Watching again, or reading once, after the catalog was invalidated or failed
while nothing watched now starts one deduplicated refresh; the observer
refetches on mount. Reconnect repair only refetches watched queries, so an
unwatched catalog stayed stale.
Each subscribe installed its own session and draft store callbacks, so N
watchers received N squared notifications, including for other agents' drafts.
One pair of subscriptions now serves all listeners, and the draft callback
ignores drafts of other agents.
A composer whose agent was removed while it was held was only released on the
next session change. The last holder letting go now checks too, and an archived
agent's idle composer is releasable.
PluginComposersSource.subscribe watched the stores directly and never the
composer model, so a component using useComposers did not keep the model's
catalog observer alive and reconnect repair skipped it. subscribe now takes the
selector and holds the model it names, following focus.
Dropping a composer when it went idle left a command a plugin had saved calling
a dead model after the tab reopened. A composer is now dropped when its agent
leaves the session and nothing watches it or has one mounted.
setAgentModel and setAgentThinkingOption checked the provider's global catalog,
while the picker reads the catalog of the agent's directory, so a model only
the project lists was rejected. The validation input now carries the directory.
The composer read the catalog once with no observer, so reconnect repair, which
refetches only active queries, never refreshed it and a failed load never
retried. While anything watches, the composer now observes the catalog query.
The shared composer map kept every agent's composer for the session. It is now
dropped when nothing watches it and no composer is mounted.
- `{ focused: true }` is the agent Command Center hands to agent items, published by its
registration through plugins/focused-agent.ts and empty outside a workspace. The composer source
reads that module instead of the remembered workspace selection, the navigation store and the
layout store, which also removes the require cycles those imports added.
- A composer the app retains in a workspace the user left has no live input. `insertText` and
`setPrompt` now store the text with `replaceDraftText`, which asks the mounted input to show it;
before, the text reached the store but the retained input never displayed it.
- `model` and `thinking` read exactly what the picker resolves before the catalog loads (no
models), then re-resolve when it lands, instead of reading null.
- The daemon validates a thinking option against the model the agent runs: the runtime model when
the provider lists it, else the configured one.
- The composer sources subscribe to query-cache updates only. A query another component creates
during its render is announced as `added`, and a subscriber that updates then made React warn
about updating a component while rendering a different one. A catalog read no longer starts its
load synchronously either.
useComposers(selector) in any plugin component and composers(selector) on every command context return the composer of a running agent, named by id or by focus, whether or not its tab is open.
A running agent now has one store-backed composer model apart from any mounted tab. The mounted composer plugs its cursor-aware insert in while mounted; otherwise insertText and setPrompt write the agent's saved prompt. model and thinking resolve through the provider catalog with the picker's own function, so a read equals what the picker shows.
The host reads one ComposerModel from context and no longer knows whether it is in an agent, a draft or New workspace. The daemon rejects an unlisted model or thinking option for a running agent, replacing the app-side guard.
The screen attached the pick side effects (remembering the isolation, syncing the picker's PR
attachment) as listeners in its own useEffect, which runs after a child item's mount effect. A
mount-time setIsolation("worktree") switched the screen but was not saved, and setBaseRef skipped
the attachment sync. The effects are now injected when the setup model is built, so a command and
a pick run the same code from the first render. Closing the picker stays in the picker's handler.
A command called from a New workspace item's mount effect rejected because the
setup model only received its inputs and the pickers' handlers from a passive
effect on the screen, which runs after a child's. The model now takes its
inputs at construction, picks the base and isolation itself, and is refreshed
in a layout effect; the screen supplies the side effects of a pick (the PR
attachment, the remembered isolation) as listeners.
The branch a plugin sets was dropped when no base was resolved, such as on a
detached HEAD, because it was sent only for an explicit branch-off. It now
applies whenever the worktree cuts a new branch, decided in one place that
mirrors the daemon.
baseRef.name was the display-shortened branch name, so setting back what was
read based the worktree on the local branch instead of the origin one. It is
now the exact ref creation uses, and host and baseRef are frozen like project
and metadata.
addNewWorkspaceItem mounts a plugin field on the New workspace screen, and
useNewWorkspace() exposes host, project, isolation, base ref, branch and
metadata with commands that follow the composer item rules: they act on the
screen's state when they run, resolve once applied, reject while the screen
is locked, and take JSON payloads. Host and project are read-only because a
command that switched them would unmount the item that called it.
The screen's pickable state (base selection, manual isolation, branch,
metadata) moves into one plain store so pickers and plugin commands write to
the same place; the screen reports derived facts to it on each render. The
branch is sent as source.branchName only for a branch-off worktree, and
metadata on workspace.create.
