Spark model rejects reasoning.summary with unsupported_parameter

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

• #12495

Powered by Codex Action

Found the same issue.

I have custom "fast review" agents based on gpt-5.3-codex-spark and since updating to v0.106.0, the subagents fail due to:

{
"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
}

Do you happen to have a model_reasoning_summary setting in the config?

Yes, I have this line in my config:
model_reasoning_summary = "detailed"

Do you happen to have a model_reasoning_summary setting in the config?

model_reasoning_summary_format="detailed"
model_reasoning_summary="detailed"

After the recent change the CLI started respecting the setting. Before that it always used auto.

After the recent change the CLI started respecting the setting. Before that it always used auto.

Is there a way to specify this config by model or is the recommended action to set it to auto?

Thanks for the quick replies

I recommend removing the config all together, then the most optimal model setting will be used.

I expect you won't get much more value out of detailed reasoning summaries.

I reproduced this again and there seems to be a narrower split in behavior.

I removed:

model_reasoning_summary = "detailed"

from ~/.codex/config.toml.

After that:

• explicit spawn_agent(..., model="gpt-5.3-codex-spark") worked in a simple read-only smoke test
• but role-based subagents still failed when their .codex/agents/*.toml had:

model = "gpt-5.3-codex-spark"
model_reasoning_effort = "high"

Affected role configs in my repo:

• .codex/agents/explorer.toml
• .codex/agents/service_owner_investigator.toml
• .codex/agents/integration_investigator.toml

Role-based spawn still fails 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
}

So removing model_reasoning_summary fixes the explicit model path for me, but not the custom-role/subagent path. It looks like role-based spawns may still be attaching reasoning.summary somewhere separately.

yes, removing it still seems to persist somewhere

got same behaviour here, gpt-5.5 is spawning a gpt-5.3-codex-spark with model_reasoning_summary=concise and as a result the agent fails to spawn.

got same behaviour here, gpt-5.5 is spawning a gpt-5.3-codex-spark with model_reasoning_summary=concise and as a result the agent fails to spawn.

How did you end up finishing your task? what you ended up doing?

I understand gpt-5.3-codex-spark is now removed thankfully (because its not optimized for the latest version) (that being said some people liked it)
(make sure your update your codex(if you are a desktop user) and , codex-cli )

if the issue keep on happening attempt to reset all of your .codex/config.toml related to the failed spawning issue
as Paval Kryimetchs @pakrym-oai suggested doing (or rm --i .codex/config.toml)

How did you end up finishing your task? what you ended up doing?

In my case the task itself was not blocked by the repo logic. The failure was in the spawned Spark path: the child request ended up carrying a reasoning summary setting, while gpt-5.3-codex-spark rejected reasoning.summary.

I finished by avoiding that failing Spark child configuration and continuing through a normal non-Spark / no-summary path for the remaining work. I did not need to delete the whole ~/.codex/config.toml; I would treat a full config reset as a last-resort diagnostic step, because it can remove unrelated settings.

The narrower workaround is to remove or disable the stale summary / spawned-agent override that causes Spark to receive reasoning summaries, then retry with an updated Codex CLI/Desktop.

So the actionable bit for this issue is still: Codex should not propagate reasoning.summary / model_reasoning_summary into Spark child requests unless the selected model actually advertises and accepts it. If Spark is removed or hidden in newer builds, stale config/profile entries should fail gracefully or be ignored, not break the whole task.

is there any workaround. I can't seem to fix this despite setting model_reasoning_summary = "none"

<img width="1676" height="362" alt="Image" src="https://github.com/user-attachments/assets/132b2a89-d7d6-4f26-90c5-ac2b6817507d" />
Still not fixed?

yep hitting this error for the first time after the newest update with the chatgpt codex thing, no such problem with the cli version

run into this error as well when switching from 5.6 to 5.3-spark

<img alt="Image" width="1676" height="362" src="https://private-user-images.githubusercontent.com/13834608/619794558-132b2a89-d7d6-4f26-90c5-ac2b6817507d.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.E2Flc9tThJwYynQUy8xrNAo1kOroGaikOMi4bx8k8e8"> Still not fixed?

