Closes#37289
The current implementation has a problem. The **`from_id` method** in
the Anthropic crate works well for predefined models, but not for custom
models that are defined in the settings. This is because it fallbacks to
using default beta headers, which are incorrect for custom models.
The issue is that the model instance for custom models lives within the
`language_models` provider, so I've updated the **`stream_completion`**
method to explicitly accept beta headers from its caller. Now, the beta
headers are passed from the `language_models` provider all the way to
`anthropic.stream_completion`, which resolves the issue.
Release Notes:
- Fixed a bug where extra_beta_headers defined in settings for Anthropic
custom models were being ignored.
---------
Signed-off-by: Umesh Yadav <git@umesh.dev>
Closes#32758
Release Notes:
- Resolved an issue with the Ollama provider that caused requests to
fail with a 400 error for models that don't support tools. The tools
object is now only sent to compatible models to ensure successful
requests.
Improves the error handling for openrouter and adds automatic retry like
anthropic for few of the status codes.
Release Notes:
- Improves error messages for Openrouter provider
- Automatic retry when rate limited or Server error from Openrouter
Closes#30188Closes#34911Closes#34906
Many OpenAI-compatible providers do not automatically filter the tool
schema to comply with the underlying model's requirements; they simply
proxy the request. This creates issues, as models like **Gemini**,
**Grok**, and **Claude** (when accessed via LiteLLM on Bedrock) are
incompatible with Zed's default tool schema.
This PR addresses this by defaulting to a more compatible schema subset
instead of the full schema.
### Why this approach?
* **Avoids Poor User Experience:** One alternative was to add an option
for users to manually set the JSON schema for models that return a `400
Bad Request` due to an invalid tool schema. This was discarded as it
provides a poor user experience.
* **Simplifies Complex Logic:** Another option was to filter the schema
based on the model ID. However, as demonstrated in the attached issues,
this is unreliable. For instance, `claude-4-sonnet` fails when proxied
through LiteLLM on Bedrock. Reliably determining behavior would require
a non-trivial implementation to manage provider-and-model combinations.
* **Better Default Behavior:** The current approach ensures that tool
usage works out-of-the-box for the majority of cases by default,
providing the most robust and user-friendly solution.
Release Notes:
- Improved tool compatibility with OpenAI API-compatible providers
Signed-off-by: Umesh Yadav <git@umesh.dev>
Co-authored-by: Peter Tripp <peter@zed.dev>
This PR fixes an issue where extension operations would never show in
the activity indicator despite this being implemented for ages. This
happened because we were always returning `None` whenever the app has a
global auto updater, which is always the case, so the code path for
showing extension updates in the indicator could never be hit despite
existing prior. Also slightly improves the messages shown for ongoing
extension operations, as these were previously context unaware.
While I was at this, I also quickly took a stab at cleaning up some
remotely related stuff, namely:
- The `AnimationExt` trait is now by default only implemented for
anything that also implements `IntoElement`. This prevents
`with_animation` from showing up for e.g. `u32` within the suggestions
(finally).
- Commonly used animations are now implemented in the
`CommonAnimationExt` trait within the `ui` crate so the needed code does
not always need to be copied and element IDs for the animations are
truly unique.
Relevant change here regarding the original issue is the change from the
`return match` to just a `match` within the activitiy indicator, which
solved the issue at hand.
If we find this to be too noisy at some point, we can easily revisit,
but I think this holds important enough information to be shown in the
activity indicator, especially whilst developing extensions.
Release Notes:
- Extension installation and updates will now be shown in the activity
indicator.
Closes #ISSUE
Initially, the `SettingsUi` trait was tied to `Settings`, however, given
that the `Settings::FileContent` type (which may be the same as the type
that implements `Settings`) will be the type that more directly maps to
the JSON structure (and therefore have the documentation, correct field
names (or `serde` rename attributes), etc) it makes more sense to have
the deriving of `SettingsUi` occur on the `FileContent` type rather than
the `Settings` type.
