Skip to content

fix: close the open issue batch and the findings its validation showed - #387

Merged
mubeda merged 16 commits into
mainfrom
claude/now-priority-bug-fixes-f937b2
Nov 19, 2026
Merged

mubeda merged 16 commits into
mainfrom
claude/now-priority-bug-fixes-f937b2

Conversation

@mubeda

@mubeda mubeda commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

Why

The goal grew from five issues triaged as "now" to every open issue of the repository: each one is either fixed here or closed with its reason. Three issues arrived while the work ran (#389, #390, #391). Findings that tests, platform lab runs or Profiler data showed on the way are fixed here too.

What changed

  • Claude approvals (Claude: "Always allow this session" behaves like "Approve once" #383). Claude is started with --permission-prompt-tool stdio when its --help lists it. A tool that needs approval opens the approval card, AskUserQuestion opens the question card, and plan review stays in BiBCode. "Always allow this session" applies to the session only. The default agent is no longer passed as --agent, so naming it does not restart the session.
  • Claude background tasks (Claude: a follow-up turn BiBCode opens for a background-task notification can stay running forever #389). A background-task notification that Claude answers inside the current turn is no longer counted as a promised extra turn, which left a turn "running" with no provider activity. Two notifications a turn did not answer each keep their own turn. An answer Claude gives with no turn open (after a background agent finishes) now completes its text, tools, reasoning, pending requests, an open compaction and its subagent rows, and refreshes the context reading, without opening or completing a turn. A notification record no longer confirms a pending steer with the same text. The working timer restarts when a new turn follows a finished one.
  • Codex default (Stop defaulting Codex to the retired gpt-5.4 model #380). Nothing names a fixed model. "No model chosen" is the slug auto, shown as "Codex default". A thread or a saved default that names a retired model is moved to the default when it is used, and says so once. The server clears a thread's retired model after a launch that dropped it, compares model spellings with one shared rule, and keeps a turn admitted under any spelling on its route.
  • Codex chat reviews (Source Control: review of uncommitted changes stops for command approval on every step #376, core). A Codex chat review runs with approval never in Codex's read-only sandbox, decided by the server from the thread's review marker, on every launch, restart and resume. Unexpected Codex approval requests and questions are declined at once. The guard line and a disabled "Read-only review" lock say so when the server advertises readOnlyCodexReviews. An unknown Codex runtime-mode string starts Supervised.
  • Server terminal (Environments: server icons get Open terminal and Dashboard actions #378). The server's own terminal and dashboard open from the Local rail icon and the server card; the client never closes that terminal, the thread id server-terminal is reserved, and removing a workspace never collects it.
  • Git Manager (Git Manager: toolbar gets New branch and Merge from the selected branch #388, PR, pipeline and Git Manager panels: first load shows no spinner, only a faint gray bar or plain text #377). New branch and "Merge into…" in the toolbar, disabled states that explain themselves, loading and refresh states. A refusal hides stale pull-request and pipeline data, on any page and across a reload, until a newer successful read.
  • Source Control and sidebar (Sidebar: collapsing a Repositories environment card leaves the open worktree visible #382, build-desktop-artifact fails with a custom CARGO_TARGET_DIR #371, AI Terminal: Send to agent button is too small and too faint to find #386). A labelled Review button; a collapsed project or environment card hides every card; the Send-to-agent button is prominent in the toolbar.
  • Sessions (Auth: pruning an expired session may skip the ClientRemoved for a revoked client #374). Pruning an expired or revoked session tells its client.
  • Connection storage (Desktop: intermittent hang at "Project data is still loading", client never opens its WebSocket #381). Every IndexedDB request has a 10 s deadline. A store that does not answer shows "Quit BiBCode and open it again" instead of loading forever.
  • Large replies (Desktop: private memory keeps climbing after large file-list and history responses #390). The Files panel reads projects.listEntries page by page (at most 960 KiB per reply) and still shows the complete tree. For clients that send fitFirstCommit, a history page stays within the same budget and an oversized first commit lists its first changed files and counts the rest.
  • Worktree removal (Worktrees: getRemovalPlan can fail with "workspace thread was not found" after a successful removal #391). A plan, a removal or a detach for a worktree that was already removed answers thread-not-found with alreadyRemoved: true; the dialog shows "This worktree was already removed." A thread that never existed still reports not found.
  • Chat timeline. Two additions to the patched list library: no size correction for a row that is gone (the timeline jumped to the top when a turn finished), and a toggled Thinking row or work step keeps its place while a lock holds a row below it. Opening a Markdown file no longer draws its diagram twice.
  • Config stream. Each section (settings, keybindings, provider list, tools) has a server-side version; a snapshot or a resync after a lag never lets a section go backward, and a refresh the client already holds sends nothing. Source-control discovery runs its tool probes together, so one slow tool no longer adds its timeout to the others.
  • Files preview. The desktop content security policy allows blob: in connect-src, which WebKit's PDF viewer needs; the PDF preview was blank on macOS.
  • Smaller fixes. JsonRpcConnection::close() always releases its writer; provider error text stays out of the capability-refresh log; the Codex inventory probe warns when it learns no sign-in type; ~ follows the home-folder environment before the platform lookup; desktop tests no longer leave data folders behind; the Profiler heartbeat treats WebKit's 1 s timer alignment on a visually idle page as an aligned run, not stalls; scripts/build-desktop-artifact.ts honours CARGO_TARGET_DIR.
  • Test harness and runbooks. Stale packaged specs corrected (the Codex default label, the read-only review lock, a right-click the driver never delivers, three Mermaid timing assumptions); two specs no longer need a 1600×1000 window; a Mermaid draw hold that exists only in the E2E build; the closing spec's group no longer reads failed after an intended quit; CI's packaged workflow gives Xvfb a 1920×1080 screen; Linux, Windows and macOS runbooks state what the lab runs found.

