

The job descriptions the embed renders still say "Superset" in their copy.
Those live in YC Work at a Startup, not this repo.

* docs(dev-doctor): read a writer task's pull request from its Paradev task (#9120)

The task columns moved off agent_sessions (#9116, #9117), so the old query fails.

* docs(dev-doctor): search Sentry in the paradev organization (#9056)

The incidents procedure named organization superset-sh, which Sentry
answers with 404. find_organizations lists only paradev.

Co-authored-by: Claude <noreply@anthropic.com>

* fix(desktop): read the load URL from the emitting webContents (#9125)

* fix(desktop): read the load URL from the emitting webContents

When a window closes during its first load, its webContents can still emit
did-finish-load. The handler read webContents through the destroyed
BrowserWindow, whose getter throws, and the uncaught error was fatal in
the main process. Hold the webContents reference and use it directly.

Refs DESKTOP-1K2

* fix(desktop): skip main-window load handlers once the window is destroyed

With the window-loader listener no longer throwing first, the main-window
did-finish-load and did-fail-load listeners run for a window closed
mid-load and call show() on it, which throws the same error.

Refs DESKTOP-1K2

---------

Co-authored-by: superset-home <338338640+superset-home@users.noreply.github.com>

* fix(devs): a message body cannot forge another message's header (#9040)

eventText wraps every prompt a Dev's main session reads in a <message from where reply_to> block, and the Dev trusts those headers for who sent a message and where to reply. A body could close its block and open one claiming another sender and reply target. Message tags in a body are now escaped, and attribute values keep quotes and newlines as entities instead of losing them. Found by the Dev doctor.

* fix(desktop): let a manual project pick unblock the new-workspace screen (#9129)

When the URL asked for a project that never appeared in the project list
(stale link, deleted project, not offered on the cloud target), the screen
waited for it forever and blocked create with "Select a project", even after
the user picked another project in the picker. A manual pick now ends that
wait. The wait itself stays, so a just-created project that is not in the
list yet still gets selected when it arrives.

* fix(agent-worker): a Dev's main session reaches the model in production (#9131)

Every DM to a Dev in production failed with "Turn has no model credentials":
a main session has no box, so it calls the router with the worker's own key,
and that key was never deployed.

- The deploy passes AGENT_ROUTER_API_KEY to the worker, which passes it to
  the container. The secret is set on the Production environment.
- The router bills a key's request to the organization the request names, so
  the worker sends x-superset-organization-id for the turn's organization,
  which turn credentials now return.
- A missing model credential retries like a provider error instead of
  failing the turn, so a deploy gap delays answers rather than dropping the
  messages.

Checked against the production router: the key alone gets a 403 ("not set
up to use the router"); the key with the organization header answers.
SUPER-2999.

* fix(desktop): let channel message text be selected (#9132)

The renderer body sets user-select: none, and channel message bodies never
opted back in, so drag-to-select and copy did nothing.

* fix(mobile): attachment upload progress accepts BigInt byte counts (#9133)

expo-modules-core 57 encodes Swift Int64 as a JS BigInt. expo-file-system
emits upload progress as Int64 while typing it as number, so on iOS
bytesSent / totalBytes was BigInt division and Math.min threw on every
native progress event, leaving the ring at 0 until completion.

Refs MOBILE-1D

Co-authored-by: superset-home <338338640+superset-home@users.noreply.github.com>

* feat(devs): routines, so a Dev wakes on a schedule or an event (#9134)

* feat(devs): routines, so a Dev wakes on a schedule or an event

A routine wakes a Dev on its own: once at a set time, on a repeating
schedule, when someone posts in a channel it watches, or when a connected
tool sends an event (GitHub, Slack, Linear, and the rest). The Dev sets
routines up from chat with a routine_update tool on its MCP server (view,
create, update, pause). A run is a turn in the Dev's main session, for the
person who asked, answering where the routine was made; the Dev replies only
if there is something worth saying.

Following Codex Dots: the Dev can pause but not delete a routine; a one-off
switches itself off after it runs; at most one run an hour; there is no run
history beyond the Dev's turns.

Routines do not touch automations. They have their own table
(agent_routines) and procedures. The one shared point is where a provider
event is recorded: ingestAutomationEvent also hands the event to
notifyRoutinesOfEvent, which matches it with the same trigger configs and
matcher and never throws.

