This PR adds a new `language_models` crate to house the various language
model providers.
By extracting the provider definitions out of `language_model`, we're
able to remove `language_model`'s dependency on `editor`, which improves
incremental compilation when changing `editor`.
Release Notes:
- N/A
This PR fixes an issue where slash commands in the output of other slash
commands were not being evaluated when configured to do so.
Closes https://github.com/zed-industries/zed/issues/20820.
Release Notes:
- Fixed slash commands from other slash commands (like `/default`) not
being evaluated (Preview only).
Meant to avoid the excessive use of "Here's a concise 3-7 word title..."
and "Title:" instances we've been seeing lately.
Follow up to: https://github.com/zed-industries/zed/pull/19530
Release Notes:
- Improve prompt for generating title summaries, avoiding preambles
This PR reworks how the Assistant Panel references slash commands,
context servers, and tools.
Previously we were always reading them from the global registries, but
now we store individual collections on each Assistant Panel instance so
that there can be different ones registered for each project.
Release Notes:
- N/A
---------
Co-authored-by: Max <max@zed.dev>
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Joseph <joseph@zed.dev>
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
This improves the experience in a few ways:
- It avoids merging slash command output sections that are adjacent.
- When hitting cmd-z, all the output from a command is undone at once.
- When deleting a pending command, it stops the command and prevents new
output from flowing in.
Release Notes:
- N/A
This PR adds support for streaming output from slash commands
In this PR we are focused primarily on the interface of the
`SlashCommand` trait to support streaming the output. We will follow up
later with support for extensions and context servers to take advantage
of the streaming nature.
Release Notes:
- N/A
---------
Co-authored-by: David Soria Parra <davidsp@anthropic.com>
Co-authored-by: Antonio Scandurra <me@as-cii.com>
Co-authored-by: David <david@anthropic.com>
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Max <max@zed.dev>
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
Co-authored-by: Will <will@zed.dev>
Before: (note the 1.3k in the upper right corner instead of 3 in the
second screenshot)
<img width="459" alt="Screenshot 2024-11-04 at 11 37 58 AM"
src="https://github.com/user-attachments/assets/64c06aff-f7d2-42a4-a767-0d7a4ba0f486">
Now:
<img width="631" alt="Screenshot 2024-11-04 at 11 38 11 AM"
src="https://github.com/user-attachments/assets/22af974d-915a-41e1-9ee0-f0622901e242">
This was intended to be a conservative estimate in case you pressed
Suggest Edits (and therefore might have an unpleasant surprise if you
were under the context limit but Suggest Edits failed with a "too much
context" error message anyway, after the Suggest Edits context got added
for you behind the scenes).
However, in retrospect this design created more [confusion in the common
case](https://github.com/zed-industries/zed/pull/19900#issuecomment-2453456569)
because it made it look like more context had been actually consumed
than what was actually consumed.
This does raise a potential design question for the future: the Suggest
Edits button adds more context at the last minute without ever
communicating that it's going to do that.
In the meantime it seems best to go back to the less-confusing way of
reporting the token counts, especially since only users of the
experimental flag could possibly press Suggest Edits anyway.
Release Notes:
- Fixed issue where initial token count was over-reported as 1.3k
instead of 3 (for the context string "You").
This PR adds a bit of telemetry for Anthropic models, in order to
understand model health. With this logging, we can monitor and diagnose
dips in performance, for example due to model rollouts.
Release Notes:
- N/A
---------
Co-authored-by: Max Brunsfeld <maxbrunsfeld@gmail.com>
This PR updates the `SlashCommand` trait to use a streaming return type.
This change is just at the trait layer. The goal here is to decouple
changing the trait's API while preserving behavior on either side.
The `SlashCommandOutput` type now has two methods for converting two and
from a stream to use in cases where we're not yet doing streaming.
On the `SlashCommand` implementer side, the implements can call
`to_event_stream` to produce a stream of events based off the
`SlashCommandOutput`.
On the slash command consumer side we use
`SlashCommandOutput::from_event_stream` to convert a stream of events
back into a `SlashCommandOutput`.
The `/file` slash command has been updated to emit `SlashCommandEvent`s
directly in order for it to work properly.
Release Notes:
- N/A
---------
Co-authored-by: Max <max@zed.dev>
This PR adds a new `SlashCommandResult` type alias.
We're going to be changing what slash commands can return in order to
support streaming, so having this type alias in place will make that
switch a bit more neat.
Release Notes:
- N/A
This changes the `/workflow` command so that instead of emitting edits
in separate steps, the user is presented with a single tab, with an
editable diff that they can apply to the buffer.
Todo
* Assistant panel
* [x] Show a patch title and a list of changed files in a block
decoration
* [x] Don't store resolved patches as state on Context. Resolve on
demand.