Scope

Closes #371
Closes #374
Closes #375
Closes #376
Closes #377
Closes #378
Closes #380
Closes #381
Closes #382
Closes #383
Closes #386
Closes #388
Closes #389
Closes #390
Closes #391

Closed as not planned, each with its reason on the issue: #384 (resume a stopped turn; three designs, the last still let a resume take another message's turn) and #385 (reuse one Codex app-server for skill reads; a kept server cannot equal a fresh launch).

#376 is delivered in part: stronger isolation than Codex's own sandbox, Claude parity and script runs are declined in a comment on the issue.

Tradeoffs and known limits

  • Claude answers that follow a background agent. When an agent finishes during a turn's last answer, Claude's own follow-up answer is not tracked as work: only the idle timeout protects it, and a revert is not refused while it runs a tool. This is the behaviour before this change. Three designs to track it were written and removed after review found defects in each (the last allowed a revert under a running tool). docs/providers/claude.md lists the remaining cases with the captured Claude Code 2.1.326 record sequences.
  • Desktop: private memory keeps climbing after large file-list and history responses #390 bounds each reply, not the panel's memory. The Files panel still holds the whole list. That desktop memory returns to baseline was not measured (the test repositories are small), and that the memory was retained at all is unproven. An older client still gets the whole list in one reply, and whole commits.
  • Desktop: intermittent hang at "Project data is still loading", client never opens its WebSocket #381 is not explained at its root. What makes WebKit leave an indexedDB.open unanswered is not established. The change turns a silent hang into an actionable error.
  • The config stream still sends a full snapshot for a capability change or a one-provider change; with a real Codex account a snapshot is about 1 MB. Sending less needs a new wire event.
  • Trees past the 25,000-entry index cap stay partial, as before; that needs per-directory loading.
  • A commit with tens of thousands of parents can still exceed the page cap.

Traced by review and not reproduced by a test, so not changed here:

  • two windows saving providerSessionDefaults at the same instant can still overwrite each other in one ordering (a server-side fix was written and withdrawn: its check found it could overwrite a model chosen in the same request and broke an older-server pairing);
  • a failed keybinding save can leave a rejected binding visible to a new subscriber (the fix was withdrawn: a disconnect at the wrong moment could erase a saved binding);
  • after a config resync, a manual refresh that changes only timestamps can leave "last checked" stale;
  • a message sent while an answer with no turn is streaming can leave that answer's text marked streaming (the fix was withdrawn: it sent a context query during the new turn).

Blast radius

Every Supervised and Auto-accept Claude chat on a Claude Code that supports the flag shows cards where tools were denied before. Every Codex thread, terminal and setting without a chosen model sends no model. Codex chat reviews no longer ask before commands. The Files panel, the history view and the settings update use new optional request fields; older clients and servers keep their behaviour. The list-library patch changes scroll anchoring for every chat timeline. The desktop content security policy is one directive wider (connect-src blob:), which lets the page read bytes it already holds and opens no host.

Verification

All on the final tree unless a line says otherwise.

