Follow-up to #62010, which added Gemini 3.6 Flash.
Google released Gemini 3.7 Flash on August 13, 2026. It keeps the 1
million token context window and the 64k output limit, and supports
thinking.
Unlike 3.5 and 3.6 Flash, it does not accept `thinking_level: MINIMAL`
(the API returns a validation error). So it exposes Low, Medium and
High, defaulting to Medium, and `disabled_thinking_level` in
`completion.rs` returns Low for it instead of Minimal when the user
turns thinking off.
The provider builds its model list from `google_ai::Model::iter()`, so
adding the enum variant is enough for it to show up in the model
dropdown.
Release Notes:
- Added Gemini 3.7 Flash to the Google AI models
Closes#62007
Adds Gemini 3.6 Flash to the Google AI provider so it can be selected in
the agent, alongside the existing Gemini 3.5 Flash.
Google released Gemini 3.6 Flash on July 21, 2026. It keeps the 1
million token context window, has a 64k output limit, and supports
thinking. The model is added with the same capabilities and thinking
levels as 3.5 Flash (Minimal, Low, Medium, High, defaulting to Medium).
The provider builds its model list from `google_ai::Model::iter()`, so
adding the enum variant is enough for it to show up in the model
dropdown. The thinking behavior in `completion.rs` keys off the
`gemini-3` id prefix, so it already applies to this model.
Release Notes:
- Added Gemini 3.6 Flash to the Google AI models
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
## Problem
When `api_url` points to a proxy or gateway that appends `data: [DONE]`
to the SSE stream, every streaming completion fails with:
```
error deserializing Anthropic API response: expected value at line 1 column 2
```
In the UI this surfaces as **"error deserializing Anthropic API response
— Retrying (Attempt 3 of 3)"** followed by **"Connection Interrupted"**,
even though the proxy returned a well-formed stream and the model's full
response was already delivered before the terminator. The failure is
100% reproducible on multi-turn tool-use conversations (Zed reads the
stream to completion after each turn, so it always reaches the trailing
`[DONE]`) and intermittent on single-turn conversations (depending on
whether the connection closes before the line is read). The same class
of error appears in the `google_ai` crate as `Error parsing JSON: ...
"[DONE]"`.
Neither side is at fault here. Zed's current parser is a correct
implementation of the Anthropic and Gemini specs, which terminate
streams by connection close and never emit `[DONE]`; at the same time,
appending a `[DONE]` terminator is a common, legitimate convention among
SSE gateways. This PR is a compatibility improvement that makes Zed
robust to that widely-used stream shape, so streaming works whether or
not the upstream emits the terminator.
## Why Zed crashes but the Anthropic SDK does not
The Anthropic TypeScript and Python SDKs handle this correctly
**without** any explicit `[DONE]` check. Their SSE parser uses a
two-layer architecture:
1. **SSE framing layer** — parses `event:` / `data:` fields into
`{event, data}` objects without inspecting the data payload.
2. **Event dispatch layer** — only calls `JSON.parse` on events whose
`event` name is in a known whitelist (`message_start`,
`content_block_delta`, etc.). A bare `data:` line with no preceding
`event:` line yields `event: null`, which is not in the whitelist and is
silently ignored.
So when a proxy sends `data: [DONE]`, the SDK produces `{event: null,
data: "[DONE]"}`, the dispatch layer drops it, and `JSON.parse` is never
run on `[DONE]`. Claude Code and Cline both use the Anthropic SDK under
the hood, so they inherit this behavior and never crash on `[DONE]`.
Zed's `anthropic` and `google_ai` crates use a simpler single-layer
parser: strip the `data:` prefix, then unconditionally
`serde_json::from_str` the remainder. There is no event-name dispatch,
so `[DONE]` goes straight into the JSON parser — `[` looks like an array
start, `D` is not a valid value, serde reports `expected value at line 1
column 2`, and the stream dies.
## Fix
Add a one-line guard after prefix stripping: if the trimmed payload is
`[DONE]`, return `None` from the `filter_map` closure. This is a
lightweight equivalent of the SDK's whitelist-ignore behavior — same
result (silently skip the terminator), without restructuring the parser
into a full event-name dispatch architecture.
