swe-1.6 model not accessible in WindsurfAPI despite working in Windsurf IDE and Devin CLI

[dmdfami]

Issue Report: swe-1.6 model not working in WindsurfAPI

Title

swe-1.6 model not accessible in WindsurfAPI despite working in Windsurf IDE and Devin CLI

Problem Description

The swe-1.6 and swe-1.6-fast models cannot be used through WindsurfAPI, even though they work perfectly in:

• Windsurf IDE
• Devin CLI (not Claude CLI)

These models are listed as free/default models and should be accessible to Pro accounts.

Environment

• WindsurfAPI version: v2.0.97
• Account type: Pro (Trial plan with 83% quota remaining, $200 extra usage)
• OS: macOS (Darwin 25.5.0)
• Windsurf session: Using devin-session-token from active Devin CLI session

Steps to Reproduce

1. Setup WindsurfAPI

git clone https://github.com/dwgx/WindsurfAPI.git
cd WindsurfAPI
bash setup.sh
node src/index.js

2. Add Account (Pro account with swe-1.6 access)

# Using devin-session-token from active Devin CLI session
curl -X POST http://localhost:3003/auth/login \
  -H "Content-Type: application/json" \
  -d '{"token": "devin-session-token$..."}'

3. Attempt to call swe-1.6

curl http://localhost:3003/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"swe-1.6","messages":[{"role":"user","content":"Hello"}],"max_tokens":10}'

Expected Behavior

• Should successfully call swe-1.6 model
• Should return response from swe-1.6 (as it does in Windsurf IDE/Devin CLI)

Actual Behavior

Error 1: Model entitlement issue

{
  "error": {
    "message": "模型 swe-1.6 在当前账号池中不可用（未订阅或已被封禁）",
    "type": "model_not_entitled"
  }
}

Error 2: After bypassing entitlement check

{
  "error": {
    "message": "unknown model UID MODEL_SWE_1_6: model not found",
    "type": "upstream_error"
  }
}

Error 3: After trying legacy flow (modelUid=null)

{
  "error": {
    "message": "neither PlanModel nor RequestedModel specified. You must specify a valid model.",
    "type": "upstream_error"
  }
}

Investigation Results

1. Account Information

From accounts.json:

• Account shows as "pro" tier but userStatus.pro = false (shows as "Trial")
• Account has valid Pro entitlements for other models (claude-opus-4.6, swe-1.5, etc.)
• swe-1.6 shows as "ok": false, "reason": "not_entitled"

2. Model Configuration

From src/models.js (line 214-220):

// Proto canonical enums: 359=MODEL_SWE_1_5 (fast), 369=THINKING, 377=SLOW, 420=1_6, 421=1_6_FAST.
'swe-1.5':     { name: 'swe-1.5',     provider: 'windsurf', enumValue: 377, modelUid: 'MODEL_SWE_1_5_SLOW', credit: 0.5 },
'swe-1.5-fast': { name: 'swe-1.5-fast', provider: 'windsurf', enumValue: 359, modelUid: 'MODEL_SWE_1_5', credit: 0.5 },
'swe-1.6':     { name: 'swe-1.6',     provider: 'windsurf', enumValue: 420, modelUid: 'MODEL_SWE_1_6', credit: 0.5 },
'swe-1.6-fast': { name: 'swe-1.6-fast', provider: 'windsurf', enumValue: 421, modelUid: 'MODEL_SWE_1_6_FAST', credit: 0.5 },

Enum values 420 and 421 are correctly assigned for swe-1.6 models.

3. Root Cause Analysis

From src/models.js (line 222-227):

// ── Adaptive (Windsurf 2026-04-06 changelog) ────────────
// Adaptive Model Router + Arena models live in the cloud catalog but their
// UIDs aren't recognized by SendUserCascadeMessage's direct-call path —
// upstream returns "unknown model UID adaptive: model not found". They only
// work through the Windsurf IDE's special routing layer that Cascade-direct
// doesn't expose. Mark deprecated so they stop showing in /v1/models. #109.

Key Finding: The comment indicates that adaptive models (like arena-fast, arena-smart) only work through "Windsurf IDE's special routing layer" that the Cascade-direct path (used by WindsurfAPI) doesn't expose.

Hypothesis: swe-1.6 may be an adaptive model that requires the same special routing layer that Windsurf IDE and Devin CLI have access to, but WindsurfAPI doesn't.