  • Gates, as CI runs them: cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings, cargo check -p bibcode-server --lib --bins, vp check, vp run typecheck, vp run check:contracts, vp run check:markdown-links, the agent-delivery check: pass. cargo test -p bibcode-server 8,037 passed and cargo test -p bibcode-desktop pass. Web 12,234, shared 858, contracts 970, client-runtime 1,335, scripts 955, the packaged-harness support tests: pass.
  • macOS, packaged app: the default suite 42 and 6 passing, both spec groups passed, exit 0.
  • macOS, real Claude Code 2.1.326 (specs/manual/claude-approval-real-provider.e2e.ts): the approval card, a session approval, the question card and plan review; a background task that ends inside its turn leaves no running turn for 60 s and the next timer starts at 0 s; a background task that outlives its turn: Claude answered by itself 25 s after the turn ended, and 30 s later no turn was running and the answer was complete. Claude's own transcript and BiBCode's provider event log of that run match the test fixtures' record order.
  • macOS, real codex-cli 0.169.0 (specs/manual/send-to-agent-real-providers.e2e.ts): 5 of 5 on the second run. On the first run the app's connection storage did not answer at launch (the condition of Desktop: intermittent hang at "Project data is still loading", client never opens its WebSocket #381); the app showed "Connection storage is not responding. Quit BiBCode and open it again" where it used to load forever. A protocol probe against the same CLI accepted the review policy of Source Control: review of uncommitted changes stops for command approval on every step #376, ran a read command with no approval request and wrote nothing; a review through the packaged app with a real Codex was not run.
  • Windows arm64 VM: cargo check -p bibcode-server --all-targets with warnings denied passes; desktop host tests 472 passed (on the tree before the last two back-outs); the default packaged suite 42 of 42, exit 0, in three of three runs, including the server-commands spec, all Mermaid cases and the PDF preview step. On an earlier tree of this branch the app exited once mid-suite in 1 of 5 full runs with no crash record and no shutdown lines; it did not recur and its cause is unknown (the test service that can restart the app does not keep its own log, and native dumps are disabled on that platform).
  • Linux, Ubuntu 22.04 arm64 container, non-root: cargo test -p bibcode-server 7,976 passed, 0 failed; cargo test -p bibcode-desktop 597 passed; the default packaged suite under Xvfb 1920x1080: 44 of 44 in one run, and 43 of 44 in another (one Mermaid case failed its setup check: the timeline's scroller was at least 83 px shorter than the case assumes, while the reader itself did not move; it passed alone 2 of 2 and in the next full run, and why the scroller was shorter is unknown).
  • Linux, Arch-based VM (Xvfb 1920x1080; the suite cannot resize its window under a tiling compositor): the default packaged suite 44 of 44, exit 0.
  • Profiler, on in every packaged run: no crash, panic or hang finding on macOS, Docker or the Arch VM. The Linux runs raised client.stall warnings (longest 343 ms and 760 ms on a quiet machine; up to 2.4 s earlier while the host was out of memory headroom) and rpc.slow warnings of 1.0 to 1.6 s; the macOS runs raised none. Windows raised rpc.slow warnings only.
  • Reviews: React changes reviewed against vercel-react-best-practices and UI.md. Testing runbooks updated (Linux prerequisites, the Xvfb screen size, the macOS window, the Windows executable path).
  • Profiler coverage: no new child process, RPC method, removal path, binary or shutdown step. New log lines go through tracing and carry no account detail, prompt or path. The existing background-work event and rpc.bytes records are unchanged in shape.
  • Not verified: Desktop: private memory keeps climbing after large file-list and history responses #390's memory criterion; x64 Windows and x64 Linux (CI only); the Windows build of the home-folder and test-folder changes beyond the compile check and the suite.

🤖 Generated with Claude Code

mubeda and others added 13 commits November 17, 2026 18:06
The wrapper looked for the bundle under the repository's own target
folder, so a build with CARGO_TARGET_DIR set finished and then failed
with TauriDesktopBuildDirectoryMissingError before copying the artifact.
The bundle, the updater app and the symbol paths now come from one
release folder that follows CARGO_TARGET_DIR. A relative value is
rejected up front, because Cargo resolves it against its own working
folder and the two could disagree.

Closes #371

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The review control was a 24 px ghost icon with no visible text, so it
was easy to miss. It now shows "Review" with a bot icon in the size and
color of Stage All. The action, the tooltip, the accessible name and the
dimmed disabled state are unchanged.