Items render through ComposerControl with a new "field" surface (badge on
desktop, full-width field when compact). Combobox ships here: Paseo owns the
loading and error states, keeps previous results, and drops superseded
searches.
Draft commands read the state captured at render, so a plugin's
`await setProvider(); await setModel()` acted on a stale provider. The draft
form and options are now plain stores that commands read live, and
`setProvider`/`setModel` reject an unavailable provider or a model the loaded
catalog does not list. Draft `options` include the plugin's pending edits.
Shared composer controls own the pending state and show a rejected command as
an error toast, like the built-in controls. Callback payloads are objects
(`{ value }`, `onPress(event)`) so they can grow without breaking plugins.
The overflow sheet's content portals to the app root and lost the composer's
agent, input, toast and query contexts, so plugin items drew nothing in it.
`useComposerSheetBridge` carries them across.
A plugin registers a component with addComposerItem and reads and changes the
composer through useComposer(): provider, model, mode, thinking, options,
prompt, attachments and agent metadata. The agent, draft and New workspace
composers provide the same ComposerInput and ComposerAgent halves, so a
command reaches the right model without the plugin knowing which composer
it sits in.
Toggle, Select and Action are drawn by one ComposerControl renderer that
the thinking select and provider option controls now use too. Controls sit in
the toolbar when there is room and become rows in the existing overflow sheet
otherwise, so on a phone they appear in the overflow. Plugin items count in
the density presence so the composer condenses to make room. Model and mode
stay specialized but share the trigger, density and overflow.
Draft metadata is kept locally, sent at creation and reset when the prompt is
sent. PluginAgentSnapshot.metadata, deferred from the metadata phase, lands
here. The unused provider badge is removed.
vitest resolves @floating-ui/react-native to its ESM build: the CJS entry
requires the real react-native, whose Flow sources do not parse.
Plugins had nowhere to keep their own data on an agent or workspace, so
Phase 5 and 6 surfaces would each have invented a side store. Both now
hold an opaque Record<string, JsonValue> that Paseo stores, snapshots and
hands to hooks without reading a key.
Set at creation or changed with agent.metadata.update.request /
workspace.metadata.update.request (top-level keys replace, null removes).
Persisted on the agent and workspace records, so it survives a restart.
Reaches agent.create.before (which may add to it), agent.open.before
(read-only), agent.create.after and workspace.create.before/after.
Gated once on server_info.features.entityMetadata; the client throws
"Update the host" instead of sending to a host without it.
A provider plugin reports `controls` and the `providerOptions` it runs in
`session.config`, and receives running-session changes as
`session.configure` `changes.providerOptions` (null returns a key to its
default), under the same apply-or-throw contract as built-in providers.
A control may declare a `defaultValue`; a value a plugin reports at that
default for a key nobody set stays unset instead of being pinned to the agent.
A plugin that still reports `settings` keeps working: the daemon reads each
setting as a control plus an option and sends changes in both shapes
(COMPAT(plugin-provider-settings)).
The app resolves a control's icon as a Lucide name when it is not one of the
built-in kebab-case names, so a plugin's icon shows live and in drafts.
Public docs and the provider-direct example use only `providerOptions` and
controls; the migration doc gains the rows.
Which options a legacy daemon exposes as features depends on the model, mode
and thinking level, not only the provider: a Haiku snapshot with no features
says nothing about Opus. The provider-keyed memory of control keys sent an
Opus create's fast_mode as providerOptions, which that daemon does not read.
A create that carries options now lists the controls for its own draft
config and splits on that answer. The memory, and the snapshot and listing
hooks that fed it, are gone.
A cold create against a daemon without providerOptionControls routed control
options (Claude fast_mode, Codex service_tier) as providerOptions, which that
daemon does not read for them. The client now lists the provider's features
before the first create or workspace create that carries options, so callers
never prime a cache.
Draft options moved into a plain model. Applying a profile replaces provider
and options in one step, so the form's provider change that follows no longer
clears a cross-provider profile's non-control options; choosing a provider by
hand still starts over. The model sends its complete options (seed, last-used
values once the provider says it still draws them, edits) to the daemon and
the controls draw only the answer, so a remembered legacy fast is drawn as
what it runs as. A toggle shows over the last answer until its own arrives.
The profile form does the same: its edits go through resolution too.
The agent snapshot's `options` are the state and `controls` only draw them.
Running and closed agents draw from the snapshot; drafts and profile forms ask
`provider.controls.list.request` with their `providerOptions`, so aliases such
as a legacy Fast profile's `service_tier: "fast"` are resolved once, by the
daemon, and the draft keeps Fast highlighted. A draft adds only the user's
edits and each provider's last-used control values; applying a profile seeds
its non-control options (e.g. `sandbox_mode`) into the draft, and a profile
form saves them back instead of pruning to the controls it can draw.
