| commit | a3e28fb8e570ab7e43f84d2f5c5fcecb81d00618 | [log] [tgz] |
|---|---|---|
| author | Josh Bleecher Snyder <josharian@gmail.com> | Fri May 30 15:53:23 2025 +0000 |
| committer | Josh Bleecher Snyder <josharian@gmail.com> | Fri May 30 12:34:24 2025 -0700 |
| tree | 5469d620466a368eae9f342bd240178e0e0cf57b | |
| parent | 15cc2e75ed6dc840c78bc7f2b2075d40d5f22470 [diff] |
llm/conversation: fix duplicate tool results in insertMissingToolResults Fix critical bug in insertMissingToolResults function that created duplicate tool_result blocks when multiple tool uses were missing results, causing Claude API 400 errors with message "each tool_use must have a single result". Problem Analysis: The function was designed to add error tool results for tool uses that were requested but had no corresponding tool_result in the user message (e.g., when hitting budget limits). However, when multiple tool uses were missing results, the function incorrectly duplicated results: 1. First iteration: adds tool1 result to prefix, sets msg.Content = [tool1] + original 2. Second iteration: adds tool2 to prefix, sets msg.Content = [tool1, tool2] + [tool1] + original This created duplicate tool_result blocks with the same tool_use_id, violating Claude's API requirements and causing 400 Bad Request errors. Root Cause: The msg.Content assignment was inside the loop that builds the prefix: Solution: Move the msg.Content assignment outside the loop, after all missing tool results have been collected: Testing: Added comprehensive test coverage for insertMissingToolResults with scenarios: - Single missing tool result - Multiple missing tool results (the problematic case) - No missing results when results already present - No tool uses in previous message The test specifically verifies no duplicate tool_use_ids are created and that all expected tool results are present exactly once. This fix resolves the "multiple tool_result blocks with id" error that was occurring in deep conversations with tool usage. Fixes #121 Co-Authored-By: sketch <hello@sketch.dev> Change-ID: s07f278af38b745aak
Sketch is an agentic coding tool. It draws the 🦉
Sketch runs in your terminal, has a web UI, understands your code, and helps you get work done. To keep your environment pristine, sketch starts a docker container and outputs its work onto a branch in your host git repository.
Sketch helps with most programming environments, but Sketch has extra goodies for Go.
go install sketch.dev/cmd/sketch@latest sketch
Currently, Sketch runs on macOS and Linux. It uses Docker for containers.
| Platform | Installation |
|---|---|
| macOS | brew install colima (or Docker Desktop/Orbstack) |
| Linux | apt install docker.io (or equivalent for your distro) |
| WSL2 | Install Docker Desktop for Windows (docker entirely inside WSL2 is tricky) |
The sketch.dev service is used to provide access to an LLM service and give you a way to access the web UI from anywhere.
Start Sketch by running sketch in a Git repository. It will open your browser to the Sketch chat interface, but you can also use the CLI interface. Use -open=false if you want to use just the CLI interface.
Ask Sketch about your codebase or ask it to implement a feature. It may take a little while for Sketch to do its work, so hit the bell (🔔) icon to enable browser notifications. We won't spam you or anything; it will notify you when the Sketch agent's turn is done, and there's something to look at.
When you start Sketch, it:
This design lets you run multiple sketches in parallel since they each have their own sandbox. It also lets Sketch work without worry: it can trash its own container, but it can't trash your machine.
Sketch's agentic loop uses tool calls (mostly shell commands, but also a handful of other important tools) to allow the LLM to interact with your codebase.
Sketch is trained to make Git commits. When those happen, they are automatically pushed to the git repository where you started sketch with branch names sketch/*.
Finding Sketch branches:
git branch -a --sort=creatordate | grep sketch/ | tail
The UI keeps track of the latest branch it pushed and displays it prominently. You can use standard Git workflows to pull those branches into your workspace:
git cherry-pick $(git merge-base origin/main sketch/foo)
or merge the branch
git merge sketch/foo
or reset to the branch
git reset --hard sketch/foo
Ie use the same workflows you would if you were pulling in a friend's Pull Request.
Advanced: You can ask Sketch to git fetch sketch-host and rebase onto another commit. This will also fetch where you started Sketch, and we do a bit of "git fetch refspec configuration" to make origin/main work as a git reference.
Don't be afraid of asking Sketch to help you rebase, merge/squash commits, rewrite commit messages, and so forth; it's good at it!
The diff view shows you changes since Sketch started. Leaving comments on lines adds them to the chat box, and, when you hit Send (at the bottom of the page), Sketch goes to work addressing your comments.
You can interact directly with the container in three ways:
ssh sketch-ilik-eske-tcha-lott. We have automatically configured your SSH configuration to make these special hostnames work.Using SSH (and/or VSCode) allows you to forward ports from the container to your machine. For example, if you want to start your development webserver, you can do something like this:
# Forward container port 8888 to local port 8000 ssh -L8000:localhost:8888 sketch-ilik-epor-tfor-ward go run ./cmd/server
This makes http://localhost:8000/ on your machine point to localhost:8888 inside the container.
You can ask Sketch to browse a web page and take screenshots. There are tools both for taking screenshots and "reading images", the latter of which sends the image to the LLM. This functionality is handy if you're working on a web page and want to see what the in-progress change looks like.
Docker images, containers, and so forth tend to pile up. Ask Docker to prune unused images and containers:
docker system prune -a
See CONTRIBUTING.md for development guidelines.
Sketch is open source. It is right here in this repository! Have a look around and mod away.
If you want to run Sketch entirely without the sketch.dev service, you can set the flag -skaband-addr="" and then provide an ANTHROPIC_API_KEY environment variable. (More LLM services coming soon!)