* fix(server): serve workspace list instantly on fetch, reconcile in background
Previously fetch_workspaces_request awaited reconcileActiveWorkspaceRecords()
before responding — this ran git operations (getCheckoutStatusLite) on every
active workspace, causing projects to appear empty for several seconds after
daemon restart or reconnect.
Now the registry snapshot is returned immediately and reconciliation runs in
the background. Any changed workspaces (e.g. stale worktrees being archived)
are pushed as workspace_update events through the existing subscription.
* fix(server): make workspace snapshot instant by removing per-workspace git calls
describeWorkspaceRecord was running two git operations per workspace:
- buildProjectPlacement (to refresh display name)
- getCheckoutShortstat (for diff indicator)
With 13 workspaces this caused fetch_workspaces_request to take 7-15s
(observed as ws_slow_request in logs).
Split into two methods:
- describeWorkspaceRecord: uses only persisted data, returns instantly
- describeWorkspaceRecordWithGitData: runs git ops, used for open_project
and create_worktree responses where fresh data is needed immediately
The snapshot path (initial load + background reconciliation) now uses the
fast variant. diffStat is null on initial load; it will be pushed later via
workspace_update once a git-aware path triggers.
* fix(server): reset session ID on query restart to prevent overwrite crash
When ensureQuery() restarts a query, the old claudeSessionId was retained.
If the new query landed on a different Claude session (e.g. after a hook or
reconnect), captureSessionIdFromMessage() would detect the mismatch and
throw a fatal "session ID overwrite" error, killing the agent mid-turn.
Reset claudeSessionId and persistence during query restart so the new
session can establish its own identity cleanly.
Closes#200
* fix(server): recover from session ID overwrite and load projects instantly
- Reset claudeSessionId unconditionally before every new query creation,
not just on explicit restarts. Fixes unrecoverable crash loop where a
pump failure left stale session ID, causing every subsequent query to
immediately throw the overwrite error again.
- Decouple fetch_workspaces_request from reconciliation: serve registry
snapshot immediately instead of awaiting git operations on all
workspaces. Background reconciliation runs after response is sent and
pushes workspace_update events for any changed workspaces (e.g. stale
worktrees getting archived).
* fix(server): auto-resume Claude session after overwrite error to preserve history
When the query pump dies (e.g. after a session ID overwrite error), preserve
claudeSessionId so the next query automatically passes resume: sessionId to
the Claude SDK. Claude responds with the same session ID, the overwrite check
passes, and the conversation history is intact.
Previously claudeSessionId was reset unconditionally before every new query,
causing the next query to start a fresh session and lose the conversation.
Now it is only reset for explicit restarts (queryRestartNeeded), where a
fresh session is the intended behavior.
* fix(server): reset session ID before throwing overwrite error to break retry loop
When the overwrite error fired, claudeSessionId was left set to the old value.
Our auto-resume fix then passed resume: oldSessionId on every retry, Claude
returned yet another new session ID each time, and the overwrite error looped
indefinitely (same "Existing" UUID across all failures in the log).
Fix: reset claudeSessionId = null in both captureSessionIdFromMessage and
handleSystemMessage before throwing, so the next query starts a fresh session
instead of looping on a dead resume target.
Auto-resume still works for legitimate pump failures (network drops etc.)
since those exit without throwing, leaving claudeSessionId intact.
* fix(server): accept session ID change mid-stream instead of failing the turn
Hooks can cause Claude to restart with a new session ID mid-turn. Previously
both captureSessionIdFromMessage and handleSystemMessage threw a CRITICAL error
when the session ID changed, failing the turn and leaving the user with an
error message on what should have been a transparent hook execution.
Now both paths log a warning and accept the new session ID, allowing the turn
to continue uninterrupted through hook-triggered subprocess restarts.
- Add CombinedModelSelector with two-level drill-down (providers → models),
favorites section, search, and single-provider auto-drill mode
- Add provider snapshot system (ProviderSnapshotManager) for cached provider/model state
- Add provider diagnostics with resolved binary paths and versions across all providers
- Add StatusBadge shared component for consistent status visual language
- Add ProviderDiagnosticSheet for detailed provider health inspection
- Add desktop auto-focus on input after any status bar interaction (model, mode,
thinking, features) using focusWithRetries
- Fix Combobox flicker on height change by switching to bottom-based CSS positioning
once initial floating-ui position resolves
- Add dynamic dropdown height: Level 1 fits content, Level 2 calculates from model count
- Remove "Other providers" section from model selector (only show available providers)
On macOS, apps launched from Finder/Dock inherit a minimal environment
(PATH is just /usr/bin:/bin:/usr/sbin:/sbin). This caused two problems:
1. Agent binaries like codex were not detected (findExecutable failed)
2. Terminals spawned by Paseo had no access to user-installed tools
(node, bun, direnv — all "command not found")
Fix: at Electron startup, spawn the user's login shell and capture its
full environment via JSON.stringify(process.env) using UUID markers.
