https://github.com/zed-industries/zed/pull/31668 renamed
`CompletionMode::Max` to `CompletionMode::Burn` which is a good change,
but this broke the deserialization for threads whose completion mode was
stored in LMDB. This adds a deserialization alias so that both values
work.
We could make a full new `SerializedThread` version which migrates this
value, but that seems overkill for this single change, we can batch that
with more changes later. Also, people in nightly already have some v1
threads with `burn` stored, so it wouldn't quite work for everybody.
Release Notes:
- N/A
The title of a (text) thread would get stuck in "Loading Summary..."
when the request to generate it failed. We now handle this case by
falling back to the default title, and letting the user manually edit
the title or retry generating it.
https://github.com/user-attachments/assets/898d26ad-d31f-4b62-9b05-519d923b1b22
Release Notes:
- agent: Handle thread title generation errors
---------
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
- Try to preserve previously selected item on update
- Do not clear list items while updating to avoid a frame with no items
rendered
Release Notes:
- agent: Preserve previously selected item in Thread History on update
---------
Co-authored-by: Danilo Leal <daniloleal09@gmail.com>
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
We were still using entry indexes to scroll, but the list now includes
the separators as items, so the indexes need to be translated
Release Notes:
- agent: Fix autoscrolling to history entry when navigating via keyboard
Fixes an issue introduced by #29959 which caused the message editor to
overflow from the agent panel bounds, making the bottom buttons
invisible when the editor was expanded (cmd+esc).
Fixing this required changing the base structure of the agent panel, but
things seem to work as expected:
https://github.com/user-attachments/assets/fc4c97fb-f7cb-4f54-a268-c30fbcb1649f
Release Notes:
- agent: Fix message editor's button positions when expanded
Turns out `naive_local` doesn't actually offset a `DateTime<Utc>` to the
local timezone before creating a `NaiveDate`.
Release Notes:
- agent: Use correct timezone for thread history separators
Adds a new `agent.model_parameters` setting that allows the user to
specify a custom temperature for a provider AND/OR model:
```json5
"model_parameters": [
// To set parameters for all requests to OpenAI models:
{
"provider": "openai",
"temperature": 0.5
},
// To set parameters for all requests in general:
{
"temperature": 0
},
// To set parameters for a specific provider and model:
{
"provider": "zed.dev",
"model": "claude-3-7-sonnet-latest",
"temperature": 1.0
}
],
```
Release Notes:
- agent: Allow customizing temperature by provider/model
---------
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
Closes#29836
The agent diff toolbar item was causing the editor toolbar to show even
when all the other elements were disabled via settings.
This PR fixes this by setting the location to
`ToolbarItemLocation::Hidden` in the states where it shouldn't show.
It also adds a new a `toolbar.agent_review` setting to hide the agent
review buttons altogether. However, if the other toolbar elements are
hidden and the file isn't under review, the editor toolbar will still be
hidden. So you only need to set this to `false` if you don't want them
to show up even under agent review.
Release Notes:
- N/A
Enables reviewing agent edits from single-file editors in addition to
the multibuffer experience we already had.
https://github.com/user-attachments/assets/a2c287f0-51d6-43a1-8537-821498b91983
This feature can be turned off by setting `assistant.single_file_review:
false`.
Release Notes:
- agent: Review edits in single-file editors
Adds popular examples of long-running commands to system prompt.
Unfortunately, I couldn't add an eval example as the new terminal tool
no longer works in `eval`. We can look into that tomorrow, but I'm
seeing improvements when manually testing this, so I'd like to merge it.
<img
src="https://github.com/user-attachments/assets/ac24e617-e068-466f-875d-c30e1f2465c4"
width=400></img>
Release Notes:
- agent: Discourage long-running commands
The `grep` tool used to include 4 lines of context around the match, but
the lines included would often be unhelpful. This PR improves this
behavior by using the range of the parent syntax node that contains the
full line(s) matched.
The match headers will also now include symbol breadcrumbs so that the
model can already gather code structure before/without reading files.
