Cody Agent

The @sourcegraph/cody-agent package implements a JSON-RPC server to interact with Cody via stdout/stdin. This package is intended to be used by non-ECMAScript clients such as the JetBrains and NeoVim plugins.

Releases

Cody Agent releases are available:

• as self-contained executables for various platforms at Cody Agent releases on GitHub
• from the @sourcegraph/cody-agent npm package (npx @sourcegraph/cody-agent help)

To build and publish a release using GitHub Actions, bump the version number in the agent's package.json and then push to the agent-vN.N.N tag (where N.N.N is that version number).

Protocol

The protocol is defined in the file protocol.ts. The TypeScript code is the single source of truth of what JSON-RPC methods are supported in the protocol.

Updating the protocol

Directly edit the TypeScript source code to add new JSON-RPC methods or add properties to existing data structures.

The agent is a new project that is being actively worked on at the time of this writing. The protocol is subject to breaking changes without notice. Please let us know if you are implementing an agent client.

Client bindings

There's no tool to automatically generate bindings for the Cody agent protocol. Currently, clients have to manually write bindings for the JSON-RPC methods.

Useful commands

The following commands assume you are in the agent directory:

| Command | What | | ------------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------- | | pnpm build | Build dist/index.js Node.js script for running the agent | | node dist/index.js | Run the agent after pnpm build. You normally do this from a client integration. | | node --inspect dist/index.js | Run the agent with debugging enabled (see chrome://inspect/, more details) | | pnpm run test | Run all agent-related tests | | (optional) src login | Make sure you are logged into your Sourcegraph instance, which is required to run the e2e test in index.test.ts | | pnpm run test src/index.test.ts | Run e2e test, requires src login to work. |

The following commands assume you are in the root directory of this repository:

| Command | What | | -------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------- | | pnpm test agent/src/index.test.ts | Run agent tests in replay mode | | source agent/scripts/export-cody-http-recording-tokens.sh | Export access tokens to enable recording mode | | pnpm update-agent-tests | Update HTTP recordings for all tests. Run this before opening a PR | | CODY_KEEP_UNUSED_RECORDINGS=true CODY_RECORD_IF_MISSING=true pnpm run test agent/src/index.test.ts | Run this when iterating on a feature and you only want to run an individual test via it.only. Remember to run pnpm update-agent-tests before sending a PR | | CODY_KEEP_UNUSED_RECORDINGS=true CODY_RECORD_IF_MISSING=true npx vitest agent/src/index.test.ts -t 'squirrel test' | Run only a single test without making changes to the code | | ./agent/scripts/reset-recordings-to-main.sh | Overwrites the local HTTP recordings with the recordings from origin/main. Useful when preparing a PR for review. | | ./agent/scripts/resolve-recordings-git-conflict.sh | Resolves git conflicts in HTTP recording files by picking the recordings on the other branch. Requires you to re-record changes in your branch. | | pnpm agent cody-bench --help | See available flags in cody-bench tool. | | pnpm agent:skip-root-build cody-bench --evaluation-config ~/dev/sourcegraph/cody-bench-data/fix-bench.json --test-count 1 --include-fixture gpt-4o | Run cody-bench against evaluation config. Only do a single test with the gpt-4o fixture. More details in the sourcegraph/cody-bench-data repo. |

Debugging the agent

• The best way to troubleshoot a problem with the agent is to run pnpm run test src/index.test.ts because it gives you the fastest feedback loop. You can easily add a new test case with custom JSON-RPC requests/notifications to reproduce the issue you are troubleshooting. One important benefit of this workflow is that you get nice stack traces point to the original TypeScript source code.
• To see all incoming/outcoming JSON-RPC traffic, set the environment variable CODY_AGENT_TRACE_PATH=/somewhere.json and use tail -f /somewhere.json to watch the trace file while running the agent. This is particularly helpful to confirm whether your client is sending the expected JSON-RPC requests/notifications and getting the expected responses.
• If you have access to stderr of the agent process, you can add console.log(...) statements throughout the TypeScript code to trace values at specific points. This is a good fallback when the other debugging workflows are not sufficient.

Client implementations

• The Sourcegraph JetBrains plugin is defined in the sourcegraph repository's client/jetbrains directory. The CodyAgentClient.java file implements the client side of the protocol.
• The Sourcegraph Neovim plugin is defined in the sourcegraph/sg.nvim repository.

Testing with the agent

The agent includes a testing mode where it can either record HTTP requests or replay from a directory of recorded HTTP request/response pairs. When running in replay mode, the agent should be suitable to use within tests because the results should be determinic and work.

To run the agent in testing mode, define the following environment variables:

• CODY_RECORDING_DIRECTORY=PATH_TO_DIRECTORY: a directory where the HTTP recordings should be stored. This directory should be committed to git so that other people who clone the repository can replay from the same recording.
• CODY_RECORDING_MODE=record|replay|passthrough: when set to record, will record HTTP requests and persist the results in the recording directory. When set to replay, will replay from the recording directory. When set to passthrough, will pass through all HTTP requests to the real Sourcegraph instance without recording the responses.
• (optional) CODY_RECORDING_MODE=TEST_NAME: if you are running multiple instances of the agent to test different features, then you should provide a unique recording name for each test. Each unique name will get a unique directory avoiding the risk of contaminating recordings between different tests.

Run pnpm run agent jsonrpc --help to get more information about all available --recording-* options.

Updating the Polly HTTP Recordings

If agent tests are failing in CI for non-agent related PRs (e.g. PollyError: [Polly] [adapter:node-http] Recording for the following request is not found and recordIfMissing is false errors) then you may need to update the HAR HTTP recordings. For example, this can happen when we make changes to the prompt the agent test to not be able to replay the autocomplete requests from old recordings.

To fix this problem, update the HTTP recordings with the following command:

# tokens are redacted in the recordings
source agent/scripts/export-cody-http-recording-tokens.sh
src login                                     # confirm you are authenticated to sourcegraph.com
pnpm update-agent-recordings                  # run tests to update recordings
# If test fails, press `u` to update the vitest snapshot assertion.
pnpm run test agent/src/index.test.ts         # validate that tests are passing successfully in replay mode

Please post in #wg-cody-agent if you have problems getting the agent tests to pass after recording. Worst case, feel free to disable the agent tests by uncommenting the block of code in index.test.ts. See comment in the code for more details about how to disable agent tests.

Iterating on agent tests

For a fast edit/test/debug feedback loop when iterating on agent tests, use the CODY_RECORD_IF_MISSING=true mode.

CODY_RECORD_IF_MISSING=true pnpm run test agent/src/index.test.ts

The benefit of this workflow is that the agent replays from the HTTP recording for tests that haven't changed, and only sends HTTP requests for new tests.

When you are happy with the result, make sure to run pnpm update-agent-recordings to clean unused recordings.

Miscellaneous notes

• By the nature of using JSON-RPC via stdin/stdout, both the agent server and client run on the same computer and there can only be one client per server. It's normal for both the client and server to be stateful processes. For example, the extensionConfiguration/didChange notification is sent from the client to the server to notify that subsequent requests should use the new connection configuration.
• Run the command git diff -- ':!*.har.yaml' to review local changes without the noisy diff in agent/recordings.
• If you get an ESBuild error about "You can mark the path "#async_hooks" as external to exclude it from the bundle, which will remove this error." then the fix is to remove dependencies on vitest from the agent bundle. Vitest depends on the p-limit npm package, which uses #async_hooks that we currently don't handle in the ESBuild config.