fix: default tokenizer to O200K when server omits it (fixes #319618)

[vs-code-engineering]

Summary

The CAPI model metadata API can return models without a tokenizer field in their capabilities. When this happens, ChatEndpoint.tokenizer is set to undefined, and TokenizerProvider.acquireTokenizer() throws Error: Unknown tokenizer: undefined. This error affects all platforms (Linux, Windows, Mac) and has been consistently hitting since v1.121.0 with 44,506 hits on the latest version alone.

Fixes #319618
Recommended reviewer: @lramos15

Culprit Commit

Not a single-commit regression. The error has been present since at least v1.121.0 (first seen 2026-05-20). The root cause is that chatEndpoint.ts line 167 trusts the server API response to always include capabilities.tokenizer, but some models omit this field.

Code Flow

graph TD
    A[CAPI /models API response] -->|capabilities.tokenizer = undefined| B[ChatEndpoint constructor]
    B -->|this.tokenizer = undefined| C[ChatEndpoint instance]
    C -->|endpoint passed to PromptRenderer| D[PromptRenderer constructor]
    D -->|tokenizerProvider.acquireTokenizer endpoint| E[TokenizerProvider.acquireTokenizer]
    E -->|switch on undefined| F[throw Error Unknown tokenizer: undefined]

Loading

Affected Files

• extensions/copilot/src/platform/endpoint/node/chatEndpoint.ts — producer of the undefined tokenizer value (line 167)
• extensions/copilot/src/platform/tokenizer/node/tokenizer.ts — crash site (line 82)
• extensions/copilot/src/extension/prompts/node/base/promptRenderer.ts — intermediate caller

Repro Steps

1. Have a Copilot model available via CAPI that does not include a tokenizer field in its capabilities metadata
2. Trigger any chat interaction that uses that model (e.g., send a message)
3. The PromptRenderer attempts to acquire a tokenizer for the endpoint, hitting the throw

How the Fix Works

Chosen approach (extensions/copilot/src/platform/endpoint/node/chatEndpoint.ts): Added a nullish coalescing default (?? TokenizerType.O200K) at line 167 where modelMetadata.capabilities.tokenizer is assigned to this.tokenizer. This fixes the issue at the data producer — where the potentially-undefined server value is first consumed — rather than at the crash site in acquireTokenizer(). O200K is the standard default used across all other endpoint implementations (BYOK, proxy endpoints, xtab, etc.).

Alternatives considered: Adding a fallback case in TokenizerProvider.acquireTokenizer() would fix at the crash site rather than the data producer, hiding the fact that server data was incomplete and potentially masking other issues downstream that depend on endpoint.tokenizer being a valid enum value.

Recommended Owner

@lramos15 — primary author of the endpoint/model-metadata infrastructure in extensions/copilot/src/platform/endpoint/.

Generated by errors-fix · ● 65.4M · ◷

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.