Summary
Add a reusable workflow (a sibling to this repo's claude.yml) that runs a small, self-hosted-model agent against a pull request --- the CI-side runnable counterpart to Morrison-Lab/wai#39 and its companion chapter section, "Small, Local Models for Autonomous Agentic Coding."
Division of labor (from wai#39): wai explains, ai-config configures, gha runs. This issue is the gha half; Morrison-Lab/ai-config#1292 is the sibling ai-config half (the skill that configures and launches the loop locally).
What the workflow should do
Run a self-hosted (or locally-reachable) small model against a PR's diff, using this repo's own composite actions as the loop's verification gates rather than a bespoke test harness:
lint-qmd / lint-markdown for Quarto/Markdown contentcheck-links for link validitycheck-phi for sensitive-data leakagetest-coverage (R package tests) where applicable
Each gate should behave the way wai#39's "Guardrails for autonomy" section describes: a non-zero result is a hard stop for that step, not advisory, and the loop should report rather than silently continue past a failed gate.
Open questions to resolve while scoping this
- How the workflow reaches a self-hosted model (a self-hosted runner with a local Ollama endpoint vs. a remote endpoint reachable from a GitHub-hosted runner) --- this has real security and secrets implications and should be resolved explicitly, not assumed.
- Whether this composes with the existing
claude.yml / claude-code-review.yml pattern (dispatch, concurrency group, permissions) or needs its own trigger shape given a self-hosted runner's different lifecycle.
Non-goals for this issue
- Model selection and guardrail rationale --- that's wai#39, link back to it rather than duplicating.
- The local/interactive launcher --- that's
ai-config#1292.
References
Summary
Add a reusable workflow (a sibling to this repo's
claude.yml) that runs a small, self-hosted-model agent against a pull request --- the CI-side runnable counterpart toMorrison-Lab/wai#39and its companion chapter section, "Small, Local Models for Autonomous Agentic Coding."Division of labor (from wai#39):
waiexplains,ai-configconfigures,gharuns. This issue is theghahalf;Morrison-Lab/ai-config#1292is the siblingai-confighalf (the skill that configures and launches the loop locally).What the workflow should do
Run a self-hosted (or locally-reachable) small model against a PR's diff, using this repo's own composite actions as the loop's verification gates rather than a bespoke test harness:
lint-qmd/lint-markdownfor Quarto/Markdown contentcheck-linksfor link validitycheck-phifor sensitive-data leakagetest-coverage(R package tests) where applicableEach gate should behave the way wai#39's "Guardrails for autonomy" section describes: a non-zero result is a hard stop for that step, not advisory, and the loop should report rather than silently continue past a failed gate.
Open questions to resolve while scoping this
claude.yml/claude-code-review.ymlpattern (dispatch, concurrency group, permissions) or needs its own trigger shape given a self-hosted runner's different lifecycle.Non-goals for this issue
ai-config#1292.References