Conversation
Extract review logic from SKILL.md into separate review-guidelines.md files. Skills now act as thin orchestrators that spawn sub-agents with clean context (no prior conversation state), ensuring unbiased reviews. - local-code-review and review-architecture accept a scope parameter - Both skills require explicit user approval before fixing findings - Evals updated to assert sub-agent delegation pattern - Smoke tests verify review-guidelines.md file existence - CLAUDE.md workflow merges review steps and notes parallel execution Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Both test_build and test_index called duckdb.execute() on the shared default in-memory connection instead of the file-scoped connection, causing "table t1 already exists" when both tests ran in the same suite. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace manual sub-agent orchestration with Claude Code's native context:fork + agent frontmatter fields. Skills now run directly in a forked sub-agent context, eliminating the need for separate review-guidelines.md files and orchestrator logic. - Add argument-hint, context:fork, agent:general-purpose to frontmatter - Use $ARGUMENTS for scope parameter substitution - Merge review guidelines back into SKILL.md (it IS the sub-agent now) - Remove review-guidelines.md files (no longer needed) - Move "ask before fixing" guardrail to CLAUDE.md workflow - Update evals to test review quality directly Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduce a general pacing guideline requiring the agent to pause for user confirmation between significant workflow phases (plan, implement, validate, commit, PR). Add explicit Plan step to the workflow. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Accept skill names as arguments (e.g. /eval-skills local-code-review). When no arguments provided, ask the user which skills to evaluate. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
catstrike
reviewed
Mar 24, 2026
| before or after significant changes. | ||
| You are reviewing the architecture of the Databao CLI project. | ||
| You have NO prior context about why these changes were made — review | ||
| purely on merit. Use your tools (Read, Grep, Glob, Bash) to inspect |
Collaborator
There was a problem hiding this comment.
Do you think it's necessary to specify a list of tools? I'm sure Claude Code will use them anyway. If you want to limit the tools list, use the allowed-tools header parameter.
Contributor
Author
There was a problem hiding this comment.
good point regarding tools - I'll better add it to the header indeed
Contributor
Author
There was a problem hiding this comment.
so I've dropped these mentions, and added a separate agent specification for that at .claude/agents/, with read-only tools available only.
ptal
Replace agent: general-purpose with agent: reviewer in both review skills. The reviewer agent restricts tools to Read, Glob, Grep, Bash, structurally preventing file edits. Remove redundant "Do NOT edit" instructions and replace verbose rewrite guardrail with concise snippet guidance. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
catstrike
requested changes
Mar 24, 2026
catstrike
approved these changes
Mar 25, 2026
gasparian
deleted the
andrei/DBA-290-refactor-review-skills-to-subagents
branch
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Refactor
local-code-reviewandreview-architectureskills to run in isolated sub-agent context using Claude Code's nativecontext: forkmechanism. Reviews execute with zero prior conversation state, ensuring unbiased findings. Also adds a pacing rule for user confirmation between workflow phases and improveseval-skillsto support multi-skill evaluation.Changes
Use native
context: forkfor review skill isolationargument-hint,context: fork,agent: general-purposefrontmatter$ARGUMENTSfor scope parameter (staged,branch,files:<path>,module:<path>,full)Files
.claude/skills/local-code-review/SKILL.md.claude/skills/review-architecture/SKILL.mdAdd pacing rule for user confirmation between phases
Files
CLAUDE.mdImprove eval-skills to support multi-skill evaluation
/eval-skills local-code-review review-architecture)argument-hintto frontmatterFiles
.claude/skills/eval-skills/SKILL.mdUpdate evals and smoke tests
review-guidelines.mdsmoke test checks (files removed)Files
.claude/skills/local-code-review/evals/evals.json.claude/skills/review-architecture/evals/evals.jsonscripts/smoke-test-skills.shFix flaky duckdb test
test_buildandtest_indexused shared default duckdb connection causing table collisionFiles
tests/test_build.pytests/test_index.pyTest Plan
make checkpassesmake testpasses (66/66, flaky test fixed)make lint-skillspasses/local-code-reviewruns in forked sub-agent contexteval-skillsiteration 1 completed — review-architecture with-skill 100% vs without-skill 85%🤖 Generated with Claude Code