qwen-code/docs/design/daemon-skill-toggle.md

Daemon Skill Toggle

Goal

Expose the CLI /skills panel's workspace enable/disable behavior through daemon REST and the TypeScript SDK, including immediate refresh of active ACP sessions.

Public contract

• POST /workspace/skills/:name/enable
• POST /workspaces/:workspace/skills/:name/enable
• Request body: { "enabled": boolean }
• SDK: DaemonClient.setWorkspaceSkillEnabled and WorkspaceDaemonClient.setWorkspaceSkillEnabled
• Capability: workspace_skill_toggle

The response contains the canonical skill name, requested state, whether persistence changed, activation state, and session refresh counts. applied means every active session refreshed, deferred means no ACP child was running, and partial means at least one session failed to refresh after persistence committed.

Semantics

The API changes workspace skills.disabled and skills.enabled as needed. Skill lookup is case-insensitive, but the canonical discovered name is persisted. Enabling a default-disabled skill writes an explicit opt-in; disabling it removes the opt-in and writes a hard workspace disable. Updating one target removes target duplicates and case variants without deleting orphan entries for unavailable skills. A second identical request is a no-op.

The route rejects states the CLI panel cannot toggle:

• unknown skill: 404 skill_not_found;
• userInvocable === false: 409 skill_not_toggleable;
• skill from an inactive extension: 409 skill_not_toggleable;
• disabled in system defaults, user, or system scope: 409 skill_not_toggleable with the locking scope;
• untrusted workspace: 403 untrusted_workspace.

The scope lock check and workspace read-modify-write happen inside the daemon's per-workspace settings lock. A failed write stops before refresh and event publication.

Skill availability versus disable-model-invocation

skills.disabled is an operator hard denylist merged as a case-insensitive union across scopes. skills.defaultDisabled supplies overridable defaults and skills.enabled supplies explicit opt-ins, with disabled > enabled > defaultDisabled precedence. Effective disables remove matching skill slash commands and model-visible skill entries, and execution-time validation rejects the skill. The daemon endpoint writes the workspace members of disabled and enabled.

disable-model-invocation is SKILL.md metadata. It hides a skill from model invocation while preserving direct user invocation. The existing managed-skill ACP operation edits that metadata and is intentionally not reused by this API.

Activation flow

1. Resolve the canonical, toggleable skill from the workspace status snapshot.
2. Under the workspace settings lock, re-read every scope, reject higher-scope locks, and commit the canonical workspace list.
3. Invalidate the daemon's cached skill status.
4. If an ACP child is live, invoke qwen/control/workspace/skills/refresh.
5. The child reloads workspace-scope settings and refreshes every active session, including busy sessions.
6. Each session reloads its own workspace settings, rebuilds and pushes available_commands_update, and notifies SkillManager consumers.
7. Publish the existing workspace settings_changed event for each changed skill-settings key.

An in-flight model request cannot be rewritten. Subsequent skill execution checks, command snapshots, and model contexts read the new state.

Downstream consumers

• Settings merge: system defaults, user, workspace, and system lists form the effective disabled-name set with disabled > enabled > defaultDisabled precedence.
• Workspace status: ACP and daemon-local skill mapping expose disabled state, disablement reason, lock scope, and false-only userInvocable.
• Slash commands: available-command construction removes disabled skills and sends updated command metadata to daemon clients.
• Model context: SkillManager change listeners refresh the Skill tool description and available-skill context.
• Execution validation: the Skill tool re-reads the disabled-name provider before invocation, so later calls are rejected immediately.
• Extension state: inactive extension skills remain non-toggleable even when they are not disabled by settings.
• Daemon cache: the cached live-child skill snapshot is invalidated after persistence so later GET requests cannot replay stale state.
• SDK consumers: both primary-workspace and workspace-qualified clients share the response and error contract.
• Events: existing settings_changed consumers observe each committed skills.disabled or skills.enabled value; there is no new event type.

Failure behavior

• Persistence failure: the HTTP request fails; no ACP refresh and no event.
• No child: persistence succeeds with deferred; the next child loads the setting at startup.
• Per-session refresh failure: persistence remains committed; successful sessions stay refreshed and the response is partial.
• Child transport race: if the child disappears after the liveness check, the response is deferred; other refresh failures are reported as partial.