[Bug] Stale default model for zai provider (glm-5.1 removed from catalog) silently falls back to outdated glm-4.7

[aniruddhaadak80]

Area

AI providers and models

Description

defaultModelPerProvider in packages/coding-agent/src/core/model-resolver.ts:20-53 maps the zai provider to "glm-5.1", but the generated model catalog (packages/ai/src/models.generated.ts, current main) no longer contains any glm-5.1 entry under provider zai. The catalog offers only:

• glm-4.7
• glm-5-turbo
• glm-5.2
• glm-5.2-highspeed
• glm-5.3

Z.AI's direct API catalog has moved on to the 5.x generation (prime-inference's default is z-ai/glm-5.2, defined at model-resolver.ts:17), so the stale reference makes every fallback path that relies on defaultModelPerProvider["zai"] miss and degrade to providerModels[0], which is glm-4.7 (catalog insertion order). Users silently end up on a much older model than intended, with no warning.

Two code paths are affected:

1. Saved default rebuild (model-resolver.ts:552-568): when the stored settings default no longer exists in this build's snapshot, buildFallbackModel (model-resolver.ts:157-171) looks up defaultModelPerProvider[provider] as the replacement template. For zai that lookup misses, so the rebuild-from-template design ("so it survives updates") fails and the session lands on glm-4.7.

2. Fresh selection with only Z.AI configured (findPreferredDefaultModel, model-resolver.ts:173-190): the loop finds no match for zai (stale id), falls through every other provider, and the caller returns availableModels[0] (model-resolver.ts:575) - again glm-4.7 for a Z.AI-only user.

Note the asymmetry: vercel-ai-gateway still serves zai/glm-5.1, so users migrating from gateway configs to a direct Z.AI API key hit this when their saved id disappears from the direct-provider snapshot.

Steps to reproduce

1. Authenticate the zai provider (set ZAI_API_KEY).
2. Set the saved default to zai/glm-5.1, or have zai as the only configured provider.
3. Start prime-agent and check the active model in /model or /status.

The fallback ordering in buildFallbackModel (... ?? providerModels[0]) deterministically selects glm-4.7 because it is the first zai entry in models.generated.ts.

Expected behavior

The zai default should reference a model that exists in the current catalog (e.g. glm-5.2, matching PRIME_INFERENCE_DEFAULT_MODEL_ID = "z-ai/glm-5.2", or glm-5.3, the newest available), and fallback should resolve to that generation instead of the oldest entry.

Actual behavior

Requests are silently served by glm-4.7. No warning is emitted; /status shows the rebuilt/fallback model only if the user inspects it.

Prime Agent version

main @ e319a66 (v0.8.0 is latest release)

Environment

Any (catalog data + resolution logic; platform-independent)

Additional context

I first filed this as issue #1676; it was auto-closed by the contribution vouch gate within seconds, so I am re-filing it here per CONTRIBUTING.md's discussion-first process.

Suggested fixes (maintainer's call on target id):

• Update defaultModelPerProvider["zai"] to an existing catalog id (glm-5.2 for parity with the prime-inference default, or glm-5.3 for newest).
• Optionally add a CI assertion (in the catalog generator or a unit test) that every defaultModelPerProvider value exists under its provider in the freshly generated snapshot, so future upstream removals fail loudly instead of degrading silently.

[aniruddhaadak80]

Prepared a reference implementation on my fork, since the vouch gate auto-closes PRs from unvouched accounts. Happy to open the upstream PR the moment a maintainer invites it.

Branch: https://github.com/aniruddhaadak80/prime-agent/tree/fix/zai-default-model-glm-5-2 (single commit 3b988eea8, based on current main e319a66d7)

Changes:

1. packages/coding-agent/src/core/model-resolver.ts - defaultModelPerProvider.zai: "glm-5.1" to "glm-5.2", matching PRIME_INFERENCE_DEFAULT_MODEL_ID = "z-ai/glm-5.2" and the newest generation available on the direct Z.AI catalog (glm-4.7, glm-5-turbo, glm-5.2, glm-5.2-highspeed, glm-5.3). Went with 5.2 over 5.3 for parity with the prime-inference default; happy to switch if maintainers prefer tracking newest.
2. packages/coding-agent/test/suite/regressions/1676-zai-stale-default.test.ts - new regression test asserting every defaultModelPerProvider entry resolves under its provider via the generated catalog (getModels). Verified red/green: fails with stale defaults: zai -> "glm-5.1" against the old value, passes with the fix.
3. packages/coding-agent/test/model-resolver.test.ts - updated the existing "zai, minimax, and cerebras defaults track current models" assertion from glm-5.1 to glm-5.2.
4. packages/coding-agent/.changes/1676-fix-zai-default-model.md - changelog fragment per CONTRIBUTING.md.

Validation: regression + full model-resolver.test.ts suites pass (41/41), and the pre-commit hook ran the complete npm run check chain clean (biome over 939 files, tsgo, installer render check, browser smoke check).