A daemon without `server_info.features.providerOptionControls` still sends
`features` and `featureValues`. `LegacyProviderOptions` in the client package
is the one place that translates snapshots, draft listings, create and apply
requests, and daemon-config profiles, so the app never branches on it.
Persisted preferences and tab layouts rename `featureValues` through tagged
COMPAT preprocessors.
The features/controls bridge moves from the server into the protocol package
so the client can reuse it.
The rename is adapter-specific, so it lives where the provider id resolves to its adapter:
AgentManager converts an agent's own options on create, resume, import, reload, apply and
draft controls, following derivedFromProviderId, so a custom provider extending codex keeps
working with a saved or requested fast_mode. Storage, MCP, schedules, hub and profiles pass
options through. The featureValues-to-options fold stays provider-agnostic at the legacy edges.
Stored records, config.json defaults, profiles, older clients' requests, MCP settings and
plugin hooks go through convertLegacyOptions as they come in. Past that point the defaults
merge, the manager and the Codex session never see fast_mode for Codex, so the alias branch
in the defaults merge and Codex's fast_mode handling are deleted.
effectiveOptions was recomputed from the current config.json defaults on every state emission, so editing a default changed the snapshot of an agent whose session never received it. It is now computed when the session opens and moves only through an apply or an options_changed. Editing a default affects agents the next time they open.
A session with no applyOptions (Pi) now rejects a patch instead of storing a value that
fails on reopen. Codex applies only service_tier, plan_mode and fast_mode live, since the
approval policy and sandbox are fixed when the thread starts. Claude also rejects a launch
option change while its first query is starting, after the options have been copied.
Replacing a started query to apply the option would end its background shells and
tasks. Once the query exists only Fast can change; before the first prompt every option
still applies, because it is read when the query is built.
A stored option has to describe what runs. OpenCode, OMP and ACP now reject a live change
to a key they apply only when the agent opens; Claude applies such a change from the next
turn and rejects it while a turn is running. Opening reports every select or setting the
session came up with, including ones nobody set, so a native default (Cursor opening with
Fast on) is stored as option state like any other.
applyOptions returns the options the session ended up with and the session
exposes its open difference as openOptionsPatch, so the manager stores the
patch and the provider's answer in one step. Removal is an ordinary patch:
the provider receives the default the key falls back to, or null, and returns
what it kept. Controls no longer take a default from a stored legacy feature
value.
Providers used to write their own reports into the agent's options while a session was being
created, resumed or updated, which needed a growing set of reconciliation, baseline and
suppression rules to stay in step with the daemon.
The contract is now transactional. A provider opens or updates with the options it is given and
emits nothing while doing so; if the result differs it says so once, afterwards, as
`options_changed`. Live apply hands the provider only the keys the patch sets and the daemon stores
them onto the agent's current options when the provider accepts, so the last write wins. A change
the provider makes on its own is stored and forwarded, not undone. Removing an option drops the key
and sends nothing live. Controls carry no state; `options` is the only option state.
A legacy `agent.create.before` hook's `featureValues` edits are applied as a patch over the
returned options, so its deletions and replacements hold. The unreleased `aliases` field on option
choices is gone; a pre-v0.10 Codex Fast preference is reported once as the model's tier id.
* feat(server): add background agents
An agent can be created background in any workspace. It stays a durable,
addressable agent, but default discovery, workspace status and attention
leave it out. Effective background is agent.background ||
workspace.background.
- Protocol: `create_agent_request.background` now marks the agent itself
and stays the workspace intent for a workspace the request creates.
Snapshots carry `background`. The daemon advertises
`features.backgroundAgents`. Clients advertise the `background_agents`
capability.
- Daemon: the flag persists in the agent record, so restart and resume
keep it. AgentDirectory owns visibility. Inclusive reads and
subscriptions from sources without `background_agents` see only agents
in background workspaces, because older apps would otherwise open a tab
for each one.
- Inheritance: omitted on agent-scoped creation, children and new
workspaces take the caller agent's background.
- MCP: create_agent accepts `background`.
- SDK: placing a background agent in an existing workspace, or under a
caller, fails fast with "Update the host to use background agents." on
older hosts.
- CLI: `ls -a` and `ls --background` add a BACKGROUND column.
* feat(app): hide background agents from tabs, track and status
* docs(plugins): point hidden helpers at background agents
* fix: keep background agents hidden across cache reload and connected hosts
* fix: keep remote creation working on older hosts and inherit foreign parents' background
* feat(app): open plugin settings from a cog with a tab per screen
The plugin row shows a settings button next to its switch when the plugin
registers settings screens. The actions menu keeps Logs, Reload and Remove.
