Conversation
…siteboon#1001) Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Keep the repository lock aligned with the latest resolution already allowed by CloudCLI's declared dependency range. This changes no source code or package manifest. Constraint: better-sqlite3 12.11.1 does not yet declare Node 26 in its engine metadata Confidence: high Scope-risk: narrow Directive: Do not describe this lock refresh as official Node 26 support until upstream package metadata includes it Tested: npm ci; native in-memory database smoke; typecheck; client/server build; lint with zero errors on Node 26.5.0 and npm 11.17.0 Not-tested: Platforms outside linux/amd64 Related: CoderLuii/HolyClaude#67 Co-authored-by: Simos Mikelatos <simosmik@gmail.com>
* fix: check CLAUDE_CODE_OAUTH_TOKEN in checkCredentials()
checkCredentials() never checked CLAUDE_CODE_OAUTH_TOKEN, so a valid
long-lived token was ignored and auth status fell through to reading
~/.claude/.credentials.json, reporting false expiry once that file's
short-lived OAuth session cycled.
Add the check in the same priority slot Claude Code itself uses:
after ANTHROPIC_AUTH_TOKEN/ANTHROPIC_API_KEY (incl. apiKeyHelper via
settings.json), before falling back to the credentials file.
Uses method: 'environment', matching the existing convention in
opencode-auth.provider.ts for env-var-sourced static credentials.
* refactor(claude-auth): use a friendlier auth status label
Finding: F5 (cosmetic nit) from cloudcli-upstream-auth-prs-2026-07 audit.
Replaces the raw env var name with a human-readable label in the auth
status email field, matching the friendlier style of the other auth
methods ('Auth Token', 'API Key Auth').
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* test(claude-auth): cover checkCredentials() OAuth token priority order
Adds coverage for the priority order in checkCredentials(): OAuth token
authenticates independent of a stale credentials file, falls back to
credentials_file when unset, reports unauthenticated on an expired
credentials file, and yields to ANTHROPIC_API_KEY per the documented
precedence order.
checkCredentials() has no dependency on the claude CLI (only env vars
and ~/.claude files), so it's called directly rather than through
getStatus()/checkInstalled() — keeps the test portable across
environments that may not have claude installed.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
* fix(claude-auth): also check CLAUDE_CODE_OAUTH_TOKEN in settings.json env block
CodeRabbit review on siteboon#979: Claude Code accepts CLAUDE_CODE_OAUTH_TOKEN
from the settings.json/settings.local.json env block, not just
process.env — same as ANTHROPIC_API_KEY/ANTHROPIC_AUTH_TOKEN, which are
already checked both ways a few lines above. Checking only process.env
missed users who configured the token there, reporting them as
unauthenticated.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
* fix(shell): recover from stale websocket connections
Guard retained PTY ownership so delayed close events cannot detach a replacement socket.
Terminate half-open sockets after a missed pong.
Cover both lifecycle paths with focused tests.
* feat(git-panel): one-click git init plus update and history UX fixes
- show a "Run git init" button in Source Control when the project is not
a git repository (new POST /api/git/init), hide the branch/refresh
header controls in that state, and stop logging the expected no-repo
state as a server error
- track commit-history loading separately so "No commits found" no longer
flashes while commits are still being fetched
- hide "Create Initial Commit" when the reposi
- auto-refresh the window after a platform update ("This will refresh the
window in N seconds") and cancel the refresh if the update fails
- handle non-JSON responses from /api/system/update so hosted/Docker
deployments no longer surface "Unexpected token '<'" when the server
restarts mid-update
* fix(shell): force restart before reconnecting
* fix: resolve session deep links to owning project and collapse chat error rows
Session routing: /session/<id> URLs could select the wrong project or show
nothing, because the frontend only searched the paginated project payloads
(first session page per project) and fell back to binding the session to
whatever project was selected.
- add GET /api/providers/sessions/:sessionId resolving a session (by app id,
falling back to the provider-native id) to its metadata and owning project
- resolve unmatched URL sessions thro correct
project, and renavigate provider-native alias ids to the canonical app id
- keep the selected-project placeholder only as a fallback for unknown ids
Chat errors and commands:
- Bash command rows no longer auto-ex and
status badge signal the failure and output stays behind the chevron
- collapsed commands truncate to one line again: the global
`.chat-message code` pre-wrap !imponcate`,
so the command now renders in a span
- non-Bash tool error results render as a collapsed red row that expands
to the full error content (new ToolErrorDisplay)
* refactor(chat): pass app session ids to provider runtimes
The chat websocket handler used to substitute the provider-native
session id into the spawn options, so runtimes keyed their process maps
by provider ids and abort/pending-approval lookups had to translate
back and forth. Now the app session id flows through everywhere, and
each runtime resolves the provider-native id itself via the new
sessionsService.resolveProviderSessionId() only where its CLI/SDK
actually needs it (resume flags, resumeThread, OpenCode's token DB).