- DB: agent_routines; agent_turns.routine_id; turn reason "routine"
  (migration 0178: a new table, one nullable column, one enum value).
- Schedules: the worker's every-minute reconcile call also runs the routines
  that are due, so there is no new workflow or schedule.
- Channels: a Dev whose routine watches a channel gets each new message
  with the routine's instructions, instead of waiting for a mention or the
  channel router.
- API: devTools.listRoutines/createRoutine/updateRoutine for the Dev;
  agentRoutine.list/update/delete/runNow for people (the creator, or whoever
  manages the Dev). mainProcedure moves to its own module so the routine
  tools can use it, and the reply target pattern is shared from turns.
- Message blocks carry the time the Dev reads them, so "in 3 minutes" has a
  start.

Tests: shared schedule rules; seven routine cases in the Postgres
integration suite (27/27 on a Neon branch of production with 0178 applied).
Checked end to end locally on the pdev worker: from a DM, Echo made a
3-minute one-off and a #roster watch; a #roster post got "SEEN" in its thread
5 s later, and KIWI-9 posted 12 s after its time, then the one-off switched
off.

* feat(desktop): a Dev's routines in its side panel, with an edit dialog

The Dev page lists the Dev's routines under Routines: a clock for a
schedule, # for a channel it watches, a bell for provider events; on the
right, when it runs ("Weekdays at 10:30 AM"), what it watches (#roster,
Monitoring), or Paused. Each row's menu has Pause or Resume, Run now, Edit
and Delete, as on the Dots scheduled-task menu.

A row opens Edit routine, laid out like the Dots Edit task dialog: title,
what it does, a Repeat routine switch, Repeat (Interval, Daily, Weekdays,
Weekly, Monthly, Custom), Time, End repeat, and under Advanced the time zone
and where it runs. A routine that only watches shows what it watches
instead (channel, or provider, repository and event). Pause and Save as in
Dots. The list refreshes on agents.changed, so routines a Dev makes in chat
appear when its turn ends.

agentRoutine.list now says where each routine answers and names what it
watches (channel names, repository names), so the panel needs no other
lookups. PanelRow moves up to the panel's components, now that two sections
use it.

Checked in the dev app: Echo's three routines listed with the right
summaries; an edit to 10:30 saved as BYHOUR=10;BYMINUTE=30 with the next run
moved; the event routine shows GitHub, ParadevAI/Paradev and the event; Run
now on a finished one-off posted again.

* fix(desktop): routines UI matches the Dots scheduled-task UI

Compared side by side with the Dots recordings and closed the gaps:

- Time is a menu of half hours labelled like "10:30 AM PT", and the time
  zone shows its name ("Pacific Time"), not the IANA id.
- Advanced rows are separate boxes; rows in a box have inset dividers.
- Save stays off until something changes.
- A paused routine's title is not dimmed; the row says Paused.
- The row menu has icons; Delete confirms with "This will permanently
  delete the routine and stop future runs" and "Delete routine".
- Events read as words ("Pull request opened"); the dialog scrolls when it
  is taller than the window; the row keeps its summary hidden while its
  menu is open.

Checked in the dev app: Repeat to Weekly adds Day and turns Save on; delete
removes the routine; the menu and dialog render as in the recordings.

* feat(desktop): pick the room a routine answers in

Runs in is now a menu, as it is in Dots: Direct messages (the DM between
the Dev and the person the routine works for) or any channel the Dev is in.
A routine made in a thread keeps that thread as a choice.

agentRoutine.update takes a room. A channel must be one the Dev can post
in, readable by both the person editing and the person the routine works
for; the DM is found or made with ensureDirectMessage.

Checked in the dev app: moved the weekday digest to #roster and back; the
integration suite covers the move and a refused channel (28/28).

* fix(desktop): encode new-workspace attachments without a per-byte string (#9137)

The new-workspace composer built the base64 payload by appending one
character per byte to a string. Each append allocates a rope node, so a
large attachment takes about 20-32 heap bytes per input byte and a file
of a few hundred MB fills the renderer heap before btoa runs. Use the
shared FileReader-based fileToBase64 helper instead.

Refs DESKTOP-1KE

Co-authored-by: superset-home <338338640+superset-home@users.noreply.github.com>

* feat(chat-ui): rich text in the agent prompt composer (#9128)

* feat(chat-ui): rich text in the agent prompt composer

The prompt composer moves from PlainTextPlugin to RichTextPlugin with lists,
quotes and code blocks.

