feat: Agent message interrupts — subagents should receive messages mid-tool-batch

[Meme-Theory]

Summary

Subagents currently cannot see SendMessage messages until ALL queued tool calls in a batch finish. If the API returns 5 serial tool calls and the coordinator sends a correction after tool #1, the agent blindly executes tools #2–#5 before seeing the message. This is the single most operationally painful gap in the coordinator/agent architecture.

This PR proposes three layers of interrupt to fix it:

                                         ┌─ Layer 1: peek between serial tools
                                         │  ┌─ Layer 2: peek between batches
API response ─► [read, read, read] ─► [edit] ─► [edit] ─► [edit] ─► [bash]
                 concurrent batch     serial ──────────────────────── serial
                                         │              │
                                    check here     check here
                                         │
                         ┌─ Layer 3: SendMessage(urgent=true)
                         │  aborts long-running individual tools

• Layer 1 — hasPendingMessages() peek between serial tool calls in runToolsSerially(). Cancel remaining tools if messages waiting.
• Layer 2 — Same check at batch boundaries in runTools().
• Layer 3 — New urgent boolean on SendMessage. Fires abortController.abort('message_interrupt') on the target. query.ts handles this by resetting the controller and falling through to the attachment phase instead of exiting.

What the agent sees after interrupt:

tool_result (tool 1): { actual result }
tool_result (tool 2): { actual result }  
tool_result (tool 3): [is_error] "Tool execution interrupted: message received..."
tool_result (tool 4): [is_error] "Tool execution interrupted: message received..."
attachment: "Stop working on X, pivot to Y instead"

Files

Since the source tree is internal, this PR includes an RFC document and 5 patch files against the internal paths:

Patch	Target	Change
01-LocalAgentTask.patch	src/tasks/LocalAgentTask/LocalAgentTask.tsx	hasPendingMessages() non-draining peek
02-messages.patch	src/utils/messages.ts	PENDING_MESSAGE_INTERRUPT + createToolResultInterruptMessage()
03-toolOrchestration.patch	src/services/tools/toolOrchestration.ts	Interrupt checks at batch + serial boundaries
04-SendMessageTool.patch	src/tools/SendMessageTool/SendMessageTool.ts	urgent field + abort on urgent
05-query.patch	src/query.ts	message_interrupt abort handling

Cost

• One getAppState() call per serial tool boundary, only for subagents (agentId guard). Zero overhead for the main thread. Zero overhead when no messages pending.

Test plan

• Spawn agent with 5+ serial tool calls, send it a message after tool Create SECURITY.md #1 — verify it sees the message after tool Create SECURITY.md #1 instead of after tool Timeout and Error After Long-Running Prompt #5
• Same test with concurrent read-only batch followed by serial batch — verify interrupt at batch boundary
• SendMessage(urgent=true) during a long-running Bash command — verify agent is interrupted and receives message
• Verify normal (non-urgent) SendMessage still queues as before when no interrupt check fires
• Verify main thread is unaffected (no agentId, checks are skipped)
• Verify message_interrupt abort resets controller and agent continues (doesn't exit query loop)

``

🤖 Generated with Claude Code