- chat.send passes the app session id; the unused `resume` flag is gone
- chat.abort addresses runtimes by aps
aborting a session's first run actually kill the runtime (previously
abort was a no-op until the provider announced its native id)
- pending permission approvals are tr,
removing the remap in chat.subscribe
- runtimes key process maps by the app session id when given, falling
back to the captured provider id fo
- resolveResumeModel now receives the app id, matching how the
active-model endpoint records session model changes
- run notifications (stopped/failed/permission) carry the app session
id, matching frontend deep links
* feat(chat): render user messages as markdown and allow image-only sends
- Render user bubbles through the shared Markdown component with
remark-breaks so typed line breaks are preserved
- Add explicit ul/ol/li styling so list markers render correctly in
both user and assistant messages
- Open composer image attachments in the fullscreen lightbox on click
- Allow sending/queueing messages with images but no text
- Remove the AI commit message generator button from the git panel
* feat(chat): markdown rendering, image-only sends, chat export, auth check, and UI fixes
User messages:
- Render user messages as markdown with remark-breaks for line break preservation
- Add explicit ul/ol/li styling so list markers render correctly in both bubbles
- Enable sending/queueing messages with images but no text
Chat export:
- Add comprehensive export to Markdown, HTML, and PDF with beautiful formatting
- Export button in top-right corner of messages pane
- Include timestamps and message metadata in exports
Provider authentication:
- Check provider authentication before sending messages
- Show error with instructions: "Not logged in to [provider]. Please log in via
Settings > Login in CloudCLI, or in the Shell tab."
Image attachments:
- Open composer image attachments in fullscreen lightbox on click (reuse existing
lightbox from sent images)
UI fixes:
- Move QuickSettingsPanel to app-level (out of MainContent overflow context) to fix
z-index layering on desktop and mobile
- Set panel z-index to z-[9999] with backdrop at z-[9998] to ensure it appears
above all content including sidebars and file trees
- Remove AI commit message generator button from git panel source control tab
* fix: ignore subagent and tool-results from scan
* fix: show mcp tool result text type correctly
* fix: disable single-dollar math
- Disable single-dollar math in remark-math to prevent dollar amounts (e.g., $20/psf)
from being parsed as LaTeX; keep double-dollar display math ($$…$$) for math expressions
* revert(chat): remove pre-send provider auth check
Drop the client-side "Not logged in to [provider]" gate that ran
before sending messages. The backend still validates provider
authentication, so this pre-flight check was redundant and blocked
sends on its own status lookup.
* feat(git): add worktree management with session switching and squash merge
Add a Worktrees tab to the git panel for Crystal-style parallel workflows:
- Create worktrees in sibling <repo>-worktrees/<branch> folders, from a new
or existing branch, with a live folder preview
- Register each worktree as a linked project ("repo · branch") so chat
sessions, shell, files, and git all work inside it; one-click switching
- Squash or --no-ff merge a worktree back into the base branch, with
optional cleanup (remove worktree + delete branch); conflicts are rolled
back server-side and reported with the conflicting files
- Safe removal: archives the linked project, blocks dirty worktrees unless
changes are explicitly discarded, never touches the main worktree
Backend: new server/modules/worktrees module (routes + list/create/open/
merge/remove services, barrel-only cross-module imports, node:test suite
with injected git runners) mounted at /api/worktrees.
* fix(git): make tabs scrollable on frontend
* fix(sidebar): keep rename input visible on mobile
* fix(codex): render wrapped tool history
* refactor(worktrees): inject module dependencies
* refactor: move auth, commands, plugins, git to their own modules
* refactor: consolidate shared server utilities
* refactor: move file-tree and voice to their own modules
* refactor(providers): centralize session token usage
Move provider-specific token calculations out of the legacy server entrypoint.
The backend now resolves storage details from the app session record.
Expose a session-id-only provider endpoint and cover every provider path
with focused tests.
* refactor(providers): centralize runtime adapters
Move provider SDK and CLI execution beside each provider implementation so
runtime ownership matches the existing auth, model, session, and MCP layout.
Dispatch runs, aborts, and Claude permission approvals through one typed
runtime service instead of wiring provider-specific functions in server/index.
Keep the stateful runtime entrypoint separate from the lower-level provider
barrel to prevent cycles through session and project services.
* refactor(system): move update endpoint into module
Keep update command selection and process execution out of the legacy server entrypoint.
The route delegates to a tested TypeScript service while preserving the existing API contract.
* refactor(providers): unify runtime ownership
Make each provider own its runtime so providerRegistry is the single resolution source.
Export dispatch through providers/index.ts and centralize structural runtime types.
Remove the redundant runtime registry and runtime facade.
* refactor(server): port entrypoints to TypeScript
Move the Browser Use MCP executable into its feature module.
Route CLI startup through the module barrel and preserve production JavaScript entrypoints.
* chore(repo): codify backend module standards
Share the backend architecture rules with forks.
Require coding agents to apply them consistently during implementation and review.
* fix(providers): discover Codex home skills
Include ~/.codex/skills in user-scoped discovery so installed Codex skills appear in the provider.
* fix: resolve coderabbit comments
* fix: preserve safety across async git workflows
Track request-owned resources so cleanup cannot delete reused checkouts.
Compensate or report post-operation failures so retries do not repeat completed Git work.
Reject unsafe inputs and stale responses before they trigger unintended actions.
Keep exported chat content literal to prevent unsafe rendering.
* fix: keep worktree actions safe and accessible
Serialize worktree operations so requests cannot overwrite each other's busy state.
Use selectable branch defaults and expose actions across touch and keyboard input.
* fix: restore deep-linked session resolution
* fix: refresh auth tokens across realtime clients
* chore: add claude usage plugin
* fix: hide gitignored files from chat mentions
Respect project .gitignore rules when building the @ mention list.
Keep the explorer and other file consumers unchanged.