In order for this to work a relatively important change had to be made
to the derive macro, that being that it now "unwraps" options into their
inner type, so a field with type `Option<Foo>` where `Foo: SettingsUi`
will treat the field as if it were just `Foo`, expecting there to be a
default set in `default.json`. This imposes some restrictions on what
`Settings::FileContent` can be as seen in 1e19398 where `FileContent`
itself can't be optional without manually implementing `SettingsUi`, as
well as introducing some risk that if the `FileContent` type has
`serde(default)`, the default value will override the default value from
`default.json` in the UI even though it may differ (but it should!).
A future PR should probably replace the other settings with `FileContent
= Option<T>` (all of which currently have `T == bool`) with wrapper
structs and have `KEY = None` so the further niceties
`derive(SettingsUi)` will provide such as path renaming, custom UI, auto
naming and doc comment extraction can be used.
Release Notes:
- N/A *or* Added/Fixed/Improved ...
This PR fixes a deserialization issue in GitHub Copilot Chat that was
causing warnings when encountering xAI models from the GitHub Copilot
API and skipping the Grok model from model selector.
Release Notes:
- Fixed support for xAI models that are now available through GitHub
Copilot Chat.
Removes excess log which got through on each start of Zed
```
ERROR [agent_ui::language_model_selector] Failed to authenticate provider: Anthropic: credentials not found
```
The `AnthropicLanguageModelProvider::api_key` method returned a
`anyhow::Result` which would convert
`AuthenticateError::CredentialsNotFound` into a generic error because of
the implicit `Into` when using the `?` operator. This would then get
converted into a `AuthenticateError::Other` later.
By specifying the error type as `AuthenticateError`, we remove this
implicit conversion and the log gets removed.
Release Notes:
- N/A
Closes#37025
This PR fixes GitHub Copilot thread summary failures by removing the
unnecessary `noop` tool insertion logic. The code was originally added
as a workaround in https://github.com/zed-industries/zed/pull/30007 for
supposed GitHub Copilot API issues when tools were used previously in a
conversation but no tools are provided in the current request. However,
testing revealed that this scenario works fine without the workaround,
and the `noop` tool insertion was actually causing "Invalid schema for
function 'noop'" errors that prevented thread summarization from
working. Removing this logic eliminates the errors and allows thread
summarization to function correctly with GitHub Copilot models.
The best way to see if removing that part of code works is just
triggering thread summarisation.
Error Log:
```
2025-08-27T13:47:50-04:00 ERROR [workspace::notifications] "Failed to connect to API: 400 Bad Request {"error":{"message":"Invalid schema for function 'noop': In context=(), object schema missing properties.","code":"invalid_function_parameters"}}\n"
```
Release Notes:
- Fixed GitHub Copilot thread summary failures by removing unnecessary
noop tool insertion logic.
## Goal
This PR creates the initial settings ui structure with the primary goal
of making a settings UI that is
- Comprehensive: All settings are available through the UI
- Correct: Easy to understand the underlying JSON file from the UI
- Intuitive
- Easy to implement per setting so that UI is not a hindrance to future
settings changes
### Structure
The overall structure is settings layer -> data layer -> ui layer.
The settings layer is the pre-existing settings definitions, that
implement the `Settings` trait. The data layer is constructed from
settings primarily through the `SettingsUi` trait, and it's associated
derive macro. The data layer tracks the grouping of the settings, the
json path of the settings, and a data representation of how to render
the controls for the setting in the UI, that is either a marker value
for the component to use (avoiding a dependency on the `ui` crate) or a
custom render function.
Abstracting the data layer from the ui layer allows crates depending on
`settings` to implement their own UI without having to add additional UI
dependencies, thus avoiding circular dependencies. In cases where custom
UI is desired, and a creating a custom render function in the same crate
is infeasible due to circular dependencies, the current solution is to
implement a marker for the component in the `settings` crate, and then
handle the rendering of that component in `settings_ui`.