Two commits, one per crate:
1. `anthropic` — guard in `stream_completion_with_rate_limit_info`
2. `google_ai` — identical guard in `stream_generate_content`
## Scope
The other ten SSE-based providers in this repo (`open_ai`, `deepseek`,
`mistral`, `open_router`, `lmstudio`, `llama_cpp`, `copilot_chat`)
already handle `[DONE]` because their upstream specs (OpenAI-compatible)
define it as a required stream terminator. `ollama` (NDJSON) and
`bedrock` (AWS SDK event stream) do not use SSE. After this PR, every
SSE parser in Zed handles `[DONE]` without error.
## Safety
- Official Anthropic and Gemini APIs never send `[DONE]`; the guard is a
no-op on direct connections.
- `line.trim() == "[DONE]"` handles both `data: [DONE]` and
`data:[DONE]`.
- Returning `None` drops the line silently; the stream concludes on the
next EOF as usual.
Release Notes:
- Fixed streaming completions failing with "error deserializing
Anthropic API response" when the SSE endpoint appends a `[DONE]` stream
terminator (affects custom `api_url` proxies for Anthropic and Google
Gemini)
# Objective
Allow zed's language model stack to express OpenAI Responses API custom
tools (freeform text-input tools with an optional lark/regex grammar),
so downstream consumers can offer tools like a freeform `apply_patch` to
GPT models.
## Solution
- `LanguageModelRequestTool` now carries a Function-vs-Custom input
variant; `LanguageModelCustomToolFormat` models text/grammar formats.
- `LanguageModelToolUse.input` becomes a typed
`LanguageModelToolUseInput::{Json, Text}`. Serialization is tagged so
persisted Text inputs round-trip losslessly; legacy plain JSON values
still deserialize as `Json`.
- `open_ai` gains the custom tool wire types (tool definition,
`custom_tool_call`/`custom_tool_call_output` input items with
string-or-content-part outputs, output item, and
`custom_tool_call_input` delta/done stream events). The Responses event
mapper accumulates raw text deltas into `ToolUse` events, and history
replay derives custom-vs-function tool results from the matching
`ToolUse` by id.
- All non-OpenAI providers and the Chat Completions path error
explicitly when a request contains custom tools — no silent drops or
empty-schema coercions.
Release Notes:
- N/A
Adds provider-side context compaction support to the Anthropic and
OpenAI API clients. Client plumbing only; not wired into the UI.
Release Notes:
- N/A
## Solution
- A simple find and replace.
## Testing
- Did you test these changes? If so, how? No, as in my opinion, a single
character change in a string does not justify the time it takes to
recompile
- Are there any parts that need more testing? No.
- How can other people (reviewers) test your changes? Is there anything
specific they need to know? Just open the models UI.
- If relevant, what platforms did you test these changes on, and are
there any important ones you can't test?
## 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 adheres to Zed's UI standards
([UX/UI](https://github.com/zed-industries/zed/blob/main/CONTRIBUTING.md#uiux-checklist)
and
[icon](https://github.com/zed-industries/zed/blob/main/crates/icons/README.md)
guidelines)
- [X] Tests cover the new/changed behavior
- [X] Performance impact has been considered and is acceptable
---
Release Notes:
- Edit name of Gemini 3.1 Flash Lite to Gemini 3.1 Flash-Lite to better
match Google's official documentation
Add a `custom_headers` setting to each HTTP-based language model
provider
(Anthropic, Bedrock, DeepSeek, Google, LM Studio, Mistral, Ollama,
OpenAI,
OpenAI-compatible, OpenCode, OpenRouter, Vercel AI Gateway, and xAI) so
users
can attach extra headers to every outgoing request. Headers managed by
Zed
(authentication, content-type, etc.) cannot be overridden and are
skipped with
a warning.
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:
- Added support for configuring custom HTTP headers on language model
providers via `language_models.<provider>.custom_headers`.
Signed-off-by: Aurabindo Pillai <mail@aurabindo.in>
Also makes sure we are properly catching and processing thinking events.
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
Release Notes:
- google: Support thinking levels for Google models.
Implements the [official upgrade
instructions](https://ai.google.dev/gemini-api/docs/whats-new-gemini-3.5#migrate-from-3-flash-preview)
for Gemini 3.5 Flash, and adds BYOK support.
The changes about thinking_level and temperature apply to our situation,
but they are only recommendations, and we have to support older models,
so I preferred not trying to force the preferred / remove the
discouraged parameters for now.
`temperature` becomes optional - we don't fill in a default anymore,
since passing it is now discouraged.
