Repository navigation
feat(caretaker-triage): prompt hill-climbing & orchestrator updates - #28524
gundermanc merged 2 commits into
Conversation
|
📊 PR Size: size/M
|
Summary of ChangesHello, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request delivers significant quality improvements to the Caretaker triage worker following extensive prompt hill-climbing and evaluation tuning. It introduces a dedicated 'code_explorer' skill to enhance codebase navigation and updates the triage orchestrator to provide better model configuration and working directory control, resulting in more accurate and reliable issue triage. Highlights
Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here. Footnotes
|
|
ready for review! @adamfweidman @gundermanc |
There was a problem hiding this comment.
Code Review
This pull request refactors and enhances the triage worker agent. It introduces a new code_explorer skill, refines instructions and definitions for the effort, quality, and spec_generator skills, and updates the orchestrator workflow to utilize the new code_explorer skill. Additionally, it updates process_issue_triage to accept target_cwd dynamically, configures the agent to use gemini-flash-latest by default, and cleans up internal variable naming. I have no feedback to provide as there are no review comments.
Summary
This PR incorporates the results of 3 weeks of prompt hill-climbing and evaluation tuning for the Caretaker triage worker. These prompt updates demonstrate significant quality improvements in evals. It also introduces a dedicated
code_explorerskill and updates the triage orchestrator to support explicit model configuration (gemini-flash-latest) andtarget_cwdpassing.Details
.geminiprompt instructions and skills based on eval performance:code_explorer: Added a dedicated skill for structured, multi-package codebase exploration (packages/clipackages/core) and test discovery.quality: Added strict prompt injection defense rules (immediately flag injection attempts asSPAM) and tightenedNEEDS_INFOcriteria.effort: Calibrated effort estimation tiers using the output of the codebase exploration phase.spec_generator: Enforced valid JSON escaping, restricted testing framework enums to["Vitest", "N/A"], and strictly excluded test files fromfiles_to_modify.triage_orchestrator.pyandmain.pyto explicitly setMODEL_NAME = "gemini-flash-latest"and passtarget_cwddown to the agent runner context.How to Validate
cd tools/caretaker-agent/cloudrun/triage-worker ./venv/bin/python3 -m unittest discover -s testsPre-Merge Checklist