GPT 5.2 thinking blocks missing #11273

Activity

1. NamedIdentity

   added

   bugSomething isn't workingSomething isn't working

   on Jan 30, 2026
2. github-actions

   assigned

   rekram1-node

   on Jan 30, 2026
3. github-actions

   added

   windows

   on Jan 30, 2026
4. rekram1-node commented on Jan 30, 2026

   

   rekram1-node

   on Jan 30, 2026

   Contributor

   More actions

   Try without plugins?
5. NamedIdentity commented on Jan 30, 2026

   

   NamedIdentity

   on Jan 30, 2026

   Author

   More actions

   Try without plugins?

   uninstalling OMO and DCP had no effect.

   Does GPT 5.2 have thinking blocks for you? Do you use OpenAI oauth or another provider?
6. THEN00P commented on Jan 30, 2026

   

   THEN00P

   on Jan 30, 2026

   More actions

   I have the same issue on Mac. Yesterday i was on 1.1.40 and the blocks were still there.
7. rekram1-node commented on Jan 30, 2026

   

   rekram1-node

   on Jan 30, 2026

   Contributor

   More actions

   

   I cannot replicate :/.
8. vosscodes commented on Jan 30, 2026

   

   vosscodes

   on Jan 30, 2026

   More actions

   1.1.36 and 1.1.44 on linux had this issue over the last 2 days pretty much constantly, no plugins installed

   I got a blip of thinking this morning around 10-12 hours ago, then it disappeared again.

   they have started working again at some point during this afternoon - this appears to be an intermittent provider issue

   @rekram1-node your screenshot is using the opencode zen provider, not the openai provider, which may be different.
9. rekram1-node commented on Jan 30, 2026

   

   rekram1-node

   on Jan 30, 2026

   Contributor

   More actions

   @vosscodes under the hood it should all function the same, here is screenshot w/ openai I can't replicate

   

   👍React with 👍1Reacted by vosscodes
10. rekram1-node commented on Jan 30, 2026

    

    rekram1-node

    on Jan 30, 2026

    Contributor

    More actions

    Note that the thinking blocks are just summary blocks so if it is thinking for a long time we don't necesarily know what its thinking unless it emits the thinking block summaries, I've seen it hang for quite some time before due to long thinking periods
11. vosscodes commented on Jan 30, 2026

    

    vosscodes

    on Jan 30, 2026

    More actions

    here is screenshot w/ openai I can't replicate

    as mentioned, things are currently working fine after thinking blocks were gone for days - it does indeed hang for quite some time while it is thinking, but the blocks were never displayed. I'd wager something was actually preventing them from emitting to us, hence them not being displayed + coming back later.

    here is a screenshot from this morning:

    
12. rekram1-node commented on Jan 30, 2026

    

    rekram1-node

    on Jan 30, 2026

    Contributor

    More actions

    hm ig lmk if it continues happening, I use gpt-5.2 and gpt-codex-5.2 every day, haven't seen this at all
13. NamedIdentity commented on Feb 17, 2026

    

    NamedIdentity

    on Feb 17, 2026

    Author

    More actions

    Update: Root Cause Found and Workaround Applied

    After investigating the OpenCode source, the issue is in how provider options flow through the LLM options merge chain for OAuth-authenticated OpenAI models.

    Root Cause

    When using OpenAI models via OAuth (Codex auth), the base options from ProviderTransform.options() in transform.ts only set include: ["reasoning.encrypted_content"] and reasoningSummary for providerID.startsWith("opencode") — not for providerID === "openai" (which is what OAuth uses). Without these parameters, the API performs reasoning internally but doesn't return thinking blocks in the response stream.

    The relevant code (transform.ts ~line 662):

    if (input.model.providerID.startsWith("opencode")) {
      result["promptCacheKey"] = input.sessionID
      result["include"] = ["reasoning.encrypted_content"]
      result["reasoningSummary"] = "auto"
    }

    OAuth models have providerID === "openai", so they skip this block entirely.

    For users with custom model definitions and variants, there's a second issue: the variant options merge on top of base options, but if the variant doesn't include include: ["reasoning.encrypted_content"], it may not be present in the final merged options unless set at the model options level.

    Workaround

    Adding options with the missing parameters directly to the model definition in opencode.json fixes thinking blocks for GPT 5.2:

    "openai": {
      "models": {
        "gpt-5.2": {
          "options": {
            "reasoningEffort": "high",
            "reasoningSummary": "auto",
            "include": ["reasoning.encrypted_content"],
            "store": false
          }
        }
      }
    }

    These model-level options merge at layer 2 of the options chain (llm.ts line 110: mergeDeep(input.model.options)), which applies to all sessions using the model — including subagents.

    Additional finding: reasoningEffort at agent level is silently ignored

    For users setting reasoningEffort at the agent config level (e.g., on custom agents), this field is not in the agent schema (agent.ts lines 22-47). It is silently ignored. The correct way to lock an agent to a reasoning level is "variant": "high" at the agent level, which IS supported (agent.ts line 210, resolved in prompt.ts lines 831-838).

    Note on Codex 5.3

    Codex 5.3 models show thinking blocks after this fix, but they are very concise (1-2 sentence summaries). This appears to be model behavior rather than a config issue — the Codex endpoint returns valid response.reasoning_summary_text.delta events, but the model produces terse summaries. See #12523 for the same report.

    Environment

    • OpenCode 1.1.34 (custom build)
    • Windows 10
    • OAuth via ChatGPT Plus
14. NamedIdentity commented on Feb 17, 2026

    

    NamedIdentity

    on Feb 17, 2026

    Author

    More actions

    Closing as workaround found for thinking blocks showing up, even though issues specific to Codex 5.3 remain.
15. NamedIdentity

    closed this as completedon Feb 17, 2026