### Foundation
This PR creates a macro and a trait both called `SettingsUi`. The
`SettingsUi` trait is added as a new trait bound on the `Settings`
trait, this allows the type system to guarantee that all settings
implement UI functionality. The macro is used to derived the trait for
most types, and can be modified through attributes for unique cases as
well.
A derive-macro is used to generate the settings UI trait impl, allowing
it the UI generation to be generated from the static information in our
code base (`default.json`, Struct/Enum names, field names, `serde`
attributes, etc). This allows the UI to be auto-generated for the most
part, and ensures consistency across the UI.
#### Immediate Follow ups
- Add a new `SettingsPath` trait that will be a trait bound on
`SettingsUi` and `Settings`
- This trait will replace the `Settings::key` value to enable
`SettingsUi` to infer the json path of it's derived type
- Figure out how to render `Option<T> where T: SettingsUi` correctly
- Handle `serde` attributes in the `SettingsUi` proc macro to correctly
get json path from a type's field and identity
Release Notes:
- N/A
---------
Co-authored-by: Ben Kunkle <ben@zed.dev>
Closes#37022Closes#36994
This update ensures all Grok models use the JsonSchemaSubset format for
tool schemas.
A previous fix for this issue was too specific, only targeting grok-4
models. This caused other variants, like grok-code-fast-1, to be missed.
We've now broadened the logic to correctly apply the setting to the
entire Grok model family.
Release Notes:
- Fix tool calling for `x-ai/grok-code-fast-1` model via OpenRouter.
Users now accept ToS from Zed's website when they sign in to Zed the
first time. So it's no longer possible that a signed in account could
not have accepted the ToS.
Release Notes:
- N/A
---------
Co-authored-by: Mikayla Maki <mikayla.c.maki@gmail.com>
This PR identifies automatic configuration options that users can select
from the agent panel. If no default provider is set in their settings,
the PR defaults to the first recommended option. Additionally, it
updates the selected provider for a thread when a user changes the
default provider through the settings file, if the thread hasn't had any
queries yet.
Release Notes:
- agent: automatically select a language model provider if there's no
user set provider.
---------
Co-authored-by: Michael Sloan <michael@zed.dev>
This removes around 900 unnecessary clones, ranging from cloning a few
ints all the way to large data structures and images.
A lot of these were fixed using `cargo clippy --fix --workspace
--all-targets`, however it often breaks other lints and needs to be run
again. This was then followed up with some manual fixing.
I understand this is a large diff, but all the changes are pretty
trivial. Rust is doing some heavy lifting here for us. Once I get it up
to speed with main, I'd appreciate this getting merged rather sooner
than later.
Release Notes:
- N/A
We'll now use the anthropic provider to get credentials for `claude` and
embed its configuration view in the panel when they are not present.
Release Notes:
- N/A
### TL;DR
* Adds `capabilities` configuration for OpenAI-compatible models
* Relates to
https://github.com/zed-industries/zed/issues/36215#issuecomment-3193920491
### Summary
This PR introduces support for configuring model capabilities for
OpenAI-compatible language models. The implementation addresses the
issue that not all OpenAI-compatible APIs support the same features -
for example, Cerebras' API explicitly does not support
`parallel_tool_calls` as documented in their [OpenAI compatibility
guide](https://inference-docs.cerebras.ai/resources/openai#currently-unsupported-openai-features).
### Changes
1. **Model Capabilities Structure**:
- Added `ModelCapabilityToggles` struct for UI representation with
boolean toggle states
- Implemented proper parsing of capability toggles into
`ModelCapabilities`
2. **UI Updates**:
- Modified the "Add LLM Provider" modal to include checkboxes for each
capability
- Each OpenAI-compatible model can now be configured with its specific
capabilities through the UI
3. **Configuration File Structure**:
- Updated the settings schema to support a `capabilities` object for
each `openai_compatible` model
- Each capability (`tools`, `images`, `parallel_tool_calls`,
`prompt_cache_key`) can be individually specified per model
### Example Configuration
```json
{
"openai_compatible": {
"Cerebras": {
"api_url": "https://api.cerebras.ai/v1",
"available_models": [
{
"name": "gpt-oss-120b",
"max_tokens": 131000,
"capabilities": {
"tools": true,
"images": false,
"parallel_tool_calls": false,
"prompt_cache_key": false
}
}
]
}
}
}
```
### Tests Added
- Added tests to verify default capability values are correctly applied
- Added tests to verify that deselected toggles are properly parsed as
`false`
- Added tests to verify that mixed capability selections work correctly
Thanks to @osyvokon for the desired `capabilities` configuration
structure!