- Typing `- `, `* `, `1. `, `> `, `| `, ``` and `code` converts as you type.
  Bold and italic come only from Cmd+B / Cmd+I, so prompt text like `*args`
  and `__init__` stays literal.
- Shift+Enter continues a list, Tab / Shift+Tab nest. Enter adds a line
  inside a code block; Cmd+Enter still sends.
- Cmd+B / Cmd+I stop propagation so the sidebar and terminal-input hotkeys do
  not also fire. Underline is blocked, so Cmd+U still attaches a file.
- @ mentions and / commands do not open inside code.
- Paste is plain text, except a copy from this composer, which keeps its list
  structure (Lexical's text/plain copy of a list has no markers).
- Escape keeps focus as before; RichTextPlugin would blur the editor.

Draft text segments now hold markdown: submit, saved drafts, failed-send
restores and history all read and write it through draftSegments. Chips are
swapped for private-use placeholders around the markdown conversion. Plain
text exports verbatim, because Lexical backslash-escapes `*_~` and the agent
would receive `\_\_init\_\_`. Empty bullets at the ends are not sent.

The large-paste card loses "Show in text field": it would now have to insert
text past the markdown parser, and the card already sends the paste as a file.

CodeFencePlugin, $isInCode and the theme move from MessageComposer to
PromptInput, since both composers use them now.

* fix(chat-ui): keep typed entities and walk past re-read history

draftText decoded every numeric entity, so a typed `&#60;` changed on send
and `&#9999999;` threw, which blocked the send. Only whitespace entities are
decoded now; Lexical writes those for the edge spaces of formatted text.

History navigation compared the editor with the raw history entry. An entry
whose markdown reads back differently, such as `  - a`, stopped ↑/↓ at that
entry. It now compares with the entry as the editor reads it back.

* refactor(chat-ui): save composer drafts as Lexical state

Drafts the composer saves were markdown, so a reload re-parsed them and lost
anything markdown cannot say exactly: italic that starts mid-word and ends on
a space came back as literal asterisks. The composer now saves one `editor`
segment, its serialized Lexical state and the markdown the agent would get,
and loads it back as is.

Drafts built outside the composer stay `text` (markdown) and `chip`
segments. linkedContextSeed now gives chips only, and a chip loads inline at
the end of the last paragraph with a space after it, so no draft mixes chips
into markdown and the chip placeholders are gone. Text and editor segments
load as blocks a blank line apart, so joinDrafts no longer adds a separator.

The markdown sent to the agent puts the edge whitespace of formatted text
outside its markers (`*x* y`, not `*x&#32;*y`), so Lexical never writes
entities there. Typed `&#N;` are shielded from Lexical's markdown import,
which decodes every one and throws past U+10FFFF.

draftChips reads chips from every segment kind, for linked context.

* fix(chat-ui): survive unreadable saved drafts; ⌘⇧V pastes text from the composer too

A saved editor state with a node type this build does not register, or a
malformed entry, threw while seeding the composer, and again on every open of
that draft. It now loads from the segment's markdown, and draftChips skips a
segment with no state.

⌘⇧V skipped the plain-text path when the clipboard came from this composer.

* refactor(chat-ui): drop an unreadable saved draft instead of rebuilding it

A saved state that fails to parse is skipped, so the chat opens with an empty composer. Rebuilding it from its markdown was more code than the case is worth.

* feat(chat-ui): paste formatted text, headings, links and tables into the composer

Pasted HTML now keeps its structure: headings, links, lists, quotes and code
blocks become composer blocks and go to the agent as markdown. A pasted table
becomes markdown table rows. ⌘⇧V still pastes plain text, and a paste of
5,000+ characters still becomes an attachment.

Copying from the composer puts markdown in the plain-text clipboard, so copied
bullets keep their markers in other apps. A copy inside one code block stays
plain code. Lexical's own paste already restores a copy from this composer,
so the namespace check is gone.

tableClipboard moves to src/utils, now that the composer uses it too.

* feat(desktop): your own sent messages render as markdown

A sent list, bold text, code block, heading, link or table now shows in your
bubble the way the agent's replies do, not as raw `- ` and `**` markers. Chips
and link chips still draw inside the formatting, a typed line break stays a
line break, and copying a selection still gives back the typed URLs.

Typed text that only looks like markdown stays as typed: the composer writes
formatting only as `*`, `**` and `~~`, so `_`, `__` and single `~` are shown
literally, and a sent message is complete, so an unclosed `*` is not closed
for it the way a streaming reply's is (new `complete` option on ChatMarkdown).

The bubble's own URL splitter is gone; markdown finds the links.

