Context compaction fails with unknown parameter 'prompt_cache_retention'

Getting this too in VS Code extension, thread id: 019d8b82-66f8-7d22-a34d-40e6f025fe4e

+1 on CLI

Getting this too in CLI

getting this codex-cli 0.120.0

Same issue

oh my god not now codex

+1 on CLI across many accounts and machines

Probably just api is down

+1 on cli, win11, 0.119.0

new model incoming

same issue, cannot compact context.

+1 cli, MacOs, v0.120.0

+1 on cli, macosx, v0.120.0

+1 on cli, v0.120.0. 019d4b5b-01b3-7533-a3c0-b4aaed5d56c8

+1 on cli, v0.120.0, 019d8d21-0988-73f1-a1aa-52f83b52d35d

+1

+1

+1

+1

good night guy , let work tmt

+1. error: Error running remote compact task: { "error": { "message": "Unknown parameter: 'prompt_cache_retention'.", "type": "invalid_request_error", "param": "prompt_cache_retention", "code": "unknown_parameter" } }

+1

Adding one more concrete reproduction from Codex Desktop on macOS.

• Date observed: 2026-04-14
• Thread id: 019d89b3-b6cf-7fd3-8b9f-292f471fc457
• Local session metadata shows: originator=Codex Desktop, cli_version=0.119.0-alpha.28, model_provider=openai
• Local config at repro time: model = "gpt-5.4", model_reasoning_effort = "xhigh"

What I found locally:

• Right before failure, token count event showed last input_tokens=236952 with model_context_window=258400, so compaction appears to be triggering near the context limit as expected.
• Immediately after that, the session logs show the inline error:

Error running remote compact task: {
  "error": {
    "message": "Unknown parameter: 'prompt_cache_retention'.",
    "type": "invalid_request_error",
    "param": "prompt_cache_retention",
    "code": "unknown_parameter"
  }
}

Possible root cause:

• OpenAI docs for POST /v1/responses include prompt_cache_retention
• OpenAI docs for POST /v1/responses/compact do not appear to accept that field
• This looks like the compaction path is forwarding a create-response-only parameter into the compact endpoint

So this does not look thread-specific. It looks like a request-shape mismatch in the compaction path.

I found a fallback

Eu encontrei um fallback.

Summary

There is an upstream Codex bug where automatic context compaction can fail with:

Unknown parameter: 'prompt_cache_retention'

I added an opt-in workaround profile to ~/.codex/config.toml so Codex can try a different compaction path without changing your default setup.

This does not change your project code.
This does not redirect traffic to a third-party provider.
This only affects Codex when you explicitly use the workaround profile.

What changed

I added this block to ~/.codex/config.toml:

[profiles.inline_compact_workaround]
model_provider = "openai_inline"

[model_providers.openai_inline]
name = "OpenAI Inline"
requires_openai_auth = true
supports_websockets = true

Current location in the file:

• ~/.codex/config.toml:40
• ~/.codex/config.toml:43

What this means

Before this change:

• your normal Codex setup used the built-in openai provider

After this change:

• your normal Codex setup is still unchanged
• there is now an additional profile called inline_compact_workaround
• that profile uses a custom provider entry named openai_inline

Important:

• if you run Codex normally, nothing changes
• the workaround is only used if you start Codex with the dedicated profile

How to use it

To resume the affected thread with the workaround:

codex -p inline_compact_workaround resume 019d8c16-d49f-72a1-b7f9-bd7ad5410a5f

To start a new session with the workaround:

codex -p inline_compact_workaround

Why this fallback may help

Confirmed facts

• As of April 14, 2026, your local CLI version is codex-cli 0.120.0.
• As of April 11, 2026, 0.120.0 is the latest published Codex release.
• As of April 14, 2026, there is an open upstream Codex issue for the same failure:

Context compaction fails with unknown parameter 'prompt_cache_retention'

• The issue is public here:

https://github.com/openai/codex/issues/17809

Confirmed facts from the workaround config

• I did not set base_url
• I did not set env_key
• I did not set http_headers
• I did not set query_params
• I set requires_openai_auth = true

That means this provider still uses the normal OpenAI authentication flow instead of sending traffic somewhere else.

Technical inference

Based on the public Codex source:

• Codex uses a special remote compaction path when the provider is recognized as built-in OpenAI
• the custom provider name OpenAI Inline may avoid that exact path
• if that happens, Codex should fall back to the regular inline/local compaction flow instead of the failing remote compact path

This is why the workaround is reasonable.

Is it safe?

