agentHost: preserve Claude tool state across steering results

[Aleksandr Logunov (Longarithm)]

Problem

A steering prompt can cause an intermediate Claude SDK result while the protocol turn is still running. The pipeline delays ChatTurnComplete until the prompt queue drains, but the mapper clears pending tool calls and foreground subagent spawns on every result. Later tool results then lose their pending state. The SDK's [ede_diagnostic] entry can also appear as a user-facing error.

Change

• Use queue state, not diagnostic text, to preserve tools and foreground subagents across intermediate results.
• Classify the result, settle its queue entry, map its signals, and emit completion synchronously. Steering cannot arrive between cleanup classification and settlement; result signals still precede completion.
• Clear foreground state on final results, abort, stream failure and disposal. Background subagent registrations remain independent of foreground cleanup.
• Filter diagnostic entries from displayed execution errors while retaining real errors in the same result.

Related work

Related to #330785, which also filters the diagnostic and proposes separate protocol turns for steering. This PR retains same-turn steering behavior and fixes pending-state cleanup independently.

Validation

Regression tests failed before the corresponding fixes and now pass:

• Foreground subagent/tool completion after an intermediate result.
• Steering queued in a result-signal microtask cannot split cleanup and completion decisions.
• Intermediate result followed by abort or stream failure clears state before recovery; late tool results cannot complete a stale tool.
• Diagnostic-only and mixed-error results, including cleanup on a final diagnostic-only result.

363 tests pass across claudeAgent, claudePromptQueue, claudeMapSessionEvents, claudeSdkMessageRouter, and claudeSdkPipeline. Three existing multi-turn tests now wait for the fake SDK to consume the next prompt before returning its result, rather than depending on microtask timing.

Targeted TypeScript checking of changed files and their imports, and repository hygiene/formatting/ESLint on all changed files, pass. Validation reused existing dependencies and Electron. No full application build or live SDK/UI reproduction was run for this upstream branch.

[vs-code-engineering]

📬 CODENOTIFY

The following users are being notified based on files changed in this PR:

TylerLeonhardt

Matched files:

• src/vs/platform/agentHost/node/claude/claudeMapSessionEvents.ts
• src/vs/platform/agentHost/node/claude/claudePromptQueue.ts
• src/vs/platform/agentHost/node/claude/claudeSdkMessageRouter.ts
• src/vs/platform/agentHost/node/claude/claudeSdkPipeline.ts

[Copilot]

🟡 Changes recommended

Result handling has a queue-state race and lacks cleanup when an intermediate result is followed by abort or failure.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Preserves Claude tool and subagent state across intermediate steering results while filtering internal diagnostics.

Changes:

• Propagates intermediate-result state through the queue, pipeline, and router.
• Defers pending tool/subagent cleanup until the final result.
• Adds regression coverage for steering and diagnostic filtering.

File summaries

File	Description
claudePromptQueue.ts	Detects prompts remaining after the current head.
claudeSdkPipeline.ts	Classifies intermediate results.
claudeSdkMessageRouter.ts	Forwards result classification.
claudeMapSessionEvents.ts	Preserves state and filters diagnostics.
claudePromptQueue.test.ts	Tests queue classification.
claudeMapSessionEvents.test.ts	Tests filtering and final cleanup.
claudeAgent.test.ts	Tests steering with foreground subagents.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 2
• Review effort level: Balanced

---

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

[Aleksandr Logunov (Longarithm)]

@microsoft-github-policy-service agree

[Copilot]

🟢 Approval recommended

The prior cleanup and race issues are addressed with focused regression coverage, and no unresolved correctness issues were found.

Review details

• Files reviewed: 7/7 changed files
• Comments generated: 0 new
• Review effort level: Balanced

[Allen Bacabis Santiago (saitanallensantiago26-beep)]

Problem

A steering prompt can cause an intermediate Claude SDK result while the protocol turn is still running. The pipeline delays ChatTurnComplete until the prompt queue drains, but the mapper clears pending tool calls and foreground subagent spawns on every result. Later tool results then lose their pending state. The SDK's [ede_diagnostic] entry can also appear as a user-facing error.

Change

• Use queue state, not diagnostic text, to preserve tools and foreground subagents across intermediate results.
• Classify the result, settle its queue entry, map its signals, and emit completion synchronously. Steering cannot arrive between cleanup classification and settlement; result signals still precede completion.
• Clear foreground state on final results, abort, stream failure and disposal. Background subagent registrations remain independent of foreground cleanup.
• Filter diagnostic entries from displayed execution errors while retaining real errors in the same result.

Related work

Related to #330785, which also filters the diagnostic and proposes separate protocol turns for steering. This PR retains same-turn steering behavior and fixes pending-state cleanup independently.

Validation

Regression tests failed before the corresponding fixes and now pass:

• Foreground subagent/tool completion after an intermediate result.
• Steering queued in a result-signal microtask cannot split cleanup and completion decisions.
• Intermediate result followed by abort or stream failure clears state before recovery; late tool results cannot complete a stale tool.
• Diagnostic-only and mixed-error results, including cleanup on a final diagnostic-only result.

363 tests pass across claudeAgent, claudePromptQueue, claudeMapSessionEvents, claudeSdkMessageRouter, and claudeSdkPipeline. Three existing multi-turn tests now wait for the fake SDK to consume the next prompt before returning its result, rather than depending on microtask timing.

Targeted TypeScript checking of changed files and their imports, and repository hygiene/formatting/ESLint on all changed files, pass. Validation reused existing dependencies and Electron. No full application build or live SDK/UI reproduction was run for this upstream branch.

[Aleksandr Logunov (Longarithm)]

Before

The SDK's [ede_diagnostic] entry is shown as a user-facing execution error after a normal follow-up prompt: