Remove the restriction that prevented users from deleting built-in
patterns (those with IDs prefixed 'pattern-'). Built-in patterns are
intended as starting points, not permanent fixtures.
Changes:
- Add deletedBuiltinPatternIds field to WorkspaceState to track which
built-in patterns the user has removed
- Update mergePatterns to skip deleted built-ins during workspace load,
preventing them from being re-added
- Remove the isBuiltinPattern guard from AryxAppService.deletePattern;
when a built-in is deleted, its ID is recorded in the tracking list
- Enable the delete button in SettingsPanel and PatternEditor for all
patterns (keep the 'Built-in pattern' label for context)
- Add 3 tests covering built-in deletion, custom deletion, and
selectedPatternId fallback
Users can still restore deleted built-ins via workspace reset.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Introduce a workspace-level agent library that allows users to define
agents once and reference them from multiple orchestration patterns.
Domain model:
- Add WorkspaceAgentDefinition type with name, model, instructions, etc.
- Extend PatternAgentDefinition with optional workspaceAgentId and overrides
- Add resolution helpers that merge workspace agent base with per-pattern overrides
- Extend WorkspaceSettings with agents array and normalize on load
IPC & main process:
- Add saveWorkspaceAgent/deleteWorkspaceAgent IPC channels and handlers
- Resolve workspace agent references in buildEffectivePattern before
sending to sidecar (no C# changes needed)
Settings UI:
- Add 'Agents' tab under Orchestration in the Settings panel
- Create WorkspaceAgentEditor component using ToolingEditorShell
- Show usage count (which patterns reference each agent)
Pattern editor integration:
- Add agent picker dropdown: 'New inline agent' or 'From library'
- Show linked badge (chain icon) on referenced agent graph nodes
- Show linked workspace agent banner in the inspector
- Add 'Save to Agent Library' action to promote inline agents
- Add 'Unlink' action to convert referenced agents back to inline
Tests:
- Add unit tests for resolution helpers (resolvePatternAgent,
resolvePatternAgents, findWorkspaceAgentUsages, normalize)
- Update existing tooling test for new agents field
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When a prompt file overrides the turn model, the original agent's
reasoning effort was carried through even when the target model does
not support it. This caused session.create failures for models like
Claude Opus 4.5 which reject reasoning effort configuration.
Now: if the target model's reasoning capabilities are unknown
(supportedReasoningEfforts undefined) or it's an unresolved model
reference, reasoning effort is stripped entirely.
Also adds a model override badge on assistant messages when the run
used a different model than the session's primary agent. The badge
reads the actual model from the run record (SessionRunRecord.agents)
rather than inferring from the prompt invocation metadata.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- parse prompt model and argument-hint metadata and persist model on prompt invocations
- expand markdown-linked file context in scanned prompt and instruction bodies
- discover .claude/rules instructions and support Claude-style paths metadata
- discover customization roots up to the nearest parent repository and watch them for changes
- apply per-turn prompt model overrides before sidecar execution
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Remove the isPackaged guard from AutoUpdateService.start() so that
background update checks run in both packaged and dev builds.
forceDevUpdateConfig already handles dev-mode config correctly via
dev-app-update.yml — the start() gate was preventing the 10s initial
check and 4h periodic checks from ever being scheduled in dev.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Sub-agent messages bypass the streaming path (turn-delta) entirely due to
SDK batching behavior. They arrive only at turn-complete time via
FinalizeCompletedMessages. Without classification, they appear as separate
chat bubbles cluttering the transcript.
Two-pronged fix:
Sidecar: FinalizeCompletedMessages now tags messages from
_reclassifiedMessageIds with MessageKind='thinking'. This covers messages
that WERE streamed and reclassified during the turn. Added MessageKind
property to ChatMessageDto.
Main process: finalizeTurn detects unstreamed assistant messages (not in
existing session.messages) and classifies them as thinking when a visible
response was already streamed. Emits message-reclassified events so the
renderer can update incrementally, though the primary path is the
workspace:updated broadcast which already includes the correct messageKind.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Wire message-reclassified sidecar events through the main process to
the renderer, where reclassified messages are filtered out of the main
transcript and collected into a collapsible ThinkingProcess component.
Main process changes:
- Route message-reclassified via dedicated onMessageReclassified callback
- Add applyMessageReclassified handler that sets messageKind and emits
the session event for the renderer
- Forward assistant-intent and reasoning-delta as turn-scoped events
Renderer changes:
- Split session.messages into visibleMessages and thinkingMessages
- Render ThinkingProcess above the last assistant message
- ThinkingProcess auto-expands during active turns, collapses on finish
- Update messagePhase to skip thinking-kind messages for final detection
Tests:
- 3 new sessionWorkspace tests for reclassification apply/dedup/ignore
- 2 new messagePhase tests for thinking-kind handling
- Updated runTurnPending test fixtures for new callback
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Position macOS traffic light buttons (trafficLightPosition) in the
BrowserWindow and add conditional left padding to the sidebar header
so the app logo and title clear the window management controls.
