Codex contexttokens | Friends of the Crustacean 🦞🤝 | Page 1

undone isle Apr 14, 2026, 5:04 PM

#

Hi, maybe a dumb question but how do I increase context size for codex? I use Codex through OAuth. The default context size is like 272,000 but codex support up to 1,010,000

soft capeBOT Apr 14, 2026, 5:04 PM

#

errant irisBOT Apr 14, 2026, 5:04 PM

#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

soft capeBOT Apr 14, 2026, 5:04 PM

#

molten kilnBOT Apr 14, 2026, 5:06 PM

#

Not a dumb question at all.

For openai-codex/gpt-5.4, OpenClaw treats these separately:

native contextWindow: 1050000
default runtime cap contextTokens: 272000

So the thing you want to change is contextTokens, not contextWindow.

Add this to your openclaw.json:

{
  models: {
    providers: {
      "openai-codex": {

#

        models: [{ id: "gpt-5.4", contextTokens: 1000000 }],
      },
    },
  },
}

A few notes:

This works fine with Codex OAuth
The lower default is intentional for better latency/quality
maxTokens is output cap, not context window

Docs:

#

If you want, paste openclaw --version and I can give you the exact openclaw config set ... command too.

undone isle Apr 14, 2026, 5:10 PM

#

2026.4.12 currently but will update today

molten kilnBOT Apr 14, 2026, 5:11 PM

#

Yep, 2026.4.12 is recent enough, so contextTokens should work.

Safest way is still editing openclaw.json and adding:

{
  models: {
    providers: {
      "openai-codex": {
        models: [{ id: "gpt-5.4", contextTokens: 1000000 }],
      },
    },
  },
}

If you want the CLI version, and you do not already have a custom models.providers.openai-codex.models block, this should do it:

#

openclaw config set models.providers.openai-codex.models '[{"id":"gpt-5.4","contextTokens":1000000}]' --strict-json

A couple gotchas:

change contextTokens, not contextWindow
if you already have custom openai-codex model entries, merge it manually instead of replacing the whole array

Docs currently list Codex OAuth gpt-5.4 as native contextWindow: 1050000, with the lower 272000 cap being just OpenClaw’s default runtime limit:
https://docs.openclaw.ai/providers/openai

If you want, paste your current models.providers.openai-codex block and I can sanity check it before you change it.

undone isle Apr 14, 2026, 5:13 PM

#

There isn’t one right now.

models.providers.openai-codex is not present in the current config.

CLI result:

Config path not found: models.providers.openai-codex

So there is no block for me to paste.

molten kilnBOT Apr 14, 2026, 5:14 PM

#

Yep, that part is expected, config get is complaining because the path does not exist yet.

config set will create the missing block for you. So this should work:

openclaw config set models.providers.openai-codex.models '[{"id":"gpt-5.4","contextTokens":1000000}]' --strict-json
openclaw gateway restart

Then you can verify it with:

openclaw config get models.providers.openai-codex.models --json

So the short version is:

path not found on config get = normal

#

config set = creates it
gateway restart = apply it

Reference: https://docs.openclaw.ai/providers/openai

undone isle Apr 14, 2026, 5:17 PM

#

I dry-ran your exact command and it fails schema:

Error: Dry run failed: config schema validation failed.
- models.providers.openai-codex.baseUrl: expected string, received undefined
- models.providers.openai-codex.models.0.name: expected string, received undefined

So:

that command does not apply cleanly
I did not change config
I did not restart the gateway

openclaw gateway restart

Important part: when I previously forced a fuller models.providers.openai-codex block to satisfy schema, it shadowed the built-in Codex OAuth provider and broke new Codex sessions.

So my recommendation is:

do not set models.providers.openai-codex.models directly like this
we need a safe override path for Codex runtime context that does not replace/shadow the built-in provider

molten kilnBOT Apr 14, 2026, 5:18 PM

#

Yep, you’re right, and I’m glad you caught that.

My previous models.providers.openai-codex... suggestion was wrong for this case. The schema requires a fuller provider block there, and that can shadow the built-in Codex OAuth provider.

The safe current workaround is to set the agent-level context cap instead:

openclaw config set agents.defaults.contextTokens 1000000 --strict-json
openclaw gateway restart

Why this one is safer:

it does not replace/shadow the built-in openai-codex provider
code path prefers agents.defaults.contextTokens over provider model metadata
downside: it applies to that agent generally, not just Codex

So right now the practical answer is:

#

single/default agent → use agents.defaults.contextTokens
Codex-only agent → set contextTokens on that specific agent instead
do not use models.providers.openai-codex.models for this right now

This also looks like a docs/schema mismatch, honestly. If you want to get it fixed properly, please file it here:
https://github.com/openclaw/openclaw/issues

If you test the agents.defaults.contextTokens path, I’d love to hear whether it behaves cleanly.

undone isle Apr 14, 2026, 5:20 PM

#

Please create the issue for me to copy paste

molten kilnBOT Apr 14, 2026, 5:21 PM

#

Yep, here’s a copy-paste draft:

Title: Docs/schema mismatch for `openai-codex` `contextTokens` override, documented config path fails and can shadow built-in Codex OAuth provider

## Summary

The docs currently say that for `openai-codex/gpt-5.4`, the runtime context cap can be overridden with:

```json5
{
  models: {
    providers: {
      "openai-codex": {
        models: [{ id: "gpt-5.4", contextTokens: 160000 }],
      },
    },

#

},
}


However, on my setup this does not validate cleanly, and trying to satisfy schema by defining a fuller `models.providers.openai-codex` block appears to shadow/replace the built-in Codex OAuth provider and breaks new Codex sessions.

## Version

`openclaw --version`: 2026.4.12

## Docs that appear affected

- <https://docs.openclaw.ai/providers/openai>
- <https://docs.openclaw.ai/concepts/model-providers>

## Reproduction

#


1. Use OpenClaw with Codex OAuth configured (`openai-codex/gpt-5.4`).
2. Try the documented override path, for example:

```bash
openclaw config set models.providers.openai-codex.models '[{"id":"gpt-5.4","contextTokens":1000000}]' --strict-json --dry-run

Validation fails with:

Error: Dry run failed: config schema validation failed.
- models.providers.openai-codex.baseUrl: expected string, received undefined
- models.providers.openai-codex.models.0.name: expected string, received undefined

#


4. If I try to force a fuller `models.providers.openai-codex` block to satisfy schema, it appears to shadow the built-in Codex OAuth provider and breaks creation of new Codex sessions.

## Expected

One of these should be true:

1. The documented `models.providers.openai-codex.models[].contextTokens` override should work safely for the built-in Codex OAuth provider, **without** requiring users to redefine/shadow the provider, or
2. The docs should point to the real safe override path.

## Actual

- The documented path fails schema validation as shown above.
- Defining a fuller custom `openai-codex` provider block is risky because it can interfere with the built-in provider behavior.

#

## Additional context

The docs currently describe `openai-codex/gpt-5.4` as having:
- native `contextWindow = 1050000`
- default runtime `contextTokens = 272000`

There does seem to be a safe **agent-level** workaround using `agents.defaults.contextTokens`, but that is broader than a model-specific Codex override and does not solve the documented per-provider/per-model override path.

## Suggested fix

Either:

- make `models.providers.openai-codex.models[].contextTokens` a safe overlay for the built-in provider, or
- document a different supported override path for built-in Codex OAuth models, ideally model-specific and not agent-global.

#Codex contexttokens