````md
### impl GitRepository for RealGitRepository › fn compare_checkpoints › L1278-1284
```rust
let result = git
.run(&[
"diff-tree",
"--quiet",
&left.commit_sha.to_string(),
&right.commit_sha.to_string(),
])
```
````
This positively impacts the `add_arg_to_trait_method` eval example with
better diff output, fewer tool failures, and reduced total turns.
Note: We have some plans to use a an "elision" approach where we would
combine all matches for a given file, skipping lines between them while
keeping symbol declaration lines. The theory is that this would be map
more closely to the expected input for edits. For now, this PR is a
significant improvement.
Release Notes:
- Agent: Enrich `grep` tool output with syntax information
#29354 introduced a bug where we would append tool uses to the last
assistant message even if it was from a previous request.
Release Notes:
- N/A
Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
We used to insert empty user messages into the `Thread::messages` `Vec`
when tools finished running and then we would attach the results when
creating the request. This approach was very easy to mess up during
state handling, leading to empty user messages displayed in the
conversation and API failures.
Instead, we will no longer insert actual user messages for tool results
to the `Thread`, and will only do this on the fly when creating the
model request. This simplifies a lot of code and show fix the mentioned
errors.
Release Notes:
- agent: Improve reliability of LLM requests when including tool results
---------
Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
Co-authored-by: Oleksiy Syvokon <oleksiy.syvokon@gmail.com>
This PR significantly improves the quality of the initial file search
that occurs when the model doesn't yet know the full path to a file it
needs to read/edit.
Previously, the assertions in file_search often failed on main as the
model attempted to guess full file paths. On this branch, it reliably
calls `find_path` (previously `path_search`) before reading files.
After getting the model to find paths first, I noticed it would try
using `grep` instead of `path_search`. This motivated renaming
`path_search` to `find_path` (continuing the analogy to unix commands)
and adding system prompt instructions about proper tool selection.
Note: I know the command is just called `find`, but that seemed too
general.
In my eval runs, the `file_search` example improved from 40% ± 10% to
98% ± 2%. The only assertion I'm seeing occasionally fail is "glob
starts with `**` or project". We can probably add some instructions in
that regard.
Release Notes:
- N/A
Instructs the model to include the fields that we display first in the
input object, so that e.g the user can see the path of a file while the
model generates the content.
Release Notes:
- N/A
Our provider code in `language_models` filters out messages for which
`LanguageModelRequestMessage::contents_empty` returns `false`. This
doesn't seem wrong by itself, but `contents_empty` was returning `false`
for messages whose first segment didn't contain non-whitespace text even
if they contained other non-empty segments. This caused requests to fail
when a message with a tool call didn't contain any preceding text.
Release Notes:
- N/A
There was recently a change that caused the Zed Edit Prediction provider
to only be assigned when the client was connected. However, this check
happened too early, resulting in restored buffers never getting
registered. We'll now subscribe to client status changes and reassign
providers accordingly.
Release Notes:
- edit prediction: Fixed bug disabling prediction in restored buffers
The UI was mistakenly using the cumulative token usage for the token
counter. It will now display the last request token count, plus an
estimation of the tokens in the message editor and context entries that
haven't been sent yet.
https://github.com/user-attachments/assets/0438c501-b850-4397-9135-57214ca3c07a
Additionally, when the user edits a message, we'll display the actual
token count up to it and estimate the tokens in the new message.
Note: We don't currently estimate the delta when switching profiles. In
the future, we want to use the count tokens API to measure every part of
the request and display a breakdown.
Release Notes:
- agent: Made the token count more accurate and added back estimation of
used tokens as you type and add context.
---------
Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
Co-authored-by: Danilo Leal <daniloleal09@gmail.com>
The `always_allow_tool_actions` setting would get overridden with the
default when we loaded each example project, leading to examples
stalling when they run a tool that needed confirmation. There's now a
separate `runner_settings.json` file where we can configure the
environment for the eval.