- Create shared platform detection utility (src/renderer/lib/platform.ts)
- Set trafficLightPosition { x: 16, y: 22 } on macOS in createMainWindow
- Apply pl-20 (80px) left padding to the sidebar header on macOS
- Consolidate navigator.platform check from keyboardShortcuts.ts
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add dev-app-update.yml so electron-updater can check GitHub releases
even when running with bun run dev. Remove the isPackaged guard from
checkForUpdates() so manual clicks always contact the update server.
Automatic periodic checks remain disabled in dev mode.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When electron-updater reports no update available, the service now
transitions to 'up-to-date' instead of reverting silently to 'idle'.
The troubleshooting UI shows a green check icon, 'Up to date' label,
and 'You are running the latest version of Aryx.' description.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add the backend contract and session-domain support for
'Branch from here':
- new branchSession IPC method and sessions:branch channel
- branchOrigin metadata on SessionRecord
- session branching helper that truncates the transcript at a
chosen user message and clears runtime state
- AryxAppService branching flow with scratchpad directory support
- persistence normalization and regression coverage
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
System tray:
- Always-visible tray icon with context menu
- Quick Scratchpad action from tray (triggers session creation
via IPC bridge to renderer)
- Running session count in tray tooltip and menu
- Click tray icon to show/focus window
- Quit option in tray menu
Minimize to tray:
- New 'Minimize to tray on close' toggle in Settings > Appearance
- When enabled, closing the window hides to tray instead of
quitting (configurable per-user, default off)
- macOS: hides dock icon when minimized to tray
- Proper force-quit handling (Cmd+Q on macOS, tray Quit)
Full IPC contract chain:
- minimizeToTray setting in WorkspaceSettings
- setMinimizeToTray channel + ElectronApi method
- Preload bridge, service handler, IPC registration
- Preserved through workspace normalization
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Show native OS notifications when a session run completes, fails, or
needs approval while the app window is unfocused. Clicking a
notification focuses the window and selects the relevant session.
Includes a toggle in Settings > Appearance to enable/disable
notifications (enabled by default).
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add three layers of usage visibility:
- ChatPane footer: premium request count, AIU consumed, and quota
remaining below the existing context-window bar
- Settings / CopilotStatusCard: on-demand account quota section with
progress bars, overage indicators, and reset dates fetched via
the new get-quota sidecar command
- Activity Panel: per-agent token/cost/duration totals on each agent
row and a Session Usage summary section between agents and timeline
Backend (sidecar):
- New get-quota command using SDK account.getQuota RPC
- New assistant-usage turn-scoped event from SDK assistant.usage
- QuotaSnapshotMapper for both typed and untyped SDK quota payloads
- DTOs: GetQuotaCommandDto, QuotaSnapshotDto, AccountQuotaResultEventDto,
AssistantUsageEventDto
Frontend:
- Shared types: AssistantUsageEvent, QuotaSnapshot, GetQuotaCommand
- IPC bridge: getQuota channel, assistant-usage event dispatch
- State: SessionRequestUsageMap accumulator with per-agent breakdown
- 8 new tests for accumulator logic and formatting helpers
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Display compact status cards in the chat transcript showing each
sub-agent's name, current activity (Thinking, Using grep, etc.),
and elapsed time. Cards appear while sub-agents are running and
clear when the session goes idle.
- Extend SessionEventRecord with description, error, toolCallId,
and model fields for subagent events
- Forward additional subagent fields in handleTurnScopedEvent
- Add subagentTracker reducer to derive active subagent state from
session events, including agent-activity correlation
- Create SubagentActivityCard component with spinner, status icon,
agent name, activity label, and live elapsed timer
- Wire activeSubagents state in App.tsx and pass to ChatPane
- Add 16 unit tests for the subagent tracker
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When the agent produces multiple intermediate responses during a turn,
each gets a unique messageId but all stay pending until finalizeTurn().
This caused multiple stacked 'Thinking' bubbles in the chat transcript.
Now, when a new messageId arrives in applyTurnDelta, any previously
pending assistant messages are marked complete and message-complete
events are emitted before the new message-delta. The renderer reducer
mirrors this logic for defensive consistency.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- add a PTY manager with platform shell resolution and streamed data/exit events
- expose terminal lifecycle and terminal height IPC through preload and app service
- persist terminal panel height in workspace settings and document the backend contract
- add backend tests and validate native packaging with bun run package
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
mergeDiscoveredToolingState now carries over probedTools from existing
servers when the config fingerprint is unchanged. The equality check in
syncProjectDiscoveredTooling also strips probedTools before comparing,
so runtime-only probe data never triggers a spurious state replacement.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The SDK's listTools() calls cacheToolMetadata() which compiles
outputSchema validators. Servers with \ in their schemas (e.g.,
icm-mcp) cause a JSON schema resolution error. Fall back to a raw
JSON-RPC tools/list request that skips schema compilation, since we
only need tool names and descriptions for the approval pill.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Some MCP servers (e.g., eschat.microsoft.com/mcp) serve their
OAuth Protected Resource Metadata at the origin without the path
suffix. Add a third candidate URL that strips the path, matching
VS Code's discovery behavior.
