Improve interleaved reasoning context by Echsecutor · Pull Request #55085 · zed-industries/zed

Echsecutor · 2026-04-28T14:29:27Z

Self-Review Checklist:

[ x] I've reviewed my own diff for quality, security, and reliability
[ x] Unsafe blocks (if any) have justifying comments
[ x] The content is consistent with the UI/UX checklist
[ x] Tests cover the new/changed behavior
[ x] Performance impact has been considered and is acceptable

Closes #51743

Added trace logging in into_open_ai for request building, SSE chunk processing, and tool-call stream events.
Added a warn log when an assistant tool-call message is created with interleaved_reasoning=true but no reasoning content was accumulated, to help diagnose proxy/API forwarding issues.
Added trace logging of the full JSON request body in stream_completion.
Added into_open_ai_interleaved_reasoning_thinking_only_before_tool_use test covering single-turn and multi-turn scenarios where the assistant responds with thinking + tool_calls and no text content.

Release Notes:

Fixed interleaved reasoning content being lost on tool-only assistant turns
When interleaved_reasoning is enabled and an assistant message contains only thinking content followed by a tool use (with no text block), the accumulated reasoning_content is now correctly attached to the generated OpenAI-format assistant message. Previously this reasoning was dropped, causing API errors with models (e.g., Kimi) that require reasoning content in conversation history.

Signed-off-by: Sebastian Schmittner <sebastian.schmittner@eecc.de>

nervgh · 2026-04-28T15:06:27Z

Thank you for your efforts!

By the way, Deepseek v4 has the same problem.

Echsecutor · 2026-04-30T12:47:09Z

Thank you for your efforts!

By the way, Deepseek v4 has the same problem.

I just tried with deepseek-v4-pro. This model works as well with the ZED binary build from this branch.

infused-kim · 2026-05-03T19:52:41Z

Thank you for your work on this issue. I tried your branch, but still received the same error if there was no accumulated reasoning.

I was able to fix it by setting reasoning_content to " " (string with a space). Setting it to an empty string didn't work because either the opencode proxy or the API requet code was removing the parameter.

You can find my branch with the commit here:
https://github.com/infused-kim/zed/tree/epc/fixing_reasoning_context

My test prompt was Can you please run the env command and show me the output?.

And my config was:

{
  "language_models": {
    "openai_compatible": {
      "OpenCode Go": {
        "api_url": "https://opencode.ai/zen/go/v1",
        "available_models": [
          {
            "name": "kimi-k2.6",
            "max_tokens": 262144,
            "max_output_tokens": 65536,
            "max_completion_tokens": 262144,
            "capabilities": {
              "tools": true,
              "images": true,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
              "interleaved_reasoning": true,
            },
          },
        ],
      },
    },
  },

infused-kim · 2026-05-03T21:57:04Z

For anyone else who wants to start using opencode go before this gets merged, I created a branch based of v1.0.0 that adds only the required commits to fix this issue:

https://github.com/infused-kim/zed/tree/pr-testing/55085-opencode-thinking-fix

This produces a release build (Zed.app) instead of (Zed Dev.app).

You can build it by checking out the branch and then running ./script/bundle-mac. You may need to install various dependencies. For more info check the documentation here:
https://github.com/zed-industries/zed/tree/28a54ce122fdd5efb2e23cc77a5efab78c07061c/docs/src/development

And my full config is:

{
  "language_models": {
    "openai_compatible": {
      "OpenCode Go": {
        "api_url": "https://opencode.ai/zen/go/v1",

