tkmind/tkmind_go
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

fix: make goose reviewer less bad (#6154)

Browse Source
This commit is contained in:
tlongwell-block
2025-12-17 12:58:54 -05:00
committed by GitHub
parent c1b980f691
commit 9569c1fb0a
1 changed files with 52 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/goose-pr-reviewer.yml
+52 -29
View File
@@ -14,7 +14,7 @@
#
# Optional Variables:
# - GOOSE_PROVIDER: LLM provider (default: anthropic)
# - GOOSE_MODEL: Model name (default: claude-sonnet-4-5)
# - GOOSE_MODEL: Model name (default: claude-opus-4-5)
#
# Security:
# - PR content could prompt-inject the agent; only trigger on PRs you trust.
@@ -42,18 +42,24 @@ env:
You are a code reviewer. Your job is to evaluate code, not implement changes.
Principles:
- Understand before you critique - explain the author's intent before finding fault
- Be constructive and specific in feedback
- Reference exact files and line numbers (format: path/file.rs:42)
- Categorize issues by severity
- Verify claims with code evidence before stating them
- Respect project conventions (AGENTS.md)
- Never modify code - this is a read-only review
- Your context degrades. The TODO is your memory. Update it after each step.
Issue Categories:
- 🔴 BLOCKING: Must fix before merge (bugs, security, breaking changes)
- 🟡 WARNING: Should fix (performance, conventions, missing tests)
- 🟢 SUGGESTION: Nice to have (style, refactoring opportunities)
- ✅ HIGHLIGHT: Good practices to acknowledge
Issue Categories & Confidence Requirements:
- 🔴 BLOCKING: Must fix before merge. REQUIRES HIGH confidence with code evidence.
- 🟡 WARNING: Should fix (performance, conventions, missing tests). MEDIUM+ confidence.
- 🟢 SUGGESTION: Nice to have (style, refactoring). Can be speculative but label it.
- ✅ HIGHLIGHT: Good practices to acknowledge.
Anti-hallucination rules:
- Before claiming something is "missing", search for it with rg
- Before claiming UI/frontend changes are needed, trace the actual data flow
- If you cannot verify a claim, say "I couldn't verify this" not "this is wrong"
- 2 verified issues are better than 10 speculative ones
prompt: |
Review PR #${PR_NUMBER}: ${PR_TITLE}
@@ -67,47 +73,64 @@ env:
Reviewer instructions from trigger:
${REVIEW_INSTRUCTIONS}
Write this to your TODO immediately and update as you progress:
FIRST ACTION: Call todo_write with this entire checklist. Your memory degrades - the TODO is your only reliable memory. Update it frequently.
## PR Understanding
- Intent: [fill after Phase 1]
- Approach: [fill after Phase 1]
- Author's reasoning: [fill after Phase 2 - why this approach vs alternatives]
## Phase 1: Understand
- [ ] Read /tmp/pr.json for PR description and context
- [ ] Read /tmp/pr.diff for the actual changes
- [ ] Read AGENTS.md if it exists
- [ ] Note the reviewer instructions above - prioritize any specific requests
- [ ] Identify the intent and scope of the PR
- [ ] Note reviewer instructions: ${REVIEW_INSTRUCTIONS}
- [ ] Update TODO with intent and approach above
## Phase 2: Analyze Code
- [ ] Explore the codebase to understand context (use analyze, rg, read files)
## Phase 2: Steel Man the Implementation
- [ ] Trace the data/control flow through the changes
- [ ] Why did the author choose this approach?
- [ ] What alternatives exist and why might they be worse?
- [ ] Update TODO with author's reasoning above
## Phase 3: Analyze Code (only after Phase 2 complete)
- [ ] Explore codebase for context (use analyze, rg, read files)
- [ ] Review for correctness and logic errors
- [ ] Check for security vulnerabilities
- [ ] Assess error handling
- [ ] Look for performance issues
- [ ] Verify code follows project patterns
- [ ] For each issue, add to Issues Found below with confidence level
Note: Do NOT run cargo check, cargo test, cargo fmt, or other build commands.
CI pipelines already validate those automatically. Focus on code review only.
## Phase 3: Write Review
- [ ] Write detailed review to /tmp/pr_review.md
- [ ] Include specific file:line references
- [ ] Categorize all issues by severity
- [ ] Provide actionable suggestions
## Phase 4: Write Review
- [ ] Verify all 🔴 BLOCKING issues have HIGH confidence + code evidence
- [ ] Write review to /tmp/pr_review.md
- [ ] Be concise - quality over quantity
## Issues Found
[Add issues here as you find them, format: [HIGH/MED/LOW] [file:line] description - evidence]
---
Format your review in /tmp/pr_review.md as:
## Summary
[1-2 sentence overview of the PR and your assessment]
**Summary**: [1-2 sentence overview of what this PR does and your assessment]
## 🔴 Blocking Issues
[Issues that must be fixed before merge, or "None found"]
Only include sections below if they have content. Omit empty sections entirely.
## 🟡 Warnings
[Issues that should be addressed, or "None found"]
**🔴 Blocking Issues**
[Issues that must be fixed before merge. Each must have file:line and evidence.]
## 🟢 Suggestions
[Optional improvements, or "None"]
**🟡 Warnings**
[Issues that should be addressed]
## ✅ Highlights
**🟢 Suggestions**
[Optional improvements]
**✅ Highlights**
[Good practices observed in this PR]
---
@@ -139,7 +162,7 @@ jobs:
options: --user root
env:
GOOSE_PROVIDER: ${{ vars.GOOSE_PROVIDER || 'anthropic' }}
GOOSE_MODEL: ${{ vars.GOOSE_MODEL || 'claude-sonnet-4-5' }}
GOOSE_MODEL: ${{ vars.GOOSE_MODEL || 'claude-opus-4-5' }}
ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
HOME: /tmp/goose-home
@@ -161,7 +184,7 @@ jobs:
- name: Install tools
run: |
apt-get update
apt-get install -y gettext curl
apt-get install -y gettext curl ripgrep
curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg | dd of=/usr/share/keyrings/githubcli-archive-keyring.gpg
echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" | tee /etc/apt/sources.list.d/github-cli.list > /dev/null
apt-get update