Compaction fails with: Error running remote compact task: Request blocked.

[AcrylicShrimp]

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.707.31428

What subscription do you have?

Pro plan

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Compaction fails with:

Error running remote compact task: Request blocked.

What steps can reproduce the bug?

Feedback ID: 019f472f-e43f-77d1-aac2-1f3ce1cfc7d7

What is the expected behavior?

No response

Additional information

[github-actions]

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

• Remote compact consistently fails for a long GPT-5.5 session and blocks thread continuation #30596

Powered by Codex Action

[Nordskog]

Experienced the same problem ( in CLI ) today with all the 5.6 models, only affecting one of my sessions.
Temporarily switching to a 5.5 model and compacting there worked fine.
Compacting with a 5.6 model afterwards also worked fine.
Edit: Until normal prompts also started getting blocked. Guess it's back to 5.5.

[cuipengfei]

Cross-referencing related gpt-5.6 Request blocked reports that may share the same root cause.

• Codex App: text-log attachment can trigger “Request blocked” and poison subsequent turns #32177 — compacted thread state / text-log attachment triggers Request blocked on gpt-5.6-sol.
• Responses Lite: changing reasoning effort can strand a large session on “Request blocked.” until compaction #32533 — changing reasoning effort → full-history request → Request blocked on gpt-5.6.
• oh-my-pi#5184 — performed marker minimization: the specific input text <|channel|>analysis is sufficient to trigger Request blocked / invalid_request_body on all gpt-5.6-* models (luna/sol/terra). Earlier models (gpt-5-mini through gpt-5.5) all accept the same payload.

Since your issue is compaction itself being blocked: compaction serializes conversation history into a text summary and sends it as input to the model. If that serialized text contains Harmony control-token markers (e.g. <|channel|>analysis from assistant reasoning/thinking), the gpt-5.6 input-side safety classifier may reject it — which would explain why compaction fails on 5.6 but works on 5.5.

Your workaround (switch to 5.5, compact, switch back) is consistent with this: the 5.5 compaction produces a summary that strips or doesn't trigger the control-token pattern, after which 5.6 can process the cleaned summary. However, your edit noting that "normal prompts also started getting blocked" suggests the trigger text may persist in the thread's context even after compaction, eventually poisoning subsequent requests as well.

Related: #32177, #32533.

[ASIeinicher]

i see this error still having 21% context left

no change on restart with codex resume