Safe in the network/authentication sense

Yes, this workaround is safe in the sense that:

• it still uses OpenAI authentication
• it does not point to a different vendor
• it does not add custom headers
• it does not inject a custom API URL

Not a guaranteed official fix

No, this is not an official upstream fix.

It is a fallback workaround, not a confirmed product patch.

What did not change

• your repo at /Users/higordedeusmatos/Work/S/Projects/mimir
• your source code
• your tests
• your shell environment
• your default Codex behavior when you do not use the workaround profile

How to revert it

If you want to undo the workaround, remove this block from ~/.codex/config.toml:

[profiles.inline_compact_workaround]
model_provider = "openai_inline"

[model_providers.openai_inline]
name = "OpenAI Inline"
requires_openai_auth = true
supports_websockets = true

After removal, Codex returns to the previous configuration.

Known limitations

Facts

• I validated that the config file still loads correctly after the change.
• I validated that the new profile is accepted by the CLI.

Unknowns

• I did not force a long enough live conversation to trigger compaction under this profile during validation.
• Because of that, the final runtime behavior is still a real-world test rather than a mathematically guaranteed outcome.

Recommended next step

Run:

codex -p inline_compact_workaround resume 019d8c16-d49f-72a1-b7f9-bd7ad5410a5f

If compaction succeeds, keep using this profile temporarily until upstream ships a proper fix.

If it still fails, remove the workaround and wait for an upstream patch in Codex.

Reference links

• Upstream bug report:

https://github.com/openai/codex/issues/17809

• Latest Codex release:

https://github.com/openai/codex/releases/tag/rust-v0.120.0

• Codex config reference:

https://developers.openai.com/codex/config-reference

• OpenAI prompt caching guide:

https://developers.openai.com/api/docs/guides/prompt-caching

+1 macOS Codex Desktop

Damm, now I have to start another tread and its going to take some time to gather context until its not dumb anymore

Seeing the same issue here on Windows 11 with codex-cli 0.120.0 using ChatGPT sign-in and gpt-5.4.

It started suddenly on April 14, 2026 after previously working reliably. Automatic compaction fails with:

Error running remote compact task: {
  "error": {
    "message": "Unknown parameter: 'prompt_cache_retention'.",
    "type": "invalid_request_error",
    "param": "prompt_cache_retention",
    "code": "unknown_parameter"
  }
}

/compact fails with the same error, so the affected thread becomes unusable once compaction kicks in.

Thread ID from my local logs: 019d8b82-60a7-77d1-b8dc-48030b8c7772.

+1 on macOS Codex Desktop

+1 on macOS codex desktop -

Error running remote compact task: { "error": { "message": "Unknown parameter: 'prompt_cache_retention'.", "type": "invalid_request_error", "param": "prompt_cache_retention", "code": "unknown_parameter" } }

+1 Error running remote compact task: { "error": { "message": "Unknown parameter: 'prompt_cache_retention'.", "type": "invalid_request_error", "param": "prompt_cache_retention", "code": "unknown_parameter" } }

+1 Error running remote compact task: stream disconnected before completion: error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)

Quit and Reopened the app on MacOS, and its working now, in the same thread.

Panic over, all working again

Working fine!