* fix: show ququed images correctly and improve sidebar archive tab
* fix: prevent repeated skills loading
Key refreshes by stable project values so settings rerenders do not restart discovery.
Keep cached results visible while an explicit refresh is running.
* fix: hide injected skill bodies from live sessions
When Claude invokes a Skill, the CLI injects the entire SKILL.md as a
synthetic user turn. Persisted transcripts tag it isMeta: true, but the
live SDK stream does not. Without a content-level check, the skill body
renders as a giant user bubble during the run and vanishes on reload.
The skill is already represented by the Skill tool call itself, so the
injected body should be filtered as internal content. Filter by the
"Base directory for this skill:" prefix that precedes the markdown.
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
* fix: discover skills in symlinked directories
* feat: support file attachments across providers
* fix(chat): dedupe image-only user messages
* fix: update chat composer UI
* fix(providers): persist the active model per session
The active model shown by /models was always the provider catalog default
— for Claude, the literal string "default" — because the command resolved
it from the catalog and never looked at the model the composer had sent.
Model selection was global too (one localStorage value per provider), so
reopening an older session showed whatever was last picked anywhere.
Make the model a property of the session, the way the Claude, Codex,
Cursor and OpenCode clients do:
- Add a `sessions.model` column, written when the user picks a model and
on every chat.send, so a session remembers what it ran with.
- Resolve it in one place. `resolveSessionModel` answers "what is this
session using?" (session row, then provider session state, then the
client default, then the catalog default) and backs /models, /cost,
/status and the composer. `resolveResumeModel` answers "what should
this run use?" and deliberately skips the provider lookup, which for
Codex reads a global config that would override the user's pick.
- Layer the session's model over the per-provider default in the
composer, so switching sessions shows and sends the right model.
Drop the session model-override sidecar (~/.cloudcli/provider-session-
active-model-changes.json), its read/write helpers, the tri-state
{supported, changed, model} shape, and changeActiveModel on all four
adapters. Session state belongs on the session row, where deleting a
session also drops its model.
Also fixes /cost and /status passing arguments shifted by one to the
model resolver (both reported "Unknown"), and the OpenCode lookup
querying its session table with the app id instead of the provider id.
Pre-existing overrides in the old sidecar file are ignored; affected
sessions resolve from provider state until their next send.
* fix(git-panel): make commit confirm message scrollable
---------
Co-authored-by: Claude Haiku 4.5 <noreply@anthropic.com>
…iteboon#1074) buildFileTree recurses into every subdirectory it lists, with no concept of which directories are legitimate project locations. listProjectFiles walks the project root to depth 10, and resolveProjectRoot returns the stored project_path without validating it — unlike the browse endpoint, which rejects system paths via validateWorkspacePath before walking. A project whose path is "/" therefore gets swept in full, including /proc. Traversing live procfs both wastes the walk on thousands of virtual per-process entries and logs an error for each unavoidable race (EINVAL on /proc/<pid>/net, ENOENT on /proc/<pid>/task/<tid> when a thread exits mid-walk), flooding the server log. Reused the existing FORBIDDEN_WORKSPACE_PATHS list — already the source of truth for which directories can never be a workspace — to skip recursion into them, rather than introducing a second, differently scoped list. The check is an exact path match, so it prunes /proc when the walk root is "/" without affecting a project that merely contains a directory named "proc".
…icker (siteboon#1036) The 'default', 'sonnet', and 'sonnet[1m]' model options in CLAUDE_FALLBACK_MODELS still describe "Sonnet 4.6" even though the underlying `sonnet` alias already resolves to Sonnet 5 today (verified: `claude --model sonnet` reports "Sonnet 5, model ID claude-sonnet-5"). The Opus entry was already updated to say "Opus 4.8" - Sonnet's labels were simply missed in that pass. This is a label-only fix: the `value` fields (`default`, `sonnet`, `sonnet[1m]`) are unchanged, since those aliases are resolved live by Claude Code CLI itself and already point at the current model. Only the human-readable description text was stale, which made it look like Sonnet 5 wasn't available in the picker at all. Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
* feat: add provider session ID copy actions * fix: address provider session ID review feedback
* feat(i18n): add missing zh-CN translations for all namespaces - Add zh-CN tasks.json (94 keys, previously missing entirely) - Add voice settings, input queue, and OpenCode translations to chat.json - Add voice settings, plugin descriptions, desktop notifications to settings.json - Add file preview and markdown editor translations to codeEditor.json - Register tasks namespace import in i18n config.js - Update .gitignore to allow zh-CN and zh-TW tasks.json Co-Authored-By: Claude <noreply@anthropic.com> * chore: simplify locale tasks.json allowlist in .gitignore --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
- Add missing src/i18n/locales/ko/tasks.json (94 keys), translated from en/tasks.json, and register it in src/i18n/config.js (ko was the only locale in config.js not importing/loading a tasks namespace bundle). - Add !src/i18n/locales/ko/tasks.json to .gitignore's task-file negation list — the blanket "tasks.json" ignore rule had no exception for ko, so the new file would otherwise be silently untracked. - Backfill settings.json (63 missing keys: voiceSettings, notifications.desktop, mainTabs.voice/browser, tabs.skills, mcpServers.managed, agents.account.opencode, and several pluginSettings entries), and remove one stray key (pluginSettings.buildYourOwn) that no longer exists in en. - Backfill smaller gaps found while auditing all locale files: chat.json (17 keys: voice.*, input.queue.*, opencode provider), codeEditor.json (5 keys: markdown preview/edit, filePreview.*), sidebar.json (11 keys: search.* and tooltips.clearSearch). All 7 ko/*.json files now have exact key parity with their en counterparts. Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
Adds the Spanish locale with all 7 namespaces (common, settings, auth, sidebar, chat, codeEditor, tasks) fully translated, registered in the i18n config and the language selector.
