[Bug]: Kimi-K2.6 and other models incorrectly receives `role: developer`, causing 422 from Azure AI Foundry

[deancallaghan451]

Description

Model incorrectly receives role: developer, causing 422 from Azure AI Foundry

Steps to reproduce

Using Microsoft AI Foundry models via OpenAI Chat API and sending any prompt.

Expected behavior

Models such as Kimi-K2.6 should use role: "system" — it is not a reasoning model.

Actual behavior

Version: v0.1.4
Provider: Azure AI Foundry (services.ai.azure.com/openai/v1)
Model: Kimi-K2.6

What happens:
Generating any design throws a 422 immediately:

Error invoking remote method 'codesign:v1:generate': CodesignError: 422 {"error":{"code":"Invalid input","status":422,"message":"invalid input error","details":[{"type":"enum","loc":["body","messages",0,"role"],"msg":"Input should be 'system', 'user', 'assistant' or 'tool'","input":"developer","ctx":{"expected":"'system', 'user', 'assistant' or 'tool'"}}]}}

Root cause:
Kimi-K2.6 and GPT 5.5 are being misidentified as a reasoning model, causing the provider layer to send role: "developer" instead of role: "system". Azure AI Foundry's Kimi deployment rejects developer as an invalid role.

v0.1.4 fixed this same issue for Qwen and DeepSeek (#183) but Kimi model names are not currently in the exclusion list.

Expected behaviour:
Kimi-K2.6 should use role: "system" — it is not a reasoning model.

Suggested fix:
Add Kimi model name patterns to the non-reasoning model exclusion list alongside Qwen and DeepSeek.

open-codesign version

0.1.4

Platform

macOS

OS version

25.3.0

Model provider (if relevant)

None

Error code

GENERATION_FAILED

Diagnostics bundle

Bundle saved locally at ~/Downloads/open-codesign-diagnostics-2026-04-29T04-47-29.zip. Attach it to this issue after submitting.

Logs

generationId: 'mojkpov4-1ld9gk',
  provider: 'custom-azure-kimik-2-6-njzu',
  modelId: 'Kimi-K2.6-2026-04-20'
}
[2026-04-29 14:47:14.573] [info]   (main:ipc) agent.turn_start { runId: 'mojkpov4-1ld9gk', generationId: 'mojkpov4-1ld9gk' }
[2026-04-29 14:47:14.574] [info]   (main:ipc) [context-prune] step=caps {
  runId: 'mojkpov4-1ld9gk',
  generationId: 'mojkpov4-1ld9gk',
  messages: 10,
  before: 10507,
  after: 10507,
  textLimit: 8192,
  toolInputLimit: 24576,
  toolResultLimit: 8192,
  window: 3
}
[2026-04-29 14:47:15.915] [info]   (main:ipc) agent.turn_end {
  runId: 'mojkpov4-1ld9gk',
  generationId: 'mojkpov4-1ld9gk',
  deltas: 0,
  tools: 0
}
[2026-04-29 14:47:15.915] [info]   (main:ipc) agent.end { runId: 'mojkpov4-1ld9gk', generationId: 'mojkpov4-1ld9gk' }
[2026-04-29 14:47:15.916] [error]  (main:ipc) [generate] step=send_request.fail {
  runId: 'mojkpov4-1ld9gk',
  generationId: 'mojkpov4-1ld9gk',
  provider: 'custom-azure-kimik-2-6-njzu',
  modelId: 'Kimi-K2.6-2026-04-20',
  ms: 1343,
  stopReason: 'error'
}
[2026-04-29 14:47:15.917] [error]  (main:ipc) generate.fail {
  runId: 'mojkpov4-1ld9gk',
  generationId: 'mojkpov4-1ld9gk',
  ms: 1358,
  provider: 'custom-azure-kimik-2-6-njzu',
  modelId: 'Kimi-K2.6-2026-04-20',
  baseUrl: '[url omitted]',
  status: 422,
  message: `422 {"error":{"code":"Invalid input","status":422,"message":"invalid input error","details":[{"type":"enum","loc":["body","messages",0,"role"],"msg":"Input should be 'system', 'user', 'assistant' or 'tool'","input":"developer","ctx":{"expected":"'system', 'user', 'assistant' or 'tool'"}}]}}\n` +
    'Please check this guide to understand why this error code might have been returned \n' +
    '[url omitted]',
  code: 'PROVIDER_ERROR'
}
[2026-04-29 14:47:15.924] [error]  (renderer:console) [store] applyGenerateError {"generationId":"mojkpov4-1ld9gk","designId":"6336f2f6-6633-4cd0-85da-42bfd16f6a5a","message":"Error invoking remote method 'codesign:v1:generate': CodesignError: 422 {\"error\":{\"code\":\"Invalid input\",\"status\":422,\"message\":\"invalid input error\",\"details\":[{\"type\":\"enum\",\"loc\":[\"body\",\"messages\",0,\"role\"],\"msg\":\"Input should be 'system', 'user', 'assistant' or 'tool'\",\"input\":\"developer\",\"ctx\":{\"expected\":\"'system', 'user', 'assistant' or 'tool'\"}}]}}\nPlease check this guide to understand why this error code might have been returned \n[url omitted]"} {}
[2026-04-29 14:47:15.926] [info]   (chat-messages-ipc) chat.append {
  designId: '6336f2f6-6633-4cd0-85da-42bfd16f6a5a',
  seq: 19,
  kind: 'error'
}