# Encontré una alternativa > Eu encontré um fallback. ## Resumen Existe un error en Codex en la versión principal donde la compactación automática del contexto puede fallar con: `` Unknown parameter: 'prompt_cache_retention' ` He añadido un **perfil de solución alternativa opcional** para ~/.codex/config.tomlque Codex pueda intentar una ruta de compactación diferente sin modificar la configuración predeterminada. Esto **no** modifica el código de tu proyecto. Esto **no** redirige el tráfico a un proveedor externo. Esto solo afecta a Codex cuando utilizas explícitamente el perfil de solución alternativa. ## ¿Qué cambió? Agregué este bloque a ~/.codex/config.toml: [profiles.inline_compact_workaround] model_provider = "openai_inline" [model_providers.openai_inline] name = "OpenAI Inline" requires_openai_auth = true supports_websockets = true Ubicación actual en el archivo: * ~/.codex/config.toml:40 * ~/.codex/config.toml:43 ## Lo que esto significa Antes de este cambio: * openaiSu configuración normal de Codex utilizaba el proveedor integrado. Después de este cambio: * Tu configuración normal del Códice sigue sin cambios. * Ahora hay un perfil adicional llamadoinline_compact_workaround * ese perfil utiliza una entrada de proveedor personalizada llamadaopenai_inline Importante: * Si ejecutas Codex normalmente, no cambia nada. * Esta solución alternativa solo se utiliza si inicias Codex con el perfil dedicado. ## Cómo usarlo Para reanudar el hilo afectado con la solución alternativa: codex -p inline_compact_workaround resume 019d8c16-d49f-72a1-b7f9-bd7ad5410a5f Para iniciar una nueva sesión con la solución alternativa: codex -p inline_compact_workaround ## Por qué esta alternativa puede ser útil ### Hechos confirmados * A partir **del 14 de abril de 2026** , su versión local de CLI es codex-cli 0.120.0. * A fecha de **11 de abril de 2026** , 0.120.0esta es la última versión publicada del Codex. * A fecha de **14 de abril de 2026** , existe un problema abierto en el Codex 2026 relacionado con el mismo fallo: Context compaction fails with unknown parameter 'prompt_cache_retention' * El asunto es público aquí: https://github.com/openai/codex/issues/17809 ### Hechos confirmados a partir de la configuración de la solución alternativa. * **No** lo configurébase_url * **No** lo configuréenv_key * **No** lo configuréhttp_headers * **No** lo configuréquery_params * Yo establecírequires_openai_auth = true Eso significa que este proveedor sigue utilizando el flujo de autenticación normal de OpenAI en lugar de enviar el tráfico a otro lugar. ### Inferencia técnica Basado en la fuente pública del Codex: * Codex utiliza una ruta **de compactación remota** especial cuando el proveedor se reconoce como integrado.OpenAI * El nombre del proveedor personalizado OpenAI Inlinepuede evitar esa ruta exacta. * Si eso sucede, Codex debería recurrir al flujo de compactación local/en línea habitual en lugar de la ruta de compactación remota que está fallando. Por eso, la solución alternativa es razonable. ## ¿Es seguro? ### Seguro en el sentido de red/autenticación Sí, esta solución alternativa es segura en el sentido de que: * Todavía utiliza la autenticación de OpenAI. * No apunta a un proveedor diferente. * no agrega encabezados personalizados * No inyecta una URL de API personalizada. ### No es una solución oficial garantizada. No, esta no es una solución oficial de la versión principal. Se trata de una **solución provisional** , no de un parche confirmado para el producto. ## Lo que no cambió * tu repositorio en/Users/higordedeusmatos/Work/S/Projects/mimir * tu código fuente * tus pruebas * tu entorno shell * Comportamiento predeterminado del Codex cuando no utiliza el perfil de solución alternativa ## Cómo revertirlo Si desea deshacer la solución alternativa, elimine este bloque de ~/.codex/config.toml`: [profiles.inline_compact_workaround] model_provider = "openai_inline" [model_providers.openai_inline] name = "OpenAI Inline" requires_openai_auth = true supports_websockets = true Tras su eliminación, Codex vuelve a la configuración anterior. ## Limitaciones conocidas ### Hechos Comprobé que el archivo de configuración se sigue cargando correctamente después del cambio. He comprobado que el nuevo perfil es aceptado por la interfaz de línea de comandos (CLI). ### Desconocidos No provoqué una conversación en directo lo suficientemente larga como para activar la compactación bajo este perfil durante la validación. Por ello, el comportamiento final en tiempo de ejecución sigue siendo una prueba en el mundo real, en lugar de un resultado garantizado matemáticamente. ## Paso siguiente recomendado Correr: codex -p inline_compact_workaround resume 019d8c16-d49f-72a1-b7f9-bd7ad5410a5f Si la compactación se realiza correctamente, siga utilizando este perfil temporalmente hasta que el desarrollador publique una solución definitiva. Si el problema persiste, elimine la solución temporal y espere a que se publique un parche en Codex. ## Enlaces de referencia Informe de error del proyecto original: La compactación del contexto falla con el parámetro desconocido 'prompt_cache_retention'  #17809 [](https://github.com/openai/codex/issues/17809) Última versión de Codex: https://github.com/openai/codex/releases/tag/rust-v0.120.0 Referencia de configuración de Codex: https://developers.openai.com/codex/config-reference Guía de almacenamiento en caché de indicaciones de OpenAI: https://developers.openai.com/api/docs/guides/prompt-caching

WORKING

We rolled back the change that caused this regression, so this should be addressed now. Apologies for the inconvenience.

Okay they said its fixed, by my tokens for this 5h sprint are all eaten now ahaha

We rolled back the change that caused this regression, so this should be addressed now. Apologies for the inconvenience.

I'm curious, what was the change?