This is the same battle-tested approach VS Code uses. The resolved
environment is merged into process.env before the daemon starts, so
all child processes — agents, terminals, git operations — inherit the
correct environment automatically.
This replaces the previous approach of invoking resolveShellEnv() (via
the shell-env npm package) at multiple scattered call sites. Now there
is a single source of truth: process.env is enriched once at startup.
Changes:
- New: packages/desktop/src/login-shell-env.ts (VS Code approach)
- Removed: resolveShellEnv(), shell-env dependency, $SHELL -lic probes
- Simplified: applyProviderEnv() and findExecutable() now trust process.env
Same fix as the Silero VAD — sherpa-onnx returns Float32Arrays backed
by native memory which Node.js rejects as "External buffers are not
allowed". Copy into JS-managed memory before passing to audio pipeline.
The bundled silero_vad.onnx lives inside Electron's app.asar which
native C++ (sherpa-onnx) cannot open via OS file I/O. On first voice
activation, copy the model to ~/.paseo/models/local-speech/silero-vad/
using Node.js fs (asar-aware) and point the native code there.
buildClientConfig in test-daemon-connection.ts never set appVersion, so
probe clients (which become the live client) sent hello without it. The
daemon's version gate then hid Pi/Copilot from all these sessions.
Also update appVersion on the Session when a client reconnects with a
newer version, so stale sessions don't stay gated forever.
worktreeRoot was added as a required field in 9154f8fc, which breaks
old clients parsing new daemon responses and new clients parsing old
daemon responses. Made it .optional() with .transform() fallbacks.
Also added a critical rule to CLAUDE.md: schema changes must always
be backward-compatible in both directions.
AgentProviderSchema was z.enum() which caused old clients to reject
session messages containing unknown providers (pi, copilot), breaking
the entire session. Changed to z.string() for future clients.
For currently deployed clients (<0.1.45), the daemon now filters out
unknown providers based on the appVersion sent in the WebSocket hello
message. Clients that don't send appVersion only see claude/codex/opencode.
* fix(app): reorder settings sections for better grouping
* feat(app): add setup hint and paseo.sh link on mobile welcome screen
New app store users land on the welcome screen with no context. Show a
brief explanation that the desktop app or server is needed, plus a link
to paseo.sh — only on mobile (iOS/Android), hidden on web/desktop.
* docs(release): add pre-release sanity check and clarify changelog scope
Add a Codex 5.4 review step before cutting releases to catch breaking
changes and backward-compatibility issues (mobile apps lag behind
desktop/daemon updates). Clarify that the changelog always covers the
delta from the previous stable release, not from the last RC.
* feat(server): add Pi agent provider and re-enable Copilot
Add Pi (pi.dev) as a new ACP-based agent provider with bundled pi-acp
adapter. Pi-acp reports thinking levels as ACP modes and bash tool calls
with kind "other", so the Pi provider uses generic ACP extension hooks
to remap these to the correct Paseo concepts:
- sessionResponseTransformer: remaps ACP modes → configOptions[thought_level]
- thinkingOptionWriter: writes thinking via setSessionMode instead of
set_config_option (which pi-acp doesn't support)
- toolSnapshotTransformer: remaps bash tool kind from "other" to "execute"
- modelTransformer: cleans slash-prefixed model labels
Also re-enables the Copilot provider (disabled since 44da0c67) and
removes the claude-acp provider.
* fix: update lockfile signatures and Nix hash
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* fix(server): deduplicate workspaces by resolving to git worktree root
Agents running in subdirectories of the same git repo were creating
separate workspace entries in the sidebar. Now workspace IDs resolve
to the git worktree root (via the already-computed `worktreeRoot` from
`inspectCheckoutContext`), so all agents in the same checkout share a
single workspace. Stale subdirectory workspace records are archived
during reconciliation.
* fix(server): fix 3 broken test files in server unit suite
- relay-reconnect: move speech mock to correct constructor position and
rename getSpeechReadiness→getReadiness after SpeechService refactor
- commands-poc: add credential/CLI availability guards matching
claude-agent.integration.test.ts pattern
- claude-sdk-behavior: pass pathToClaudeCodeExecutable via findExecutable
and add credential guards matching real provider configuration
* fix(server): fix race condition in loop-service test mock
Replace setTimeout(..., 0) with queueMicrotask() in ScriptedAgentSession
so turn_completed events fire before the verify check runs. Fixes flaky
CI failure where the file write hadn't completed before verification.