Release Notes:
- OpenAI-compatible models now have configurable capabilities (#36370;
thanks @calesennett)
---------
Co-authored-by: Oleksiy Syvokon <oleksiy@zed.dev>
Some APIs fail when they get this parameter
Closes#36215
Release Notes:
- Fixed OpenAI-compatible providers that don't support prompt caching
and/or reasoning
Release Notes:
- Added `reasoning_effort` support to custom models
Tested using the following config:
```json5
"language_models": {
"openai": {
"available_models": [
{
"name": "gpt-5-mini",
"display_name": "GPT 5 Mini (custom reasoning)",
"max_output_tokens": 128000,
"max_tokens": 272000,
"reasoning_effort": "high" // Can be minimal, low, medium (default), and high
}
],
"version": "1"
}
}
```
Docs:
https://platform.openai.com/docs/api-reference/chat/create#chat_create-reasoning_effort
This work could be used to split the GPT 5/5-mini/5-nano into each of
it's reasoning effort variant. E.g. `gpt-5`, `gpt-5 low`, `gpt-5
minimal`, `gpt-5 high`, and same for mini/nano.
Release Notes:
* Added a setting to control `reasoning_effort` in OpenAI models
## Summary
Enable image processing capabilities for GPT-5 series models by updating
the `supports_images()` method.
## Changes
- Add vision support for `gpt-5`, `gpt-5-mini`, and `gpt-5-nano` models
- Update `supports_images()` method in
`crates/language_models/src/provider/open_ai.rs`
## Models with Vision Support (after this PR)
- gpt-4o
- gpt-4o-mini
- gpt-4.1
- gpt-4.1-mini
- gpt-4.1-nano
- gpt-5 (new)
- gpt-5-mini (new)
- gpt-5-nano (new)
- o1
- o3
- o4-mini
This brings GPT-5 vision capabilities in line with other OpenAI models
that support image processing.
Release Notes:
- Added vision support for OpenAI models
Tested prompt:
John is one of 4 children. The first sister is 4 years old. Next year,
the second sister will be twice as old as the first sister. The third
sister is two years older than the second sister. The third sister is
half the age of her older brother. How old is John? Return your thinking
inside <think></think>
Release Notes:
- Add thinking to Mistral Provider
---------
Signed-off-by: Umesh Yadav <git@umesh.dev>
Co-authored-by: Peter Tripp <peter@zed.dev>
This uses the `current_user` watch in the `UserStore` instead of looping
every 100ms in order to detect if the user had signed in.
We are changing this because we noticed it was causing the deterministic
executor in tests to never detect a "parking with nothing left to run"
situation.
This seems better in production as well, especially for users who never
sign in.
/cc @maxdeviant
Release Notes:
- N/A
Co-authored-by: Ben Brandt <benjamin.j.brandt@gmail.com>
This is really just a small beginning, as there are many other icons to
be revised and cleaned up. Our current set is a bit of a mess in terms
of dimension, spacing, stroke width, and terminology. I'm sure there are
more non-used icons I'm not covering here, too. We'll hopefully tackle
it all soon leading up to 1.0.
Closes https://github.com/zed-industries/zed/issues/35576
Release Notes:
- N/A
This pull request should be idempotent, but lays the groundwork for
avoiding to connect to collab in order to interact with AI features
provided by Zed.
Release Notes:
- N/A
---------
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
This PR updates the Agent panel to work with the `CloudUserStore`
instead of the `UserStore`, reducing its reliance on being connected to
Collab to function.