        // For more models:
        // https://models.dev/?sort=model&order=asc&search=opencode+go
        // https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models
        "available_models": [
          {
            "name": "kimi-k2.6",
            "max_tokens": 262144,
            "max_output_tokens": 65536,
            "max_completion_tokens": 262144,
            "capabilities": {
              "tools": true,
              "images": true,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
              "interleaved_reasoning": true,
            },
          },
          {
            "name": "glm-5.1",
            "max_tokens": 204800,
            "max_output_tokens": 131072,
            "max_completion_tokens": 204800,
            "capabilities": {
              "tools": true,
              "images": false,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
            },
          },
          {
            "name": "deepseek-v4-pro",
            "max_tokens": 1000000,
            "max_output_tokens": 384000,
            "max_completion_tokens": 1000000,
            "capabilities": {
              "tools": true,
              "images": false,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
              "interleaved_reasoning": true,
            },
          },
          {
            "name": "deepseek-v4-flash",
            "max_tokens": 1000000,
            "max_output_tokens": 384000,
            "max_completion_tokens": 1000000,
            "capabilities": {
              "tools": true,
              "images": false,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
              "interleaved_reasoning": true,
            },
          },
          {
            "name": "minimax-m2.7",
            "max_tokens": 204800,
            "max_output_tokens": 131072,
            "max_completion_tokens": 204800,
            "capabilities": {
              "tools": true,
              "images": false,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
            },
          },
          {
            "name": "qwen3.6-plus",
            "max_tokens": 262144,
            "max_output_tokens": 65536,
            "max_completion_tokens": 262144,
            "capabilities": {
              "tools": true,
              "images": true,
              "parallel_tool_calls": true,
              "prompt_cache_key": true,
              "chat_completions": true,
            },
          },
        ],
      },
    },
  },

Echsecutor · 2026-05-04T10:17:17Z

Hi @infused-kim ! Thanks for testing!

I am using zed build form this branch for actual coding with a lot of tool calling for a couple of days now with kimik 2.6 and deepseek 4 pro routing through an openai_compatible conversation API. This is working fine for me.

In prinziple, sending " " instead of none/empty string sounds a bit hacky, but I am happy to check and maybe integrate into this PR if that helps in your provider setup. Can you point me to the specific commit/change? Trying to Pull your branch into mine creates a huge diff, that's probably not intended.

There is also some discussion about this PR in #51743 , btw.

commit c6a0aac Author: Kim Streich <kim@infusedinsight.com> Date: Sun May 3 13:46:43 2026 -0600 Set reasoning_content to “ “ if no content has accumulated commit 202be6e Author: Sebastian Schmittner <sebastian.schmittner@eecc.de> Date: Sat May 2 21:34:00 2026 +0200 commit message generation/context summary with interleved reasoning commit fbcb79c Author: Sebastian Schmittner <sebastian.schmittner@eecc.de> Date: Tue Apr 28 16:29:11 2026 +0200 Simplify reasoning content logic and add test helpers commit 54c9448 Author: Sebastian Schmittner <sebastian.schmittner@eecc.de> Date: Tue Apr 28 15:34:19 2026 +0200 improved error logging for reasoning context Signed-off-by: Sebastian Schmittner <sebastian.schmittner@eecc.de> commit 557ece0 Author: Bennet Bo Fenner <bennet@zed.dev> Date: Mon Apr 27 14:36:11 2026 +0200 agent: Allow tools to output multiple content parts (zed-industries#54518) Self-Review Checklist: - [x] I've reviewed my own diff for quality, security, and reliability - [x] Unsafe blocks (if any) have justifying comments - [x] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [x] Tests cover the new/changed behavior - [x] Performance impact has been considered and is acceptable Closes #ISSUE Release Notes: - N/A

infused-kim · 2026-05-04T21:52:19Z

@Echsecutor yeah, it could be an issue with OpenCode's implementation of the API. I am not sure, but without setting reasoning_content to " " it didn't work for me.

I also agree that it feels dirty, but for now I just wanted to get it working.

I rebased my branch on top of yours, so you should now be able to fast forward my commit:
infused-kim@76556c0
https://github.com/infused-kim/zed/commits/pr-orig/55085-opencode-thinking-fix/

…ies#55574) OpenCode API endpoints for DeepSeek were [moved from Anthropic-compatible to OpenAI-compatible](anomalyco/opencode#24500) and DeepSeek requires interleaved reasoning enabled to work. I ran a _"rename this variable to potato"_ test and I can confirm DeepSeek V4 Flash and Pro both work now 🎉 Some other OpenCode Go models were marked [on models.dev](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models) as supporting `interleaved_reasoning` so they too got that enabled. Kimi K2.5 and Kimi K2.6 continue to fail with zed-industries#51743 (zed-industries#55085 seems to hint at this being [an OpenCode issue](zed-industries#51743 (comment))?), but all other models seem to work fine both with `interleaved_reasoning` and without it 🤷 I assume it's better to have that turned on? Again, the intersection of OpenAI Chat Completions API, different models, different inference providers, how they all work together is something I know nothing about! Self-Review Checklist: - [X] I've reviewed my own diff for quality, security, and reliability - [X] Unsafe blocks (if any) have justifying comments - [X] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [ ] Tests cover the new/changed behavior - [X] Performance impact has been considered and is acceptable Release Notes: - OpenCode Go: use correct DeepSeek endpoints - OpenCode: add support for interleaved_reasoning