* feat(marketing): add founder group photos to media kit (#9150)

* feat(marketing): add founder group photos to media kit

Adds two high-res photos of the three founders (YC Demo Day and a YC
talk) to the media kit's Founders section, next to the headshots.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(i18n): more idiomatic fr/ko founder photo captions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* ci(desktop): raise macOS build timeout to 90 minutes (#9105)

The desktop-v2.0.2 macOS x64 build hit the 60-minute limit in its post-cache step after every artifact had uploaded. GitHub marked the job cancelled and skipped Create GitHub Release, so the release needed a full x64 rebuild. Compile took 32 min on that runner against ~21 min normally; a normal x64 run already ends 15-20 min under the limit.

* fix(desktop): move the "New" line to what arrived while you were away (#9156)

The line was frozen where it was when the channel opened, so after a
conversation went on for a while it sat above messages you had long
read. Coming back to the window with unread messages now moves the line
to the first of them, in channels and in threads. Messages that arrive
while you watch are still marked read without moving it, and scrolling
keeps the old rule of placing a line only when there is none.

* fix(devs): stronger channel routing — by-name asks, thread replies via Jev, Jev retries (#9159)

* fix(devs): channel router answers a Dev spoken to by name

Jev treated a message addressed to a named person as not for the Dev,
even when the name was the Dev's own. A Dev with no description,
called by name, scored ~0.4 against the 0.8 threshold and never
replied.

The prompt now counts speaking to the Dev by name, and only stays out
for names other than the Dev's. A Dev with no description still does
not pick up open "can someone" asks. Eval set gains a #roster
channel with an area-less Dev: 100% exact on 99 cases (old prompt
missed both by-name cases).

* chore(devs): split channel router evals into real-world and generated sets

The cases move next to the runner that uses them, in
scripts/evalChannelRouter/. real-world/ holds cases from real channel
messages; generated/ holds the written eval and holdout sets. The
runner scores each set on its own. AGENTS.md says where a case goes
and how to run it.

* fix(devs): route thread replies through the channel router

A Dev that had written in, followed, or been mentioned in a thread got
every later reply in it, and its prompt tells it to answer each
message. Two people talking to each other under a Dev's thread each
got a reply from the Dev.

Thread replies now go to the channel router like any unmentioned
channel message; @mentions and DMs still go straight to the Dev. A Dev
the router picks in a thread it is already in keeps the thread_reply
reason. The router's context also drops join and leave rows, which
reached it as empty messages from the Dev.

Seven thread cases in the generated set (people talking past the Dev,
thanks, info, status, redirect by name): 105/105 exact, twice; holdout
unchanged at 31/32.

* fix(devs): retry Jev, and keep thread replies flowing when it cannot answer

Jev's endpoint returned 503 on about half of calls on 10-30, and any
failure routed the message to nobody. With thread replies now routed
by Jev, an outage would make a Dev ignore replies in its own thread.

routeChannelMessage retries a 5xx or network error within its 2 s
budget, logs the failure, and returns null when it has no answer.
On null, a thread reply goes to the Devs already in the thread, as it
did before Jev routed threads; a top-level message still goes to
nobody. The eval runner counts no-answer cases.

Checked on a local stack against this branch's database: in a thread
under an @mention, three replies between people did not reach the
Dev, a reply naming it did, and with Jev unreachable a thread reply
reached the Dev while a top-level message did not.

* fix(agents): sync Codex models with its live catalog (#9160)

Codex's live catalog (codex app-server -> model/list, codex 0.162.1) lists
GPT-6.1 Terra, which the picker was missing, and no longer lists GPT-5.5,
which retired 2026-10-14. The picker now matches the catalog's ids and order.
Every model takes max; ultra is offered on all but the two Luna models.

Fixes SUPER-3030.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* fix(host-service): treat a check run that failed to start as a failure (#8963)

mapCheckRunStatus had no STARTUP_FAILURE case, so a COMPLETED check run
whose workflow never started fell through to default and reported pending.
computeChecksStatus then kept the pull request pending forever.

Map it to failure next to TIMED_OUT and ACTION_REQUIRED, and cover both
the per-check status and the aggregate in tests.

