feat(frontend): add sub-agent observability with orchestrator thinking and execution chain display

[genni613]

Summary

Improve sub-agent observability in agent swarm mode by surfacing the orchestrator's thinking process and each subagent's internal execution chain (reasoning + tool calls with parameters).

Changes

• Message history accumulation: Extended Subtask type with messageHistory field and implemented _appendMessage protocol in task context for deduplicated message accumulation from task_running SSE events
• Orchestrator thinking display: New OrchestratorThinking component showing the lead agent's reasoning before splitting subtasks, with task decomposition count
• Subagent execution chain: Enhanced SubtaskCard to render internal reasoning and tool calls with parameters in expanded state, using shared convertToSteps + ToolCall modules extracted from message-group.tsx
• Reasoning fallback: When model doesn't output reasoning_content, fall back to text content as thinking indicator
• Tool args display: Generic tool fallback now shows key string parameters instead of just tool name
• Markdown line breaks: Added remark-breaks plugin so single newlines in reasoning text render as line breaks

Screenshots

Test plan

• Run pnpm check — lint + typecheck pass
• Run pnpm test — 19/19 unit tests pass
• In ultra mode, send a complex prompt that triggers subagent delegation
• Verify OrchestratorThinking shows lead agent's analysis and task count
• Expand SubtaskCard and verify internal reasoning + tool calls with parameters are visible
• Switch to zh-CN locale and verify new i18n strings display correctly

Closes #2315

[Copilot]

Pull request overview

This PR improves agent-swarm UX by surfacing additional execution details: the orchestrator’s pre-delegation “thinking” and each subagent’s internal message/tool-call chain, plus better markdown rendering for reasoning text.

Changes:

• Extend Subtask to accumulate messageHistory from task_running SSE events (deduped).
• Add OrchestratorThinking and enhance SubtaskCard to render reasoning + tool calls (via extracted shared message-steps utilities/components).
• Improve markdown rendering of reasoning text by enabling single-newline line breaks (remark-breaks).

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
frontend/src/core/threads/hooks.ts	Appends task_running messages into subtask state via _appendMessage.
frontend/src/core/tasks/types.ts	Adds messageHistory?: AIMessage[] to Subtask.
frontend/src/core/tasks/context.tsx	Implements _appendMessage protocol for deduped history accumulation.
frontend/src/core/streamdown/plugins.ts	Adds remark-breaks to Streamdown plugin sets.
frontend/src/core/i18n/locales/zh-CN.ts	Adds new subtask/orchestrator i18n strings (zh-CN).
frontend/src/core/i18n/locales/types.ts	Extends Translations.subtasks typing with new keys.
frontend/src/core/i18n/locales/en-US.ts	Adds new subtask/orchestrator i18n strings (en-US).
frontend/src/components/workspace/messages/subtask-card.tsx	Renders subagent execution chain from messageHistory.
frontend/src/components/workspace/messages/orchestrator-thinking.tsx	New component to display orchestrator analysis + decomposition count.
frontend/src/components/workspace/messages/message-steps.tsx	Extracted shared “steps” conversion + ToolCall renderer.
frontend/src/components/workspace/messages/message-list.tsx	Integrates OrchestratorThinking into subagent message groups.
frontend/src/components/workspace/messages/message-group.tsx	Refactors to consume shared message-steps module.
frontend/package.json	Adds remark-breaks dependency.
frontend/pnpm-lock.yaml	Locks remark-breaks and transitive deps.

Files not reviewed (1)

• frontend/pnpm-lock.yaml: Language not supported

[WillemJiang]

@genni613 Here are some additional comments for the code

1. _appendMessage convention is fragile (context.tsx:43)

Using an underscore-prefixed _appendMessage field that's destructured out and never makes it into the Subtask type is a side-channel pattern. It works but is easy to break accidentally — anyone reading Subtask type won't know this protocol exists.

Suggestion: Consider making updateSubtask accept a second parameter or use a discriminated union for the update payload instead of smuggling data through underscore-prefixed fields.

2. CoTToolCallStep.result type mismatch with old code (message-steps.tsx:521)

In the old message-group.tsx, result was string | Record<string, unknown> (union). In the new message-steps.tsx, the interface declares it the same way, but convertToSteps assigns step.result = json where json is any (from JSON.parse). This means arrays also pass through — e.g., web_search results which are Array<...>. The type says Record<string, unknown> but arrays flow through fine at runtime. This isn't a bug, but the types are misleading.

Suggestion: Either broaden the type to unknown or add explicit handling.

3. Redundant Array.isArray check (message-steps.tsx:167-168)

{Array.isArray(results) && (

{Array.isArray(results) &&

The outer guard already confirmed that the results are an array. The inner Array.isArray(results) is redundant.

[WillemJiang]

@genni613 Please run the below comments to fix the lint error.

cd frontend
make format

[WillemJiang]

@genni613 Did you have a chance to try the latest frontend in your local box? My browser blocked when the backend started the subagent.

[genni613]

@WillemJiang Already fixed. updateSubtask calls were moved from the render-phase groupMessages callback into a useEffect with a syncedTaskIdsRef dedup guard, breaking the render → setState → re-render loop.