…erry-pick to preview) (#55777) Cherry-pick of #55574 to preview ---- OpenCode API endpoints for DeepSeek were [moved from Anthropic-compatible to OpenAI-compatible](anomalyco/opencode#24500) and DeepSeek requires interleaved reasoning enabled to work. I ran a _"rename this variable to potato"_ test and I can confirm DeepSeek V4 Flash and Pro both work now 🎉 Some other OpenCode Go models were marked [on models.dev](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models) as supporting `interleaved_reasoning` so they too got that enabled. Kimi K2.5 and Kimi K2.6 continue to fail with #51743 (#55085 seems to hint at this being [an OpenCode issue](#51743 (comment))?), but all other models seem to work fine both with `interleaved_reasoning` and without it 🤷 I assume it's better to have that turned on? Again, the intersection of OpenAI Chat Completions API, different models, different inference providers, how they all work together is something I know nothing about! Self-Review Checklist: - [X] I've reviewed my own diff for quality, security, and reliability - [X] Unsafe blocks (if any) have justifying comments - [X] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [ ] Tests cover the new/changed behavior - [X] Performance impact has been considered and is acceptable Release Notes: - OpenCode Go: use correct DeepSeek endpoints - OpenCode: add support for interleaved_reasoning Co-authored-by: Vlad Ionescu <Vlaaaaaaad@users.noreply.github.com>

…ies#55574) OpenCode API endpoints for DeepSeek were [moved from Anthropic-compatible to OpenAI-compatible](anomalyco/opencode#24500) and DeepSeek requires interleaved reasoning enabled to work. I ran a _"rename this variable to potato"_ test and I can confirm DeepSeek V4 Flash and Pro both work now 🎉 Some other OpenCode Go models were marked [on models.dev](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models) as supporting `interleaved_reasoning` so they too got that enabled. Kimi K2.5 and Kimi K2.6 continue to fail with zed-industries#51743 (zed-industries#55085 seems to hint at this being [an OpenCode issue](zed-industries#51743 (comment))?), but all other models seem to work fine both with `interleaved_reasoning` and without it 🤷 I assume it's better to have that turned on? Again, the intersection of OpenAI Chat Completions API, different models, different inference providers, how they all work together is something I know nothing about! Self-Review Checklist: - [X] I've reviewed my own diff for quality, security, and reliability - [X] Unsafe blocks (if any) have justifying comments - [X] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [ ] Tests cover the new/changed behavior - [X] Performance impact has been considered and is acceptable Release Notes: - OpenCode Go: use correct DeepSeek endpoints - OpenCode: add support for interleaved_reasoning

cla-bot · 2026-05-07T08:51:47Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

cla-bot · 2026-05-07T08:52:06Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

Echsecutor · 2026-05-07T08:53:26Z

@infused-kim ok, I have merged your contirbution, many thanks! But, as the bot says, you need to accept that cla for this PR to be mergeable now 😉

cla-bot · 2026-05-07T10:00:28Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

infused-kim · 2026-05-07T15:41:13Z

@cla-bot check

@Echsecutor thank you! Actually, I’m not sure if that’s needed anymore. Version 1.1.5 that was published yesterday includes official open code go support.

If this was an issue that was only happening with open code, then we might not even need this.

cla-bot · 2026-05-07T15:41:16Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

cla-bot · 2026-05-07T15:41:17Z

The cla-bot has been summoned, and re-checked this pull request!

Vlaaaaaaad · 2026-05-07T16:12:06Z

actually, I’m not sure if that’s needed anymore. Version 1.1.5 that was published yesterday includes official open code go support.

If this was an issue that was only happening with open code, then we might not even need this

Kimi K2.6 from OpenCode Go still fails with the error from #51743 🙂 As I mentioned in this comment, I have no idea if the fix needs to be in Zed or in OpenCode 🤷

I don't know if other providers are still impacted by this or not — Moonshot was mentioned as impacted on the issue discussion?

raphaelluethy · 2026-05-07T17:10:44Z

actually, I’m not sure if that’s needed anymore. Version 1.1.5 that was published yesterday includes official open code go support.