* perf(desktop): stop hidden right-pane and background work on every git change (#9149)

* fix(desktop): tell panes in a closed right pane that they are hidden

Since #9008 the right pane stays mounted when closed. KeptTab provided
TabVisibilityContext = true to its active tab regardless of its parent,
so a hidden Changes pane kept re-rendering on every git change. SidebarSlide
now provides !hidden, and KeptTab combines it with its own visibility.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* perf(desktop): build the base-branch list only while its popover is open

BaseBranchSelector built one row per branch on every render, closed or
not, because the list was evaluated before PopoverContent decided not to
mount. With thousands of branches that ran on every git change.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* perf(desktop): persist the query cache at most every 10s

The async persister re-serializes the whole persisted set with superjson
on every cache event, throttled to 1s, even for queries that are never
persisted. Git churn kept that running about once a second.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(desktop): remember a missing GitHub login for 30s

viewerGithub.status runs before every PR read. Without a login, each call
spawned `git credential fill` and `gh auth token` again.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(desktop): time the missing-login window with a monotonic clock

A backward wall-clock correction made the elapsed time negative and kept
PR panes on the cloud login until the clock caught up.

Co-Authored-By: Claude <noreply@anthropic.com>

* test(panes): cover kept-tab visibility under a hidden parent

TabVisibilityScope holds the rule KeptTab applies: a pane is visible only
while its tab and every parent are. The test fails if the scope goes back
to overriding its parent.

Co-Authored-By: Claude <noreply@anthropic.com>

---------

Co-authored-by: Claude <noreply@anthropic.com>

* fix(plugins): use Resend's official logo (#9163)

The Simple Icons SiResend glyph is Resend's old R mark. Swap in the current
icon from resend.com/brand, cropped to the mark's bounds, as a currentColor
glyph so it still follows the theme in desktop and marketing.

* refactor: share check status mapping between host-service and desktop (#9161)

Host-service and the desktop pull request view each had their own copy of
the GitHub check run to status mapping and of the overall checks rollup.
The copies drifted: the desktop handled STARTUP_FAILURE, host-service did
not (#8962, fixed in #8963). Move both functions to
@superset/shared/pull-request-checks so a new GitHub conclusion is added
in one place.

The v2 workspace computeChecksRollup and the mobile effectiveCheckStatus
stay as they are: they count checks and treat skipped and neutral runs
differently, so sharing them would change what the UI shows.

* feat(db): add user_settings table for account-wide settings (Chat) (#9155)

* feat(db): add users.experimental_chat, off by default

Account-level switch for opening new agent sessions as chat, so desktop,
mobile and server-side launches read one value. Constant-default ADD
COLUMN: metadata-only on Postgres 11+.

* feat(db): keep the Chat setting in a new user_settings table

Per review, account settings stay off auth.users, which every
authenticated request reads. user_settings follows
channel_notification_prefs: user_id primary key, typed columns with
defaults, and no row means the defaults, so Chat stays off for every
account without a backfill.

Replaces the users.experimental_chat column from the previous commit,
which never shipped.

---------

Co-authored-by: Harshith Mullapudi <17528887+harshithmullapudi@users.noreply.github.com>

* feat: account-level Chat setting (user_settings), out of PostHog (#9162)

* feat(db): add users.experimental_chat, off by default

Account-level switch for opening new agent sessions as chat, so desktop,
mobile and server-side launches read one value. Constant-default ADD
COLUMN: metadata-only on Postgres 11+.

* feat(db): keep the Chat setting in a new user_settings table

Per review, account settings stay off auth.users, which every
authenticated request reads. user_settings follows
channel_notification_prefs: user_id primary key, typed columns with
defaults, and no row means the defaults, so Chat stays off for every
account without a backfill.

Replaces the users.experimental_chat column from the previous commit,
which never shipped.

* feat(api): open automations as chat by default, take chat surface from clients

Automations no longer read the acp-chat PostHog flag: they ask for a chat
whenever the agent supports ACP, and the host still answers with a
terminal when it can't chat with that agent.

cloudWorkspace.create takes an optional `surface`. Desktop and mobile send
their Chat experimental setting; callers that omit it (released builds,
CLI, MCP, SDK) keep the old flag lookup, so the input is additive.

* feat(desktop): move chat into Experimental settings, off by default

The Chat switch is shown to everyone and is the only gate; the acp-chat
PostHog flag is no longer read. A stored null (never touched) now means
off. ACP agents open as chat when it is on; every other agent, and every
agent when it is off, opens a terminal.

- Show the presets bar on chat tabs again; agent presets already pick
  chat or terminal from the setting and ACP support.
- Hide "Fork into chat" while Chat is off.
- Cloud workspace creation sends the setting as `surface`.

* feat(mobile): add Chat experimental switch, off by default

Device-local switch in Settings → Experimental replaces the acp-chat
PostHog flag in every mobile chat gate. Cloud workspace creation (composer
and voice tool) sends it as `surface`.