[github-actions]

Thanks for reporting this. I could verify the failure class in the current code, but I could not verify from the repository that Kimi-K2.6 is classified as a reasoning model by name alone on third-party openai-chat endpoints.

What is verified in code today:

• For third-party openai-chat providers, reasoning is enabled by an allowlist in packages/providers/src/wire-policy.ts:50-58 and applied in packages/providers/src/wire-policy.ts:75-105. That allowlist currently includes :thinking, Claude 4, OpenAI o* / gpt-5, MiniMax M, deepseek-r, and qwen/qwq; there is no Kimi-specific matcher there.
• The generated Pi model then inherits that reasoning flag in packages/providers/src/index.ts:197-225 and packages/providers/src/index.ts:266-279. That is the path that can lead pi-ai to emit role: developer on OpenAI-style wires.
• A separate verified path can also force reasoning on: if the provider has a saved reasoning-depth override, resolved capabilities set supportsReasoning: true in packages/shared/src/config.ts:211-215, and inferReasoning() immediately returns true for that provider in packages/providers/src/wire-policy.ts:82-84.
• Your note that connection test passes also matches the current code: apps/desktop/src/main/connection-ipc.ts:547-576 prefers /models, and the fallback probe in apps/desktop/src/main/connection-ipc.ts:732-745 sends only a minimal user message, so it would not catch a system/developer role mismatch during generation.

This looks related to #183 rather than an exact duplicate. #183 fixed the broad "custom OpenAI-compatible endpoints reject developer role" path, and #207 added the explicit wire/role/reasoning policy, but I do not see a Kimi-specific heuristic in the current matcher.

Relevant code:

• packages/providers/src/wire-policy.ts:50-58 — third-party reasoning model allowlist
• packages/providers/src/wire-policy.ts:75-105 — inferReasoning() decision path
• packages/providers/src/index.ts:197-225 — synthesized PiModel gets reasoning
• packages/providers/src/index.ts:266-279 — registry-backed PiModel reasoning override
• packages/shared/src/config.ts:211-215 — reasoning-depth override implies supportsReasoning: true
• apps/desktop/src/main/connection-ipc.ts:547-576 — connection test uses /models when available
• apps/desktop/src/main/connection-ipc.ts:732-745 — fallback inference probe uses only a user message

Need more info:

• In Settings for this provider, is Reasoning depth still Default, or did you set it to a specific value?
• Is the provider saved with wire openai-chat or openai-responses?
• If you can attach the diagnostics bundle or paste the redacted provider config fields, the minimum useful fields are baseUrl, wire, reasoningLevel, and capabilities.

---

open-codesign Bot

[rsxdalv]

Same for minimax models

[github-actions]

Thanks for the additional report, @rsxdalv. The same underlying mechanism affects any model that is incorrectly identified as a reasoning model by name-based pattern matching.

As noted in the previous response, the allowlist in packages/providers/src/wire-policy.ts:50-58 currently includes Minimax M, but this may not cover all minimax model IDs (e.g., MiniMax-Text-01 or others). If you can share the exact model ID that fails for minimax, that would help confirm whether the pattern needs extension.

For both Kimi and minimax cases, the fix would involve adding appropriate model name patterns to that allowlist (or excluding them if they are not reasoning models). The root cause is that non-reasoning models hosted on OpenAI-compatible endpoints that reject developer role should fall back to system.

If you haven't already, please also answer the questions from the previous bot response about reasoning depth setting and wire type — that will help isolate whether the issue is purely name-based or involves a saved override.

---

Open-CoDesign Bot

[Sun-sunshine06]

PR #291 has been merged into main and this issue is now closed.

What landed:

• Kimi/Moonshot and MiniMax model IDs now short-circuit to non-reasoning on OpenAI Chat-compatible wires.
• MiniMax namespaced IDs no longer match the broad third-party reasoning allowlist.
• Regression coverage includes Kimi-K2.6-2026-04-20, moonshotai/kimi-k2-instruct, MiniMax-M2.7, and minimax/minimax-m2.7.

Validation on the PR:

• CI Lint, typecheck, test passed
• CodeQL passed
• Codex PR review found no blocker or major issues
• Dependency review passed