4. Comparison with Working Models

• ✅ swe-1.5-fast - Works perfectly (enumValue: 359, modelUid: 'MODEL_SWE_1_5')
• ✅ swe-1.5 - Works perfectly (enumValue: 377, modelUid: 'MODEL_SWE_1_5_SLOW')
• ✅ claude-opus-4.6 - Works perfectly
• ❌ swe-1.6 - Fails with "unknown model UID"
• ❌ swe-1.6-fast - Fails with "unknown model UID"

5. Architecture Analysis

From src/windsurf.js (line 6-8):

* Two flows:
*   Legacy  → RawGetChatMessage (streaming, simpler)
*   Cascade → StartCascade → SendUserCascadeMessage → poll GetCascadeTrajectorySteps

From src/models.js (line 4-6):

* Routing logic:
*   modelUid present  → Cascade flow (StartCascade → SendUserCascadeMessage)
*   only enumValue>0  → RawGetChatMessage (legacy)

Both routing methods fail for swe-1.6, suggesting the issue is with the underlying protocol/backend support rather than routing logic.

Attempted Workarounds

1. Manual entitlement override

Modified accounts.json to set swe-1.6 entitlement to true - Result: Auto-reverted by WindsurfAPI's probe mechanism

2. Model UID mapping

Attempted to map swe-1.6 → MODEL_SWE_1_5 (swe-1.5-fast UID) - Result: Works but calls wrong model (swe-1.5 instead of swe-1.6)

3. Legacy flow (modelUid=null)

Set modelUid to null to force legacy RawGetChatMessage flow - Result: "neither PlanModel nor RequestedModel specified" error

4. Entitlement check bypass

Modified src/handlers/chat.js to bypass entitlement check for swe-1.6 - Result: Still fails with "unknown model UID" error

Impact

• Users cannot access swe-1.6 through WindsurfAPI despite having valid Pro accounts
• swe-1.6 is reportedly a free/default model in Windsurf IDE/Devin CLI
• Creates inconsistency between WindsurfAPI and official Windsurf tools
• Limits the utility of WindsurfAPI for users who specifically want swe-1.6

Suggested Solutions

1. Implement Special Routing Layer (Recommended)

• Reverse-engineer or obtain documentation for Windsurf IDE's special routing layer
• Implement this routing mechanism in WindsurfAPI
• This would enable adaptive models like swe-1.6, arena models, etc.

2. Direct Cloud API Integration

• Investigate if there's a direct cloud API endpoint (like the GetCascadeModelConfigs endpoint in src/windsurf-api.js)
• Implement alternative calling mechanism that bypasses the local language server limitation
• The cloud API might support swe-1.6 where the local LS doesn't

3. Backend Registration

• Work with Windsurf team to register MODEL_SWE_1_6 and MODEL_SWE_1_6_FAST in the Cascade-direct path
• This would make the model UIDs recognizable by the current SendUserCascadeMessage flow

4. Documentation and Deprecation

• If swe-1.6 truly cannot be supported through the current architecture:
  • Clearly document this limitation
  • Remove swe-1.6 from the model catalog or mark it as unsupported
  • Provide guidance on using swe-1.5-fast as an alternative

Additional Information

Cloud API Endpoints

From src/windsurf-api.js:

• server.self-serve.windsurf.com/exa.api_server_pb.ApiServerService/GetCascadeModelConfigs
• This endpoint successfully fetches model configs including swe-1.6
• Suggests the cloud backend recognizes swe-1.6 even if the local LS doesn't

Account Verification

Account used for testing has:

• Email: dmd.fami@gmail.com
• Pro tier access (despite showing as "Trial" in some fields)
• 83% daily quota remaining
• $200 extra usage balance
• Successfully uses swe-1.6 in Windsurf IDE and Devin CLI

Related Issues

• Issue [Feature] 这个没用5.5模型 #109: Similar adaptive model routing limitation mentioned in code comments
• The comment explicitly mentions arena models having the same "special routing layer" requirement

Priority

High - This affects a core model that is:

• Advertised as free/default in official Windsurf tools
• Working in Windsurf IDE and Devin CLI
• Not accessible through WindsurfAPI despite valid account credentials
• Creates significant inconsistency in the Windsurf ecosystem