This commit also adds support for `thinking_level`, since it is now
preferred to `thinking_budget`.
`FunctionCall` and `FunctionResponse` now support passing an `id` to
properly maintain chain-of-thought preservation and match execution IDs
across turns. When resolving incoming tool uses, the mapper prefers the
execution ID returned by Gemini, falling back to sequential naming in
other scenarios.
Release Notes:
- Added support for Gemini 3.5 Flash in the Google AI model provider.
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 (none)
Release Notes:
- Google: Added Gemini 3.1 Flash Lite
---------
Co-authored-by: Ben Brandt <benjamin.j.brandt@gmail.com>
Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com>
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
Drop the `count_tokens` API and related implementations across
providers, and remove the unused `tiktoken-rs` dependency.
I was going to update the dependency becuase they finally released a fix
we needed. But then I realized we only used this api in one place, the
Rules library. And for most models it would have been wildly incorrect
becuase we use tiktoken, i.e. OpenAI tokenizers, for almost every model,
which is going to give incorrect results.
Given that, I just removed these because the difference in how we get
these has caused plenty of confusion in the past.
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
Release Notes:
- N/A
- `language_model` no longer depends on provider-specific crates such as
`anthropic` and `open_ai` (inverted dependency)
- `language_model_core` was extracted from `language_model` which
contains the types for the provider-specific crates to convert to/from.
- `gpui::SharedString` has been extracted into its own crate (still
exposed by `gpui`), so `language_model_core` and provider API crates
don't have to depend on `gpui`.
- Removes some unnecessary `&'static str` | `SharedString` -> `String`
-> `SharedString` conversions across the codebase.
- Extracts the core logic of the cloud `LanguageModelProvider` into its
own crate with simpler dependencies.
Release Notes:
- N/A
---------
Co-authored-by: John Tur <john-tur@outlook.com>
Gemini 3 Pro Preview has been deprecated in favor of Gemini 3.1 Pro.
This removes the `Gemini3Pro` variant from the `Model` enum and all
associated match arms, updates eval model lists, docs, and test
fixtures.
A serde alias (`"gemini-3-pro-preview"`) is kept on `Gemini31Pro` so
existing user settings gracefully migrate to the replacement model.
Closes AI-66
Release Notes:
- Removed deprecated Gemini 3 Pro Preview model; existing configurations
automatically migrate to Gemini 3.1 Pro.
The change simplifies the `max_token_count` and `max_output_tokens`
methods by grouping Gemini models with identical token limits.
Release Notes:
- N/A
Closes#43040
Release Notes:
- Remove the end-of-support Gemini 1.5 model from the options.
- Remove the older Gemini 2.0 model from the options.
- Please let me know if you think it's better to keep it, as it is still
a usable model.
- Update the incorrect amounts for some input/output tokens.
- Update the default model to Gemini 2.5 Flash-Lite.
- Rename variant `Gemini3ProPreview` to `Gemini3Pro`
When this PR is merged, users will be able to select the following
Gemini models.
- 2.5 Flash
- 2.5 Flash-Lite
- 2.5 Pro
- 3 Pro
We've been considering removing workspace-hack for a couple reasons:
- Lukas ran into a situation where its build script seemed to be causing
spurious rebuilds. This seems more likely to be a cargo bug than an
issue with workspace-hack itself (given that it has an empty build
script), but we don't necessarily want to take the time to hunt that
down right now.
- Marshall mentioned hakari interacts poorly with automated crate
updates (in our case provided by rennovate) because you'd need to have
`cargo hakari generate && cargo hakari manage-deps` after their changes
and we prefer to not have actions that make commits.
Currently removing workspace-hack causes our workspace to grow from
~1700 to ~2000 crates being built (depending on platform), which is
mainly a problem when you're building the whole workspace or running
tests across the the normal and remote binaries (which is where
feature-unification nets us the most sharing). It doesn't impact
incremental times noticeably when you're just iterating on `-p zed`, and
we'll hopefully get these savings back in the future when
rust-lang/cargo#14774 (which re-implements the functionality of hakari)
is finished.
