Context

ZenMux had two related reliability issues:

1. Model metadata/caching could produce invalid or incomplete model capabilities (notably around context window and native tool support), which led to unstable behavior.
2. Native tool-call flow in ZenMux could trigger repeated "tool not used" retry loops because tool capability and request-shaping were not consistently preserved through runtime paths.

This PR hardens ZenMux model normalization/caching and fixes the native tool/message pipeline so tool-enabled tasks behave consistently.

Implementation

This PR combines two branch commits:

• first commit: Context window/cache hardening

  • Maps ZenMux context_length into contextWindow.
  • Adds fallback behavior when optional model metadata is missing.
  • Adds stale ZenMux cache self-healing for invalid contextWindow <= 0 records.
  • Adds/extends ZenMux fetcher and model cache tests for stale cache rejection and valid cache acceptance.

• second commit: Native tool reliability + pipeline cleanup

  • Adds explicit native-tool defaults for ZenMux model info (supportsNativeTools, defaultToolProtocol).
  • Ensures fetched ZenMux models include native-tool metadata.
  • Merges native-tool defaults at runtime to self-heal stale model entries missing native tool fields.
  • Respects task-locked toolProtocol and gates tool params accordingly.
  • Refactors ZenMux stream creation to use already-transformed OpenAI messages (instead of rebuilding from raw messages), preserving protocol/model-specific transforms.
  • Adds dedicated regression tests covering:
    • stale metadata default merge
    • native vs xml tool gating
    • DeepSeek R1 transformed-message passthrough
  • Adds user-facing patch changeset for release notes.

How to Test

1. Run ZenMux-targeted backend tests from src:

• cd src && pnpm test api/providers/fetchers/__tests__/zenmux.spec.ts
• cd src && pnpm test api/providers/__tests__/zenmux.spec.ts
• cd src && pnpm test api/providers/__tests__/zenmux-native-tools.spec.ts

2. Run extension-wide backend suite:

• cd src && pnpm test

3. Validate repo gates:

• pnpm check-types
• pnpm lint
• pnpm build

4. Manual behavior check (ZenMux):

• Configure provider: zenmux
• Use model: z-ai/glm-5
• In Ask mode, send a simple prompt like hi
• Confirm it does not enter repeated "did not use a tool" retry loops due to missing native-tool wiring.

@simonleung963 reported the problem on Discord and tested my custom builds.