Genericize run-evals prompts and enforce lint cheating rubric - #196
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request updates the skill evaluation framework to support dynamic target package roots (<target-package-root>) and adds a new lint_cheating rubric rule to prevent subagents from bypassing lint configurations or adding file-level ignores. Feedback suggests extending the lint_cheating rule to also forbid line-level ignores, and clarifying that the <target-package-root> placeholder should be replaced in all prompt templates, including the baseline execution prompt.
| 3. **Determine Agent Configuration**: Check the `agent_config` field in the target target JSON file to determine the environment/harness to spawn. If `agent_config` is `"bare-agent"`, spawn a subagent with the `bare-agent` profile. If it is a specific contributor profile (e.g. `"reidbaker-agent"`), use that profile to provide the necessary contributor context. | ||
| 4. **Orchestrate**: By default, run an Integration Test by spawning a single **With-Skill** subagent using `Workspace: branch` and the identified `agent_config`. | ||
| - Provide the task prompt. See `resources/with_skill_execution_prompt.md` for the template. When filling in `<path-to-skill>`, you MUST use a relative path from the repository root, not an absolute path. If you are running a cross-skill evaluation, fill in `<path-to-skill>` with `"none (cross-skill meta-eval)"`. | ||
| - Provide the task prompt. See `resources/with_skill_execution_prompt.md` for the template. When filling in `<path-to-skill>`, you MUST use a relative path from the repository root, not an absolute path. If you are running a cross-skill evaluation, fill in `<path-to-skill>` with `"none (cross-skill meta-eval)"`. Also, replace `<target-package-root>` with the actual directory path. |
There was a problem hiding this comment.
The prompt template resources/baseline_execution_prompt.md also contains the <target-package-root> placeholder. It would be clearer to instruct the agent to replace this placeholder in any of the prompt templates used, rather than just mentioning with_skill_execution_prompt.md.
…s evals.json and clarify prompt template placeholder instructions
| @@ -1,6 +1,5 @@ | |||
| { | |||
| "repo_criteria": [ | |||
| "evals/code_quality_rubric.json" | |||
There was a problem hiding this comment.
I removed it because of the feedback from the code review bot on the upstream flutter/packages PR (#12285). The bot flagged that when downstream consumers (like camera_android_camerax) adopt the run-evals skill, they typically don't have the evals/code_quality_rubric.json file in their repository root, which causes the eval runner to crash with a file not found error.
Additionally, according to evals/README.md (## Cross-Cutting Rules), only "Skills that author or modify code MUST adhere to the universal code quality expectations". Since run-evals is an orchestrator skill that only spawns subagents and doesn't write Dart code itself, it seemed safe to remove it from this specific skill's criteria.
Should I put it back and find a different way to resolve the downstream missing file issue?
There was a problem hiding this comment.
That works for now. My gut says that we will end up moving both the code_quality_rubric.json and its evals into the run-evals skill so the rubric can be shared with consumers.
2 participants
This came up as part of feedback in flutter/packages#12285. My goal was to reduce the duplication of llm prompts to the minimum while still enabling evals to run in the package repo.
Like in dart skills lint there is no automated way to run these tests. They run as a skill inside whatever harness you are using. The evals authored here are all the internal version of agy running against gemini 3.1 pro.
Agent authored description
Addresses PR feedback to make run-evals generic and prevents lint cheating by subagents.
Evaluation Results
Code Quality Rubric Meta-Evals Results
Eval 1 (
bad_script.dart): PASSEval 2 (
good_script.dart): PASSWorkspace: branchenvironment by explicitly addingdart pub getto the eval execution prompts (commitb8202c9), the evaluator correctly recognized that all 5 code quality criteria (compilation_and_health,effective_dart_and_idioms,cross_platform_compatibility,directory_and_placement_hygiene,lint_cheating) passed successfully without any failures.Skill Evaluator (
run-evalsMeta-Evals) ResultsOverall Grade: PASS
run-evalsskill passed successfully. Independent subagents were spawned for each scenario. They accurately navigated the skill instructions (branching to evaluate thedefinition-of-doneskill as instructed) and correctly aborted when warned about workspace limitations (Eval 4).Code Review Findings
The code review subagent flagged that
evals/README.mdand*_evals.jsonare internal totool/dart_skills_lintand shouldn't have been genericized. However, this feedback was dismissed because downstream packages likecamera_android_cameraxdo maintain their own copies of these files when they adopt therun-evalsskill infrastructure.