feat: install agent-ci for local GitHub Actions validation (issue #20)#21
feat: install agent-ci for local GitHub Actions validation (issue #20)#21peterp wants to merge 13 commits into
Conversation
|
kindling — task complete Phases (4/4)
Completion reportStatus: Done What happenedIssue 20 was resolved across two developer passes and two reviewer passes. The agent-ci tool was installed as a project development dependency, the agent skill was registered with symlinked integration for Claude Code, and agent instructions were added requiring local validation before declaring work complete. A revise cycle was triggered when the Reviewer found that command examples in the skill definition were using a registry-fetching form instead of invoking the locally installed binary — a violation of the project's pnpm-only convention. The Developer corrected all three affected commands, and the Reviewer passed the revision. PR 21 is open and ready for review. Decisions made
Worth noting
Last updated: 2026-04-16T11:29:45.067Z |
Installs @redwoodjs/agent-ci@^0.10.7 as a devDependency via pnpm. Adds .gitignore to exclude node_modules/.
Runs `pnpm dlx skills add redwoodjs/agent-ci --skill agent-ci --yes`. Creates .agents/skills/agent-ci/SKILL.md (universal skill definition), .claude/skills/agent-ci symlink (Claude Code integration), and skills-lock.json locking the skill at commit eb25909.
Instructs agents to run agent-ci before declaring work done. References issue #9 for the future --workflow flag once the GitHub Actions deploy workflow is in place.
The project mandates pnpm exclusively. Since @redwoodjs/agent-ci is already installed as a devDependency, pnpm exec agent-ci uses the locally installed binary and respects the locked version — correct over npx (which would fetch from the registry) or pnpm dlx (same).
|
Addressed review feedback:
All three |
Problem
The project lacked a continuous integration skill for Claude agents, meaning there was no standardized, reusable way to run quality gates — such as type checks, linting, and tests — within an automated agent workflow. Without this capability, agents completing implementation work had no reliable mechanism to verify their changes before finalizing a pull request. This gap created risk of undetected regressions and inconsistent code quality across agent-driven contributions.
Solution
A new agent CI skill was introduced and registered, providing a well-defined interface for running the project's quality gate steps from within an agent session. The necessary package dependencies were added and locked to ensure the skill operates in a reproducible environment. Supporting configuration and ignore rules were updated to keep the repository tidy, and an active work log was created to track the progress of this change through the standard investigation-to-finalization workflow.