I recomend you DO NOT USE 5.3 , attempt to move to Terra or Luna(luna is the cheapest standing at even cheaper than 5.3)
(I also think its cool , that its not called like the 5.4mini with the word "mini"... people gave it unjust - considering how much fun you can have with 5.4mini! even of analysing documents and pdfs! )

I wouldnt even put much weight if the prompts you give it be used in the next generation of models of training

I recomend you try the smaller 5;6 models, as I guess you , use 5.3 for cost and speed reason (5.6 will also solve things with less steps - so its very -very good :) )

<img width="859" height="346" alt="Image" src="https://github.com/user-attachments/assets/20f072c4-9444-4172-900b-d10ab2f093d4" />

Thanks for reading))))

are you a bot? 5.3-spark is not 5.3, 5.3 is not even available in codex.

are you a bot? 5.3-spark is not 5.3, 5.3 is not even available in codex.

thats a good thing if its not in codex anymore

<img width="1558" height="550" alt="Image" src="https://github.com/user-attachments/assets/d2d84d4e-0f22-4361-b008-9415406a80ee" />
this guy commented it
so he clearly tried to use 5.3

Codex Spark is still there, but requires a Pro subscription
<img width="410" height="450" alt="Image" src="https://github.com/user-attachments/assets/92fad96a-6248-4ec0-9df2-b3e408c434f1" />

Codex Spark is still there, but requires a Pro subscription <img alt="Image" width="410" height="450" src="https://private-user-images.githubusercontent.com/11856987/619963057-92fad96a-6248-4ec0-9df2-b3e408c434f1.png?jwt=eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.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.DCa2CBew0RoX2aj771e43MXRLfr7d5dkymWaJ1xieo0">

I recomend not to use spark, because codex has changed a lot since then - and things can be broken fast with gpt 5.3 -- there are COUTLESS behavior issues that got fix in 5.4 (5,5 I consider already perfect, but so many "model behaviors" got fixed)

as per what you might want it, the answer is "economics" and "get more token usage"

as per what prompts are the best,,, its up to openai to see 80% of the prompts and what users try to use it for

i personally think using SOL is pointless, but you should try it
as for usage Terra is great, and luna is very cheap .....

I recomend trying to get work done with Terra , and if you can, I recomend trying to get things done even with luna , I get with high-luna(the 6$;per;mill output token model)

I understand how annoying it is hitting your usage limit (as 5,5 was released and it was a heavier models [either if openai has a bigger model , or the models has an higher cpu usage? not sure what drove the cost higher--- but you clearly get more usage out of it other than 5.4 which if you choose you get more usage] )

if your issues is not ratelimits,, but you like things get faster
know for a fact 5.6luna (($1.00 input / $6.00 output, $0.10 cached input reads)) is literally better than 5.3 ... also dont forget my point bro, 5.3 had a lot of behavioral issues

so please dont use 5.3.... openai has released luna and terra,,,, for the reason so people dont uses older models -- but I think as things go now --- models are better [[adjusted for more and more use cases ]] which is the same intellegance you expected from a human to be good at few tasks at once, and use concepts for other ideas --- the fact that its not neurons doing that, buy logic via words and logic is very coool (despite some how still relaying on the concept of predicting the next token ,, and being limited by tokens-per-second and its not diffused the same way you diffusion and iteration for the generative image models --- and how everything is being viewed from a realestate betting on the idea the moorse law is dead ,,, and now chips become cars that you use for 6 years ---- not that xeon cpus cant be used for 20 years but okay)

<img width="1353" height="336" alt="Image" src="https://github.com/user-attachments/assets/7e710f5d-9257-4df2-a261-524f4d33a184" />

for the 1.2million context window , do you rather serve a single user; or 5 users ?

I am having the same issue. no such setting in my config. no such issue in cli. this is annoying as we literally cannot use the service I pay £200 for...

I am having the same issue. no such setting in my config. no such issue in cli. this is annoying as we literally cannot use the service I pay £200 for...

I recomend , please try giving a shot the , gpt Terra model (i understand why you wouldnt want SOL, because 5.3 give you so called higher usage) I think Terra is right for you

> I am having the same issue. no such setting in my config. no such issue in cli. this is annoying as we literally cannot use the service I pay £200 for... I recomend , please try giving a shot the , gpt Terra model (i understand why you wouldnt want SOL, because 5.3 give you so called higher usage) I think Terra is right for you

this doesnt help at all. I want to use spark

> > I am having the same issue. no such setting in my config. no such issue in cli. this is annoying as we literally cannot use the service I pay £200 for... > > > I recomend , please try giving a shot the , gpt Terra model (i understand why you wouldnt want SOL, because 5.3 give you so called higher usage) I think Terra is right for you this doesnt help at all. I want to use spark

I recomend not get too attached to any models, I recomend trying to use Terra or Luna (I am having a lot of fun with Luna)

see it this way, the newer models seen a lot more use cases where older models fail and repeat themselves

the newer models do a better job at even understanding intent and executing what you want- so you wont be disapointed!
<img width="885" height="327" alt="Image" src="https://github.com/user-attachments/assets/5a19a5dc-7412-4313-82fa-5116636067a2" />

@2yxh , see message above