…red (siteboon#1085) Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
…iteboon#1115) The set/restore block around query() was inert for two reasons: the child's env is snapshotted earlier in mapCliOptionsToSDK() as { ...process.env }, so a later mutation of process.env never reaches the spawned CLI; and the Claude CLI does not read CLAUDE_CODE_STREAM_CLOSE_TIMEOUT at all. The real knob for how long the CLI waits on still-running background agents after a turn ends is CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS (default 600000 ms, 0 = wait indefinitely). It already passes through via the env snapshot, so document it instead of setting anything in code. No behavior change.
* feat(sidebar): add recent conversation feed * fix(sidebar): address recent conversation review --------- Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
* chore: upgrade codex sdk * fix: add rename to mobile session renaming * fix(sessions): name CloudCLI sessions from initial message * fix(file-tree): cap project tree traversal at 10,000 entries buildFileTree walked project roots to depth 10 with no total node budget, so a broad but valid root such as a populated home directory could retain gigabytes of tree nodes, starve other API requests, and crash-loop the server. Share one remaining-entry budget across every recursive branch and fail oversized trees with an HTTP 413 FILE_TREE_TOO_LARGE error that asks for a narrower root or ignore rules. The budget is charged after ignore filtering, so generated directories and .gitignore exclusions do not consume it. Fixes siteboon#1082. * fix(file-tree): surface file tree load failures in the UI A failed /files request only logged to the console and left the tree empty, so the 413 FILE_TREE_TOO_LARGE refusal added in 700c786 rendered as "No files found / Check if the project path is accessible" — the one message that does not explain what went wrong. useFileTreeData now exposes the API's error message, and FileTreeBody renders it as a warning state ahead of the empty-tree branch. Requests aborted by a project switch stay silent. * feat(models): add customizable provider model library - replace model caching with source-controlled predefined catalogs - update predefined models for each provider - persist only custom models in auth.db - add custom model create, edit, and delete APIs - add model management to the picker and /models - keep predefined models read-only - remove model refresh controls * fix(chat): persist reasoning effort per session Prevent stale state from overriding effort selections, restore effort when reopening sessions, and record the selected effort on every send. * fix(models): prevent model library content from being clipped * fix(codex): do not ignore non-bash tool calls * fix: resolve coderabbit comments --------- Co-authored-by: Simos Mikelatos <simosmik@gmail.com>
…n#1078) Co-authored-by: Haile <118998054+blackmammoth@users.noreply.github.com>
* fix(claude): default background agent wait ceiling to 30 minutes * fix: respect gitignore in the Files view Use project .gitignore rules as the authoritative file-tree filter so legitimate build directory names remain visible. Fall back to the conventional ignored-directory list when no .gitignore exists. * feat(claude): keep background work alive after turn completion - Hold Claude CLI input open for deferred background tasks - Notify users when background work completes - Prevent duplicate completion events - Clean up obsolete background wait configuration * fix(chat): prevent hidden and unbounded history refreshes Gate message-history requests by chat visibility, coalesce deferred refreshes, and reconcile bounded latest pages while preserving pagination, realtime messages, and scroll position. * feat(editor): make file paths easy to copy Copy the complete path even when its displayed value is truncated. Provide accessible success feedback after copying. * refactor(frontend): simplify JSON parsing utilities Keep malformed tool-input handling at its call site so the fallback is explicit. Remove generic or unused storage helpers that no longer justify shared abstractions. * refactor: rename LLM provider logo component * chore(docs): move readme to docs * feat(git): allow force deleting unmerged branches * refactor: relocate shared frontend types * fix(chat): polish composer navigation and layout Keep keyboard-selected file mentions visible in long result lists. Align the queued-message card with the composer edge. * fix(chat): hide Codex plan envelope tags * feat(search): prioritize session title matches Show database-backed session matches before transcript content. This lets users find conversations by name, even before a transcript is indexed. * refactor: centralize shared platform constants * chore: remove unused util * fix(ui): restore French locale and responsive tabs Register French resources so language selection no longer falls back to English. Keep large plugin tab sets reachable on desktop and mobile. Anchor the tab rail to the right whenever it fits. * fix: resolve markdown formatting issues * feat: support mermaid diagram * feat: add upload to specific directories * fix(claude): stop orphaned query instances when a run supersedes an active session addSession now interrupts a different live instance before overwriting its map entry, and run-loop cleanup only removes/emits for entries it still owns. Prevents unreachable spinning generators (CPU leak) when an abort races run setup and a new send takes over the session id. * fix: addressgithub code quality suggestions * feat: show session title for browser title * fix: address review findings
# Conflicts: # server/modules/providers/tests/claude-sessions.test.ts
…oon#1265) The OpenCode provider authenticated any credentials file entry, so a subscriber who connected OpenCode Go (`opencode-go` in auth.json) counted as logged in - but the curated catalog had no `opencode-go/*` entries at all. The connected-provider filter then found no Go options, fell back to the full catalog, and left Go users staring at a list of models their CLI rejects while every model they pay for was missing. See siteboon#840, which was closed without a fix while the bug persists on main. Add the 27 `opencode-go/<model-id>` entries `opencode models --verbose` reports, with the reasoning-variant values from the same output wired as effort metadata so the existing effort picker resolves them through `--variant`. Labels follow the catalog's existing style; the CLI's marketing suffixes ("(2x usage)", "(New)") are dropped. `DEFAULT` stays on the Zen default - the filter already relocates it for Go-only installs. The docs also list `minimax-m2.5`, but the CLI does not report it, so it is deliberately left out; the catalog mirrors CLI output.
