[Bug]: GENERATION_FAILED (fp: ebd4e21f)

[rsxdalv]

Description

Custom OAI provider not working

Steps to reproduce

1. Add OAI provider - minimaxi.com/api/v1
2. Try to generate anything

Expected behavior

No error. The API works in other clients, such as OpenCode and VSCode, with the exact same configuration.

Actual behavior

Error invoking remote method 'codesign:v1:generate': CodesignError: 400 invalid params, invalid role: developer (2013) — user notes: Happens when trying to use www,minimaxi.com/v1 API, any prompt

Something does go through as the project gets named: The user wants me to create a

open-codesign version

0.1.4

Platform

Windows

OS version

Windows 11 (10.0.26200)

Model provider (if relevant)

Other

Error code

GENERATION_FAILED

Diagnostics bundle

Bundle saved locally at ~\Downloads\open-codesign-diagnostics-2026-04-26T09-10-08.zip. Attach it to this issue after submitting.

Logs

[2026-04-26 12:06:17.615] [info]   (main:ipc) [generate] step=send_request {
  runId: 'mofjna12-n9uy4h',
  generationId: 'mofjna12-n9uy4h',
  provider: 'custom-minimaxi-gy2s',
  modelId: 'MiniMax-M2.7'
}
[2026-04-26 12:06:17.615] [info]   (main:ipc) agent.turn_start { runId: 'mofjna12-n9uy4h', generationId: 'mofjna12-n9uy4h' }
[2026-04-26 12:06:17.615] [info]   (main:ipc) [context-prune] step=caps {
  runId: 'mofjna12-n9uy4h',
  generationId: 'mofjna12-n9uy4h',
  messages: 2,
  before: 557,
  after: 557,
  textLimit: 8192,
  toolInputLimit: 24576,
  toolResultLimit: 8192,
  window: 3
}
[2026-04-26 12:06:18.728] [info]   (main:ipc) agent.turn_end {
  runId: 'mofjna12-n9uy4h',
  generationId: 'mofjna12-n9uy4h',
  deltas: 0,
  tools: 0
}
[2026-04-26 12:06:18.728] [info]   (main:ipc) agent.end { runId: 'mofjna12-n9uy4h', generationId: 'mofjna12-n9uy4h' }
[2026-04-26 12:06:18.728] [error]  (main:ipc) [generate] step=send_request.fail {
  runId: 'mofjna12-n9uy4h',
  generationId: 'mofjna12-n9uy4h',
  provider: 'custom-minimaxi-gy2s',
  modelId: 'MiniMax-M2.7',
  ms: 1113,
  stopReason: 'error'
}
[2026-04-26 12:06:18.729] [error]  (main:ipc) generate.fail {
  runId: 'mofjna12-n9uy4h',
  generationId: 'mofjna12-n9uy4h',
  ms: 1117,
  provider: 'custom-minimaxi-gy2s',
  modelId: 'MiniMax-M2.7',
  baseUrl: '[url omitted]',
  status: 400,
  message: '400 invalid params, invalid role: developer (2013)',
  code: 'PROVIDER_ERROR'
}
[2026-04-26 12:06:18.732] [error]  (renderer:console) [store] applyGenerateError {"generationId":"mofjna12-n9uy4h","designId":"3811e65a-59ac-41b8-af9e-d7901c46682e","message":"Error invoking remote method 'codesign:v1:generate': CodesignError: 400 invalid params, invalid role: developer (2013)"} {}
[2026-04-26 12:06:18.733] [info]   (chat-messages-ipc) chat.append {
  designId: '3811e65a-59ac-41b8-af9e-d7901c46682e',
  seq: 3,
  kind: 'error'
}

[github-actions]

Thank you for the report. I could verify the provider layer does have special handling for this exact class of failure: the app stores the custom provider's wire and resolved capabilities, then generation always uses that saved provider config (apps/desktop/src/main/provider-settings.ts:237, apps/desktop/src/main/index.ts:790). For custom providers on openai-responses, the default capability profile enables Responses API semantics and reasoning (packages/shared/src/config.ts:208, packages/shared/src/config.ts:211), while openai-chat is supposed to keep reasoning off for third-party gateways unless an explicit capability or known reasoning-model heuristic turns it back on (packages/providers/src/wire-policy.ts:20, packages/providers/src/wire-policy.ts:75). The MiniMax-specific heuristic currently only matches minimax/minimax-m... style model IDs (packages/providers/src/wire-policy.ts:50), and I did not find a dedicated MiniMax provider adapter beyond the generic OpenAI-compatible path.

Based on that, your 400 invalid params, invalid role: developer is consistent with the app sending a reasoning-enabled request, but for a custom OpenAI-compatible provider that should only happen on certain saved wire/capability combinations. If your provider was saved as openai-chat, this looks like a real bug. Open CoDesign is still pre-alpha, so custom-provider edge cases are still being hardened.

Relevant code:

• apps/desktop/src/main/index.ts:790 — generation resolves the canonical active provider from saved config before making the request
• apps/desktop/src/main/provider-settings.ts:237 — saved provider entry supplies wire, baseUrl, and resolved capabilities
• packages/shared/src/config.ts:207 — openai-chat defaults to chat completions
• packages/shared/src/config.ts:208 — openai-responses defaults to Responses API semantics
• packages/shared/src/config.ts:211 — reasoning defaults on for openai-responses
• packages/providers/src/wire-policy.ts:20 — comment documents that developer is injected when reasoning is enabled and can be rejected by third-party gateways
• packages/providers/src/wire-policy.ts:75 — inferReasoning(...) gate for custom OpenAI-compatible endpoints
• packages/providers/src/wire-policy.ts:50 — MiniMax heuristic currently matches minimax/minimax-m...
• packages/providers/src/index.test.ts:808 — test coverage that proxy endpoints should not inherit official OpenAI/OpenRouter reasoning heuristics after being repointed

Need more info:

• Which wire did you select when adding the custom provider: OpenAI Chat or OpenAI Responses?
• If you open the provider in Settings, what exact Base URL and Default model are saved there now?
• Please attach the diagnostics bundle mentioned in the issue so we can inspect the saved provider metadata for this run.

---

open-codesign Bot

[rsxdalv]

OpenAI Chat

Base URL https://api.minimaxi.com/v1

Model: MiniMax-M2.7

open-codesign-diagnostics-2026-04-26T09-10-08.zip