The settings page shows one tab per registered screen; a plugin with one
screen shows no tabs. Tabs swap screenId in place, so existing
plugins/[pluginId]/[screenId] links and openSettings keep working.
* fix(app): render settings tab icons inside the plugin's runtime
Component icons now mount under the installation provider, so plugin SDK
hooks such as useTranslation work. Each icon has its own error boundary;
an icon that throws leaves its tab with the label alone.
* feat(app): name only the plugin in its settings page header
The header shows the plugin's manifest name, or its id when it has none,
for single- and multi-screen plugins alike. The selected tab names the
screen. The Plugins page and the header share one plugin list query.
* feat(app): show plugin display names and a Settings tooltip on the cog
Plugin rows on Settings → Plugins show the manifest name, falling back to
the id; when a name is set, the id stays visible beneath it. The row's
controls, logs sheet, confirmations and feedback use the same name.
The cog is a ToolbarButton labeled Settings, with a matching tooltip.
* fix(claude): show thinking options for models from settings.json
Models appended from Claude's settings.json were pushed to the catalog
without thinkingOptions, so the composer had no effort selector for them.
Resolve them like profile and additionalModels entries: manifest options for
built-in IDs, the standard Claude effort levels otherwise.
* docs(claude): clarify effort selection for settings models
Describe explicit effort selection without promising how every client initializes an unsaved value.
Co-authored-by: Okan YILDIRIM <63511519+ragokan@users.noreply.github.com>
Co-authored-by: wangyahao <wangyahao@fxiaoke.com>
---------
Co-authored-by: Okan YILDIRIM <63511519+ragokan@users.noreply.github.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: wangyahao <wangyahao@fxiaoke.com>
* fix(android): correct inverted chat wheel direction
* test(android): run wheel regressions with current React Native
Provide the native-only test context with React Native’s event dispatcher for an absent JavaScript instance, retaining real scroll handling.
Co-authored-by: Seth Jeon <paseo@sethjeon.com>
---------
Co-authored-by: Seth Jeon <paseo@sethjeon.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* feat(keyboard): support hardware keyboards on iPad
Register app shortcuts as iOS UIKeyCommands so a paired hardware keyboard
drives the same bindings as web and desktop. Overlays (menus, comboboxes,
modals, command center) get first refusal on bare keys, and a focused
terminal keeps the keys it needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(keyboard): record iOS hardware key observation spike
With a UITextView focused on iPad Pro M4/iOS 18.6, window and controller observers received Cmd+B, Cmd+P and Escape. Escape also invoked UIKeyCommand when registered, so the stream must be the sole emitter. Use observation for shortcuts and commands only for text-control ownership. Simulator intercepted Cmd+K as HID184; input capture blocked synthetic typing entirely. Keep capture off and verify the toggle via a rebind during simulator QA.
* refactor(keyboard): give overlays one key registration API
Overlay ownership now follows one layer-ordered registry on every runtime. Web retains focus trapping and restoration through its adapter; native registers the same key handler without a DOM scope. Picker navigation and dismissal no longer require a parallel native stack.
Co-authored-by: jeroenfrenken <16974549+jeroenfrenken@users.noreply.github.com>
* feat(keyboard): route native hardware keys through the shared resolver
Observe each physical press once, including native modal windows, and derive focus from declared composer and terminal surfaces. Native bindings and rebinds now use the same resolver and overlay precedence as browser and Electron input.
Co-authored-by: jeroenfrenken <16974549+jeroenfrenken@users.noreply.github.com>
* feat(keyboard): show shortcuts when hardware keyboards are present
Keep routing independent of compact layout and detect hardware capability for hints and settings. Rebind capture uses the shared overlay interface on native too. Keep the filtered sidebar model available for keyboard navigation while its panel is hidden.
Co-authored-by: Les Stroud <public@lesstroud.com>
Co-authored-by: jeroenfrenken <16974549+jeroenfrenken@users.noreply.github.com>
* test(keyboard): allow shortcut hints in compact menu assertions
Compact menus now display hardware keyboard hints. Assert action labels, order and uniqueness without treating a trailing shortcut badge as part of the action label.
* fix(keyboard): retain an overlay's opening layer in state
Capture the registry's current top when an overlay opens and retain that layer through rerenders. React Compiler removed the previous useMemo cache after the registry getter moved across modules, allowing a parent modal to rise above its child and start an infinite focus loop.
Compile this hook in Vitest as Expo does. The new hook regression failed first with layer 220 instead of 20, then passed with explicit opening state. Focused unit and browser tests pass, and the previously failing onboarding E2E completes through the real Expo compiler. Typecheck, lint and format pass.