…on#1249) Question/answer cards split a multi-select answer on ", " unconditionally, so a single option whose own label contains ", " ("Yes, always") rendered as two chips, neither matching an option, both marked (custom). Match one exact option first and split only when no option matches, which leaves genuine multi-select answers unchanged.
…eboon#1238) * feat(chat): recall sent messages with arrow keys in the composer ArrowUp in an empty composer recalls messages previously sent in the open chat, newest first; ArrowDown walks forward and finally restores the draft. History is kept per chat scope (the same session-or-project key drafts use) in localStorage, capped per scope and across scopes, and recorded on send, queue, and slash-command execution. The arrows keep their normal meaning while editing text or navigating the command/mention menus. Signed-off-by: Matthew McClintock <matthew@mcclintock.net> * fix(chat): walk a stable history snapshot during recall An append from another tab mid-recall shifted what ArrowDown landed on and could displace the draft restore. Navigation now snapshots the entries when recall starts and walks that array until recall ends. Signed-off-by: Matthew McClintock <matthew@mcclintock.net> * fix(chat): record queued messages before the session-switch return Switching sessions while a queued message's attachments uploaded hit the early return after persistence, so the message dispatched later without ever entering input history. Record it right after persistence, under the session it was queued for. Signed-off-by: Matthew McClintock <matthew@mcclintock.net> --------- Signed-off-by: Matthew McClintock <matthew@mcclintock.net> Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
* feat: collapsible branches in the model picker
The picker lists every model of every installed provider in one flat
run - an OpenCode installation alone contributes ninety - so reaching
the provider below means scrolling past all of them.
Each branch now collapses by its heading, with the model count beside
it, and the state is remembered in localStorage. A search opens every
branch for as long as it runs, because a hit inside a collapsed one
would be invisible. A collapsed branch keeps one row ("N hidden - click
to show"): cmdk drops a group whose items are all gone, and a heading
that is no longer rendered cannot be clicked open again.
The rendering moved into its own component rather than growing the
440-line empty state further.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: persist the collapsed branches from an effect
React may run a state updater more than once or discard its result, so a
`localStorage` write inside one is not tied to committed state. The
updater only computes the next set now; an effect on it does the write.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix: clear the model search when the picker closes
The search box is controlled, and closing the dialog only flipped `dialogOpen`.
A query stayed behind and came back the next time the picker opened - with
every provider branch expanded, because searching expands them. That undoes
the collapsing this picker exists for.
Closing now clears the query, through one function the four call sites share.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…oon#1223) Node's explicit module extensions fell through to the default arm of the language switch, so a .mts or .cts file opened in the editor with no highlighting at all. The existing `ext.includes('ts')` check already picks TypeScript for the two new TS extensions and JavaScript for the other two, so the four cases are all that is missing. Co-authored-by: Malte Buttjer <claude@buttjer.net> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…eboon#1220) * fix(sessions): keep an archived session archived across a rescan createSession resets isArchived on every re-index, without checking whether anything was written. The full rescan runs on each project-list fetch and selects transcripts by birthtime, so it re-indexes every session created since the last fetch -- and each of those loses its archive flag. Archiving a session started since the sidebar was last loaded therefore never sticks, which covers every session a user archives right after finishing it. Reset only when the caller reports newer activity: an explicit timestamp that is newer than the stored one, or no timestamp at all, which the statement already treats as "now" when it writes updated_at. julianday() parses both the column-default and ISO forms that occur in the column. * fix(sessions): treat an omitted timestamp as activity on the upsert path too Review catch: the ON CONFLICT branch compared excluded.updated_at, which is already COALESCE(?, CURRENT_TIMESTAMP), so a call that passes no timestamp was compared as "now" -- and CURRENT_TIMESTAMP resolves to whole seconds, so it is not newer than a row written in the same second. An archived app-created row, indexed for the first time without timestamps, therefore stayed archived while its own updated_at said it was live. Same reasoning as the UPDATE branch, same `? IS NULL` arm. Two tests, one per direction, both checked against the unpatched query first: the omitted-timestamp one fails without this change, and the stale-transcript one pins that an older transcript is still not activity. --------- Co-authored-by: Malte Buttjer <claude@buttjer.net> Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