Release Notes:
- N/A
---------
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
Closes#26030
*Note: This is my first contribution to Zed*
This addresses a second streaming bottleneck in Bedrock that remained
after the initial fix in #28281 (released in preview 194).
The issue is in the mechanism used to convert Zed's internal `AsyncBody`
into the `SdkBody` expected by the Bedrock language provider. We are
using a non-streaming converter that buffers responses.
**How the fix works:**
The AWS SDK provides streaming-compatible converters to create `SdkBody`
instances, but these require the input body to implement the `Body`
trait from the `http-body` crate.
This PR enables streaming by implementing the required trait and
switching to the streaming-compatible converter.
**Changes (2 commits):**
* 1st Commit - **Implement http-body Body trait for AsyncBody:**
- Add `http-body = 1.0` dependency (already an indirect dependency)
- Implement the `Body` trait for our existing `AsyncBody` type
- Uses `poll_frame` to read data chunks asynchronously, preserving
streaming behavior
* 2nd Commit - **Use streaming-compatible AWS SDK converter:**
- Create `SdkBody` using `SdkBody::from_body_1_x()` with the new `Body`
trait implementation
**Details/FAQ:**
**Q: Why add another dependency?**
A: We tried to avoid adding a dependency, but the AWS SDK requires the
`Body` trait and `http-body` is where it's defined. The crate is already
an indirect dependency, making this a reasonable solution.
**Q: Why modify the shared `http_client` crate instead of just
`aws_bedrock_client`?**
A: We considered implementing the `Body` trait on a wrapper in
`aws_bedrock_client`, but since `AsyncBody` already uses `http` crate
types, extending support to the companion `http-body` crate seems
reasonable and may benefit other integrations.
**Q: How was this bottleneck discovered?**
A: After @5herlocked's initial streaming fix in #28281, I tested preview
194 and noticed streaming still had issues. I found a way to reproduce
the problem and chatted with @5herlocked about it. He immediately
pinpointed the exact location where the issue was occurring, his
diagnosis made this fix possible.
**Q: How does this relate to the previous fix?**
A: #28281 fixed buffering issues higher in the stack, but unfortunately
there was another bottleneck lower-down in the aws-http-client. This PR
addresses that separate buffering issue.
**Q: Does this use zero-copy or one-copy?**
A: The `Body` implementation includes one copy. Someone more
knowledgeable might be able to achieve a zero-copy approach, but we
opted for a conservative approach. The performance impact should not be
perceptible in typical usage.
**Testing:**
Confirmed that Bedrock streaming now works without buffering delays in a
local build.
Release Notes:
- Improved Bedrock streaming by eliminating response buffering delays
---------
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
TODO
- [x] OpenAI Compatible API Icon
- [x] Docs
- [x] Link to docs in OpenAI provider section about configuring OpenAI
API compatible providers
Closes#33992
Related to #30010
Release Notes:
- agent: Add support for adding multiple OpenAI API compatible providers
---------
Co-authored-by: MrSubidubi <dev@bahn.sh>
Co-authored-by: Danilo Leal <daniloleal09@gmail.com>
Closes#32675
Exactly the same changes as in #33640 by @sviande
The PR has been in WIP state for 3 weeks with no activity, and the issue
basically makes Mistral models unusable. I have tested the changes
locally, and it does indeed work. Full credit goes to @sviande, I just
want this feature to be finished.
Release Notes:
- agent: Fixed an issue with tool calling with the Mistral provider
(thanks [@sviande](https://github.com/sviande) and
[@armyhaylenko](https://github.com/armyhaylenko))
Co-authored-by: sviande <sviande@gmail.com>
This includes making sure that both the agent panel and Zed's edit
prediction have a consistent narrative when it comes to onboarding users
into the AI features, considering the possible different plans and
conditions (such as being signed in/out, account age, etc.)
Release Notes:
- N/A
---------
Co-authored-by: Bennet Bo Fenner <53836821+bennetbo@users.noreply.github.com>
Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