Release Notes:
- N/A
Co-Authored-By: Ben K <ben@zed.dev>
Co-Authored-By: Anthony <anthony@zed.dev>
Co-Authored-By: Mikayla <mikayla@zed.dev>
Release Notes:
- settings: Major internal changes to settings. The primary user-facing
effect is that some settings which did not make sense in project
settings files are no-longer read from there. (For example the inline
blame settings)
---------
Co-authored-by: Ben Kunkle <ben@zed.dev>
Co-authored-by: Mikayla Maki <mikayla.c.maki@gmail.com>
Co-authored-by: Anthony <anthony@zed.dev>
This prevents the common footgun of copy/pasting an API key
starting/ending with extra newlines, which would lead to a "bad request"
error.
Closes#37038
Release Notes:
- agent: Support pasting language model API keys that contain newlines.
Updates google_ai to use latest model information from the respective
model cards: https://ai.google.dev/gemini-api/docs/models
Release Notes:
- google: Update to latest Gemini 2.5 models
Previously we were using a mix of `u32` and `usize`, e.g. `max_tokens:
usize, max_output_tokens: Option<u32>` in the same `struct`.
Although [tiktoken](https://github.com/openai/tiktoken) uses `usize`,
token counts should be consistent across targets (e.g. the same model
doesn't suddenly get a smaller context window if you're compiling for
wasm32), and these token counts could end up getting serialized using a
binary protocol, so `usize` is not the right choice for token counts.
I chose to standardize on `u64` over `u32` because we don't store many
of them (so the extra size should be insignificant) and future models
may exceed `u32::MAX` tokens.
Release Notes:
- N/A
Closes#31243
As described in my issue, the [thinking
budget](https://ai.google.dev/gemini-api/docs/thinking) gets
automatically chosen by Gemini unless it is specifically set to
something. In order to have fast responses (inline assistant) I prefer
to set it to 0.
Release Notes:
- ai: Added `thinking` mode for custom Google models with configurable
token budget
---------
Co-authored-by: Ben Brandt <benjamin.j.brandt@gmail.com>
https://github.com/zed-industries/zed/issues/30972 brought up another
case where our context is not enough to track the actual source of the
issue: we get a general top-level error without inner error.
The reason for this was `.ok_or_else(|| anyhow!("failed to read HEAD
SHA"))?; ` on the top level.
The PR finally reworks the way we use anyhow to reduce such issues (or
at least make it simpler to bubble them up later in a fix).
On top of that, uses a few more anyhow methods for better readability.
* `.ok_or_else(|| anyhow!("..."))`, `map_err` and other similar error
conversion/option reporting cases are replaced with `context` and
`with_context` calls
* in addition to that, various `anyhow!("failed to do ...")` are
stripped with `.context("Doing ...")` messages instead to remove the
parasitic `failed to` text
* `anyhow::ensure!` is used instead of `if ... { return Err(...); }`
calls
* `anyhow::bail!` is used instead of `return Err(anyhow!(...));`
Release Notes:
- N/A
* `CountTokensRequest` now takes a full `GenerateContentRequest` instead
of just content.
* Fixes use of `models/` prefix in `model` field of
`GenerateContentRequest`, since that's required for use in
`CountTokensRequest`. This didn't cause issues before because it was
always cleared and used in the path.
Release Notes:
- N/A
Sometimes Gemini would report `Content` without a `parts` field.
Release Notes:
- Fixed a bug that would sometimes cause Gemini models to fail streaming
their response.
This is to enable alternative streaming solutions at the application
layer. I'm not sure we really should have performed parsing of the input
at this layer. Either way I want to experiment with streaming approaches
in a separate crate on a branch, and this will help.
/cc @maxdeviant @bennetbo @rtfeldman
Closes #ISSUE
Release Notes:
- N/A
* Adds a fast / cheaper model to providers and defaults thread
summarization to this model. Initial motivation for this was that
https://github.com/zed-industries/zed/pull/29099 would cause these
requests to fail when used with a thinking model. It doesn't seem
correct to use a thinking model for summarization.
* Skips system prompt, context, and thinking segments.
* If tool use is happening, allows 2 tool uses + one more agent response
before summarizing.
Downside of this is that there was potential for some prefix cache reuse
before, especially for title summarization (thread summarization omitted
tool results and so would not share a prefix for those). This seems fine
as these requests should typically be fairly small. Even for full thread
summarization, skipping all tool use / context should greatly reduce the
token use.
Release Notes:
- N/A
See #28793, the name of the field is actually `systemInstruction` not
`systemInstructions`.
Release Notes:
- Fixed an issue where Gemini requests would fail
Arnesh