* feat: make Chat an account setting and always show existing chats

The Chat experimental switch now writes users.experimental_chat through
user.updateExperimentalChat and every surface reads it from the session,
so desktop, mobile and the server agree. It only decides how new agent
sessions open; chats that already exist (from automations, another
device, or a released build) are always adopted and listed.

- Drop the device-local stores; desktop's acp-chat key moves to DEAD_KEYS.
- Drop the unreleased cloudWorkspace.create `surface` input: the server
  reads the account setting for every caller.
- Desktop: chat sessions are adopted and automation run links focus chats
  regardless of the setting; the switch moves into AcpChatSetting.
- Mobile: chats are listed, managed and stoppable regardless of the
  setting; only new sessions follow it.

* feat(api): automations follow the owner's Chat setting

Automations open as chat only when the owner has Chat on and the agent
supports ACP; otherwise a terminal. Everyone starts with Chat off, so
released builds, which hide chats from accounts outside the acp-chat
flag, keep seeing every run.

* feat: read the Chat setting from user_settings, not the session

Per review, the setting moves off auth.users into user_settings, so the
session and auth stay untouched.

- New userSettings router (get/update), mirroring channelNotificationPrefs:
  no row means the defaults, update upserts only the given fields.
- Server reads it through loadUserSettings for automations and cloud
  workspace starts.
- Desktop and mobile read it with a userSettings.get query; the switch
  writes the returned settings back into that query.
- Desktop: the "Track file changes per turn" switch moves into its own
  component, alongside AcpChatSetting.

* fix: address review on the account-level Chat setting

- Launch paths wait for the setting (useResolveAcpChatEnabled) instead of
  reading "off" while userSettings.get loads: desktop launcher, new agent
  session, new workspace; mobile new terminal workspace and new session.
- Desktop sign-out resets userSettings.get, so the next account never sees
  the previous one's value. Mobile already clears its query cache.
- The Chat switch is disabled while a save is in flight, so quick toggles
  cannot land out of order (desktop and mobile).
- "Track file changes per turn" stays visible, disabled while Chat is off,
  so settings search never points at a hidden control.
- Comments name user_settings.experimental_chat and the Chat setting.

---------

Co-authored-by: Harshith Mullapudi <17528887+harshithmullapudi@users.noreply.github.com>

* feat(plugins): add Sazabi plugin (#9169)

* feat(plugins): add Sazabi plugin

* docs(plugins): explain one Sazabi connection per organization

* feat(plugins): add four more Sazabi skills

* feat(plugins): document the new Sazabi skills and regenerate manifests

* fix(plugins): browse Sazabi operations by namespace and tighten connection docs

* fix(connectors): request Autumn rewards scopes (#9176)

The Autumn connector's OAuth grant didn't include rewards:read or
rewards:write, so the Autumn plugin's reward and promo-code tools
(listRewards, createReward, updateCatalog with rewards) returned 403
"Insufficient scopes. Missing: rewards:read". Existing connections need
to reconnect to pick up the new scopes.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* feat(marketing): full-screen footer with globe grid and closing CTA (#9179)

The footer now fills the viewport below the header, Pax-style: a large
closing headline with the download and GitHub buttons on top, link
columns and a bottom bar below.

- Move CTASection into the footer so every page ends with the CTA; drop
  the per-page copies and the CLI install box so it fits one screen
- Projected globe grid in dark lines that lights up around the cursor,
  reusing the EdgeGlare pointer hook (promoted to app/[lang]/hooks)
- Split Product links into two columns and compress spacing on short
  viewports so 1280x720 through 1920x1080 fit exactly
- Legal links move to a bottom bar with the copyright and language switcher
- New headline "Smash your backlog." translated for all locales

* feat(i18n): ship Ukrainian (#9190)

Adds uk to SUPPORTED_LOCALES, the lazy catalog map and lingui.config,
with a complete catalog of 7,053 messages, plus the mobile side: iOS
permission strings, composer and attachments-sheet .lproj resources,
and the Hermes PluralRules and RelativeTimeFormat locale data.

The code is uk (the language), not ua (the country). The picker shows
"Українська". Plurals carry all four CLDR categories (one, few, many,
other), and the native test now expects few/many for uk as it does
for pl, ru and cs. Spelling follows the 2019 orthography (проєкт).

