Compaction fails with: Error running remote compact task: Request blocked. #32028

Activity

1. AcrylicShrimp

   added

   appIssues related to the Codex desktop appIssues related to the Codex desktop app

   on Jul 10, 2026
2. github-actions

   added

   bugSomething isn't workingSomething isn't working

   contextIssues related to context management (including compaction)Issues related to context management (including compaction)

   on Jul 10, 2026
3. github-actions commented on Jul 10, 2026

   

   github-actionsBot

   on Jul 10, 2026 – with GitHub Actions

   Contributor

   More 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
4. TheRoughy commented on Jul 10, 2026

   

   TheRoughy

   on Jul 10, 2026

   Last edited by TheRoughy

   More actions

   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.

   👍React with 👍3Reacted by Grallen, Wei-Chih and Junyu Wang
5. cuipengfei commented on Jul 12, 2026

   

   cuipengfei

   on Jul 12, 2026

   Last edited by cuipengfei

   More actions

   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.
6. Yeachan-Heo

   added 2 commits that reference this issue on Jul 14, 2026

   fix(ai): neutralize leaked control tokens at the Responses request bo…

   ...

   632b8a4

   fix(ai): neutralize leaked control tokens at the Responses request bo…

   ...

   5ca557e
7. ASIeinicher commented on Jul 16, 2026

   

   ASIeinicher

   on Jul 16, 2026

   Last edited by ASIeinicher

   More actions

   i see this error still having 21% context left

   no change on restart with codex resume
8. jsgreenawalt commented on Jul 19, 2026

   

   jsgreenawalt

   on Jul 19, 2026

   More actions

   I also received this error when working with codex on modifications to a llama.cpp jinja template for an open weights model, that channel marker was definitely in my context, as we'd been discussing the template logic. I asked for a change to the template at some point later and then the session died with every request blocked. I tried to compact, the compact operation was also blocked.
9. giladgd commented on Aug 11, 2026

   

   giladgd

   on Aug 11, 2026

   More actions

   I also received this error when working with codex on modifications to a llama.cpp jinja template for an open weights model, that channel marker was definitely in my context, as we'd been discussing the template logic. I asked for a change to the template at some point later and then the session died with every request blocked. I tried to compact, the compact operation was also blocked.

   I second this. My requests consistently get blocked when I'm using codex to help me with a jinja template
10. jason931225

    added 2 commits that reference this issue on Sep 18, 2026

    fix(ai): neutralize leaked control tokens at the Responses request bo…

    ...

    9c23f83

    fix(ai): neutralize leaked control tokens at the Responses request bo…

    ...

    e9f7e01