Tried in order:
1. RFC 9728: {origin}/.well-known/{suffix}{path}
2. Appended: {origin}{path}/.well-known/{suffix}
3. Origin-only: {origin}/.well-known/{suffix}
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Tokens are in-memory and empty at startup, so HTTP servers requiring
OAuth return 401 during initial probing. After the user completes
OAuth (either via session auth prompt or proactive auth), re-probe
the authenticated server so its tools appear in the approval pill
without requiring an app restart.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When SSE fallback gets 405 (confirming a Streamable HTTP server),
surface the original Streamable HTTP error instead of the misleading
SSE 405. Extract HTTP status codes from SDK error objects to produce
clearer log messages like 'HTTP 401: Streamable HTTP error: ...'.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Many MCP servers only support legacy SSE despite being configured as
generic HTTP endpoints. Follow the MCP SDK's recommended fallback
pattern: try Streamable HTTP first, then retry with SSE on failure.
This matches VS Code's behavior for these servers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Increase default probe timeout from 10s to 30s to handle slow
package managers (uvx, npx) that install on first run
- Add console logging for probe success/failure to aid debugging
- Probe manually configured MCP servers (not just discovered ones)
so all servers with wildcard tools get their tools discovered
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When MCP server configs declare wildcard tools (empty tools array),
the Copilot SDK has no API to list individual tool names. Add direct
MCP protocol probing using @modelcontextprotocol/sdk to discover
available tools from each accepted server.
- Add McpToolProber service supporting stdio/SSE/HTTP transports
- Probe accepted MCP servers on project load, acceptance, and rescan
- Store probed tools on DiscoveredMcpServer and McpServerDefinition
- Use probed tools in listApprovalToolDefinitions when declared tools
are empty, so the approval pill shows individual tool toggles
- Remove unused isMcpServerApprovalKey helper
- Fix effectiveAutoApprovedCount Math.max workaround in ChatPane
- Add comprehensive tests for probed tool behavior
- Update ARCHITECTURE.md tooling integration section
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use the repository's default Electron import pattern in the MCP OAuth service, make the attachment-path test platform-neutral, and ensure the hook runner cwd/env test drains stdin before asserting shell output.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Give each scratchpad session its own working directory, migrate existing scratchpad sessions on workspace load, and route sidecar turns through the session-specific cwd. Add regression coverage for create, duplicate, delete, and migration flows, and document the new persistence model.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add mode-aware graph mutation helpers for pattern topology so backend
code can add and remove agents without forcing a full graph rebuild.
Also preserve an existing pattern graph in savePattern instead of
re-syncing it from defaults, and cover the new behavior with shared
and main-process tests.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When multiple tool calls of the same permission category (e.g. three
concurrent view/read calls) arrive simultaneously, approving one now
cascades to all queued approvals sharing the same category key. This
eliminates repeated approval prompts for parallel tool calls.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
resetLocalWorkspace() deleted workspace.json and the scratchpad
directory but never broadcast the fresh state to the renderer.
The renderer also discarded the IPC return value. This left stale
projects, sessions, and activities visible in the UI after reset.
- Emit workspace-updated after reloading the workspace in the backend
- Use the returned fresh workspace to immediately update renderer state
- Clear sessionActivities and close the settings panel on reset
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Previously, every tooling update re-probed all enabled HTTP MCP servers
for OAuth requirements. Now it compares the new selection against the
previous one and only probes servers that were just enabled.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use provider-aware OAuth handling instead of a single hardcoded client ID.
GitHub now follows VS Code's GitHub auth setup with the GitHub client ID,
GitHub authorize/token endpoints, prompt=select_account, vscode.dev/redirect,
and JSON-compatible token exchange. Entra keeps the Copilot client ID via a
known provider match, while unknown providers still fall back to metadata
and dynamic registration.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Use vscode.dev/redirect as redirect URI when using the Copilot client
ID, with the local callback URL encoded in the state parameter
- Use PRM resource scopes (e.g. api://icmmcpapi-prod/mcp.tools) instead
of generic OIDC scopes, always include offline_access
- Listen on root path instead of /callback to match vscode.dev/redirect
behavior
- Extract PRM scopes from protected resource metadata discovery
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Use the well-known Copilot client ID (aebc6443-996d-45c2-90f0-388ff96faa56)
when no static client config is provided and the auth server doesn't
support dynamic client registration. This matches how VS Code authenticates
with Entra ID-backed MCP servers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Microsoft Entra ID and other OIDC providers expose metadata at
.well-known/openid-configuration rather than RFC 8414's
.well-known/oauth-authorization-server. Now tries both suffixes.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>