* [ ] Better presentation of patches in the panel
* [ ] Show a spinner while patch is streaming in
* Patches
* [x] Preserve leading whitespace in new text, auto-indent insertions
* [x] Ensure patch title is very short, to fit better in tab
* [x] Improve patch location resolution, prefer skipping whitespace over
skipping `}`
* [x] Ensure patch edits are auto-indented properly
* [ ] Apply `Update` edits via a diff between the old and new text, to
get fine-grained edits.
* Proposed changes editor
* [x] Show patch title in the tab
* [x] Add a toolbar with an "Apply all" button
* [x] Make `open excerpts` open the corresponding location in the base
buffer (https://github.com/zed-industries/zed/pull/18591)
* [x] Add an apply button above every hunk
(https://github.com/zed-industries/zed/pull/18592)
* [x] Expand all diff hunks by default
(https://github.com/zed-industries/zed/pull/18598)
* [x] Fix https://github.com/zed-industries/zed/issues/18589
* [x] Syntax highlighting doesn't work until the buffer is edited
(https://github.com/zed-industries/zed/pull/18648)
* [x] Disable LSP interaction in Proposed Changes editor
(https://github.com/zed-industries/zed/pull/18945)
* [x] No auto-indent? (https://github.com/zed-industries/zed/pull/18984)
* Prompt
* [ ] make sure old_text is unique
Release Notes:
- N/A
---------
Co-authored-by: Marshall Bowers <elliott.codes@gmail.com>
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Richard <richard@zed.dev>
Co-authored-by: Marshall <marshall@zed.dev>
Co-authored-by: Nate Butler <iamnbutler@gmail.com>
Co-authored-by: Antonio Scandurra <me@as-cii.com>
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
collab: Remove dependency on X11
I'm not sure if this is the best solution (perhaps pulling
`LanguageName` into a separate `language_types` crate would be
better...?) - but it massively reduces build time / dependencies / size
and means that the collab server no longer requires X11 libraries to be
installed.
tl;dr: `telemetry_events` requires the `language` crate, and the
language crate requires a whole ton of extra stuff. Since
telemetry_events only uses `language` for a single type definition
(`LanguageName`, aka `String`), we can cut all of these out by using the
base `String` type (This doesn't seem too terrible, given that all other
telemetry fields are using basic datatypes like String as opposed to
more strongly-typed variants).
FYI the dependency tree for "why does collab need X11 libraries??" looks
like this:
```
collab
\- telemetry_events
\- language
|- gpui
|- fuzzy
| \- gpui
|- git
| \- gpui
|- lsp
| |- gpui
| \- release_channel
| \- gpui
|- settings
| |- fs
| | \- gpui
| \- gpui
|- task
| \- gpui
\- theme
\- gpui
```
Release Notes:
- N/A
This PR adds support to the assistant for display billing-related
errors.
Pulling this out of #19081 to make it easier to cherry-pick.
Release Notes:
- N/A
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Richard <richard@zed.dev>
This PR adds a bit more metadata for assistant logging.
Release Notes:
- Assistant: Added `language_name` and `model_provider` fields to
telemetry events.
---------
Co-authored-by: Marshall Bowers <elliott.codes@gmail.com>
Co-authored-by: Max <max@zed.dev>
This PR makes the `report_assistant_event` method take an
`AssistantEvent` struct instead of all of the struct fields as
individual parameters.
Release Notes:
- N/A
This PR introduces functionality for creating *branches* of buffers that
can be used to preview and edit change sets that haven't yet been
applied to the buffers themselves.
Release Notes:
- N/A
---------
Co-authored-by: Marshall Bowers <elliott.codes@gmail.com>
Co-authored-by: Marshall <marshall@zed.dev>
* [x] put this slash command behind a feature flag until we release
embedding access to the general population
* [x] choose a name for this slash command and name the rust module to
match
Release Notes:
- N/A
---------
Co-authored-by: Jason <jason@zed.dev>
Co-authored-by: Richard <richard@zed.dev>
Co-authored-by: Jason Mancuso <7891333+jvmncs@users.noreply.github.com>
Co-authored-by: Richard Feldman <oss@rtfeldman.com>
Release Notes:
- Allow Anthropic custom models to override "temperature"
This also centralized the defaulting of "temperature" to be inside of
each model's `into_x` call instead of being sprinkled around the code.
This PR makes the `Buffer::apply_ops` method infallible for
`text::Buffer` and `language::Buffer`.
We discovered that `text::Buffer::apply_ops` was only fallible due to
`apply_undo`, which didn't actually need to be fallible.
Release Notes:
- N/A
Release Notes:
- Added a new `/delta` command to re-insert changed files that were
previously included in a context.
---------
Co-authored-by: Roy <roy@anthropic.com>
This PR simplifies how images are inserted into the context editor.
We don't need to hold the `images` in a `HashMap` on the `Context`, as
we were only inserting them to pull them out again.
Release Notes:
- N/A
This PR makes it so we pass up the tool results in the `tool_results`
field in the request message to the LLM.
This required reworking how we track non-text content in the context
editor.
We also removed serialization of images in context history, as we were
never deserializing it, and thus it was unneeded.