Release Notes:
- N/A
---------
Co-authored-by: Oleksiy <oleksiy@zed.dev>
Release Notes:
- agent: Replace `bash` tool with `terminal` tool which uses the current
shell
---------
Co-authored-by: Bennet <bennet@zed.dev>
Co-authored-by: Antonio <antonio@zed.dev>
The bash tool will now truncate its output to 8192 bytes (or the last
newline before that).
We also added a global limit for any tool that produces a clearly large
output that wouldn't fit the context window.
Release Notes:
- agent: Truncate bash tool output
---------
Co-authored-by: Michael Sloan <mgsloan@gmail.com>
I found a few more cases where the UI wasn't updated immediately after
an interaction.
Release Notes:
- agent: Fixed delay after removing threads from "Past Interactions"
- agent: Fixed delay after adding/remove context via keyboard
This would cause the history view not to get refreshed immediately when
a thread was deleted
Release Notes:
- agent: Fixed a bug where the history view wouldn't refresh after
deleting a thread
#28102 introduced a bug where thread summaries wouldn't get generated
because they would get set to the default title instead of `None`.
Not adding a release note because the bug didn't make it to Preview.
Release Notes:
- N/A
Closes: https://github.com/zed-industries/zed/issues/20582
Allows users to select a specific model for each AI-powered feature:
- Agent panel
- Inline assistant
- Thread summarization
- Commit message generation
If unspecified for a given feature, it will use the `default_model`
setting.
Release Notes:
- Added support for configuring a specific model for each AI-powered
feature
---------
Co-authored-by: Danilo Leal <daniloleal09@gmail.com>
Co-authored-by: Bennet Bo Fenner <bennetbo@gmx.de>
This makes context essentially work the same way as `read-file`,
increasing the likelihood of cache hits.
Just like with `read-file`, we'll notify the model when the user makes
an edit to one of the tracked files. In the future, we want to send a
diff instead of just a list of files, but that's an orthogonal change.
Release Notes:
- agent: Improved caching of files in context
---------
Co-authored-by: Antonio Scandurra <me@as-cii.com>
### Context
The Anthropic API fails if a request message contains a tool use and no
`Text` segments or it only contains empty `Text` segments. These are
cases that the model itself produces, but the API doesn't support
sending them back.
#27917 fixed this by appending "Using tool..." in the thread's message,
but this causes the actual conversation to include it, so it would
appear in the UI (we would actually display a gap because we never
rendered its markdown, but "Using tool..." would show up when the thread
was restored).
### Solution
We'll now only append this placeholder when we build the request, so the
API still sees it, but the UI/Thread doesn't.
Another issue we found is that the model starts mimicking these
placeholders in later tool uses which is undesirable. So unfortunately,
we had to add logic to filter them out.
Release Notes:
- agent: Improved rendering of tool uses without text
---------
Co-authored-by: Bennet <bennet@zed.dev>
This happened because of two reasons:
- `Workspace::recent_navigation_history` didn't include the current file
- The context picker added the current file to a exclude list
The latter was actually intentional because we already show the file in
the suggested context, but now that we actually have mentions, it's just
inconvenient not to have it there.
Release Notes:
- N/A
We were including the context at the end which meant it never got
cached. We'll now include it with the first message that introduced it
so it's cached as long as it doesn't change.
This is an improvement, but we probably still need to think of ways to
optimize caching for cases where files in context change.
Release Notes:
- N/A
We'll now prompt the user to start a new thread when the active one gets
too long.
<img width=500
src="https://github.com/user-attachments/assets/91445bc0-3e81-422f-aa4a-b8f0741f9d9a"></img>
When they click "Start New Thread", will create a new one with the
previous one added as context.
<img width=500
src="https://github.com/user-attachments/assets/c3b4223f-5bdd-4ba4-956f-5a5880d5e2c3"></img>
Instead of including the full thread text, we'll now add summarized
versions of threads to the context, allowing you to continue the
conversation even if it was near the token limit.
- Thread summaries are cached and persisted.