Product vocabulary translates per the glossary: робочий простір,
агент, панель, завдання, сеанс, гілка. Git verbs keep the forms
Ukrainian developers use (коміт, закомітити, запушити, злити,
перебазувати); worktree, diff, stash and issue stay in Latin script.
Paywall is пейвол, added to the glossary row.

* fix(marketing): keep the cookie banner buttons inside the box (#9194)

The banner is max-w-xs and its action row did not wrap, so in fr, es, uk,
ja and other locales the Accept button rendered past the right edge. The
privacy link now takes its own line and the two buttons share a two-column
grid, so the banner keeps one shape in every locale.

* fix(marketing): prose lint reads CTASection from its new Footer path (#9185)

#9179 moved CTASection under Footer/components, so vale stopped on a missing
argument and the Vale prose lint check fails on every PR.

* fix(mobile): native chat streams replies token by token (#9184)

* fix(mobile): native chat shows a streaming reply token by token

The live text was read from the snapshot inside renderItem, but LegendList
re-renders a cell only when its data or extraData changes. The rows do not
change while a turn streams, so the reply showed only when the turn ended.
Each row now carries its text in the list data, and a row keeps the same
entry until its text changes, so only the streaming row re-renders. The data
change also lets maintainScrollAtEnd follow the stream while the reader is at
the bottom.

* fix(mobile): chat stops following the stream while the reader drags

Every streamed chunk is now a data change, and maintainScrollAtEnd scrolled
to the end on each one while the reader was within its threshold, so a slow
drag up was pulled back down. Following is off from drag start to drag end.

* docs(marketing, docs): name Paradev where answer engines copy, and fix the /cli docs 404 (#9148)

* docs(marketing, docs): name Paradev where answer engines copy, and fix the /cli docs 404

Peec shows our compare pages retrieved, and often cited first, for prompts we
lose, while answers name other tools. Put the Paradev answer in the parts
engines lift: a tools row in the where-to-run table, the nightly recipe first,
flaky tests and phone follow-ups in titles and first screens, Paradev first in
the Design Mode sentence, and a review summary on the remote-machine page.
Mark Paradev cloud workspaces as early access on where-to-run.

Docs: redirect /cli (404, hit by AI crawlers through docs.superset.sh/cli) to
/cli/getting-started, and add a dispatch section there.

* docs(marketing, docs): shorter sentences in the new first-screen copy

* docs(marketing, docs): make the dispatch CLI examples run on a local-only host

Every example now passes --local and --project PROJECT_ID, shows where each
id comes from (projects list, terminals list, the create output), and uses a
dated branch, because a branch that already has a workspace reuses it.
Replace prj_/aut_/h_ id prefixes that do not match real ids, fix
tasks list --status in_progress (needs a status UUID), and correct the
host-default rules for workspaces create. Align the visual-UI summary with
its table: stagewise sends no screenshot.

* docs(marketing, docs): keep CI dispatch on one host, get terminal IDs from JSON, revert an unproven flaky fix

* docs(marketing): scope worktree and port claims on the scheduled and remote pages

* docs(marketing): guides for running agents with Linear, GitHub, Sentry and more (#9152)

Six tutorial pages: Claude Code on Linear issues, Claude Code and Codex on GitHub pull request events, fixing Sentry issues with coding agents, running agents from an iPhone, Design Mode for frontend teams, and starting agents from Slack. Vendor claims are cited from primary sources in the PR description; the GitHub review prompt never checks out or runs code from fork pull requests.

* docs: correct how worktrees and the app are described (#9151)

Product copy across marketing, docs and README now says each task gets a separate Git worktree and branch (worktrees do not sandbox processes, and local processes share the host's ports; Paradev shows which ports each workspace opens) and describes Paradev as a desktop app rather than a terminal. The marketing site description is translated in every locale, including Ukrainian. Historical changelog entries and dated posts are unchanged.

* docs(marketing): name Paradev in the terms and subprocessors descriptions (#9205)

The legal entity is Superset Inc. and the product's trade name is Paradev. The terms and subprocessors page descriptions now say Paradev; the legal text and lastUpdated dates are unchanged.

* feat(cli): let agents drive the right sidebar with superset panes (#9204)

* feat(cli): let agents drive the right sidebar with superset panes

The panes CLI only saw the main area. Now that the workspace has a right
sidebar, agents can read and arrange it too.