* refactor(keyboard): let overlays own native hardware input
Remove the TextInput key handlers that duplicate declared overlay arrows and Escape in Add Project and Command Center. Remove Add Project's hidden responder now that native window sources observe keys without it. Keep soft-keyboard submit editing behavior.
* fix(keyboard): close Add Project outside its state updater
Root Escape previously called the parent close action from a React state updater, producing a cross-component render warning when navigation was still queued. Compute the previous page in the event handler and close outside the updater.
The existing Escape/Back E2E now batches navigation and dismissal through the DOM key source. It failed first with the AddProjectFlowHost render warning, then passed after the correction. Typecheck, lint and format pass; native Escape checks pass on Android and iPad.
* fix(keyboard): preserve Android dead-key composition
* fix(keyboard): recognize focused native text inputs
* test(keyboard): replace JSDOM coverage with browser tests
* refactor(menu): spell out arrow navigation branches
* fix(composer): claim fullscreen Escape on native
* refactor(composer): one fullscreen Escape module for every platform
The shared registration lived in fullscreen-exit.shared.ts with two one-line
re-exports (fullscreen-exit.ts and fullscreen-exit.native.ts). Metro needs no
native override now that the Android back button is its own adapter, so the
implementation moves into fullscreen-exit.ts and the browser test imports the
module every platform resolves.
---------
Co-authored-by: Ubuntu <jeroenfrenken@icloud.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: jeroenfrenken <16974549+jeroenfrenken@users.noreply.github.com>
* fix(workspaces): keep merged pull requests visible
A remote update before merge can leave the checkout behind the PR head. Match terminal requests by branch for display and require an exact fresh head match before automatic archive.
* fix(forges): retain pull requests targeting a fork
A fork checkout can have its own pull request as well as one targeting its parent. Compare both candidates before choosing which request to display.
* fix(forges): rank closed drafts as terminal
GitHub retains the draft flag after closing a pull request. Selection must use the request state so a closed draft does not override an open or exact-head request.
* feat(app): link #123 issue and pull request references in Markdown
Agents cite issues and pull requests by number; reading one meant looking
up the URL by hand. Markdown in a workspace now links forge shorthand
against the workspace's origin remote: `#N`, GitLab and Gitea `!N`, and
cross-repository `owner/repo#N` on the same host.
Each forge's URL grammar declares which sigil maps to which path, so the
syntax stays per forge: GitHub has no `!`, and its `#` targets /issues/,
which GitHub redirects to the pull request. Recognition goes through the
Markdown engine's text link regex so code spans, fences and authored links
stay plain without a second Markdown parse while streaming. The reference
alternative sits before path recognition so `group/project#12` is not read
as a file path.
* fix(app): keep http:// for self-hosted forge web links
A self-hosted forge cloned over http:// serves its web UI over http on
the same origin. Forcing https made reference, blob and tree links fail
against an HTTP-only port. Cloud hosts and SSH remotes stay on https.
The library's web renderer marked an image inline only when it was a
direct child of a paragraph with other content. A linked image sits
under a Link node, so badges such as a review bot's
`[](url) Confidence Score` header rendered as
100%-wide, 320px-tall block images in the PR panel and file preview.
Images nested in links (or emphasis) now count as inline when their
paragraph has other content; a paragraph made only of images or
image-only links still renders them as blocks. Inline images now take
the line height with automatic width instead of a line-height square,
so wide badges and photos keep their aspect ratio instead of squashing.
* feat(app): add first-run setup modal
New users landed on an empty "Add a project" screen and never found
providers, Paseo tools or phone pairing, which all live in Settings. A
one-time setup modal now walks them through Appearance, Providers,
Projects, Orchestration and Phone. Every step renders the same component
Settings uses, so changing something in setup changes it in Settings too.
The modal opens only when the setup host is connected, its workspaces have
loaded and it has no projects. Existing users already have projects, so
they never see it after upgrading and no migration is needed. Finishing or
skipping writes `@paseo:onboarding` on this client. It uses its own key
because it records feature state, and e2e specs replace app settings
wholesale. Settings > General > "Show setup again" reopens the modal on
step 1. Native phones skip the modal for now.
Shared pieces moved out so setup and Settings render the same thing:
- ThemePicker is a tile grid and replaces the Settings theme dropdown.
- PaseoToolsCard is extracted from the host page.
- The skeleton pulse moves to components/ui. Provider rows show it while
the first snapshot loads instead of "Loading…".
- The pair-device direct-connection hint links to the Tailscale docs.
The e2e fixture marks setup completed by default so existing specs keep
their starting state. The desktop smoke and lifecycle scripts, which start
from a fresh daemon, mark it before they continue.