If this was an issue that was only happening with open code, then we might not even need this

Kimi K2.6 from OpenCode Go still fails with the error from #51743 🙂 As I mentioned in this comment, I have no idea if the fix needs to be in Zed or in OpenCode 🤷

I don't know if other providers are still impacted by this or not — Moonshot was mentioned as impacted on the issue discussion?

From my local testing on my own pr, the kimi models fail due to the reasoning content field being discarded when the string is empty. Simply adding an empty string with a space fixes the issue. Opencode seems to do the same in their cli, so i.g. this fix is the only (or easier) way.

Vlaaaaaaad · 2026-05-08T08:10:08Z

@raphaelluethy yup the " " hack fixes the Kimi K2.6 from OpenCode Go issue (which has been known for a while)! In my comment I was talking about Zed 1.1.5 still being broken — I should've made that clearer, sorry!

raphaelluethy · 2026-05-08T14:02:05Z

As this is an issue on opencodes side, I published my testing branch that puts the fix in the opencode go provider. I think this this would be the cleaner version, as we dont pollute the global implementation of the openai compat provider. #56182
I leave this up to the maintainers which approach they would like to follow.

cla-bot · 2026-05-12T08:38:17Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

Echsecutor · 2026-05-12T08:39:43Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

@infused-kim did you click that agreement? since its like two lines of code I could also rewrite the history for this PR, but I would appreciate if your contribution is honored 😉

cla-bot · 2026-05-12T08:39:45Z

We require contributors to sign our Contributor License Agreement, and we don't have @infused-kim on file. You can sign our CLA at https://zed.dev/cla. Once you've signed, post a comment here that says '@cla-bot check'.

cla-bot · 2026-05-12T08:39:46Z

The cla-bot has been summoned, and re-checked this pull request!

infused-kim · 2026-05-12T17:38:25Z

@cla-bot check

cla-bot · 2026-05-12T17:38:30Z

The cla-bot has been summoned, and re-checked this pull request!

infused-kim · 2026-05-12T17:39:20Z

Sorry for the delay @Echsecutor. I thought I already did it, but must have been mistaken.

Thank you for your patience and for the effort to make sure my contribution is included. I really appreciate it.

infused-kim · 2026-05-12T17:41:25Z

@cla-bot check

cla-bot · 2026-05-12T17:41:29Z

The cla-bot has been summoned, and re-checked this pull request!

benbrandt · 2026-05-17T19:19:26Z

Hi given we reworked the responses api and other areas of the openai providers this week, I assume there is a decent amount of conflicts (github won't even show them to me, just that they exist)

Mind doing another pass if needed off main?

Echsecutor · 2026-05-21T12:51:45Z

It seems like your reworking fixed the bug @benbrandt , I did not encounter it any more in the recent release versions, so this PR has probably outlived its purpose 😉

infused-kim · 2026-05-22T06:12:00Z

I believe I saw in another issue that this was fixed on OpenCode's side. Thank you for all your work on it @Echsecutor .

OpenCode API endpoints for DeepSeek were [moved from Anthropic-compatible to OpenAI-compatible](anomalyco/opencode#24500) and DeepSeek requires interleaved reasoning enabled to work. I ran a _"rename this variable to potato"_ test and I can confirm DeepSeek V4 Flash and Pro both work now 🎉 Some other OpenCode Go models were marked [on models.dev](https://github.com/anomalyco/models.dev/tree/dev/providers/opencode-go/models) as supporting `interleaved_reasoning` so they too got that enabled. Kimi K2.5 and Kimi K2.6 continue to fail with zed-industries/zed#51743 (zed-industries/zed#55085 seems to hint at this being [an OpenCode issue](zed-industries/zed#51743 (comment))?), but all other models seem to work fine both with `interleaved_reasoning` and without it 🤷 I assume it's better to have that turned on? Again, the intersection of OpenAI Chat Completions API, different models, different inference providers, how they all work together is something I know nothing about! Self-Review Checklist: - [X] I've reviewed my own diff for quality, security, and reliability - [X] Unsafe blocks (if any) have justifying comments - [X] The content is consistent with the [UI/UX checklist](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist) - [ ] Tests cover the new/changed behavior - [X] Performance impact has been considered and is acceptable Release Notes: - OpenCode Go: use correct DeepSeek endpoints - OpenCode: add support for interleaved_reasoning

