0.153.4 is unusable

[Regression][0.153.4] Sol, Terra, and Astra repeatedly lose task context and produce extremely slow, low-quality work

What version of Codex CLI is running?

codex-cli 0.153.4

What subscription do you have?

[Enter subscription]

Which models were you using?

• gpt-5.6-sol
• gpt-5.6-terra
• gpt-6-astra

The problem persists across multiple reasoning-effort settings.

What platform is your computer?

Linux / WSL

What terminal emulator are you using?

[Enter terminal and version]

What issue are you seeing?

Codex CLI became extremely slow and materially less capable after updating to version 0.153.4.

This is not isolated to one model. I tested Sol, Terra, and Astra with multiple reasoning-effort settings. All exhibit similar behaviour:

• Codex repeatedly reads the same files, sometimes multiple times within the same task context.
• It does not reliably preserve requirements, findings, decisions, or completed work between model/tool cycles.
• It repeatedly reconstructs repository context instead of continuing from the established execution frontier.
• It ignores or loses explicit requirements that were already established in the thread.
• It spends a long time exploring without converting the findings into implementation.
• It produces locally plausible actions that do not compose into the requested system.
• It claims progress without implementing the required behaviour completely.
• Corrections often cause another exploration cycle rather than a focused correction.
• Execution is dramatically slower than before the update.
• The final output is incomplete and of significantly lower quality.

The observable behaviour feels like each model/tool cycle receives a weakened or partially reconstructed version of the task context. The model repeatedly re-derives information it had already established and fails to maintain continuity across the implementation.

This makes Codex effectively unusable for nontrivial engineering work. The problem is not merely response style or increased verbosity. It affects reasoning continuity, implementation correctness, execution time, and the resulting software.

What steps can reproduce the bug?

1. Open a nontrivial repository in Codex CLI 0.153.4.
2. Select GPT-5.6 Sol, GPT-5.6 Terra, or GPT-6 Astra.
3. Assign a task requiring several connected steps: inspect existing architecture, preserve stated constraints, implement a change, and verify it.
4. Allow Codex to inspect the relevant files.
5. Continue the same task or provide a correction after its first implementation attempt.
6. Observe that Codex reads the same files again, loses previously established requirements, repeats analysis, and fails to continue from

Reproduce the regression in Codex CLI 0.153.4 on Linux/WSL using a nontrivial repository and one of the listed models, then compare behavior with the previous version. Investigate task-context continuity across model/tool cycles; done means requirements and prior findings persist, repeated file reads and exploration are reduced, and the connected task completes correctly.