- panes list and every mutation result carry a sidebar block: shown,
  full screen, and its tabs and panes.
- split, resize, focus, close, swap, equalize and move find a pane in
  either area; move --to and swap work across areas, move --sidebar and
  move --main give a pane its own tab in that area.
- panes open changes|file|files|pages|browser shows the sidebar with that
  tab in front (--background to add it quietly). open file PATH shows a
  file in the editor, open changes PATH points Changes at one file, both
  with --line. The host turns the path into the form each pane needs.
- panes sidebar show [--full-screen] and panes sidebar hide.
- new-tab --sidebar opens a terminal there.
- Right pane full screen moves from component state to a small store so
  the CLI can set it; both pane layout stores register for live edits.
- The panes skill covers the sidebar.

* fix(host-service): keep panes open file inside the workspace

Follow symlinks and reject files outside the worktree, so open file only
shows workspace files. move --main no longer needs sidebar support, and
panes list says the main area is empty instead of the whole workspace.

* fix(desktop): keep sidebar viewer panes out of the main area

The app moves any pane that is not a terminal or chat out of the main area
as soon as it lands there, so a CLI move of a browser, file or Changes pane
into it came back to the sidebar and was then dropped by the sidebar write.
move and swap now refuse that with a clear error, and the sidebar is written
before the main area.

Also open the code view for open file without --line, and accept names
like ..notes.md in open changes.

* fix(connectors): register a new dynamic client when a connector's scopes grow (#9203)

* fix(connectors): register a new dynamic client when a connector's scopes grow

A dynamic connector registered its OAuth client once, with the scopes it
asked for at the time, and reused that client forever. When a connector's
scope list later grew, the provider rejected the authorize request: Autumn
answered invalid_scope for rewards:read and rewards:write after #9176,
although its metadata lists both, because the stored client was never
registered for them.

plugin_oauth_clients now records the scopes each client was registered
for and allows several clients per issuer and redirect URI. A new grant
uses a client whose scopes cover the request, or registers one. A grant
already issued stays on its client, since its refresh token is bound to
that client_id: connections.oauth_client_id records it at token exchange,
carried from authorize to callback in the signed OAuth state. Connections
from before this column existed keep refreshing with the oldest client
for the issuer, which is the one they were issued to.

A row from before scopes were recorded covers no scope, so each dynamic
connector that asks for scopes registers one new client on its next
connect. Treating it as covering would leave Autumn broken.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* fix(connectors): keep legacy grants on their original client

A connection from before client ids were recorded fell back to the
first unexpired client, so once its original client expired or was
forgotten it moved to a newer, scoped one. Its refresh token can't work
there, and the resulting invalid_client would forget that healthy client
and every grant issued to it. Legacy grants now use only the row from
before scopes were recorded, and throw ForgottenClientError (needs
reauth) instead of switching or registering.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

* perf(desktop): agent events and workspace switches stop re-rendering and refetching the whole UI (#9202)

* perf(desktop): read agent bindings outside the workspace page

V2WorkspaceContent read the workspace's agent bindings through
useClearActivePaneAttention (twice) and useAutoAdoptChatSessions. Every
hook event refetches the bindings with a new lastEventAt, so each event
re-rendered the whole page, about 800 components. The hooks now run in
WorkspaceAgentEffects, which renders nothing.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* perf(desktop): stop every sidebar row re-rendering on a workspace switch

Four values changed identity on each switch and re-rendered every row:
- useHostWorkspaces and SandboxAccessProvider built their context values
  from useQueries arrays, which are new on every render. Both now use a
  stable module-level combine, and the host value is memoized.
- Each row called useMatchRoute, which subscribes to every navigation.
  It now selects only its own active flag from the router state.
- selectWorkspaceFromEvent depended on the active workspace id, which it
  only reads on click. It now reads the id from a ref.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* perf(desktop): one bindings refetch per agent event, not one per observer

Every component reading a workspace's agent bindings subscribed to the
agent events and ran its own cancel-then-invalidate. Each call cancelled
the previous fetch, so one hook event caused about 36 invalidations and
cancelled requests with a few tabs open. refreshQuery now collapses calls
for the same key in one task into a single cancel-then-invalidate, which
still starts after the event.

Refs #9136

Co-Authored-By: Claude <noreply@anthropic.com>

* fix(desktop): match the sidebar highlight and selection to the page on screen

- The active-row selector read the pending location, so a slow navigation
  lit the destination row before its page showed. It now reads the
  resolved location wh…

This branch had an error being deployed

1 failed deployment
preview — 91bc4d13 Deployed Oct 30, 2026 by akshatcoder-hash via Create Database (Neon) #12957
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Desktop: a check run that failed to start shows the checks summary as pending

2 participants