- A cached summary is invalidated if the thread is continued.
- We start generating the thread summary as soon as it's selected from
the picker. Most times, the summary will be ready by the time the user
sends the message.
- If the summary isn't ready by the time a message is sent, the user
message will be displayed in the thread immediately, and a "Summarizing
context..." indicator will appear. After the summaries are ready, we'll
start generating the response and show the usual "Generating..."
indicator.
Release Notes:
- N/A
---------
Co-authored-by: Danilo Leal <daniloleal09@gmail.com>
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
Models seem to do this ever so often and get very confused. Failing here
helps them recover.
Release Notes:
- N/A
Co-authored-by: Richard Feldman <richard@zed.dev>
Adds some guidance for the assistant on how to respond to tool results
and other interactions
Release Notes:
- N/A
Co-authored-by: Richard Feldman <richard@zed.dev>
Edit action markers look like git conflicts and can trip up tooling used
to resolve git conflicts. This PR creates them programmatically so that
they don't appear in source code.
Release Notes:
- N/A
This required adding scrollbar support to `list`. Since `list` is
virtualized, the scrollbar height will change as more items are
measured. When the user manually drags the scrollbar, we'll persist the
initial height and offset calculations accordingly to prevent the
scrollbar from moving away from the cursor as new items are measured.
We're not doing this yet, but in the future, it'd be nice to budget some
time each frame to layout unmeasured items so that the scrollbar height
is as accurate as possible.
Release Notes:
- N/A
- Add instructions in description to read before editing
- Add instructions in edit prefix to explicitly ask for reads
- Fix `connection error: delay between messages too long` by processing
chunks off a channel
Release Notes:
- N/A
Sometimes the editor model returns no search/replace blocks. This
usually happens when the architect model calls the edit tool before
reading any files. When this happens, we'll now return the raw response
from the editor model to the architect model so it can recover
accordingly.
Release Notes:
- N/A
Instead of just reporting a search match failure, we'll now indicate
whether the file is empty or exists to help the model recover better
from bad edits.
Release Notes:
- N/A
This is surprising, but the Anthropic API returns a 400 if a tool output
is an empty string because it thinks we're attaching a `tool use`
without a corresponding `tool result`, but we are not, it's just empty
(which seems totally reasonable) 🙃
Release Notes:
- N/A
Anthropic API doesn't allow `tool_use` messages without a corresponding
`tool_result`, so we'll skip those when building a request. I'll
separately investigate why we are sending request before the tool result
as that might lead to separate issues, but that might take a while and
this is currently very frustrating.
Release Notes:
- N/A
Sometimes the model produces SEARCH queries that don't match the
indentation of the source file exactly.
When we can't find an exact match, we'll now attempt to match the lines
while being more flexible about the leading whitespace as long as all
lines are consistently offset from the source, and extend the leading
whitespace in the REPLACE string accordingly.
Release Notes:
- N/A
When we failed to match a search string, we were reporting the replace
string as not found, this confuses the model and can make it go into a
doom loop. This PR fixes that improves the error output in general to
help it recover faster.
Release Notes:
- N/A
#26935 is leading to bad edits, so let's revert it for now. I'll bring
back a version of this, but it'll likely just focus on indentation
instead of making the whole search fuzzy.
Release Notes:
- N/A
When the model reads file, we'll track the version it read, and let it
know if the user makes edits to the buffer. This helps prevent edit
failures because it'll know to re-read the file before.
Release Notes:
- N/A
We'll now include the search/replace block that got applied as part of
the tool output. We think this will help the model have a better idea of
how the file changed and prevent later edit failures.
Release Notes:
- N/A
Without this, we running into the following error:
```
Running into this when running tests. Is this
dyld[45041]: Library not loaded: @rpath/WebRTC.framework/WebRTC
Referenced from: <B2EA63A5-994E-3FB0-A74B-C9C4F7E5C1EF> /Users/aguz/zed/zed/target/debug/deps/assistant_tools-522d7745dd439dfb
Reason: no LC_RPATH's found
```
Thanks Piotr!