Release Notes:
- N/A
---------
Co-authored-by: Antonio <antonio@zed.dev>
This PR adds the initial groundwork for invoking tools in response to
tool uses from the model.
Tool uses are run when the model responds with a `stop_reason` of
`tool_use`.
Currently the tool results are just inserted as text into the user
message. We'll want to include these as `tool_result` content on the
message, but Claude seems to understand it regardless.
Release Notes:
- N/A
This PR makes it so we propagate the `stop_reason` from Anthropic up to
the Assistant so that we can take action based on it.
The `extract_content_from_events` function was moved from `anthropic` to
the `anthropic` module in `language_model` since it is more useful if it
is able to name the `LanguageModelCompletionEvent` type, as otherwise
we'd need an additional layer of plumbing.
Release Notes:
- N/A
This PR adds a tool registry to hold tools that can be called by the
Assistant.
Currently we just have a `now` tool for retrieving the current datetime.
This is all behind the `assistant-tool-use` feature flag which currently
needs to be explicitly opted-in to in order for the LLM to see the
tools.
Release Notes:
- N/A
This PR adjusts the approach we use to encoding tool uses in the
completion response to use a structured format rather than simply
injecting it into the response stream as text.
In #17170 we would encode the tool uses as XML and insert them as text.
This would require then re-parsing the tool uses out of the buffer in
order to use them.
The approach taken in this PR is to make `stream_completion` return a
stream of `LanguageModelCompletionEvent`s. Each of these events can be
either text, or a tool use.
A new `stream_completion_text` method has been added to `LanguageModel`
for scenarios where we only care about textual content (currently,
everywhere that isn't the Assistant context editor).
Release Notes:
- N/A
This PR updates the Assistant with support for receiving tool uses from
Anthropic models and capturing them as text in the context editor.
This is just laying the foundation for tool use. We don't yet fulfill
the tool uses yet, or define any tools for the model to use.
Here's an example of what it looks like using the example `get_weather`
tool from the Anthropic docs:
<img width="644" alt="Screenshot 2024-08-30 at 1 51 13 PM"
src="https://github.com/user-attachments/assets/3614f953-0689-423c-8955-b146729ea638">
Release Notes:
- N/A
## Todo
* [x] Parse and present new XML output
* [x] Resolve new edits to buffers and anchor ranges
* [x] Surface resolution errors
* [x] Steps fail to resolve because language hasn't loaded yet
* [x] Treat empty `<symbol>` tag as None
* [x] duplicate assists when editing steps
* [x] step footer blocks can appear *below* the following message header
block
## Release Notes:
- N/A
---------
Co-authored-by: Mikayla <mikayla@zed.dev>
Co-authored-by: Peter <peter@zed.dev>
Co-authored-by: Marshall <marshall@zed.dev>
Co-authored-by: Antonio <antonio@zed.dev>
Co-authored-by: Antonio Scandurra <me@as-cii.com>
We've noticed performance issues in long conversations with assistants;
the profiles pointed to slowiness in WrapMap (and indeed there were some
low hanging fruits that we picked up in
https://github.com/zed-industries/zed/pull/16761). That however did not
fully resolve the issue, as WrapMap still cracked through in profiles;
basically, the speedup I've landed has just moved the post elsewhere.
The higher level issue is that we were trying to refresh message headers
for all messages, irrespective of whether they've actually needed to be
updated. This PR fixes that by using `replace_blocks` API where
possible.
Release Notes:
- Improved performance of Assistant Panel with long conversations.
Release Notes:
- Adds UI affordances to the assistant panel to show which messages have
been cached
- Migrate cache invalidation to be based on `has_edits_since_in_range`
to be smarter and more selective about when to invalidate the cache and
when to fetch.
<img width="310" alt="Screenshot 2024-08-16 at 11 19 23 PM"
src="https://github.com/user-attachments/assets/4ee2d111-2f55-4b0e-b944-50c4f78afc42">
<img width="580" alt="Screenshot 2024-08-18 at 10 05 16 PM"
src="https://github.com/user-attachments/assets/17630a60-7b78-421c-ae39-425246638a12">
I had originally added the lightening bolt on every message and only
added the tooltip warning about editing prior messages on the first
anchor, but thought it looked too busy, so I settled on just annotating
the last anchor.
Release Notes:
- Adds support for Prompt Caching in Anthropic. For models that support
it this can dramatically lower cost while improving performance.
* Improve the tab title: give it an icon, and indicate the step index.
* Display the line number ranges that the symbols resolve to.
* Don't open duplicate tabs for the same step
Release Notes:
- N/A
Now that there's a dedicated, user-facing view for each workflow step,
we don't need the inspector functionality. This PR also cleans up some
naming around workflow steps and step resolutions.
Release Notes:
- N/A
Marshall Bowers
Danilo Leal
Antonio Scandurra
Richard Feldman
Boris ChernyandMax Brunsfeld
Adam Wolff
Shish
Piotr Osiewicz
Roy Williams
Conrad Irwin
Kirill Bulatov
Kyle Kelley
Roy Williams