What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.707.72221
What subscription do you have?
Codex Pro Model GPT 5.6
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
During QA, agents repeatedly retrieve documents they already read completely, even when the content remains in context and hashes confirm the files have not changed. This causes substantial avoidable token use, latency, and context pressure, especially in controller/subagent workflows.
With 5.5 I never had to tell an agent not to re-read an unchanged file that was already in context. But by simply adding a rule in agents.md to not re-read documents that are already in context and are proven unchanged via hash or other means. I kept wondering why the token burn was so bad with sub-agents and even the controlling agents that it would do this. It is so good at following instructions in 5.6 that it completely loses it's common sense for not re-reading something it just read because an instruction says to read it. That should be obvious that it means read it if you have not. So it's a double-edged sword. Very good at following instructions, almost to a fault. So not trash talking it, but thought I'd bring it to your attention or for others that may have been running into this issue. It's a super simple fix and helped greatly adding that rule in agents.md
What steps can reproduce the bug?
Have an agent fully read several documents and record their hashes.
Ask it to run cross-document QA.
Observe whether it rereads the unchanged documents.
Repeat with an explicit instruction to reuse complete reads unless coverage is missing, content changed, or a conflict appears.
With that instruction, a seven-document cross-QA used only about 8% of available context and completed much faster.
What is the expected behavior?
A complete verified read should remain reusable while its source is unchanged. Agents should retrieve only missing, changed, conflicting, or post-compaction content.
Additional information
This suggests a context-reuse efficiency heuristic may not be activating reliably by default in GPT-5.6. The suspected cause is unconfirmed, but the explicit reuse instruction produced a large observable improvement.
What version of the Codex App are you using (From “About Codex” dialog)?
Version 26.707.72221
What subscription do you have?
Codex Pro Model GPT 5.6
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
During QA, agents repeatedly retrieve documents they already read completely, even when the content remains in context and hashes confirm the files have not changed. This causes substantial avoidable token use, latency, and context pressure, especially in controller/subagent workflows.
With 5.5 I never had to tell an agent not to re-read an unchanged file that was already in context. But by simply adding a rule in agents.md to not re-read documents that are already in context and are proven unchanged via hash or other means. I kept wondering why the token burn was so bad with sub-agents and even the controlling agents that it would do this. It is so good at following instructions in 5.6 that it completely loses it's common sense for not re-reading something it just read because an instruction says to read it. That should be obvious that it means read it if you have not. So it's a double-edged sword. Very good at following instructions, almost to a fault. So not trash talking it, but thought I'd bring it to your attention or for others that may have been running into this issue. It's a super simple fix and helped greatly adding that rule in agents.md
What steps can reproduce the bug?
Have an agent fully read several documents and record their hashes.
Ask it to run cross-document QA.
Observe whether it rereads the unchanged documents.
Repeat with an explicit instruction to reuse complete reads unless coverage is missing, content changed, or a conflict appears.
With that instruction, a seven-document cross-QA used only about 8% of available context and completed much faster.
What is the expected behavior?
A complete verified read should remain reusable while its source is unchanged. Agents should retrieve only missing, changed, conflicting, or post-compaction content.
Additional information
This suggests a context-reuse efficiency heuristic may not be activating reliably by default in GPT-5.6. The suspected cause is unconfirmed, but the explicit reuse instruction produced a large observable improvement.