Echsecutor added 2 commits April 28, 2026 15:34

improved error logging for reasoning context

ce1b045

Signed-off-by: Sebastian Schmittner <sebastian.schmittner@eecc.de>

Simplify reasoning content logic and add test helpers

753e59b

cla-bot Bot added the cla-signed The user has signed the Contributor License Agreement label Apr 28, 2026

Merge branch 'main' into fixing_reasoning_context

83c22ba

zed-community-bot Bot added the first contribution the author's first pull request to Zed. NOTE: the label application is automated via github actions label Apr 28, 2026

SomeoneToIgnore added the area:ai Related to Agent Panel, Edit Prediction, Copilot, or other AI features label Apr 28, 2026

Echsecutor mentioned this pull request Apr 28, 2026

Kimi k2.5 error "thinking is enabled but reasoning_content is missing in assistant tool call message" #51743

Closed

Echsecutor changed the title ~~improved error logging for reasoning context~~ improved interleaved reasoning context Apr 28, 2026

maxdeviant changed the title ~~improved interleaved reasoning context~~ Improve interleaved reasoning context Apr 28, 2026

Merge branch 'main' into fixing_reasoning_context

fca056d

commit message generation/context summary with interleved reasoning

7335e67

Vlaaaaaaad mentioned this pull request May 3, 2026

opencode: Support interleaved_reasoning and fix DeepSeek #55574

Merged

5 tasks

Merge branch 'main' into fixing_reasoning_context

72c26bd

Echsecutor mentioned this pull request May 4, 2026

WIP: compare branches european-epc-competence-center/zed#1

Draft

Set reasoning_content to “ “ if no content has accumulated

76556c0

zed-zippy Bot mentioned this pull request May 5, 2026

opencode: Support interleaved_reasoning and fix DeepSeek (#55574) (cherry-pick to preview) #55777

Merged

5 tasks

restore warning logs when using hacky blank string work around

f62be9d

cla-bot Bot removed the cla-signed The user has signed the Contributor License Agreement label May 7, 2026

Merge branch 'main' into fixing_reasoning_context

66085c7

Merge branch 'main' into fixing_reasoning_context

c6170eb

Vlaaaaaaad mentioned this pull request May 8, 2026

opencode: Add support for OpenCode Go #53651

Merged

5 tasks

raphaelluethy mentioned this pull request May 8, 2026

opencode: Ensure non-empty reasoning content for interleaving #56182

Closed

4 tasks

Merge branch 'main' into fixing_reasoning_context

a03e158

cla-bot Bot added the cla-signed The user has signed the Contributor License Agreement label May 12, 2026

benbrandt closed this May 17, 2026

Uh oh!

Conversation

Echsecutor commented Apr 28, 2026

Uh oh!

nervgh commented Apr 28, 2026

Uh oh!

Echsecutor commented Apr 30, 2026

Uh oh!

infused-kim commented May 3, 2026

Uh oh!

infused-kim commented May 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Echsecutor commented May 4, 2026

Uh oh!

infused-kim commented May 4, 2026

Uh oh!

cla-bot Bot commented May 7, 2026

Uh oh!

cla-bot Bot commented May 7, 2026

Uh oh!

Echsecutor commented May 7, 2026

Uh oh!

cla-bot Bot commented May 7, 2026

Uh oh!

infused-kim commented May 7, 2026

Uh oh!

cla-bot Bot commented May 7, 2026

Uh oh!

cla-bot Bot commented May 7, 2026

Uh oh!

Vlaaaaaaad commented May 7, 2026

Uh oh!

raphaelluethy commented May 7, 2026

Uh oh!

Vlaaaaaaad commented May 8, 2026

Uh oh!

raphaelluethy commented May 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cla-bot Bot commented May 12, 2026

Uh oh!

Echsecutor commented May 12, 2026

Uh oh!

cla-bot Bot commented May 12, 2026

Uh oh!

cla-bot Bot commented May 12, 2026

Uh oh!

infused-kim commented May 12, 2026

Uh oh!

cla-bot Bot commented May 12, 2026

Uh oh!

infused-kim commented May 12, 2026

Uh oh!

infused-kim commented May 12, 2026

Uh oh!

cla-bot Bot commented May 12, 2026

Uh oh!

benbrandt commented May 17, 2026

Uh oh!

Echsecutor commented May 21, 2026

Uh oh!

infused-kim commented May 22, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

infused-kim commented May 3, 2026 •

edited

Loading

raphaelluethy commented May 8, 2026 •

edited

Loading