agentHost: don't subscribe to subagent chats that never spawn

[roblourens]

Found while scanning recent Insiders logs for actionable errors.

Issue

A task tool call that fails without spawning a subagent left the workbench subscribed to a child chat URI the host never created. The host waited out its pending-subagent window (15s) and then failed the subscription:

12:52:30.201 [trace] [AgentService] subscribe: ahp-chat://subagent/.../toolu_01FE...
12:52:30.223 [info] [Copilot:...] Tool completed: toolu_01FE...
12:52:30.226 [info] [AgentSideEffects] ... success=false
12:52:45.263 [error] [ProtocolServer] Request 'subscribe' failed Resource not found: ahp-chat://subagent/.../toolu_01FE...

The client saw the matching failure:

[RemoteAgentHostProtocol] Request 2891 failed: {"code":-32001,"message":"Resource not found: ahp-chat://subagent/.../toolu_01FE..."}

The underlying tool failure was legitimate and self-describing — the SDK rejected the nested delegation:

Maximum sub-agent depth of 4 reached. Complete this task without spawning further sub-agents.

So the agent behaved correctly; the client just kept observing a child chat that was never going to exist, producing a misleading Resource not found error.

Root cause

AgentHostSessionHandler._tryObserveSubagentToolCall began observing any recognized subagent tool once it reached Running or Completed, and never reconsidered that decision. A subagent-spawning tool briefly passes through Running, so the observation started before the depth-limit failure arrived and was never released.

Fix

• Added shouldObserveSubagentChat in stateToProgressAdapter.ts: observe while the tool is Running, and after completion only when it actually has a subagent content block or succeeded (preserving legacy/restored snapshots that carry no content).
• AgentHostSessionHandler now releases the child-chat observation when a subagent tool completes unsuccessfully without a subagent content block, instead of holding a subscription to a chat that will never exist.
• Per-observation disposables moved into a DisposableMap keyed by tool call, so releasing an observation also tears down its autoruns and chat subscription.

Live and nested subagents are unaffected: they keep their subscription from the moment the tool starts running.

Tests

Two regressions, both at the layer that owns the behavior (no E2E fixture needed — this is a workbench-side decision, not host/AHP behavior):

• stateToProgressAdapter.test.ts — the predicate observes running/successful tools and a failed tool that produced a child, but not a failed tool without one.
• agentHostChatContribution.test.ts — end-to-end through the handler: the child chat subscription is live while the tool runs, and released after the failed completion. This one fails without the handler change.

Validation

• ./scripts/test.sh --grep "subagent" over both files: 21 passing
• Full stateToProgressAdapter.test.ts: 131 passing
• npm run typecheck-client: clean

(Written by Copilot)

[Copilot]

Pull request overview

Prevents stale subscriptions when subagent tool calls fail without creating child chats.

Changes:

• Adds a predicate for determining whether a subagent chat should remain observed.
• Tracks observations with disposable maps for immediate cleanup.
• Adds predicate and subscription-lifecycle regression tests.

Show a summary per file

File	Description
stateToProgressAdapter.ts	Adds the observation predicate.
agentHostSessionHandler.ts	Releases invalid child-chat observations.
stateToProgressAdapter.test.ts	Tests predicate outcomes.
agentHostChatContribution.test.ts	Tests subscription cleanup after failure.

Review details

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

• Files reviewed: 4/4 changed files
• Comments generated: 0
• Review effort level: Balanced