›_ Codex Issues
Codex Issues / #10858

incorrect context window % left in /status in oss mode

Open 💬 2 comments Opened Feb 6, 2026 by dkhokhlov
💡 Likely answer: A maintainer (dkhokhlov, contributor) responded on this thread — see the highlighted reply below.

What version of Codex CLI is running?

codex-cli 0.98.0

What subscription do you have?

Team

Which model were you using?

gpt-oss:20b

What platform is your computer?

Ubuntu 22

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

_No response_

What issue are you seeing?

Incorrect ctx window "XX% left" in /status: "Context window: 89% left (14.9K used / 38K) ".

╭─────────────────────────────────────────────────────────────────────────────────╮
│  >_ OpenAI Codex (v0.98.0)                                                      │
│                                                                                 │
│ Visit https://chatgpt.com/codex/settings/usage for up-to-date                   │
│ information on rate limits and credits                                          │
│                                                                                 │
│  Model:                gpt-oss:20b (reasoning high, summaries off)              │
│  Model provider:       gpt-oss - http://x.x.x.x:11434/v1                        │
│  Directory:            /lib                                                     │
│  Approval:             never                                                    │
│  Sandbox:              danger-full-access                                       │
│  Agents.md:            ../AGENTS.md                                             │
│  Account:             xxxxxxx@xxxxxx (Team)                                     │
│  Collaboration mode:   Default                                                  │
│  Session:              019c3193-7aaf-70d0-a0b7-f13340366d63                     │
│                                                                                 │
│  Context window:       89% left (14.9K used / 38K)                              │
╰─────────────────────────────────────────────────────────────────────────────────╯

What steps can reproduce the bug?

  • ollama
  • version is 0.15.4
  • on macOS, M1 32GB
  • export OLLAMA_CONTEXT_LENGTH=40000
> ollama ps
NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL
gpt-oss:20b    17052f91a42e    14 GB    100% GPU     40000      Forever
  • codex --oss
  • on Ubuntu
  • codex config:
model = "openai/gpt-oss-20b"
model_temperature = 0.0
model_context_window = 40000
model_reasoning_effort = "high"
model_verbosity = "medium"
model_reasoning_summary = "none"
model_supports_reasoning_summaries = false
model_provider = "ollama"
oss_provider = "ollama"
  • try to use context
  • /status

What is the expected behavior?

Correct percentage of context used.

Additional information

I suspect this also affects auto-compact.

View original on GitHub ↗

2 Comments

dkhokhlov contributor · 6 months ago

Findings:

The status card's "XX% left" value comes from percent_of_context_window_remaining, which subtracts a 12,000-token baseline before calculating the percentage, creating an apparent mismatch with raw numbers

The method computes an "effective" window by subtracting BASELINE_TOKENS (12,000) from both the total window and used tokens:

  • effective_window = context_window - 12,000
  • used = (tokens_in_context_window - 12,000).max(0)
  • percent = (effective_window - used) / effective_window * 100

Why 89% left appears with 40K window and ~14.9K used

With a 40,000-token window and ~14,900 tokens used:

  • Effective window: 40,000 - 12,000 = 28,000
  • Effective used: 14,900 - 12,000 = 2,900
  • Remaining: 28,000 - 2,900 = 25,100
  • Percentage: 25,100 / 28,000 ≈ 89.6% → rounded to 89%

Looks like it needs to be fixed in UI and auto-compact.

swma24 · 4 months ago

I can reproduce this in current Codex CLI and confirmed the root cause is a mixed metric in /status.

Example observed output:

  • Context window: 99% left (14.5K used / 258K)

Why this happens:

  • % left is computed on the effective context window (window - 12,000 baseline)
  • but used / window in the same line was shown with raw values

So the line combines two different denominators, which makes it look incorrect.

I prepared a fix that aligns both values to the same effective-window basis in the status card:

  • used -> effective used (total_tokens - 12,000, floored at 0)
  • window -> effective window (context_window - 12,000)
  • % left still uses the same baseline logic, now consistent with the displayed absolute values

Patch branch (fork):

Validation run locally:

  • just fmt
  • cargo test -p codex-protocol
  • cargo test -p codex-tui

If useful, I can open an invited upstream PR from this branch.