| commit | a7cb6766385de0fbe8a6cf38d6e9c69db9594a7c | [log] [tgz] |
|---|---|---|
| author | Matthias Sohn <matthias.sohn@gmail.com> | Sun Nov 03 22:06:42 2024 +0000 |
| committer | Gerrit Code Review <noreply-gerritcodereview@google.com> | Sun Nov 03 22:06:42 2024 +0000 |
| tree | 71c356efd0cf2b87f75041973e218d2af4f04d4c | |
| parent | fa30a85d426a5e220c669eb5abbdf0d30765fb65 [diff] | |
| parent | fc789491fb11984eb2cc1c58357d5b6a198588d7 [diff] |
Merge changes I2c8adb06,I43613aa6,I3dde13c0 * changes: Use jgit @NonNull annotation Add license headers Format source code using google-java-format 1.24.0
This plugin has been based on the initial chatGPT plugin available on github: chatgpt-code-review-gerrit-plugin
A full history of the previous plugin is in the branch: gerrithub-chatgpt-code-review/main.
The requirement of the updated plugin was to abstract the chatGPT specific nature of the plugin and to provide a more generic naming point, and set of configuration and API naming that used the openAPI specification that is common between chatGPT and other private AI service providers.
This plugin allows you to use different AI Chat services, e.g. ChatGPT or OLLAMA for code review in Gerrit conveniently. After submitting a Patch Set, OpenAI will provide review feedback in the form of comments and, optionally, a vote. You can continue to ask the AI Chat by @{gerritUserName} or @{gerritEmailAddress} (provided that gerritEmailAddress is in the form “gerritUserName@<any_email_domain>”) in the comments to further guide it in generating more targeted review comments. Reviews can be also triggered by directing a comment with the /review command which will be interpretted and sent on by the plugin.
Build: Requires JDK 11 or higher, Maven 3.0 or higher.
mvn -U clean package
If the user needs to disable test just run
mvn -U -DskipTests=true clean package
Install: Upload the compiled jar file to the $gerrit_site/plugins directory.
Configure: First, you need to create an AI Code Review user in Gerrit. Then, set up the basic parameters in your $gerrit_site/etc/gerrit.config file under the section
[plugin "ai-code-review"]:
aiToken: Authentication token, for aiType selected (Default: chatGPT bearer token).
gerritUserName: Gerrit username of AI Code Review user.
globalEnable: Default value is false. The plugin will only review specified repositories. If set to true, the plugin will by default review all pull requests.
For enhanced security, consider storing sensitive information like aiToken in a secure location or file. Detailed instructions on how to do this will be provided later in this document.
Verify: After restarting Gerrit, you can see the following information in Gerrit's logs:
INFO com.google.gerrit.server.plugins.PluginLoader : Loaded plugin ai-code-review, version ...
You can also check the status of the ai-code-review on Gerrit's plugin page as Enabled.
If you wish to test a local installation of ‘Ollama’ llm, then you can follow the instructions on the ollama site directly, or the official docker image instructions. Using the docker instructions allows you to start various different models, and you can start and stop the image as much as you need to and the manifest data is associated with the container, so only when it is finally deleted with it remove the volume associated with that image. This allows you to clean up all the associated test images very easily with one command.
docker run -d -v ollama:/root/.ollama -p 11434:11434 --name ollama ollama/ollama
You can then attach to this running ollama service and run any model/manifest you wish.
docker exec -it ollama ollama run qwen2.5-coder
In the following example, a Patch Set receives a score of “-1” indicating a recommendation.
NOTE: Voting is disabled by default. To use this feature, it needs to be activated either across all projects or on a per-project basis via the enabledVoting configuration option, as described below.
In the example below, the AI Chat initially posits a potential unintended behavior in the code, assigning a “-1” score. Upon receiving clarification, it resets the score to “0”.
More examples of ChatGPT's use within code reviews and inline discussions based on the origin ChatGPT version of this plugin are available at https://wiki.amarulasolutions.com/opensource/products/chatgpt-gerrit.html
You have the option to establish global settings, or independently configure specific projects. If you choose independent configuration, the corresponding project settings will override the global parameters.
To configure these parameters, you need to modify your Gerrit configuration file (gerrit.config). The file format is as follows:
[plugin "ai-code-review"]
# Required parameters
aiToken = {aiToken}
...
# Optional parameters
aiType = {aiType}
aiModel = {aiModel}
aiSystemPrompt = {aiSystemPrompt}
...
Note aiType is only optional as by default the system will assume an aiType of chatGPT so as to maintain compatibility with the previous chat-gpt version of the plugin. If you require a different service, update the optional parameters such as aiType.
[plugin "ai-code-review"]
# Required parameters
aiType = OLLAMA
aiModel = qwen2.5-coder
aiDomain = http://localhost:11434/
...
