IMPORTANT drain usage on /review task [x2 times]

Resolved 💬 12 comments Opened Mar 14, 2026 by YoloMark98 Closed Mar 14, 2026
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

codex-cli 0.114.0

What subscription do you have?

Pro

Which model were you using?

gpt-5.4 high

What platform is your computer?

Linux 6.8.0-101-generic x86_64 x86_64

What terminal emulator and version are you using (if applicable)?

Tabby

What issue are you seeing?

I tried using /review (again), a new feature offered by OpenAI, and whilst using it, my weekly limits dropped to 0%, whilst my code review usage is still at 100%... I can’t explain how this is possible, as I haven’t used the following:

  • I don’t use fast mode
  • I don’t use the context window expansion to 1M tokens (I’m using the default 256k)
  • I haven’t used multi-agents
  • I haven’t used the xhigh version, only medium and high in the last few days
  • I only use this in config.toml: '[mcp_servers.context7]

args = [‘-y’, ‘@upstash/context7-mcp’, ‘--api-key’, ‘ctx7sk-a5851794-f6d1-4435-b17f-fa9306021567’]
command = ‘npx’
enabled = true

[mcp_servers.playwright]
command = ‘npx’
args = [‘-y’, ‘@playwright/mcp@latest’, ‘--headless’, ‘--no-sandbox’, ‘--isolated’]
enabled = false # disables the server

[notice]
hide_gpt5_1_migration_prompt = true
‘hide_gpt-5.1-codex-max_migration_prompt’ = true
hide_rate_limit_model_nudge = true

[notice.model_migrations]
‘gpt-5.1-codex-max’ = ‘gpt-5.2-codex’

approval_policy = ‘on-request’
sandbox_mode = ‘workspace-write’
‘gpt-5.1-codex-mini’ = ‘gpt-5.3-codex’

[sandbox_workspace_write]
network_access = true

web_search = ‘live’

[features]
unified_exec = true
shell_snapshot = true
steer = true
personality = true
prevent_idle_sleep = true

[[skills.config]]
path = ‘~/.codex/skills’
enabled = true'

<img width="1229" height="410" alt="Image" src="https://github.com/user-attachments/assets/bc894340-40cc-4064-bbc8-7d7eed865bf6" />

What steps can reproduce the bug?

Uploaded thread: 019cebca-6ed2-78d3-aece-11548c592648

You can use /review – I reported this issue a few days ago; I thought it had been resolved, but it hasn’t been yet

What is the expected behavior?

I would simply like to see only the appropriate charges applied, as that is what I expect from a $200 subscription.

Additional information

Here is my user ID so you can check more easily: user-grEQqq7dokHb56wxlYbBeUWO

Tell me if im doing something wrong.

View original on GitHub ↗

11 Comments

github-actions[bot] contributor · 4 months ago

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

  • #14322
  • #14628
  • #14023
  • #13186

Powered by Codex Action

DeanStr · 4 months ago

Codex review usage is for in GitHub with "@codex review", using /review in CLI comes out of normal usage. It would be nice if we could use /review in CLI and it came out of this usage, doing it in GitHub does not suit my workflow.

Sidenote, 5.4 does use a lot of usage for /review in my experience but it is quite thorough. Better to maybe use a different model for /review if you don't need it to be so thorough.

etraut-openai contributor · 4 months ago

As @DeanStr said, code review usage is for the cloud-based @codex review. Local /review comes out of your normal local usage budget. So the behavior you're seeing here is expected, not a bug.

miraclebakelaser contributor · 4 months ago

@YoloMark98 not sure if you're aware you exposed your context7 api key, you should rotate it.

YoloMark98 · 4 months ago
@YoloMark98 not sure if you're aware you exposed your context7 api key, you should rotate it.

Yes, thanks.

YoloMark98 · 4 months ago
As @DeanStr said, code review usage is for the cloud-based @codex review. Local /review comes out of your normal local usage budget. So the behavior you're seeing here is expected, not a bug.

Yes, but I still can’t figure out how I managed to use up almost the entire context window of my Pro subscription in a single day, isn’t that strange? Especially as I don’t use fast mode and haven’t made any changes to the context window tokens.

etraut-openai contributor · 4 months ago

@YoloMark98, it depends on the nature of the source base and the code review. Some code reviews require significant work on the part of the model, so token consumption can be very high.

YoloMark98 · 4 months ago
@YoloMark98, it depends on the nature of the source base and the code review. Some code reviews require significant work on the part of the model, so token consumption can be very high.

I’d been reviewing some uncommitted files that were due to be ready for committing. Now I’ll have to wait five days before I can get back to work; I’ll just have to make do somehow or buy another subscription, I don’t know, building open-source project is getting expensive XD

etraut-openai contributor · 4 months ago

@YoloMark98, if you're an OSS maintainer, you may be interested in our Open Source Program, which provides six months of Codex usage for qualifying maintainers / projects.

YoloMark98 · 4 months ago
@YoloMark98, if you're an OSS maintainer, you may be interested in our Open Source Program, which provides six months of Codex usage for qualifying maintainers / projects.

Unfortunately, my open-source project is still in development, so it hasn’t been officially released yet, is the application still considerated?

raye-deng · 4 months ago

The /review token drain is a real issue, and it highlights a broader problem: AI code review tools need to be smart about what they analyze. Reviewing an entire large PR from scratch every time burns tokens and produces diminishing returns.

One approach that works well: deterministic pre-filtering. Before sending code to an LLM for review, run fast rule-based checks that catch known AI-code patterns — hallucinated imports (packages that do not exist on the registry), deprecated API calls masked by version pinning, cross-file inconsistencies from multi-session generation. These catch maybe 30-40% of AI-code defects with zero token cost.

We took this approach with Open Code Review — deterministic checks handle the known patterns, and optional AI analysis handles the nuanced stuff. For CI pipelines this makes the economics much more viable.

If anyone wants to try: npx @opencodereview/cli scan . --sla L1

Would be great if Codex /review could do a lightweight deterministic pass first and only invoke the model on the diff segments that actually need AI judgment.

Showing cached comments. Read the full discussion on GitHub ↗