Channel P0 Identity And Task Lifecycle Design

Goal

Implement the first P0 foundation for channel-resident multiplayer agents: channel-scoped identity and memory-boundary metadata, plus a shared task lifecycle hook in @qwen-code/channel-base.

This intentionally does not add a Slack adapter, daemon event stream, adapter UI changes, proactive scheduling, cross-channel context, or real core-memory path isolation.

Background

qwen channel already supports messaging adapters, shared sessions, sender attribution, dispatch modes, streaming chunks, tool-call callbacks, cancellation, and platform-specific progress surfaces such as Feishu cards. The missing P0 product layer is a stable way to say "this channel has its own resident agent identity" and "this prompt turn has a lifecycle adapters can observe."

Issue #6103 tracks this focused slice. It builds on the broader qwen tag roadmap in #5887, but keeps this PR small enough to review and ship independently.

Scope

In scope:

• Add optional channel identity metadata to ChannelConfig.
• Add optional memory-scope metadata to ChannelConfig.
• Derive safe defaults when the new config is omitted.
• Inject a concise channel boundary note into the first prompt for each agent session, together with existing channel instructions.
• Add a protected onTaskLifecycle(event) hook on ChannelBase.
• Emit lifecycle events from the shared channel flow for prompt start, text chunks, tool calls, cancellation, completion, and errors.
• Add focused package-local tests in packages/channels/base.

Out of scope:

• Core memory storage changes or file-path namespace isolation.
• Daemon/SSE event publication.
• Feishu, DingTalk, Telegram, WeChat, or QQ UI changes.
• New platform adapters.
• Token budgets, tool ACLs, or cross-channel context sharing.

Design

Channel Identity

Add a small optional config object:

export interface ChannelIdentityConfig {
  id?: string;
  displayName?: string;
  description?: string;
}

ChannelConfig gains identity?: ChannelIdentityConfig.

At runtime, ChannelBase derives:

• id: config.identity.id or channel:<name>
• displayName: config.identity.displayName or <name>
• description: config.identity.description, if present

The runtime identity is metadata only. It does not change session routing, access control, or platform adapter behavior.

Memory Scope Metadata

Add:

export type ChannelMemoryScopeMode = 'metadata-only';

export interface ChannelMemoryScopeConfig {
  namespace?: string;
  mode?: ChannelMemoryScopeMode;
}

ChannelConfig gains memoryScope?: ChannelMemoryScopeConfig.

At runtime, ChannelBase derives:

• namespace: config.memoryScope.namespace or channel:<name>
• mode: always 'metadata-only' for this PR

This is deliberately not a real core-memory namespace. It is an explicit, inspectable boundary marker and prompt instruction so later work can wire the same namespace into core memory paths without changing channel config shape.

Prompt Boundary Injection

ChannelBase already prepends config.instructions once per session; that behavior is unchanged. The generated boundary note below is added to the same first-message injection only when a channel configures identity or memoryScope (instructions-only channels keep the existing prompt shape). It is appended after custom instructions so the boundary takes recency precedence:

Channel identity:
- id: channel:ops
- display name: Ops Bot
- description: Helps the ops group coordinate repository maintenance.

Memory scope:
- namespace: qwen-tag:ops
- mode: metadata-only
- data from other channels must not be shared.

The exact wording should be concise and stable enough for tests, but avoid over-promising isolation. If no description exists, omit that line.

This note is injected once per agent session, like existing instructions (a transient channel-memory read failure retries the whole context block on the next turn, so consecutive turns may repeat it). When the bridge reports a session death, the existing instructedSessions cleanup continues to allow reinjection for the next session.

For compatibility, channels with no instructions, identity, or memoryScope configuration keep the existing raw prompt shape. Runtime identity and memory metadata are still derived for lifecycle events and status commands.

Status Visibility

Extend /who and /status with identity and memory metadata:

• /who should include identity display name and memory namespace.
• /status should include the identity id and memory mode.

Keep the output short. Do not expose absolute paths or hidden configuration.

Task Lifecycle Hook

Add a discriminated union:

export type ChannelTaskLifecycleEvent =
  | {
      type: 'started';
      channelName: string;
      chatId: string;
      sessionId: string;
      messageId?: string;
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    }
  | {
      type: 'text_chunk';
      channelName: string;
      chatId: string;
      sessionId: string;
      messageId?: string;
      chunk: string;
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    }
  | {
      type: 'tool_call';
      channelName: string;
      chatId: string;
      sessionId: string;
      toolCall: ToolCallEvent;
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    }
  | {
      type: 'cancelled';
      channelName: string;
      chatId: string;
      sessionId: string;
      messageId?: string;
      reason: 'cancel_command' | 'clear' | 'steer' | 'timeout';
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    }
  | {
      type: 'completed';
      channelName: string;
      chatId: string;
      sessionId: string;
      messageId?: string;
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    }
  | {
      type: 'failed';
      channelName: string;
      chatId: string;
      sessionId: string;
      messageId?: string;
      error: string;
      identity: ChannelRuntimeIdentity;
      memoryScope: ChannelRuntimeMemoryScope;
    };

ChannelBase adds:

protected onTaskLifecycle(_event: ChannelTaskLifecycleEvent): void {}

Default behavior is no-op. Adapters can opt in later without changing the prompt execution path.

Lifecycle Emission Points

Emit from shared ChannelBase flow:

• started: immediately after activePrompts.set() and before onPromptStart().
• text_chunk: when the prompt's textChunk listener accepts a non-cancelled chunk.
• tool_call: in the existing bridge tool-call listener after resolving the session target.
• cancelled: when /cancel succeeds, when /clear cancels or evicts an active prompt, and when steer marks the active turn cancelled.
• completed: after bridge.prompt() resolves and before or after onResponseComplete(), as long as the turn was not cancelled.
• failed: when bridge.prompt() or response delivery throws.

Lifecycle hook failures should be caught and logged to stderr. A platform adapter's lifecycle UI must not break prompt execution or cleanup.

Error Handling

• Invalid identity or memory fields are not fatal in this PR; config parsing should preserve the existing permissive shape and only accept string fields where explicit parsing already exists.
• Lifecycle hook exceptions are swallowed after a stderr diagnostic.
• Memory scope mode is constrained to 'metadata-only'; omitted or unknown config should resolve to 'metadata-only' rather than enabling behavior that does not exist.

Tests

Focused tests in packages/channels/base/src/ChannelBase.test.ts should cover:

• Default identity and memory metadata are derived from channel name.
• Custom identity and memory namespace are included in the first prompt.
• Boundary metadata is injected once per session and re-injected after sessionDied.
• /who and /status include the new metadata without leaking cwd.
• onTaskLifecycle sees started, text_chunk, tool_call, completed.
• onTaskLifecycle sees cancelled for /cancel, /clear, and steer.
• onTaskLifecycle sees failed when bridge.prompt() rejects.
• A throwing lifecycle hook does not reject handleInbound().

Use package-local commands:

cd packages/channels/base
npx vitest run src/ChannelBase.test.ts

Final verification before PR:

npm run build
npm run typecheck

Open Decisions

None for this PR. Real core-memory namespace enforcement, daemon publication, adapter UI, tool/data ACLs, budgets, and proactive follow-up are explicitly future work.