[BOT ISSUE] Mistral: fix codestral-2508 max_input_tokens regression (256k → 128k)

[github-actions]

Stale metadata (recurring regression)

codestral-2508 has max_input_tokens: 256000 in the catalog but Mistral's official docs confirm the context window is 128,000. The parent alias codestral-latest correctly has max_input_tokens: 128000, creating an inconsistency.

Current vs correct

Field	Current (wrong)	Correct	Source
max_input_tokens	256,000	128,000	Mistral model card

Current catalog entry (line 10903)

"codestral-2508": {
  "format": "openai",
  "flavor": "chat",
  "input_cost_per_mil_tokens": 0.3,
  "output_cost_per_mil_tokens": 0.9,
  "displayName": "Codestral 2508",
  "parent": "codestral-latest",
  "max_input_tokens": 256000,
  "available_providers": [
    "mistral",
    "openrouter"
  ]
}

The parent codestral-latest (line 10930) correctly has max_input_tokens: 128000.

Cross-source verification

1. Mistral model card (docs.mistral.ai/models/model-cards/codestral-25-08): States "Context 128k"
2. Mistral known limitations (docs.mistral.ai/resources/known-limitations): Lists Codestral context window as "128k tokens"

Regression history

This gap has been filed and fixed twice before:

• Issue [BOT ISSUE] Mistral: codestral-2508 stale max_input_tokens (256000 → 128000) and voxtral-small-2507 stale output pricing ($0.4 → $0.3) #862 (closed June 26, 2026) — fixed 256k → 128k
• Issue [BOT ISSUE] Mistral: codestral-2508 has stale context window (256k → 128k) #936 (closed July 9, 2026) — fixed the same regression

The fix keeps being reverted by the automated LiteLLM sync because codestral-2508 is not listed in SYNC_PRESERVED_FIELDS in packages/proxy/scripts/sync_models.ts. The downstream fix should also add "codestral-2508": ["max_input_tokens"] to SYNC_PRESERVED_FIELDS to prevent future regressions.

Local Files Inspected

• packages/proxy/schema/model_list.json — codestral-2508 at line 10903 with max_input_tokens: 256000; parent codestral-latest at line 10930 with max_input_tokens: 128000 (correct)
• packages/proxy/scripts/sync_models.ts — no entry for codestral-2508 in SYNC_PRESERVED_FIELDS

{
  "kind": "stale_metadata",
  "provider": "mistral",
  "models": ["codestral-2508"],
  "status": "active",
  "model_specs": {
    "codestral-2508": {
      "format": "openai",
      "flavor": "chat",
      "input_cost_per_mil_tokens": 0.3,
      "output_cost_per_mil_tokens": 0.9,
      "displayName": "Codestral 2508",
      "parent": "codestral-latest",
      "max_input_tokens": 128000,
      "available_providers": [
        "mistral",
        "openrouter"
      ]
    }
  },
  "source_urls": [
    "https://docs.mistral.ai/models/model-cards/codestral-25-08",
    "https://docs.mistral.ai/resources/known-limitations"
  ]
}

Summary: Fix codestral-2508 context window from 256k to 128k (third occurrence of this regression) and add to SYNC_PRESERVED_FIELDS to prevent future sync reversions.

[github-actions]

Prepared catalog updates for codestral-2508. Updated models: codestral-2508.

Verification sources: https://docs.mistral.ai/models/model-cards/codestral-25-08, https://docs.mistral.ai/resources/known-limitations
Verified metadata applied:

• codestral-2508: format=openai, flavor=chat, displayName=Codestral 2508, parent=codestral-latest, max_input_tokens=128000, max_output_tokens=not provided, pricing(input/output per 1M)=0.3/0.9, providers=mistral|openrouter

Included in the daily batch PR: #1081