…iteboon#1192) * feat(i18n): connect remaining hardcoded UI strings to translations Wire user-facing strings that never went through i18n so they respond to the language setting. Adds a new git namespace for the source control panel and extends the settings, auth, chat and common namespaces for skills, browser use, onboarding, the command palette, MCP server views, the setup form, the version upgrade modal, the folder browser and a few smaller screens. Template strings become interpolated t() calls, plural counts use i18next _one/_other forms, labels stored in constant files are resolved through t() at render time, and commit dates in the history view follow the active locale instead of hard-coded en-US. English keys cover every namespace touched here; other languages fall back to English through the existing fallbackLng setting until their translations are added. * fix(i18n): address CodeRabbit full-review findings - AuthContext: use the auth namespace with stable errors.* keys for session/status/login/registration/network messages - CommandResultModal: translate the modal footer hint and Close button - MergeWorktreeModal: refresh the untouched default message when the locale changes (tracked per worktree) - gitPanelUtils: FILE_STATUS_LABELS now returns git:status.* translation keys, matching FileChangeItem's t() rendering - BrowserUseSettingsTab: always show localized operation-failure messages instead of raw API error text; render the translated runtime-required fallback instead of the server's English status message * feat(browser-use): error-code contract for browserUse API Server now replies with the structured AppError envelope { success:false, error:{ code, message, details } } on every browser-use endpoint (status, settings get/save, runtime install, sessions list/stop/delete). Stable BROWSER_USE_* codes let clients map errors to localized copy instead of rendering raw English API text. - readApiJson throws ApiRequestError carrying code/details/status and accepts both legacy string and structured error envelopes - BrowserUseSettingsTab and BrowserUsePanel translate failures through code->i18n-key tables with localized per-operation fallbacks - add en browserUse.installFailed key used by the panel mapping * fix(api): preserve legacy top-level details in readApiJson Legacy envelopes ({ success:false, error:'msg', details }) lost their top-level details because payload was empty for string errors. Fall back to data.details and cover structured, legacy and bare envelopes with tests. * chore: align package-lock.json and package.json with upstream main The PR carries no package-lock.json changes (reviewer request): the lockfile now matches origin/main exactly. --------- Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
…eboon#1162) The project and session delete dialogs in `SidebarModals.tsx` render five translation keys that exist in none of the locale catalogs: - deleteConfirmation.archiveProject - deleteConfirmation.archiveSession - deleteConfirmation.archiveSessionNotice - deleteConfirmation.archivedSessionNotice - deleteConfirmation.deleteSessionPermanently Each call site passes an inline English default, so the keys never surface as missing at runtime — they silently render English in all 11 languages. The effect is worst where it matters most. Both dialogs offer a safe action and a destructive one side by side. The destructive button uses `deleteConfirmation.deleteAllData`, which *is* translated everywhere, while the safe "Archive" button falls back to English. A non-English user reading the project dialog sees a localized "Delete all data permanently" next to an English "Archive project", which is precisely the choice the dialog exists to make clear. Adds the five keys to all 11 locales, following the pattern of siteboon#896. Note: `removeFromSidebar`, `allConversationsDeleted` and `cannotUndo` are translated in every catalog but no longer referenced by any component — they appear to be the pre-refactor names of these same strings. Left untouched here to keep this change additive; happy to remove them in a follow-up.
…oon#1274) listPluginSkills read a plugin's commands/ or its skills/, whichever it found first: the existence of a commands folder ended that plugin's turn, so a plugin shipping both contributed only its commands. The commands reader also takes .md only, so a plugin whose commands are in another agent's format lost both halves at once -- nothing matched, and the `continue` walked past the skills sitting beside them. The CLI offers both halves for the same plugin, so those skills work in the terminal and are absent from the slash menu, with nothing to say why. Read both folders. The skills branch already skips a plugin with no skills/, and the menu dedupes by command, so a name present in both places is still listed once. Fixes siteboon#1273 Signed-off-by: Liran Funaro <liran.funaro@gmail.com>
…on#1159) `getProviderModels` resolves to a `ProviderModelsDefinition` - `OPTIONS` and `DEFAULT` - but the agent route read `.models` off it, so both catalog handles came back undefined. A codex or opencode run that named no model then died on `codexModels.DEFAULT` with a TypeError before the provider was ever started, and the route reported it as a 500. The route's own test mock encoded the same imaginary wrapper, which is why nothing caught it. The mock now returns the real shape, and a new test drives the codex path with no model and asserts the catalog default reaches the runtime.