* feat(onboarding): suggest projects from agent history
The setup's Projects step lists the repos the user already ran agents in,
with busy recent ones preselected, so a new user adds their projects in one
click instead of browsing for folders. With rows checked, the footer's main
button adds them and only moves on once every one was added; a checked row
always means "this will be added".
The daemon resolves each session folder to its project root (main checkout
for worktrees, repo root for subfolders) because only it can read the git
state; the app asks once per provider so a slow or broken provider never
holds up the others and can be retried on its own. Hosts without the
projectSuggestions feature show only the Add project button.
Onboarding e2e hosts read an empty Claude history with other providers
off, so suggestions never depend on the machine running the tests.
* fix(onboarding): keep late project adds on their step and show Retry early
- A finished add only advances setup while the Projects step is still open.
- A failed provider shows its Retry as soon as any provider answers.
- The desktop lifecycle check no longer reloads the window: setup never opens
there because no host is connected.
* fix(onboarding): bind late project adds to their step visit and freeze rows while adding
- Each step visit has an id; a finished add only advances the visit that started it,
so leaving and returning to Projects, or reopening setup, ignores it.
- Suggestion rows are disabled while their batch is being added.
* test(app): drive theme selection through the tile picker
Settings → Appearance no longer has a theme dropdown. The Escape test opens the
syntax theme dropdown instead, and the Android system-bars check taps theme tiles.
* fix(server): return project suggestion roots in the form project.add stores
Git reports Windows repo roots with forward slashes, so an added suggestion kept
showing because its path never matched the new project's root.
* feat(app): two control sizes, orchestration switches and mobile connect choices in setup
Controls had four sizes picked per call site, so a button in a dialog, a
settings row and a setup step came out at different heights. Controls now
have two: md (default; 32px with a pointer, 44px in the compact layout so
phone controls are full touch targets) and sm (28px, for tight chrome and
nested settings rows). The breakpoint picks md's height, not the caller.
Setup:
- Orchestration is three switches (Paseo tools, agent skills, CLI), all on,
applied on Continue. Turning one off undoes it; the desktop CLI uninstall
removes only the file this app installed.
- The Mobile step has store buttons and a "Choose how to connect" section:
Relay opens the existing pairing modal, Tailscale opens the docs.
Settings section docs links are one pattern, SettingsSection's `link` prop,
and modals sit on surface0 like a screen.
The vitest unistyles stub now resolves breakpoint values by window width,
as Unistyles does at runtime; without it md controls had no height in
browser tests.
* fix(app): finish setup on its last step and confirm undoing orchestration
In the compact layout Orchestration is the last step, and Finish closed
setup without applying its switches. The last step's work now runs before
Finish, the same as Continue.
Turning off skills now asks first, as Settings does. Removing the CLI keeps
a paseo command this app didn't install and says so. A failed Paseo tools
update shows a toast. Adding suggested projects no longer advances when a
row was checked while the add ran.
Chat sender attribution and the message preview toggle go back to the sm
button: md padding is a breakpoint value, and on web its media-query rule
beat the zero right padding that aligns the sender with its bubble.
* fix(app): ignore a late Finish once the user left the last setup step
Finish called complete(), which ignored the step visit, so pressing Back
while Orchestration applied its choices still closed setup and recorded it
as done. next(visit) now finishes setup from the last step, so a late
Finish is dropped the same way a late Continue already was.
* fix(app): hold the skills switch while setup applies its choices
The Paseo tools and CLI switches were disabled while Continue or Finish
applied them, but the skills switch only locked while a skills operation
ran. With skills needing no change, it stayed editable during a pending
tools update and the new value was silently dropped.
* fix(app): read setup's skills choice as off when no skill is picked
Saving an empty selection in the skills picker left setup's skills switch
on. Continue then asked the host to install nothing, the host reported
skills as not installed, and the step stayed put with no error. Setup now
reads skills as off with nothing picked, as Settings does.
* feat(plugins): stream events and bytes between plugins and the app
Plugins can declare a subscription with defineSubscription, handle it with
server.handle(contract, (input, { push, signal })), and listen from a surface
with useSubscription or client.subscribe. A top-level Uint8Array in an event,
an RPC output, or an RPC input crosses the wire as a binary frame (opcodes
0x20-0x2f) without JSON or base64; a nested one fails validation on the sender.
Each subscriber runs its own handler, owned by the session that subscribed,
so release, disconnect, plugin stop and reload need no plugin code. The
runtime buffers pushes until the subscribe response has gone out, so a client
never sees an event for a subscription it does not know yet.
Gated on server_info.features.pluginSubscriptions; binary outputs reach only
clients that advertise plugin_binary_payloads.
