providerContextCaps.openai is ignored for native GPT-5.6 catalog rows

[mkarolczyk]

Client or integration

Codex App

Area

Catalog / models

Summary

The supported per-provider context cap is persisted and reported as active for openai, but native GPT-5.6 catalog rows ignore it.

On OpenCodex 2.12.0, both contextCapValue and providerContextCaps.openai are set to 272000. After restarting the managed proxy and running a full opencodex sync, GPT-5.6 Sol, Terra, and Luna are still generated with a 372,000-token context window and a 334,800-token auto-compaction threshold.

I expected the enabled OpenAI provider cap to constrain native GPT-5.6 models to 272,000 tokens, with auto-compaction at 244,800.

This is intentionally narrower than #1051: regardless of whether 372k is the desired default, the documented/configurable provider cap currently has no effect on the native GPT-5.6 overrides.

Reproduction

1. Install or update to OpenCodex 2.12.0.
2. Configure the global and OpenAI provider caps:

{
  "contextCapValue": 272000,
  "providerContextCaps": {
    "openai": 272000
  }
}

The same result occurs when the OpenAI cap is enabled through the management API. Readback confirms providerContextCaps.openai = 272000 and the API reports the cap without degraded mode.

3. Restart the managed OpenCodex service so the installed version is active.
4. Run a full sync:

opencodex sync

5. Verify the persisted configuration:

jq '{contextCapValue, providerContextCaps: {openai: .providerContextCaps.openai}}' \
  ~/.opencodex/config.json

Result:

{
  "contextCapValue": 272000,
  "providerContextCaps": {
    "openai": 272000
  }
}

6. Inspect the generated catalog or runtime model metadata:

codex debug models | jq -r '
  .models[]
  | select(.slug | test("^gpt-5\\.6-(sol|terra|luna)$"))
  | [.slug, .context_window, .max_context_window, .auto_compact_token_limit]
  | @tsv
'

Actual result:

gpt-5.6-sol    372000    372000    334800
gpt-5.6-terra  372000    372000    334800
gpt-5.6-luna   372000    372000    334800

Expected result while the OpenAI provider cap is enabled:

gpt-5.6-sol    272000    272000    244800
gpt-5.6-terra  272000    272000    244800
gpt-5.6-luna   272000    272000    244800

The proxy is healthy and opencodex doctor confirms that the Codex app-server catalog is current with the on-disk catalog, which rules out a stale proxy or stale model cache.

Suspected cause

The native GPT-5.6 rows appear to bypass the configured provider cap because two source-level overrides remain fixed at 372,000:

src/codex/catalog/metadata.ts
NATIVE_GPT56_CONTEXT_WINDOW = 372_000

src/providers/registry.ts
OPENAI_CODEX_GPT56_CONTEXT_WINDOW = 372_000

Changing only those two values to 272_000, restarting the managed service, and running opencodex sync immediately produces the expected 272000 / 272000 / 244800 values for Sol, Terra, and Luna. Spark remains unchanged at 100000 / 100000 / 90000, confirming that the failure is scoped to the native GPT-5.6 override path.

Suggested behavior

When providerContextCaps.openai is enabled, apply it as a ceiling to native OpenAI catalog rows as well as routed/provider-discovered models. The 372k native value can remain the default when no custom cap is enabled.

Related reports:

• #1051 — original Codex App report; closed after 372k was described as intentional.
• [Bug] Claude Code shows 200k context for native OpenAI models (gpt-5.6/5.5): model-info.ts omits contextWindow; NATIVE_GPT56_CONTEXT_WINDOW stale at 372k #1218 — related Claude Code metadata report; also confirms the upstream 272k value, though the maintainer kept 372k as the OpenCodex default.

Version

2.12.0

Operating system

macOS 27.0 (26A5388g), arm64

Provider and model

OpenAI (Codex login) / gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna

Logs or error output

gpt-5.6-sol    372000    372000    334800
gpt-5.6-terra  372000    372000    334800
gpt-5.6-luna   372000    372000    334800

Screenshots and supporting files

• OpenCodex removes model_context_window and changes Codex App context from 272k to 372k #1051
• [Bug] Claude Code shows 200k context for native OpenAI models (gpt-5.6/5.5): model-info.ts omits contextWindow; NATIVE_GPT56_CONTEXT_WINDOW stale at 372k #1218
• Backport refreshed bundled model metadata to 0.144 openai/codex#33972
• Narrow 0.144 hotfix to GPT-5.6 prompts and context openai/codex#34009

Redacted configuration

{
  "providers": {
    "openai": {
      "adapter": "openai-responses",
      "baseUrl": "https://chatgpt.com/backend-api/codex",
      "codexAccountMode": "pool",
      "authMode": "forward"
    }
  },
  "contextCapValue": 272000,
  "providerContextCaps": {
    "openai": 272000
  }
}

Checks

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

[Ingwannu]

Confirmed against current dev: the report is valid, and the fixed 372k constants are defaults rather than the real configuration contract. providerContextCaps.openai should act as a ceiling across every native OpenAI catalog surface.

PR #1435 is close and now handles synchronized/preserved rows and routing evidence correctly, but its current head still omits the cap on the live Codex discovery route at src/server/index.ts:914 (GET /v1/models?client_version=...). I requested that final endpoint fix plus an endpoint-level regression and a refresh onto current dev before merge.

Keeping this issue open until the corrected exact head passes CI and is integrated into dev.

[Wibias]

Closed by #1435