GPT-5.3 Codex Spark fails because Codex sends unsupported reasoning.summary

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #31846
• #31969
• #32370

Powered by Codex Action

These all need to be merged by a maintainer.

same issue. I am unable to use 5.3spark in Mac GUI app. CLI works fine. I have delved around and played with my config/s and nothing fixes

"Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model."

Adding another reproduction from the Codex Desktop multi-agent/subagent path.

Environment:

• codex-cli 0.144.0-alpha.4
• macOS / Codex Desktop thread
• User config has no explicit model_reasoning_summary; relevant config is only:
• model = "gpt-5.6-sol"
• model_reasoning_effort = "low"

What I tested:

1. Forced a Codex subagent with model: "gpt-5.3-codex-spark", agent_type: "explorer", and reasoning_effort: "low".
2. The agent spawned, but wait_agent failed immediately with:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "code": "unsupported_parameter",
    "message": "Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.",
    "param": "reasoning.summary"
  },
  "status": 400
}

1. Retried without reasoning_effort; same reasoning.summary error.
2. Control test with gpt-5.4-mini subagent completed successfully, so the subagent mechanism itself was working.
3. Direct local CLI path with codex exec --ephemeral --ignore-rules --skip-git-repo-check --sandbox read-only --model gpt-5.3-codex-spark ... succeeded and returned the expected response. In that direct path, the CLI displayed reasoning summaries: none.
4. codex debug models for Spark locally reports:

{
  "slug": "gpt-5.3-codex-spark",
  "default_reasoning_level": "high",
  "supported_reasoning_levels": ["low", "medium", "high", "xhigh"],
  "supported_in_api": false,
  "supports_reasoning_summaries": true,
  "default_reasoning_summary": "none"
}

Also notable: in the current multi-agent tool schema exposed to the parent session, Spark is no longer listed as an available model override, but a forced model: "gpt-5.3-codex-spark" override is still accepted by spawn_agent and then fails at request time with the unsupported reasoning.summary parameter.

This supports the existing hypothesis that the app/subagent/shared-session request path is resolving or carrying a reasoning summary field that is not valid for Spark, while at least some direct CLI paths can avoid sending it.

Adding another macOS Codex App reproduction.

Environment

• macOS
• Codex App with bundled Codex core 0.144.2
• Model: gpt-5.3-codex-spark
• Reproduces in a newly created task/session, not only in an older failed thread

Observed behavior

Every Spark request fails before generation with:

Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

Other models work in the same installation, so this does not appear to be an account-access or general connectivity problem.

Diagnostic evidence and workaround

The App-side model capability/catalog data treats Spark as supporting reasoning summaries. As a result, the request includes reasoning.summary, while the service endpoint rejects that field for Spark.

Adding the following as a top-level entry in ~/.codex/config.toml, fully quitting/reopening the App, and creating a fresh task makes Spark return normally (verified with an APP_OK test):

model_supports_reasoning_summaries = false

high / xhigh reasoning effort still works with this workaround; only reasoning-summary display is disabled. In the current App, the override appears to need to be global rather than scoped reliably to Spark.

Likely fix

Before request serialization, validate reasoning fields against the final selected model and omit reasoning.summary entirely for gpt-5.3-codex-spark. The Spark capability declaration/model catalog should also be corrected so the App does not reintroduce the unsupported field.

Same failure on Windows (not only macOS).

Environment

• Codex desktop app: OpenAI.Codex package 26.707.8479.0 (runs as ChatGPT.exe), Windows 11
• Bundled core: codex-cli 0.144.2; standalone CLI on PATH: 0.144.1
• Plan hit its usage limit, so the app auto-fell back to gpt-5.3-codex-spark

Symptom — every turn on gpt-5.3-codex-spark fails with:
Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.

The model_supports_reasoning_summaries = false workaround does NOT help in the Windows app. Added it to the top of ~/.codex/config.toml, fully quit + reopened the app, started a fresh task — the key stays in the file but the app still sends reasoning.summary. There is also no in-app toggle to disable reasoning summaries: the model/effort menu only exposes model + reasoning level, the View menu only has layout toggles, and a Settings search for summary/reasoning/thinking surfaces nothing summary-related.

CLI contrast (same machine/account):

# fails (summary forced on):
codex exec -c model=gpt-5.3-codex-spark -c model_reasoning_summary=detailed "hi"
# succeeds (field omitted):
codex exec -c model=gpt-5.3-codex-spark -c model_reasoning_summary=auto "hi"
codex exec -c model=gpt-5.3-codex-spark -c model_reasoning_summary=none  "hi"

So codex exec correctly omits reasoning.summary for this model, but the desktop app requests it unconditionally and ignores both the model capability and the config override. This is especially painful because it only triggers when falling back to Spark after the main quota is exhausted — exactly when there's no other model to switch to.