Closes #375

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A collapsed project or environment card kept the open chat's worktree
card visible. When that card was the only child, the group looked as if
it had not collapsed, apart from the chevron. The collapsed flag alone
now decides what renders, in the Environments and the Repositories
view, and the jump labels follow it. The collapsed header is marked as
current while it holds the open chat.

Closes #382

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
The Claude driver answered "Always allow this session" with the same
one-time allow as "Approve once". The reply now carries the rules Claude
suggested as updatedPermissions. Every rule is written to the session,
never to a settings file, and a suggested mode change is dropped, so the
thread keeps the mode it shows. With no usable rule the approval is a
one-time allow and the resolved event says why.

Measured against Claude Code 2.1.326: with this reply a second identical
command runs without a prompt and no settings file is written.

Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
BiBCode named gpt-5.4, and gpt-5.4-mini for Git text, whenever nobody
chose a Codex model. Codex retired both for ChatGPT sign-in, so such a
thread or AI Terminal stopped at Codex's "no longer available" question.

The default is now "auto", which names no model, so Codex runs its own
recommended one. The model list starts with that choice, named after
the model that Codex's model/list marks with isDefault. The Plan
collaboration mode, which requires a name, takes the model Codex
reported for the thread and gpt-6.1-sol only when it reported none.

A thread that still stores a retired model starts without one when the
published inventory says Codex is signed in with ChatGPT, and the thread
shows one warning line. An API-key session keeps the stored model.

Measured on codex-cli 0.169.0: model/list marks gpt-6.2-sol with
isDefault and lists gpt-6.1-sol.

Refs #380

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Codex. Switching a thread from a named model to "Codex default" now
restarts the session. A reported or custom model named "auto" no longer
duplicates the default entry. The fallback picks the "auto" entry
wherever a saved picker order put it. The sidebar card names Codex's
default instead of showing the slug. A send that replaces a stored
retired model shows one notice.

Claude approvals. A session rule is kept only when it is an allow rule
or a readable folder with content. An allow echoes the input Claude sent,
which Claude Code before 2.1.207 requires, and a deny carries a message.

Sidebar. Next and previous thread work from a chat hidden by a collapsed
card. Source Control. The base-branch label truncates before it can push
Review out of a narrow panel. Scripts. The CARGO_TARGET_DIR test runs
through the copy step.

Refs #371
Refs #375
Refs #380
Refs #382
Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
…default

The recorded settings stream named gpt-5.4-mini as the Git text model.

Refs #380

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A Supervised Claude chat thread never showed an approval card. BiBCode
started Claude Code without a permission prompt target, so Claude denied
by itself every tool that needed approval, and "Always allow this
session" had nothing to answer.

Claude is now started with --permission-prompt-tool stdio when its help
lists --permission-prompts. A tool that needs approval opens the
approval card, which names the command or the file. AskUserQuestion
opens the question card, and its answers go back as the tool's input.
ExitPlanMode is denied at once, so plan review stays in BiBCode.

A decision, Claude's cancel request, an interrupt and session shutdown
each claim the pending request, and the first claim wins. The driver
publishes the resolved event after the reply is written, which closes
the card. A later answer is refused and nothing is written.

Measured on Claude Code 2.1.326: the session reply stops a second prompt
for the same command and writes no settings file, an interrupt makes
Claude withdraw a pending request, and a read-only command in plan mode
does not prompt.

Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude approvals. A request still open when its turn ends is closed with
the turn. A reply that cannot be written ends the session instead of
leaving Claude waiting, and shutdown no longer waits on a full event
queue. A session rule keeps only its valid entries, the one-time
fallback reason is the detail the thread shows, and a question's options
always carry the description the question card requires.

Codex. A thread resumed with no chosen model names the default Codex
reports, because Codex keeps a resumed thread's previous model when none
is given (measured on codex-cli 0.169.0). Switching to "Codex default"
restarts the session when the next message starts, not under a running
turn. Every casing of "auto" is one entry.

Web. The "auto" fallback applies to Codex only, so other providers keep
their model order. The retired-model notice shows only when Codex no
longer offers the stored model.

Refs #380
Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
It drives the installed Claude Code in the packaged app: a Supervised
Claude chat opens an approval card that names the command, "Always
allow this session" runs it, the same command then runs with no card,
and no settings file is written. The runbook has the procedure.

Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Codex runs the model its own configuration names when a request names
none, and model/list marks another one as default. Measured on
codex-cli 0.169.0 with a configured model: a thread started with no
model reported the configured one, while the list marked a newer model.

