The WorkflowValidator rejected loop edges with an 'always' condition,
but maxIterations already guarantees termination at runtime (enforced by
WorkflowConditionEvaluator). This made built-in templates like
Collaborative Group Chat fail validation immediately.
Relax the check so 'always' conditions are accepted when maxIterations
provides the termination guarantee. The condition error now only fires
when neither a non-default condition nor a maxIterations cap is present.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add first-class backend support for workflow orchestration modes,
including mode settings, graph scaffolding, mode-aware validation,
and real Agent Framework handoff/group-chat execution paths.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add MVP code-executor and function-executor runtime support, request-port
bridging through Aryx user input, state-scope helpers, lockstep execution
mode handling, and validation/tests for new workflow node kinds.
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>
Move handoff routing and ownership rules out of per-agent system
prompts and into the Agent Framework handoff builder guidance.
This keeps AgentInstructionComposer focused on Aryx-owned system
prompt content while letting WithHandoffInstructions supply the
workflow-level handoff semantics.
- remove handoff-mode runtime guidance from AgentInstructionComposer
- expand HandoffWorkflowGuidance with the triage/specialist rules
- update tests to pin the new split of responsibilities
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Explicitly configure AIAgentHostOptions for the workflow modes that
host agents directly instead of relying on Agent Framework defaults.
- add a shared host-options factory that preserves Aryx's current
behavior
- use custom sequential, concurrent, and round-robin group-chat
builders so all host options are set intentionally
- add workflow-level tests asserting the configured host options
Keep EmitAgentResponseEvents disabled because Aryx still projects
streaming transcript state itself and enabling response events would
require a separate reconciliation change.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Explicitly configure handoff workflows to use
HandoffToolCallFilteringBehavior.HandoffOnly instead of relying on
Agent Framework's current default.
This keeps normal tool-call history visible across handoffs while
still stripping handoff plumbing, and adds a regression test to pin
that behavior.
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>
When the pre-tool-use hook returns 'ask', the Copilot CLI creates
PermissionRequestHook instead of categorized PermissionRequestRead/
Write/Shell. This caused 'Permission: hook' labels and broke category-
based auto-approval ('Always approve read' wouldn't cover grep/glob).
Add ResolveHookToolCategory mapping in CopilotApprovalCoordinator to
map known tool names (view/grep/glob→read, edit/create→write, etc.)
to their permission categories. Wire into GetFallbackToolName,
BuildPermissionApprovalEvent, and CreateApprovalPolicyOutput so:
- Approval banner shows 'Permission: read' instead of 'Permission: hook'
- 'Always approve' stores the category key, covering all tools in it
- Hook short-circuits when category is already auto-approved
Unknown tools (MCP, custom) keep existing 'hook' behavior.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Allow SDK-managed orchestration tools to bypass pre-tool approval prompts so Aryx matches Copilot CLI behavior for non-side-effectful meta tools.
Keep store_memory under the existing memory approval category.
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>
MCP servers configured with empty tools arrays (wildcard) now appear in the
auto-approval pill with a server-level toggle. When toggled, a server-level
approval key (mcp_server:<name>) is added to autoApprovedToolNames. The
sidecar matches this key against PermissionRequestMcp.ServerName to auto-
approve all tools from that server without needing individual tool names.
- Add buildMcpServerApprovalKey/isMcpServerApprovalKey helpers
- Create approval groups for all MCP servers including empty-tools ones
- Add serverApprovalKey to ApprovalToolGroup for server-level toggles
- Update sidecar RequiresToolCallApproval to check server-level keys
- Include server-level keys in listApprovalToolNames for pruning safety
- Add tests for both shared domain and sidecar approval matching
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Remove the packaging-only create-dmg dependency in favor of native hdiutil, and make the hook runner working-directory test assert behavior without depending on macOS temp-path canonicalization.
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>
Stop forwarding hook lifecycle events when a project has no configured
.github/hooks/*.json commands so the UI only receives meaningful hook
activity. Also expose the configured-hook state on the bundle, cover the
suppression wiring with tests, and document the updated event semantics.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add an AlwaysApprove flag to resolve-approval commands and cache
approved tool keys for the current request so repeated runtime
permissions are auto-approved immediately within the same turn.
The cache is cleared when the turn finishes to avoid leaking across
future requests.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
AryxCopilotAgent was converting ALL Copilot tool requests (ask_user,
MCP tools, etc.) into FunctionCallContent. The Agent Framework's
AIAgentHostExecutor tracks every FunctionCallContent as an outstanding
request. Since SDK-handled tools like ask_user never produce a
matching FunctionResultContent, HasOutstandingRequests stayed true and
TurnToken was never emitted — stalling group-chat advancement after
an ask_user interaction.
Now only handoff_to_* tool requests are projected as FunctionCallContent.
All other tool calls remain invisible to the Agent Framework executor,
matching the upstream GitHubCopilotAgent behaviour.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace the broken forced-tool workaround with a repo-local Copilot
agent adapter that merges runtime handoff instructions and tool
declarations into Copilot sessions and projects Copilot tool requests
back into Agent Framework function-call updates.
Also add regression coverage for the adapter and document the runtime
integration detail in the architecture guide.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Constrain the initial handoff triage agent so it cannot use regular
session tooling and must use a handoff tool on its first invocation.
This prevents the model from narrating delegation in plain text while
never actually transferring control.
Also add regression tests for the new entry-agent constraints.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add backend support for plan-mode turn shaping and exit-plan review
checkpoints. Run-turn commands now accept an interaction mode,
plan-mode prompt guidance tells agents to produce a plan and call
exit_plan_mode, and the sidecar emits a structured
exit-plan-mode-requested event when the SDK raises that session event.
For now the backend intentionally treats exit_plan_mode as a turn
boundary and graceful-degradation review checkpoint. The current
Agent Framework GitHubCopilotAgent wrapper does not expose a clean way
to bridge the live CopilotSession back into same-turn exit-plan
resolution, so that follow-up remains documented for the frontend and a
future backend slice.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Exclude internal runtime tools from sidecar capability discovery so they do not surface in the approval tool catalog.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Implement the Copilot SDK user input round-trip in the backend:
- add user-input-requested and resolve-user-input protocol DTOs
- add a CopilotUserInputCoordinator that mirrors approval flow with
pending TaskCompletionSource state and explicit resolution
- wire SessionConfig.OnUserInputRequest through CopilotAgentBundle
and CopilotWorkflowRunner
- extend SidecarProtocolHost to emit user input events, accept
resolve-user-input commands, and filter ask_user from runtime
approval tools
- add regression tests for the new coordinator and protocol flow
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Expose SDK permission request details on approval-requested events via a
new permissionDetail payload. This preserves the existing title/detail
summary while carrying the command, URL, diff, args, and other
kind-specific data needed for richer approval UI in the renderer.
Also adds sidecar coverage for all supported permission request kinds and
verifies the new payload serializes over the protocol.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
David Kaya