It is highly recommended to store sensitive information such as aiToken in the secure.config file. Please edit the file at $gerrit_site/etc/secure.config and include the following details:
[plugin "ai-code-review"]
aiToken = {aiToken}
If you wish to encrypt the information within the secure.config file, you can refer to: https://gerrit.googlesource.com/plugins/secure-config
To add the following content, please edit the project.config file in refs/meta/config:
[plugin "ai-code-review"]
# Required parameters
gerritUserName = {gerritUserName}
...
# Optional parameters
aiModel = {aiModel}
aiSystemPrompt = {aiSystemPrompt}
...
Please ensure strict control over the access permissions of refs/meta/config since sensitive information such as aiToken is configured in the project.config file within refs/meta/config.
Starting from version 3, the plugin introduces a new Stateful interaction mode with ChatGPT is available alongside the traditional Stateless mode.
In Stateless mode, each request to ChatGPT must include all contextual information and instructions necessary for ChatGPT to provide insights, utilizing the Chat Completion API calls made available by OpenAI.
Conversely, Stateful mode uses the Assistant resource, recently made available in beta by OpenAI, to maintain a richer interaction context. This mode is designed to:
The advantages of the Stateful mode over the Stateless are twofold:
aiMode: Select whether requests are processed in Stateless or Stateful mode. For backward compatibility, the default value is stateless. To enable Stateful mode, set this parameter to stateful.
aiModel: The default model is gpt-4o. You can also configure it to gpt-3.5-turbo or gpt-4-turbo.
aiDomain: The default if not specified is the ChatGPT domain https://api.openai.com. This correlates with the aiType field which defaults to CHATGPT if not specified.
aiSystemPrompt: You can modify the default system prompt (“Act as a PatchSet Reviewer”) to your preferred prompt.
aiReviewTemperature: Specifies the temperature setting for ChatGPT when reviewing a Patch Set, with a default setting of 0.2. Higher values like 0.8 will make the output more random, while lower values like 0.2 will make it more focused and deterministic.
aiCommentTemperature: Specifies the temperature setting for ChatGPT when replying to a comment, with a default setting of 1.0.
aiReviewPatchSet: Set to true by default. When switched to false, it disables the automatic review of Patch Sets as they are created or updated.
aiReviewCommitMessages: The default value is true. When enabled, this option also verifies if the commit message matches with the content of the Change Set.
gptFullFileReview: Enabled by default. Activating this option sends both unchanged lines and changes to ChatGPT for review, offering additional context information. Deactivating it (set to false) results in only the changed lines being submitted for review.
aiStreamOutput: The default value is false. Whether the response is expected in stream output mode or not.
aiType: Allows the selection of different ai chat services the current list is ChatGPT, Ollama, AzureOpenAI and Generic. For compatibility the default is ChatGPT when not specified. Note: Gemeric is used to allow the specification of new uri endpoints, and different authorization headers to allow new service testing more quickly.
aiAuthHeaderName: When using the GENERIC aiType, you can specify any authorization header, which can allow you to test non bearer token authorization e.g. api-key used for AZUREOpenAI.aiChatEndpoint: When using the GENERIC aiType, you can specify any endpoint you wish to test a different ai service, or new API. e.g. /api/someapi. Although it is best to ensure the API is compatible with the openAPI specification or additional response parsing will be needed to correctly read the data into the new API data structures.maxReviewLines: The default value is 1000. This sets a limit on the number of lines of code included in the review.
maxReviewFileSize: Set with a default value of 10000, this parameter establishes a cap on the file size that can be included in reviews.
enabledUsers: By default, every user is enabled to have their Patch Sets and comments reviewed. To limit review capabilities to specific users, list their usernames in this setting, separated by commas.
disabledUsers: Functions oppositely to enabledUsers.
enabledGroups: By default, all groups are permitted to have their Patch Sets and comments reviewed. To restrict review access to certain groups, specify their names in this setting, separating them with commas.
disabledGroups: Operates in reverse to enabledGroups, excluding specified groups from reviews.
enabledTopicFilter: Specifies a list of keywords that trigger ChatGPT reviews based on the topic of the Patch Set. When this setting is active, only Patch Sets and their associated comments containing at least one of these keywords in the topic are reviewed.
disabledTopicFilter: Works in contrast to enabledTopicFilter, excluding Patch Sets and comments from review if their topics contain specified keywords.
enabledFileExtensions: This limits the reviewed files to the given types. Default file extensions are “.py, .java, .js, .ts, .html, .css, .cs, .cpp, .c, .h, .php, .rb, .swift, .kt, .r, .jl, .go, .scala, .pl, .pm, .rs, .dart, .lua, .sh, .vb, .bat”.
enabledVoting: Initially disabled (false). If set to true, allows ChatGPT to cast a vote on each reviewed Patch Set by assigning a score.
