devin adapter ignores provider output budget — every Codex turn is silently capped at the encoder's 8192 default #5190

Client or integration

Codex CLI

Area

Provider adapter

Summary

The devin adapter never forwards a configured output budget. When the caller omits max_output_tokens — which Codex always does — the cloud-direct encoder falls back to DEFAULT_MAX_OUTPUT_TOKENS = 8192 (src/adapters/devin/cloud-direct/chat.ts), so every devin/* turn is capped at 8192 output tokens. Any turn that legitimately needs more (e.g. writing a whole source file in one tool call) ends as an upstream incomplete / max_output_tokens, which Codex surfaces as stream disconnected before completion: Incomplete response returned, reason: max_output_tokens after several full-size retries hit the same deterministic wall.

There is also no config escape hatch: applyOAuthPresetCatalog in src/oauth/index.ts iterates OAUTH_RECONCILE_FIELDS (which includes defaultMaxOutputTokens and modelMaxOutputTokens) and deletes any field the registry preset doesn't define, so a hand-edited value in config.json is wiped on every startup. The devin registry entry defines neither field.

Reproduction

1. ocx start with a signed-in devin provider (authMode: oauth, adapter: devin, https://server.codeium.com).
2. Route a Codex turn to devin/swe-2 that produces a long single response (e.g. "write this entire C file").
3. Observe ~/.opencodex/usage.jsonl: outputTokens: 8192 exactly, terminalStatus: "incomplete", terminalSource: "upstream", upstreamError: "Output reached the requested token limit (max_output_tokens)". Codex retries the turn verbatim (identical inputTokens across consecutive entries) and each retry dies at the same bound.
4. Separately: curl /v1/responses with "max_output_tokens": 64 truncates at exactly 64 tokens — proving the wire field reaches the encoder — while any request omitting it cannot exceed 8192.

Version

2.58.0 (still present in 2.59.0 — maxInputTokens was wired to config there, maxOutputTokens was not)

Operating system

Windows 11 x64

Provider and model

devin / swe-2 (affects every devin model)

Logs or error output

{"provider":"devin","model":"swe-2","status":200,"terminalSource":"upstream","terminalStatus":"incomplete","closeReason":"terminal","upstreamError":"Output reached the requested token limit (max_output_tokens)","usage":{"inputTokens":136830,"outputTokens":8192,"totalTokens":145022}}

Codex client side:

stream disconnected before completion: Incomplete response returned, reason: max_output_tokens

Suggested fix

• In adapters/devin.ts runTurn, resolve the cap like the other adapters do:

  const maxOutputTokens = typeof parsed.options.maxOutputTokens === "number"
    ? parsed.options.maxOutputTokens
    : modelRecordValue(provider.modelMaxOutputTokens, parsed.modelId) ?? provider.defaultMaxOutputTokens;

  and pass it through when > 0.
• Give the devin registry entry a defaultMaxOutputTokens (the Anthropic entry already does this for the same reason), and/or exempt defaultMaxOutputTokens/modelMaxOutputTokens from the OAuth reconcile delete branch, the way requiresReasoningPlaceholderModels is already exempted.

Verified locally with those two changes: a request that previously truncated at exactly 8192 completed with 16,327 output tokens.

Redacted configuration

{
  "providers": {
    "devin": {
      "adapter": "devin",
      "baseUrl": "https://server.codeium.com",
      "authMode": "oauth",
      "defaultModel": "swe-2",
      "liveModels": true
    }
  }
}

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.