* fix(server): use /bin/sh for loop verify checks instead of /bin/zsh
More portable across CI environments. Also use queueMicrotask in test
mock to fix race condition between event emission and waiter setup.
* fix(server): correct import paths for isCommandAvailable and findExecutable
Import from utils/executable.js (where they're exported) instead of
provider-launch-config.js (which only imports them internally).
Introduce a generic feature system where providers declare dynamic
features (toggles/selects) and the app renders controls automatically.
One message pair (set_agent_feature_request/response) handles all
feature mutations. Feature values persist and restore on agent resume.
First consumer: Codex fast mode (service_tier) — gated to supported
model families, with proper cleanup on model switch.
Electron apps on macOS inherit a minimal PATH (/usr/bin:/bin:/usr/sbin:/sbin)
that doesn't include Homebrew or other user-installed tools. This caused
GitHub features (PR status, Create PR) to show as unavailable in the
desktop app even when gh CLI was installed and authenticated.
Extract findExecutable, resolveShellEnv, and related utilities from
provider-launch-config.ts into a shared utils/executable.ts module.
Use findExecutable("gh") with lazy caching so the login-shell resolution
happens once per daemon lifetime. Pass resolveShellEnv() as the process
env for all gh subprocess calls.
Update all consumers to import directly from utils/executable.ts —
no re-exports through provider-launch-config.ts.
Remove the hardcoded mode whitelist so user-defined agents (from
opencode.json or global config) appear in the mode picker. Add a
listProviderModes API (client → session → provider) so the draft
agent form fetches dynamic modes instead of relying on the static
manifest. Fix slash command dispatch to accept optional arguments.
Closes#185
Stop auto-archiving workspaces when all their agents are archived —
only archive when the directory no longer exists on disk. This prevents
projects from disappearing when users archive their last agent.
Add a kebab dropdown menu to all project rows (desktop) with a
"Remove project" action that archives all workspaces in the project.
When the desktop app starts and finds a running daemon with a different
version, it now restarts it automatically. Only applies to daemons the
desktop app itself started, identified by a PASEO_DESKTOP_MANAGED flag
written into the PID lock file.
Also rewrites resolveStatus() to use the CLI daemon status command
instead of manually parsing the PID file, and surfaces daemonVersion
in the CLI status output from the WebSocket hello message.
The daemon is a detached process that survives Electron restarts,
so after an auto-update the old daemon version would keep running.
Now we stop it before quitAndInstall so the new app instance
starts a fresh daemon with the updated binary.
These providers cause old mobile clients (<=0.1.40) to fail parsing
the list_available_providers_response because their AgentProviderSchema
enum rejects unknown provider IDs, dropping the entire message and
leaving the model picker empty.
The provider implementations remain in the codebase — only the manifest
entries and factory registrations are removed until the updated mobile
app (0.1.43) is live in the App Store.
* feat(server): add ACP base provider with Claude ACP integration
Implement a generic Agent Client Protocol (ACP) base provider that any
ACP-compatible agent can extend with minimal code. Includes a concrete
Claude ACP implementation via @agentclientprotocol/claude-agent-acp
with full parity to the existing Claude Code provider.
The base handles subprocess lifecycle, streaming translation, permission
bridging, terminal/fs callbacks, listModels, loadSession/resume, and
mode/model management. The concrete class only specifies the command,
modes, and availability check.
* fix: update lockfile signatures and Nix hash
* feat: add Copilot ACP provider, eliminate hardcoded provider unions, fix ACP streaming bugs
Add Copilot as an ACP provider (copilot --acp), with real modes and models
discovered from the ACP server. Fix two ACP base bugs: duplicate assistant
text (emit deltas not cumulative) and idle→running→stuck (fire-and-return
startTurn). Replace all hardcoded provider string unions with string/manifest-
derived values so adding a provider only requires: impl class, manifest entry,
registry factory, icon, and E2E config. Add provider docs and Copilot icon.
* fix: update lockfile signatures and Nix hash
* feat(app): add OpenCode provider icon
---------
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
The SDK's supportedModels() API hides the 1M context variant behind a
"default" alias and doesn't expose it as a selectable model. Replace
dynamic SDK discovery with a hardcoded catalog that exposes both
claude-opus-4-6[1m] (1M context) and claude-opus-4-6 (200k) as
distinct models. Delete sdk-model-resolver which parsed SDK descriptions.
Add listCommands() to OpenCodeAgentSession via the OpenCode SDK's
command.list() API, and route recognized /commands through
session.command() instead of promptAsync(). This fixes issue #168
where slash commands didn't load when OpenCode was the selected harness.