* fix(plugins): close subscription gaps found in review
- Losing daemon.manage now releases the session's plugin subscriptions;
binary events bypass the outbound permission check, so streams would
otherwise outlive the revocation.
- A subscribe that times out sends unsubscribe, since the plugin may
still start the handler after the daemon gave up.
- The byte limit leaves room for the largest frame header and routing,
and lives in assertPluginWireValue so events, outputs and inputs share it.
- A release right after the host accepted keeps `ready` resolved, so a
handler that returns at once reports closed rather than an error.
- Session trace logging tolerates loggers without isLevelEnabled.
* fix(plugins): withhold binary plugin results after plugin access is revoked
A binary RPC output or event skipped the outbound permission check its
JSON form passes, so an RPC in flight when daemon.manage was revoked still
delivered its bytes. Each frame now checks the permission of the JSON
message it stands in for.
* test(plugins): cover subscriptions in a real browser instead of JSDOM
The app subscription tests mounted hooks in JSDOM, which docs/testing.md
rules out. The adapter and contract behavior is now a plain unit test,
and a Playwright spec drives the diff-watch and media-stream examples:
JSON events following file changes, streamed bytes, and an upload.
* test(plugins): add the plain subscription unit test and the browser spec
* fix(plugins): apply subscription schema transforms once
Both ends parsed each subscription value, and the sender sent its parse
output, so the receiver parsed a schema's output as fresh input. An event
schema like z.string().transform((s) => s.length) closed the stream as
invalid. The sender now validates and sends the value as given; the
receiver's parse produces what the handler or callback sees.
* fix(plugins): parse a subscription's input once in the plugin process
server.handle wrapped subscription handlers in a second input parse, so a
transforming input schema rejected its own parsed output before the
handler ran.
* fix(app): add hosts without leaving the current screen
A retained Settings route replayed its addHost query after appearance remounts. Own the connection flow outside screen lifetimes and let each consumer choose its completion navigation.
* fix(app): release the project dialog before adding a host
* fix(app): size the QR scanner viewport on wide layouts
* test(app): align creation and desktop setup with in-place host flow
* fix(server): retry a connected host that failed the capability check
Every failure in a connected host's acceptance step was terminal: the
row went to "error" with no retry, and the status listener ignored
later reconnects. A host rejected with unsupported_capability stayed
unusable after it was updated, until the registration was saved again
or the managing daemon restarted. Capability and other non-identity
acceptance failures now go "offline" with nextRetryAt and retry with
capped backoff (1.5 s doubling to 30 s), so an updated host comes
online by itself. Self, duplicate, changed-identity and authentication
failures stay terminal.
Fixes#7971
* fix(server): keep a retried host's relay security floor and update guidance
A scheduled capability retry built a new transport factory, which
dropped the relay security floor learned from a protected handshake, so
the retry could negotiate legacy encryption and send credentials.
Retries now reuse the registration's transport factory (createHostClient
accepts one). The update guidance also stays visible while a retry
connects, the backoff resets after a successful acceptance, and
disabling or removing a host cancels its pending retry.
* fix(terminal): render emoji at their intended width
The native grid scales glyphs to the columns assigned by the terminal buffer. Use grapheme widths across current hosts and clients so emoji fit their slots and replay preserves cursor positions.
* fix(terminal): make width regression test portable
* test(terminal): wait for replay fixture output to finish
A trailing newline could arrive after the readiness marker and change the retained history between replay captures.
* fix(server): report a Codex thread only once Codex can resume it
Codex writes a new thread's rollout at its first turn, but the adapter
starts a thread when the agent is created and reported its ID at once.
AgentManager stored that ID as the persistence handle, so a Codex agent
that was reopened before its first prompt (idle runtime release, daemon
restart, plugin tool reload) failed every prompt with "no rollout found".
The adapter now reports a thread's ID, as runtime info and as a handle,
only once it is resumable: after Codex announces the thread's first
turn, or when the thread was resumed or forked. Until then a reopened
agent starts a new session, as an unprompted Claude agent does.
Document that a persistence handle must be resumable.
* test(server): reopen promptless Codex agents without model requests
Exercise the real app-server before any first turn, so publishing an unsaved persistence handle fails at the actual resume operation.
Co-authored-by: Jeong-Uk,Kim <dnrtoki@gmail.com>
---------
Co-authored-by: Jeong-Uk,Kim <dnrtoki@gmail.com>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
* fix(codex): start custom providers with reserved names
Codex rejects custom provider keys such as ollama before a turn starts.
Namespace the generated keys without changing user-facing provider ids.
Keep older sessions discoverable and select the new key explicitly when
resuming them, so the startup fix does not strand existing conversations.