filterNegativeComments: Activated by default (true), ensuring only negative review comments (scored below the filterCommentsBelowScore threshold outlined further) are displayed. Disabling this setting (false) will also show positive and neutral comments.
filterCommentsBelowScore: With filterNegativeComments active, review comments with a score at or above this setting's value will not be shown (default is 0).
filterRelevantComments: This setting is enabled by default (true) to display only those review comments considered relevant by ChatGPT, which means they have a relevance index at or above the filterCommentsRelevanceThreshold specified below. Turning off this option (false) allows the display of comments ChatGPT marks as irrelevant.
filterCommentsRelevanceThreshold: When filterRelevantComments is enabled, any review comment assigned a relevance score by ChatGPT below this threshold will not be shown. The default threshold is set at 0.6.
aiRelevanceRules: This option allows customization of the rules ChatGPT uses to determine the relevance of a task.
patchSetCommentsAsResolved: Initially set to false, this option leaves ChatGPT‘s Patch Set comments as unresolved, inviting further discussion. If activated, it marks ChatGPT’s Patch Set comments as resolved.
inlineCommentsAsResolved: Initially set to false, this option leaves ChatGPT‘s inline comments as unresolved, inviting further discussion. If activated, it marks ChatGPT’s inline comments as resolved.
votingMinScore: The lowest possible score that can be given to a Patch Set (Default value: -1).
votingMaxScore: The highest possible score that can be given to a Patch Set (Default value: +1).
ignoreResolvedChatGptComments: Determines if resolved comments from ChatGPT should be disregarded. The default setting is true, which means resolved ChatGPT comments are not used for generating new comments or identifying duplicate content. If set to false, resolved ChatGPT comments are factored into these processes.
ignoreOutdatedInlineComments: Determines if inline comments made on non-latest Patch Sets should be disregarded. By default, this is set to false, meaning all inline comments are used for generating new responses and identifying repetitions. If enabled (true), inline comments from previous Patch Sets are excluded from these considerations.
enableMessageDebugging: This setting controls the activation of debugging functionalities through messages (default value is false). When set to true, it enables commands and options like --debug for users as well as the Dynamic Configuration commands.
forceCreateAssistant: In Stateful mode, forces the creation of a new assistant with each request instead of only when configuration settings change or Changes are merged.
NOTE: This option may increase OpenAI API usage and should be used for testing or debugging purposes only.
globalEnable: Set to false by default, meaning the plugin will review only designated repositories. If enabled, the plugin will automatically review all pull requests by default (not recommended in production environments).enabledProjects: The default value is an empty string. If globalEnable is set to false, the plugin will only run in the repositories specified here. The value should be a comma-separated list of repository names, for example: “project1,project2,project3”.isEnabled: The default is false. If set to true, the plugin will review the Patch Set of this project.Reviewing a Change Set or the last Patch Set can occur automatically upon submission or be manually triggered using the commands outlined in this section.
/review: when used in a comment directed at ChatGPT on any Change Set, triggers a review of the full Change Set. A vote is cast on the Change Set if the voting feature is enabled and the ChatGPT Gerrit user is authorized to vote on it./review_last: when used in a comment directed at ChatGPT on any Change Set, triggers a review of the last Patch Set of the Change Set. Unlike /review, this command does not result in casting or updating votes.--filter=[true/false]: Controls the filtering of duplicate, conflicting and irrelevant comments, defaulting to “true” to apply filters.
--debug: When paired with /review or /review_last commands, this option displays useful debug information in each ChatGPT reply, showing all replies as though the filter setting were disabled.
NOTE: The usage of --debug option is disabled by default. To enable it, enableMessageDebugging setting must be set to true.
Directives are mandatory instructions written in plain English that ChatGPT must adhere to during its reviews. They can be specified using the following command.
/directive <DIRECTIVE_CONTENT>: This command, when included in a comment with a subsequent directive description “<DIRECTIVE_CONTENT>”, specifies a directive that ChatGPT must adhere to.
You can now dynamically alter the plugin configuration via messages sent to the ChatGPT user, primarily for testing and debugging purposes. This feature becomes available when the enableMessageDebugging configuration setting is enabled.
/configure displays the current settings and their dynamically modified values in a response message.
/configure --<CONFIG_KEY_1>=<CONFIG_VALUE_1> [... --<CONFIG_KEY_N>=<CONFIG_VALUE_N>] assigns new values to one or more configuration keys.
NOTE: Values that include spaces, such as aiSystemPrompt, must be enclosed in double quotes.
The reset option can be employed to restore modified settings to their original defaults. Its usage is detailed below:
/configure --reset restores all modified settings to their default values./configure --reset --<CONFIG_KEY_1> [... --<CONFIG_KEY_N>] specifically restores the indicated key(s) to their default values.Apache License 2.0