Release Notes:
- N/A
When the user attached context in the thread, the editor model request
would fail because its tool use wouldn't be removed properly leading to
an API error.
Also, after an edit, we'd keep the old file snapshot in the context.
This would make the model think that the edits didn't apply and make it
go in a loop.
Release Notes:
- N/A
Instead of getting the whole text from the buffer, replacing with
`String::replace`, and getting a whole diff, we'll now use `SearchQuery`
to get a range, diff only that range, and apply it (all in the
background).
When we match zero strings, we'll record a "bad search", keep going and
report it to the model at the end.
Release Notes:
- N/A
---------
Co-authored-by: Max <max@zed.dev>
Adds an `debug: edit tool` action that opens a new view which will help
us debug the edit tool internals. As the edit tool runs, the log
displays:
- Instructions provided by the main model
- Response stream from the editor model
- Parsed edit blocks
- Tool output provided back to main model
The log automatically records all edit tool interactions for staff, so
if you notice something weird, you can debug it retroactively without
having to open the debug tool first. We may want to limit the number of
recorded requests later.
I have a few more ideas for it, but this seems like a good starting
point.
https://github.com/user-attachments/assets/c61f5ce8-08b1-4500-accb-db2a480eb3ab
Release Notes:
- N/A
Assertions on the parsed system prompt should use CRLF on Windows. I
didn't see it before because I was testing on my Windows VM from a
shared folder I cloned on macOS.
Release Notes:
- N/A
#26538 fixed part of the issue, but it would keep trailing carriage
returns in the old/new strings. The model is unlikely to produce those,
but we might as well support them.
Release Notes:
- N/A
Exposes a new "edit files" tool that the model can use to apply
modifications to files in the project. The main model provides
instructions and the tool uses a separate "editor" model (Claude 3.5 by
default) to generate search/replace blocks like Aider does:
````markdown
mathweb/flask/app.py
```python
<<<<<<< SEARCH
from flask import Flask
=======
import math
from flask import Flask
>>>>>>> REPLACE
```
````
The search/replace blocks are parsed and applied as they stream in. If a
block fails to parse, the tool will apply the other edits and report an
error pointing to the part of the input where it occurred. This should
allow the model to fix it.
Release Notes:
- N/A
---------
Co-authored-by: Antonio Scandurra <me@as-cii.com>
This PR makes `io.open` use our own implementation again, but instead of
the real filesystem, it will now use the project's to check file
metadata and perform read and writes using project buffers.
This also cleans up the `io.open` implementation by splitting it into
multiple methods, adds tests for various File I/O patterns, and fixes a
few bugs in read formats.
Release Notes:
- N/A
We decided to expose scripting as tools again. We are aware of the UX
downsides of doing so, but we want to focus on getting it working well
first, and the model seems to make better use of it as an actual tool.
In the future, the tools API might support streaming. If it doesn't and
we need to ship, we can consider reverting this.
Release Notes:
- N/A
This PR makes refactors the scripting functionality to be a first-class
concept of the assistant instead of a generic tool, which will allow us
to build a more customized experience.
- The tool prompt has been slightly tweaked and is now included as a
system message in all conversations. I'm getting decent results, but now
that it isn't in the tools framework, it will probably require more
refining.
- The model will now include an `<eval ...>` tag at the end of the
message with the script. We parse this tag incrementally as it streams
in so that we can indicate that we are generating a script before we see
the closing `</eval>` tag. Later, this will help us interpret the script
as it arrives also.
- Threads now hold a `ScriptSession` entity which manages the state of
all scripts (from parsing to exited) in a centralized way, and will
later collect all script operations so they can be displayed in the UI.
- `script_tool` has been renamed to `assistant_scripting`
- Script source now opens in a regular read-only buffer
Note: We still need to handle persistence properly
Release Notes:
- N/A
---------
Co-authored-by: Marshall Bowers <git@maxdeviant.com>
Agus Zubiaga