Closes#6874
Co-authored-by: Egg3901 <rainfordmason@gmail.com>
* fix(codex): keep native sessions out of custom provider imports
* fix(providers): align custom provider example with generated config
Correct the documented internal provider key and simplify import filtering while preserving native and legacy session selection.
* fix(codex): exclude native GCP sessions from custom imports
* test(codex): seed sessions through the portable launcher
---------
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Co-authored-by: Egg3901 <rainfordmason@gmail.com>
Restore configured terminal history on reconnect, retaining 10,000 rows plus the screen. Serialize the worker-owned xterm buffer directly to ANSI and forward negotiated bounded replay chunks while preserving ANSI-era clients and per-viewer live ordering.
Preserve captured geometry through replay, bound pressure recovery to 500 rows, and keep the omitted visible-history default at 200. Remove the full cell-grid copy and obsolete JSON rendering.
All 21 CI checks pass. Real browser, native Debug emulator/simulator and Linux Electron coverage and platform limits are documented in the PR. The Android Debug invalid native delegate dispatch also reproduces on upstream main; its cause and any change in frequency remain unproven.
Credit Manav Dhindsa (DrakenKor), #6139 and #6148, and Lennart Schoch, #6743, for prior ANSI replay and retention work informing this change.
Closes#6137.
Closes#6147.
Shape rows in native Canvas and CoreText, then place clusters on xterm columns so fallback font advances cannot separate text from its cursor. Wait for measured dimensions before attachment and repaint overhanging ink with affected rows.
Use Android platform ICU to avoid bundling approximately 14 MB of data; retain and document the Android 10 conjunct limitation.
* fix(server): bound every timeline item so one fetch cannot close the socket
A timeline page is one WebSocket message. Only three tool-call fields were
capped, so MCP results, file contents, errors and merged assistant rows could
make a 20-entry or even a 1-entry fetch pass the outbound high-water mark.
limitAgentTimelineItemContent is now one budgeted walk over any item: 64 KiB
for tool calls, 256 KiB for other items. Over budget, the largest strings are
cut to a shared length with a visible marker and oversized data URLs become a
marker. Fetches also apply it to merged assistant and reasoning rows.
* fix(server): keep short strings whole and typed arrays valid when limiting items
Strings up to 1 KiB are never cut, so a tool call keeps its callId and an
item limits to the same result twice. Arrays drop trailing elements
without inserting a marker element, which typed arrays such as todo items
would reject.
* fix(server): respect environment proxies across daemon traffic
Install proxy routing at daemon and plugin process startup, with a separate HTTP agent for outbound WebSockets. Always exempt loopback so local provider bridges remain reachable.
* test(server): run usage source regressions in CI
* fix(server): validate proxy settings without blocking startup
Normalize scheme-less proxies and share HTTP-only validation across fetch and WebSockets. Record ignored variables without credentials in the daemon log. Separate built-in and directory plugin regressions through a typed fixture.
* refactor(server): treat ignored proxy variables as unset
* fix(server): read Windows proxy environment entries once
* fix(server): apply proxy policy to connected host links
* test(server): isolate proxy fixture credential directories
* fix(server): keep proxy routing and authentication consistent
Normalize Windows environment names and wildcard bypass rules at the network boundary. Compare bypass ports numerically and keep Basic proxy credentials on the CONNECT pool, including empty credential fields.
Backport the Unistyles 3.5.3 Activity configuration fix while keeping 3.2.4 and its existing fixes. Application resources could leave the responsive shell on the previous fold size after resume. Cover the real device-state sequence with a native UI regression.
* fix(usage): fetch only when shown and add a host-wide off switch
* fix(usage): restore policy defaults and isolate usage test state
* refactor(usage): give the sidebar footer sole ownership of its summary
* fix(app): preserve adaptive theme events while hidden
Read the current scheme through fresh matchMedia lists so hidden runtime
reads do not consume the pending events on listener-owned lists.
Carry Dan Sabin's fix from #7172 into the current combined Unistyles patch,
and cover both transition directions with a real Electron regression in
the existing desktop browser suite.
Co-authored-by: Dan Sabin <dan@freethinking.it>
* test(desktop): run adaptive theme regression on Linux
Electron retains --no-sandbox in argv, so the positional lookup used the script filename as the artifact directory. Read the final argument so the same real Electron regression exercises Linux as well as macOS.
Co-authored-by: Nurşat Bülbül <85260051+Nurshot@users.noreply.github.com>
Co-authored-by: Dan Sabin <dan@freethinking.it>
---------
Co-authored-by: Nurshot <85260051+Nurshot@users.noreply.github.com>
Co-authored-by: Dan Sabin <dan@freethinking.it>
Co-authored-by: paseo-bot[bot] <266920839+paseo-bot[bot]@users.noreply.github.com>
Mohamed Boudra