fix: mark 'Language model unavailable' as expected chat error (fixes #326356)

[vs-code-engineering]

Summary

The chatAgentError telemetry bucket chatagenterror-Language model unavailable fires whenever ExtHostChatAgents2.getModelForRequest cannot resolve a language model for a chat request — thrown from $invokeAgent (and $detectChatParticipant). This is an expected operational condition (the user is not signed in, or no chat models have been contributed yet), not a code bug, but it was thrown as a generic Error and therefore published to error telemetry as noise. The fix marks it as a ChatExpectedError so the existing suppression path keeps it out of chatAgentError telemetry while still surfacing the user-facing message.

Fixes #326356
Recommended reviewer: @bryanchen-d

Culprit Commit

Field	Value
Commit	Not identified — pre-existing / re-bucketing
Author	n/a
PR	n/a
Message	n/a
Why	The throw new Error('Language model unavailable') in getModelForRequest is a long-standing pattern. The isExpectedError / ChatExpectedError suppression mechanism was added later (3727d0ce, "add isExpectedError flag to suppress expected agent errors from telemetry", by @bryanchen-d, 2026-04-27). This throw was simply never opted into that mechanism, so it continued surfacing as a chatAgentError bucket. No single regression commit introduced the error path.

No culprit commit — this is a pre-existing expected-error condition that was never registered with the ChatExpectedError telemetry-suppression channel.

Code Flow

sequenceDiagram
    participant Agent as Chat request
    participant Invoke as extHostChatAgents2.$invokeAgent
    participant Producer as getModelForRequest
    participant Main as mainThreadChatAgents2

    Agent->>Invoke: invoke agent
    Invoke->>Producer: getModelForRequest(request, extension)
    Note over Producer: ⚠️ No user-selected model and<br/>getDefaultLanguageModel() returns undefined<br/>(user not signed in / no models)
    Producer-->>Invoke: throw new Error('Language model unavailable')
    Note over Invoke: 💥 Generic Error → errorCallstack packed,<br/>isExpectedError NOT set
    Invoke->>Main: rpcResult { errorCallstack, errorName }
    Note over Main: publishes chatAgentError telemetry<br/>(bucket: chatagenterror-Language model unavailable)

Loading

Affected Files

File	Role	Evidence
src/vs/workbench/api/common/extHostChatAgents2.ts	root cause / producer	L948-L949 (pre-fix): if (!model) { throw new Error('Language model unavailable'); } — expected condition thrown as generic Error
src/vs/workbench/api/common/extHostChatAgents2.ts	crash site	L900/L950 (from shipped-commit stack): getModelForRequest / $invokeAgent
src/vs/workbench/api/browser/mainThreadChatAgents2.ts	telemetry emitter	L380-L397: if (rpcResult?.errorCallstack && !rpcResult.errorDetails?.isExpectedError) { ... publicLogError2('chatAgentError', ...) } — suppresses when isExpectedError is set

Repro Steps

1. Launch the Agents window / chat in a profile that is not signed in to Copilot (or that has no contributed chat models).
2. Send a chat request to an agent, or trigger participant detection.
3. getModelForRequest resolves no user-selected model and getDefaultLanguageModel() returns undefined, so throw new Error('Language model unavailable') fires.
4. Before this fix, the generic error is packed with a callstack and published as the chatAgentError telemetry event, producing the chatagenterror-Language model unavailable bucket.

How the Fix Works

Chosen approach — src/vs/workbench/api/common/extHostChatAgents2.ts, getModelForRequest: the fix is applied at the data producer (where the invalid/expected state is first detected), not at the crash site or in a downstream try/catch. Instead of throwing a generic Error, the "no model available" case now throws an error whose name is set to ChatExpectedError. $invokeAgent already maps e.name === 'ChatExpectedError' to errorDetails.isExpectedError = true, and mainThreadChatAgents2 already skips publishing the chatAgentError event when that flag is set (the existing suppression channel added for rate-limit / quota / connectivity / "cloud agent not enabled" conditions). This removes the bucket from error telemetry while preserving the user-facing message and the logService.error log entry — no logging is removed and no error is swallowed in a try/catch.

This is the intended use of the ChatExpectedError mechanism: a genuinely expected, user-actionable operational condition opts out of bug telemetry rather than being silenced.

Alternatives considered:

• Gating the logService.error(e, ...) call or the chatAgentError emission on the message string 'Language model unavailable' — rejected because string-matching in the consumer is brittle and puts the decision in the wrong place; the producer already knows the condition is expected.
• Wrapping the getModelForRequest call in a try/catch that returns a benign result — rejected because it hides the condition from the user-facing error surface and masks the state instead of classifying it.

Recommended Owner

@bryanchen-d — owner of the chatAgentError telemetry event (owner: 'bryanchen-d' in mainThreadChatAgents2.ts) and author of the isExpectedError / ChatExpectedError suppression mechanism this fix uses. Active VS Code team member; already assigned to the issue.

Generated by errors-fix · 2K AIC · ⌖ 34.2 AIC · ⊞ 71K · ◷

[Copilot]

Copilot can't review bot-authored pull requests automatically. A user with Copilot access can request a review manually.

[Copilot]

Copilot can't review bot-authored pull requests automatically. A user with Copilot access can request a review manually.