…ion model (siteboon#1207) * fix(claude): ignore <synthetic> model placeholder when resolving session model Claude Code stamps locally-synthesized rows in session JSONL files (API-error placeholders and similar) with model: "<synthetic>". The session-model scan in readClaudeSessionModelFromJsonl walks the JSONL backwards and takes the first event carrying a model field, so when the last row happens to be such a placeholder, "<synthetic>" is surfaced as the session's active model. The UI then adopts it, sends it back as the model for the next turn, and the SDK issues a real API request with model "<synthetic>", which the upstream API rejects with 404 model_not_found. Treat angle-bracketed values as placeholders and skip them, falling back to earlier events or the provider default. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(claude): skip placeholder hits inside content-part scan Move the placeholder filter into extractClaudeModelFromMessageContent so a placeholder in an earlier content part no longer stops the scan before a later part carrying the real model tag. Add regression tests. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * fix(claude): filter placeholders per candidate inside text extraction A placeholder "set model to <synthetic>" stdout hit no longer shadows a real <model> tag later in the same text. Filtering now lives at each candidate site in extractClaudeModelFromTextContent, so the content-part scan needs no extra checks. Regression tests cover both content shapes. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
…wn (siteboon#1157) A Conversations row was a link and a chevron: you could open the session and nothing else. Renaming it, copying its provider session id, forking it or archiving it meant finding it again under its project. This gives that list the controls the Projects list already has, from one component rather than a second copy of the markup. Re-ported onto the layout siteboon#1206 landed, which moved src/components/sidebar to src/modules/sidebar and flattened view/subcomponents. Nothing of the previous version survived as a diff, so this is that change rewritten against the new tree, with two differences worth naming: - fork is included. It did not exist when this was first written, and it is the third action the Conversations row was missing. The fork path reads only a session's id, provider and owning project, which is exactly what a recents row knows, so no session object is invented. - PROVIDER_LABELS moves to utils/sidebarProjectFormatting, beside the other row formatting, rather than being exported from a component. The shape is unchanged from the version that was reviewed down to ~116 net lines: the mechanism is extracted, not duplicated. SessionOptions the options menu and the inline rename that replaces it, with the outside-click dismissal that belongs to it useProviderSessionIdCopy the id behind "copy id": fetched when the menu opens rather than per row, dropped when it closes, sequenced so a late reply is discarded SessionRowActions the nine members a row needs, named once; SidebarProjectListProps composes it rather than restating it, and the call site passes one prop SidebarSessionItem is 208 lines lighter for giving those up; the Conversations row spends 103 to gain all four actions and two states. Callers keep only what is genuinely theirs. The Conversations row shows the amber attention dot and the processing spinner, but not the green "touched recently" dot — in a list ordered by recency it would be on nearly every row and say nothing. A row whose project is unknown gets no rename, since a rename is keyed by project; it is withheld rather than guessed. Also fixes a bug the second list exposed: nothing refetched the recents feed after a rename or a delete, so the row you had just acted on kept its old title or stayed in the list. Both now patch in place, which also preserves the pages loaded past the first that refetching page zero would discard. Build, both typechecks and lint are clean; 380 client tests and 394 server tests pass, including six new ones asserting what the Conversations row resolves for the shared controls. Signed-off-by: Liran Funaro <liran.funaro@gmail.com> Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
* feat(codex): expose GPT-6 Astra model * chore(codex): upgrade runtime for GPT-6 Astra
* feat(codex): expose GPT-6 Astra model * chore(codex): upgrade runtime for GPT-6 Astra * fix(codex): replace unsupported default approval policy Use on-request to prevent startup failures while preserving workspace sandboxing. Update permission descriptions and test both new and resumed Codex sessions. --------- Co-authored-by: Simos Mikelatos <simosmik@gmail.com>
Adds realjustinwu/CloudCLI-GLM-Usage to the unofficial plugin recommendations: live GLM/Zhipu coding-plan quota gauges plus 30-day token history parsed from local Claude Code transcripts. Co-authored-by: Claude <noreply@anthropic.com>
…eboon#1312) Make the project header sticky at the top of the scroll area when a project is expanded, so the header and its session list are visually isolated: the header stays fixed below the search bar and only the sessions scroll beneath it. - Wrap the project header in a sticky top-0 z-10 container - Give the sticky header a solid bg-background so scrolling sessions do not show through - Restructure desktop Button background logic so the selected, starred and default states each provide a solid background for sticky mode Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
…iteboon#1277) * fix(codex): restore user prompts from canonical typed rollout rows Recent Codex releases write real user prompts as typed `item_completed` rows carrying `UserMessage` items; the legacy `event_msg/user_message` fan-out no longer exists in those files. The history reader only created user rows from the legacy event, so after a refresh the transcript showed assistant and tool activity without the user's own prompts. The reader now also builds user rows from typed `UserMessage` items, keeping the same per-turn anchoring rules and carrying local/pre-encoded images through to the transcript. Wire-level `response_item` rows are still deliberately ignored for prompts because they can carry internal content. When one file mixes typed and legacy copies of the same prompt (for example across a format boundary), the typed row wins and the legacy row is dropped so no duplicate bubbles appear. Verified against a real 0.153 rollout: user prompts from typed rows render identically to legacy prompts, alongside unchanged assistant/tool history. * fix(codex): dedupe typed and legacy prompts per turn The previous cleanup removed every legacy user row once a file contained any typed item. A rollout spanning a format boundary can have legacy-only turns before the typed era and typed-only turns after it; deleting all legacy rows lost the earlier prompts even though no typed copy existed. Track each legacy user row's turn id and drop it only when a typed UserMessage row exists for that same turn. Legacy-only turns and rows without an attributable turn are preserved. Regression test covers the mixed-era file with both prompt sources. * fix(codex): simplify typed user prompt restoration Codex legacy and paginated rollouts are distinct formats and never mix in one file, so the typed-vs-legacy deduplication was defensive code that could only be exercised by synthetic fixtures. Remove that state and post-read cleanup, and tighten the typed item reader and its integration test to the minimal behavior needed: restore UserMessage text and image inputs while ignoring wire-level user response items. --------- Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
…teboon#1333) * fix(opencode): compose the provider prefix into session model ids * test(opencode): keep the cleanup boundary around database initialization
…iteboon#1303) Both runtimes forward a stderr chunk to the chat verbatim as an error message. The CLIs style that output for a terminal, so the escapes reach the browser as literal text: ! Error �[93m�[1m! �[0m�[0mpermission requested: external_directory Clean the text before it is surfaced, and drop a chunk that turns out to have been styling only rather than showing an empty error bubble. Four private copies of an ANSI stripper had already grown in the backend, each with a different regex and none of them exported. Move one implementation to `server/shared/utils.ts` and point every caller at it. The shared pattern is the union of what those copies covered: OSC, CSI in both the `ESC [` and 8-bit introducer forms, and the remaining ECMA-48 escape sequences. Covered by unit tests. Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
Model replies that reference an image by workspace path, such as  after generating or screenshotting a file, rendered as a broken image: react-markdown emitted a bare <img src> that the browser resolved against the web origin, and the project files route needs the auth header anyway. Add an img override for chat markdown that fetches workspace paths through api.readFileBlob (the same authenticated route the file tree and attachment thumbnails use), shows the blob inline and expands it in the existing ImageLightbox. Web, data and blob URLs pass through untouched; a missing file falls back to the alt text. ChatInterface provides the selected projectId via a small context so every Markdown call site in the transcript resolves paths against the right project. Co-authored-by: pmoncadaisla <pmoncadaisla@users.noreply.github.com> Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
…lded in (siteboon#1295) * feat(chat): say a compaction out loud, as one row with its summary folded in Compaction is the most expensive thing a long session does without being asked, and the conversation said nothing about it: both records that describe one are `system` events, which normalize to nothing. What was left was the summary itself, a 15-25 KB assistant bubble arriving with nothing in front of it to say what it was. The server turns both into one ordinary assistant row -- everything they say is in `content`, so a client that knows nothing of the new `compact` field still reads the sentence: Compacted · manual · 335k → 10k tokens · 2m 22s Compacting conversation… Compaction failed: context still too large The client draws that row, folds the summary into it behind a disclosure, and reconciles the three rows a compaction actually produces: - the boundary (the numbers) and the flagged summary arrive in opposite orders live and from history, so the fold works from either direction; - the CLI writes the summary twice, once flagged and once into the live stream, and the unflagged copy is dropped wherever it landed; - the CLI also acknowledges compaction with the bare word `Compacted`, which is dropped only when a real row exists to replace it -- against a server that reports none it is the only trace there is. Closes siteboon#1292 Signed-off-by: Liran Funaro <liran.funaro@gmail.com> * fix(chat): fold the summary into a copy of the row, not the row itself The row a summary folds into can be one the projection cache handed back, which is shared across renders -- and a memoized row that keeps its identity while its content changes does not redraw. Replace it instead. Signed-off-by: Liran Funaro <liran.funaro@gmail.com> * fix(chat): supersede the row directly above, and never a user row Three things the review caught, all real. A `running` row was left standing under the row that said how the compaction ended, because only a summary-only row was being replaced. Both were then visible, live. A terminal row now replaces a running one — and carries over the summary if one had already folded into it, which is the order running → summary → boundary that live delivery produces and that would otherwise have dropped the summary on the floor. The stray-copy scan could take a *user* row that happened to repeat the summary text. Assistant rows only. The orphan-summary scan reached backwards past intervening rows, so a compaction that arrived later could adopt a summary belonging to an earlier one. Only the row directly above is superseded now; a compaction further back belongs to itself. Four cases added: running → done, running → failed, running → summary → boundary keeping the summary, an orphan left alone behind an ordinary row, and a user echo of the summary text surviving. Two of them fail without the first fix. Signed-off-by: Liran Funaro <liran.funaro@gmail.com> --------- Signed-off-by: Liran Funaro <liran.funaro@gmail.com> Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
Co-authored-by: blackmammoth <118998054+blackmammoth@users.noreply.github.com>
* fix(opencode): hide child sessions Signed-off-by: Jesper Derehag <jderehag@hotmail.com> * perf(opencode): prune child sessions once --------- Signed-off-by: Jesper Derehag <jderehag@hotmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
20 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This targets
fix/session-title-priorityrather thanmain, so merging it here makes siteboon#982 mergeable again without you having to redo the rebase yourself.It is a merge of current
main(c1be241) into your branch, so your two commits stay exactly as they are and remain ancestors — your branch fast-forwards onto this. I added one commit on top for the outstanding review comment.What conflicted
Only the test file.
claude-session-synchronizer.provider.tsmerged automatically.Since 10 July,
maingrew its ownserver/modules/providers/tests/claude-sessions.test.tscovering subagent transcripts, agent notifications and prompt editing, which collides add/add with yours. The two sets are disjoint — no shared test names — so I kept both and reconciled only the setup at the top of the file:main'swithIsolatedDatabaseand dropped the near-identical copy from your version, which differed only in the temp-directory prefixpatchHomeDirandwriteSessionJsonlhelpersClaudeSessionSynchronizerandbuildLookupMapimports your tests needThe file now runs 22 tests:
main's 10 plus your 12.The extra commit
CodeRabbit noted twice that
extractSessionAiTitleFromEndno longer describes what the function does — after your fix it scans forward, and it returns acustom-title,ai-titleorlast-promptrather than only an AI title. I renamed it toextractSessionTitleand rewrote the JSDoc in the prose style the rest of the file uses. It is a separate commit, so drop it if you would rather keep the original name.Verification
claude-sessions.test.tstsc --noEmit -p server/tsconfig.jsonoxlint server/The one failure is
conversation search streams title matches before transcript results, which fails at the same assertion inprovider.routes.test.ts:119on unmodifiedmain. It is pre-existing and unrelated to this change.I also reproduced the original bug independently on the published 1.37.2 before finding your PR, and confirmed it is still live on current
main. Your diagnosis and precedence chain match what I arrived at separately.