codex
Facts
- Openness
- Open source
- Licence
- Apache-2.0
- Language
- Rust
- Pinned version
a956835(2026-10-01)- Topics checked
- 32 of 33
- Kind of agent
- Coding
What it does notably well
- Budgets
Goal budget exhaustion marks the goal budget_limited and injects a steering message once per goal telling the model to stop substantive work and leave a clear next step, rather…
One RolloutBudget counter is shared across a whole root thread tree and fed by server-reported budget units when present, so spawned agents cannot escape the limit.
- Code mode
The model catalog, not a user flag, selects the tool mode: tool_mode code_mode_only on 10 of 11 bundled models makes exec and wait the only model-visible tools, overriding the…
Fail-closed exclusive mode: with no host process the model gets an error and the code never runs in the harness process.
Yielding cells: exec and wait return after 10,000 ms with the program still alive, so a long program does not block the turn and the model can wait, poll or terminate.
- Compaction
With provider openai and API-key sign-in (bundled catalog, token_budget off), compaction is a Responses-API round trip that returns an encrypted Compaction item; beside it the…
See how codex does on each topic
The topic-by-topic verdicts, the code lines and a side-by-side comparison are in the app. Sign in with Google; new accounts are approved by hand.
Independent analysis, not affiliated with codex.