The "Codex default" entry and a resume with no chosen model now take the
configured model from config/read first and the listed default second.
The lookup is bounded to 5 seconds, so a Codex that does not answer
cannot hold the provider supervisor.

Also from the third review: a queued switch to "Codex default" defers
its options with the switch, a session rule with a non-string
ruleContent is dropped, the fallback list skips every casing of "auto",
and the retired-model notice waits for Codex's own model list.

Refs #380
Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
… spec

The spec now also checks that Claude's question opens the question
card and that the chosen answer reaches Claude, and that a plan-mode
turn shows the plan with no approval card.

Refs #383

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Codex's input reads the packaged driver's Enter and Return as a new
line, so the manual real-provider spec never submitted its first prompt
and the steps that need a finished turn recorded nothing. macOS now
sends Return through System Events, as the Mermaid spec does for PageUp.

Refs #380

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 18, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-18T02:33:48.517548Z 16e00f8 PR opened
🔒 Security Review ✅ Completed 2026-11-18T02:36:06.273074Z 16e00f8 PR opened
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@github-actions github-actions Bot added vouch:trusted PR author is trusted by repo permissions or the VOUCHED list. size:XXL labels Nov 18, 2026
A session approval given during a new Claude chat's first message was
lost. The second message names the default agent `claude`, which the
first one left out, and that restarted the Claude process. The default
agent named or left out is now the same launch: no restart, and the
delivery route check accepts either spelling on every path.

Jump shortcuts numbered every unarchived thread, including the primary
card and panel threads, and ignored pins. They now number the rows a
project card renders.

The collapsed approval card on a narrow screen showed no command. It
now shows the approval's detail, scrollable when long.

The Profiler raised `client.stall` while WebKit delivered a visible,
focused page's timers 1 s apart with frames still rendering. Two
heartbeats in a row that arrive about 500 ms late are now an aligned
run and count as no stalls. A single such beat, a stall of 1 s or more,
and a beat held when focus is lost still count.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@mubeda mubeda added the ci:full Run full macOS + seeded upgrade CI label Nov 19, 2026
Issues: Claude background-task turns (#389), bounded file-list and history
replies (#390), removal plans for an already removed worktree (#391), the
server terminal and dashboard (#378), Git Manager toolbar actions, loading
states and refusals (#388, #377), Codex chat reviews in the read-only sandbox
(#376, core), session pruning (#374), connection-store deadlines (#381), the
Send-to-agent button (#386), and the remaining parts of the retired Codex
model handling (#380).

Claude: a notification answered inside its turn opens no follow-up turn; an
answer given with no turn open completes its text, tools, reasoning, pending
requests, compaction and subagent rows and refreshes the context reading
without completing a turn; a notification record never confirms a steer; the
default agent is not passed as --agent.

Codex: one shared model equality; a session remembers the model its launch
dropped; a server-only engine command clears a thread's retired model on the
instance that dropped it; the inventory probe warns when it learns no sign-in
type.

Chat timeline: the patched list library no longer corrects the size of a row
that is gone and holds a toggled row in place under a lock below it; a
Markdown file's diagram is drawn once.

Config stream: per-section versions keep a snapshot or a resync from going
backward; a refresh the client already holds sends nothing; a per-provider
update of the saved session defaults.

Also: source-control discovery probes run together; the desktop content
security policy allows blob: in connect-src for the PDF preview; the home
folder follows the environment first; desktop tests keep their data in
temporary folders; the packaged UI harness, its specs and the testing
runbooks are corrected for what the macOS, Windows and Linux lab runs found.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@mubeda mubeda changed the title fix: Claude approvals, Codex default model, Review button, sidebar collapse and artifact paths fix: close the open issue batch and the findings its validation showed Nov 19, 2026
…llery build

The gallery's renderer-unavailable state replaces the renderer loader in a
production build. Gating that seam to unit tests left the state without its
text, so the Chromium core capture refused both of its states.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@mubeda
mubeda merged commit b5afb5e into main Nov 19, 2026
36 checks passed
mubeda added a commit that referenced this pull request Nov 19, 2026
Raises the Claude Terminal hook ceiling to Claude Code 2.1.328 after the
live PreToolUse guard proof passed on that build, sets every version
source to 0.17.0 and writes the curated changelog entry for PRs #370,
#373, #379 and #387.

Verified locally: cargo check --workspace --all-targets --locked, the
hook ceiling tests, cargo fmt --all --check, vp check, vp run typecheck,
vp run release:smoke and the Markdown link check.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment