notque/vexjoy-agent / skills/process/workflow
workflow
VerifiedStructured work: multi-phase tasks, feature builds, planning, objective loops, hill climbing.
references/comprehensive-review/references/wave-2-deep-dive.md
View on GitHub# Wave 2: Deep-Dive Agents (10 agents)
These agents receive Wave 0+1 aggregated findings as input. They perform targeted deep-dives informed by per-package analysis (Wave 0) and cross-cutting analysis (Wave 1).
**ALL 10 agents MUST be dispatched in ONE message for true parallel execution.**
**Dispatch.** Each roster row is one fresh `general-purpose` worker (Agent tool `subagent_type: "general-purpose"`), so every lens reviews independently with no shared context. The reference column names the review reference that worker loads. Start each worker prompt with: "Call the Skill tool with `review`. Then read ~/.claude/skills/review/references/<reference> and follow it; load every file its Reference Loading Table matches, and apply only the <lens> lens."
Review agents run on `opus` (the `work` role in `workflow-helpers.js`); pass the owner's model instead when the owner names one. Never use haiku or sonnet for code review.
## Agent Roster
| # | Reference (worker loads) | Focus Area | Wave 1 Context Used |
|---|-------|------------|---------------------|
| 11 | `reviewer-code.md` (performance lens) | Performance | Architecture findings → focus on hot paths |
| 12 | `reviewer-system.md` (concurrency lens) | Concurrency | Silent-failure + architecture findings → concurrent paths |
| 13 | `reviewer-system.md` (api-contract lens) | API Contracts | Business-logic + type-design findings → contract-sensitive code |
| 14 | `reviewer-system.md` (dependency-audit lens) | Dependencies | Docs-validator findings → dependency documentation gaps |
| 15 | `reviewer-system.md` (error-messages lens) | Error Messages | Silent-failure + code-quality findings → error paths |
| 16 | `reviewer-code.md` (dead-code lens) | Dead Code | Code-quality + docs-validator findings → abandoned artifacts |
| 17 | `reviewer-code.md` (naming-consistency lens) | Naming | Code-quality + language-specialist findings → convention baselines |
| 18 | `reviewer-system.md` (observability lens) | Observability | Silent-failure findings → observability gaps at failure points |
| 19 | `reviewer-code.md` (config-safety lens) | Config Safety | Security + docs-validator findings → config security gaps |
| 20 | `reviewer-system.md` (migration-safety lens) | Migration Safety | API-contract + business-logic findings → migration-sensitive changes |
## Standard Agent Prompt Template
Each Wave 2 agent prompt includes the standard review scope PLUS the Wave 0+1 context:
```
REVIEW SCOPE:
- Files to review: [list of changed files]
- Change context: [what was changed and why, if known]
- Repository: [current directory]
WAVE 0+1 CONTEXT (use this to focus your analysis):
[Insert $WAVE01_SUMMARY — loaded from $REVIEW_DIR/wave01-summary.md]
MCP TOOL DISCOVERY (do this FIRST, before any file reads):
- Use ToolSearch to check for available MCP tools that can enhance your analysis:
a. Run ToolSearch("gopls") — if Go files are in scope, this loads type-aware
analysis tools (go_file_context, go_diagnostics, go_symbol_references, etc.)
b. Run ToolSearch("context7") — loads library documentation lookup tools for
verifying dependency usage and API correctness
- If gopls tools are available AND this is a Go repository:
* Use go_file_context after reading any .go file to understand intra-package dependencies
* Use go_symbol_references before flagging unused or misused symbols
* Use go_diagnostics on files you flag to confirm real vs false-positive issues
- If Context7 tools are available:
* Use resolve-library-id + query-docs to verify library API usage in flagged code
INSTRUCTIONS:
1. Read the CLAUDE.md file(s) in this repository first
2. Run MCP TOOL DISCOVERY steps above
3. Review the Wave 0 per-package context for package-level issues already found
4. Review the Wave 1 cross-cutting context for foundation issues already found
5. Use Wave 0+1 findings to FOCUS your deep-dive analysis:
- Prioritize packages flagged as NEEDS_ATTENTION or CRITICAL by Wave 0
- Prioritize files and paths flagged by Wave 1
- Look for issues in YOUR domain that neither Wave 0 nor Wave 1 would catch
- Cross-reference your findings with both waves to add depth
6. Do NOT simply repeat Wave 0 or Wave 1 findings — add NEW insights
7. Use MCP tools (gopls, Context7) during analysis where they add precision
8. Use structured output format with severity classification
9. Include file:line references for every finding
OUTPUT FORMAT:
Return findings as:
### [CRITICAL|HIGH|MEDIUM|LOW]: [One-line summary]
**File**: `path/to/file:LINE`
**Issue**: [Description]
**Impact**: [Why this matters]
**Fix**: [Concrete code fix]
**Wave 0+1 Cross-Ref**: [Which earlier finding this relates to, if any]
---
```
## Agent-Specific Context Instructions
| Reference (lens) | Extra Context Instructions |
|-------|--------------------------|
| `reviewer-code.md` (performance) | Use Wave 0 per-package findings to identify packages with complexity issues. Use Wave 1 architecture findings to identify hot paths. Focus on algorithmic complexity, N+1 queries, allocation waste. **MCP**: For Go, use gopls `go_symbol_references` to trace hot path call chains |
| `reviewer-system.md` (concurrency) | Use Wave 0 per-package findings for concurrent patterns within packages. Use Wave 1 silent-failure + architecture findings for cross-package concurrent paths. Focus on races, goroutine leaks, deadlocks. **MCP**: For Go, use gopls `go_diagnostics` to detect race condition warnings |
| `reviewer-system.md` (api-contract) | Use Wave 0 per-package findings to understand internal API surfaces. Use Wave 1 business-logic + type-design findings for contract-sensitive endpoints. Focus on breaking changes, status codes. **MCP**: Use Context7 to verify API contract claims against library docs |
| `reviewer-system.md` (dependency-audit) | Use Wave 1 docs-validator findings to cross-reference documented vs actual dependencies. Run govulncheck/npm audit/pip-audit. Focus on CVEs, licenses, deprecated packages. **MCP**: Use Context7 `resolve-library-id` + `query-docs` to verify dependency API usage. For Go, use gopls `go_vulncheck` for vulnerability scanning |
| `reviewer-system.md` (error-messages) | Use Wave 0 per-package error handling patterns. Use Wave 1 silent-failure + code-quality findings. Focus on error message quality, actionability, consistency. |
| `reviewer-code.md` (dead-code) | Use Wave 0 per-package findings to identify unused internal APIs between files. Use Wave 1 code-quality + docs-validator findings. Focus on unreachable code, unused exports. |
| `reviewer-code.md` (naming-consistency) | Use Wave 0 per-package naming patterns to detect intra-package drift. Use Wave 1 code-quality + language-specialist findings. Focus on cross-package naming consistency. |
| `reviewer-system.md` (observability) | Use Wave 0 per-package findings for packages missing instrumentation. Use Wave 1 silent-failure findings for error paths missing observability. Focus on RED metrics gaps. |
| `reviewer-code.md` (config-safety) | Use Wave 0 per-package findings for hardcoded values within packages. Use Wave 1 security + docs-validator findings. Focus on secrets, missing env var validation. |
| `reviewer-system.md` (migration-safety) | Use Wave 1 api-contract + business-logic findings to identify migration-sensitive changes. Focus on reversible migrations, deprecation paths, rollback safety. |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |