notque/vexjoy-agent / skills/process/workflow
workflow
VerifiedStructured work: multi-phase tasks, feature builds, planning, objective loops, hill climbing.
references/explore-pipeline.md
View on GitHub---
description: "Systematic codebase exploration with parallel subagents."
user-invocable: false
allowed-tools:
- Read
- Write
- Bash
- Grep
- Glob
- Task
context: fork
command: /explore
routing:
triggers:
- understand codebase
- explore repo
- how does this work
- codebase exploration
- understand this code
- map architecture
- analyze quality
- assess consistency
- evaluate patterns
- "explore codebase"
- "understand this repo"
- "map the codebase"
- "what is in this repo"
pairs_with:
- codebase-overview
- writing
complexity: medium
category: meta
---
# Exploration Pipeline
## Overview
This skill performs systematic codebase exploration using parallel subagents and tiered depth selection. It is read-only (never modifies files) and saves structured artifacts at every phase. Depth is determined by the query type: **Quick** (single question, Phase 1 only), **Standard** (subsystem understanding, 4 phases), or **Deep** (full quality assessment with recommendations, 8 phases).
The pipeline implements three core constraints:
1. **Scope discipline**: Answer the question asked, stay focused on the target subsystem and deliver only requested recommendations
2. **Artifact-first**: Save findings to files at each phase; context is ephemeral
3. **Gate enforcement**: Do not skip phases within the selected tier. Each phase has defined exit criteria and cannot be omitted
Optional behaviors are disabled by default: use `--deep` for comprehensive analysis, `--quick` for overview, `--focus [area]` for targeted exploration, or `--tier quick|standard|deep` for explicit depth selection.
## Instructions
### Phase 1: SCAN (Parallel Subagents)
**Goal**: Gather broad structural understanding through parallel investigation.
Launch 3 scanners in parallel using Task:
**Structure Scanner:**
- List all top-level directories and their purposes
- Identify language(s) and frameworks used
- Find configuration files (build, CI, linting, env)
- Locate test directories and test infrastructure
- Identify build, deploy, and infrastructure files
**Entry Point Scanner:**
- Find main executables (main.py, main.go, index.ts, etc.)
- Identify CLI entry points and argument parsers
- Locate API route definitions and server startup
- Find worker, background job, or queue entry points
- Return list with file paths and brief descriptions
**Pattern Detector:**
- Detect naming conventions (files, functions, variables)
- Identify directory organization patterns
- Catalog import and dependency patterns
- Analyze error handling approach
- Identify testing patterns and coverage strategy
**Gate**: All 3 scanners have returned results. If a scanner times out after 5 minutes, proceed with available data. Minimum 2 of 3 scanners MUST complete. Proceed only when gate passes.
### Phase 2: MAP
**Goal**: Create an architecture map from scan findings.
**Step 1: Identify layers**
Determine the architectural layers present (presentation, application, domain, infrastructure, or whatever pattern the codebase uses).
**Step 2: Map component relationships**
For each major component discovered in Phase 1:
- What does it depend on?
- What depends on it?
- How does data flow between components?
**Step 3: Save architecture artifact**
Save `architecture-map.md` with layer diagram, component relationships, and data flow.
**Gate**: Architecture map saved as artifact with component relationships documented. Proceed only when gate passes.
### Phase 3: ANALYZE
**Goal**: Deep investigation of key areas identified in Phases 1-2.
**Step 1: Analyze core abstractions**
For each key component identified in the architecture map:
1. What is its single responsibility?
2. What patterns does it implement?
3. How is it configured?
4. How is it tested?
**Step 2: Trace critical paths**
Follow the most important execution paths end-to-end (e.g., request lifecycle, data pipeline, event processing).
**Step 3: Identify conventions**
Document the implicit rules: how errors propagate, how state is managed, how components communicate.
**Gate**: Key components analyzed with evidence from actual code. Critical paths traced. Proceed only when gate passes.
### Phase 4: COMPILE (Deep tier only)
**Goal**: Structure raw findings from Phases 1-3 into a structured corpus for evaluation.
**Step 1: Categorize findings**
Group all Phase 1-3 findings by evaluation dimension:
- **Consistency**: Same pattern used everywhere, or mixed approaches?
- **Quality**: Meets best practices? Error handling complete? Tests adequate?
- **Coverage**: Any gaps? Components without tests, docs, or error handling?
- **Patterns**: What patterns recur? What's the dominant approach?
**Step 2: Build comparison matrix**
Create a structured table comparing components against each dimension. Save as `analysis-compilation.md`.
**Gate**: Findings structured into evaluation dimensions with per-component data. Compilation artifact saved.
### Phase 5: ASSESS (Deep tier only)
**Goal**: Evaluate the compiled findings against quality criteria.
**Step 1: Score each dimension**
For each evaluation dimension from Phase 4:
- Rate compliance as a percentage (X/Y components follow the pattern)
- Identify outliers (components that deviate from the norm)
- Assess severity of deviations (cosmetic vs functional vs safety)
**Step 2: Identify root causes**
For each deviation found:
- Is it intentional (documented exception) or accidental (drift)?
- Is it a one-off or a recurring pattern?
- What's the fix complexity (trivial, moderate, significant)?
**Gate**: All dimensions scored with evidence. Deviations cataloged with root causes.
### Phase 6: SYNTHESIZE (Deep tier only)
**Goal**: Combine assessment findings into actionable recommendations.
**Step 1: Rank findings by impact**
| Severity | Criteria |
|----------|----------|
| Critical | Safety issue, data loss risk, session-blocking |
| High | Functional gap, inconsistency that causes confusion |
| Medium | Quality gap, missing best practice |
| Low | Cosmetic, style preference |
**Step 2: Group by theme**
Cluster related findings into themes (e.g., "error handling consistency", "test coverage gaps", "naming drift").
**Gate**: Findings ranked by severity, grouped by theme.
### Phase 7: REFINE (Deep tier only)
**Goal**: Verify findings against actual code before reporting.
**Step 1: Spot-check top findings**
For the top 5 findings by severity, re-read the actual source code to confirm the finding is accurate. Adjust if the initial analysis was wrong.
**Step 2: Remove false positives**
If spot-checking reveals a finding was incorrect (e.g., the pattern is actually intentional), remove it from the report.
**Gate**: Top findings verified against source. False positives removed.
### Phase 8: REPORT (final phase for all tiers)
**Goal**: Produce a structured, reusable exploration report.
**Standard tier**: This is effectively Phase 4 (Phases 4-7 were skipped). Report covers architecture and patterns.
**Deep tier**: This is Phase 8. Report includes everything from Standard PLUS quality assessment, consistency scores, and ranked recommendations from Phases 4-7.
Generate and save `exploration-report.md` following this structure:
```markdown
# Codebase Exploration Report
## Executive Summary
[2-3 sentence overview of what this codebase does and how]
## Quick Facts
- Language: [primary language(s)]
- Framework: [main framework(s)]
- Architecture: [pattern name, e.g., layered, hexagonal, microservices]
- Test Strategy: [how tests are organized]
## Architecture Overview
[Layer diagram and component relationships from Phase 2]
## Key Components
### [Component Name]
- Purpose: [single responsibility]
- Location: [file path(s)]
- Dependencies: [what it uses]
- Dependents: [what uses it]
## Patterns and Conventions
- [Pattern]: [example with file reference]
## Critical Paths
- [Path name]: [entry point] -> [intermediate steps] -> [result]
## Next Steps for Understanding
1. Read [file] for [reason]
2. Trace [flow] to understand [concept]
```
**Gate**: Report saved to file with all sections populated from evidence gathered in prior phases. Report references actual file paths. Proceed only when gate passes.
---
## Error Handling
### Error: "Repository Too Large for Parallel Scan"
Cause: Monorepo or very large codebase overwhelms scanners
Solution:
1. Constrain scan to top 2 directory levels
2. Focus scanners on specific subdirectories if `--focus` provided
3. Use Glob patterns to sample representative files rather than exhaustive listing
### Error: "Scanner Subagent Timed Out"
Cause: Subagent stuck on large directory traversal or slow file reads
Solution:
1. Proceed with results from completed scanners (minimum 2 of 3). Do not wait for all three to complete; minimum 2 is the gate.
2. Fill gaps with targeted manual investigation in Phase 3
3. Note incomplete coverage in the final report
### Error: "Architecture Pattern Not Clear"
Cause: Codebase lacks conventional structure or uses unfamiliar patterns
Solution:
1. Document what IS present rather than forcing a known pattern
2. Trace entry points to understand actual flow
3. Report "unconventional structure" with observed organization
---
## Tiered Depth Model
The tier is determined by query type, not guessed. Matching depth to question scope prevents waste (not spending 30 minutes on a 2-minute fact-check) while ensuring completeness (not running a shallow scan when full understanding is needed). **Default to Standard if depth is unspecified.** All phases within the selected tier must run; phases cannot be skipped or reordered.
### Quick Verify (2-5 minutes)
**Purpose**: Confirm a specific fact about the codebase.
**Scope**: Answer one question. Read only the files necessary to answer it. No document generation — the answer IS the output. Stay focused on the specific fact; adjacent subsystems are out of scope.
**Phases used**: Phase 1 (SCAN) only — single targeted scanner, not parallel.
**Exit criteria**: Question answered with file path evidence, or "could not determine" with a list of what was checked.
**Examples**: "Does this project use dependency injection?" "What ORM does the API use?" "Is there a CI pipeline configured?" "What version of React is this using?"
**Output format**: Direct answer in conversation (no saved report file).
### Standard (15-30 minutes)
**Purpose**: Understand a subsystem or map one area of the codebase. This is the default tier when the user hasn't specified a depth preference.
**Scope**: All 4 phases execute. Parallel scanners in Phase 1 (minimum 2 of 3 must complete). Produce a single structured document covering the targeted area. Do not jump to analysis without creating an architecture map in Phase 2.
**Phases used**: All 4 phases (SCAN, MAP, ANALYZE, REPORT). Phases cannot be skipped.
**Exit criteria**: Report covers the subsystem's boundaries, key patterns, and integration points. Saved as `exploration-report.md`.
**Examples**: "How does authentication work in this app?" "Map the payment processing flow." "Explain how the event system works." "What's the testing strategy for this repo?"
**Output format**: Saved `exploration-report.md` with all sections.
### Deep Dive (1+ hour)
**Purpose**: Full analysis — architectural understanding PLUS quality evaluation, consistency assessment, and pattern analysis. Use this tier when the user asks to "analyze quality of", "assess consistency of", or "evaluate patterns in" the codebase.
**Scope**: All 8 phases. Phases 1-3 explore the codebase. Phases 4-7 compile findings, assess quality, synthesize recommendations, and verify against source code to remove false positives. Phase 8 produces a comprehensive report. Do not generate improvement recommendations unless findings have been verified against actual source code in Phase 7.
**Phases used**: All 8 phases: SCAN → MAP → ANALYZE → COMPILE → ASSESS → SYNTHESIZE → REFINE → REPORT. All phases are mandatory; none can be skipped.
**Exit criteria**: Comprehensive report covers full architecture, quality assessment with scores, consistency evaluation, pattern analysis, and ranked recommendations. Multiple artifact files produced. Top 5 findings verified against source in Phase 7 to confirm accuracy.
**Examples**: "I'm new to this codebase, give me the full picture." "We're considering a major refactor — what do we need to know?" "Analyze the quality and consistency of our error handling." "Assess which patterns are used consistently vs inconsistently." "Full architectural review before we plan next quarter." "Evaluate the test coverage patterns across all modules."
**Output format**: Saved `exploration-report.md` plus supplementary files (`architecture-map.md`, `analysis-compilation.md`, component-specific documents as needed).
### Tier Selection Logic
```
If user specifies --tier or --quick or --deep:
Use the specified tier.
Else if user asks a single specific question (what/which/does/is):
Use Quick Verify.
Else if user asks to analyze/assess/evaluate/audit quality or patterns:
Use Deep Dive (requires all 8 phases including COMPILE, ASSESS, SYNTHESIZE, REFINE for verification).
Else if user asks about a specific subsystem or flow:
Use Standard.
Else if user asks for full picture / onboarding / comprehensive:
Use Deep Dive.
Else:
Default to Standard.
```
---
## Source Hierarchy for Research
When exploration requires looking up external information (framework conventions, library APIs, configuration options), follow this source hierarchy strictly. Lower-quality sources introduce hallucination risk -- a wrong framework convention applied confidently is worse than admitting uncertainty.
### Hierarchy (highest to lowest priority)
1. **Context7 MCP** (preferred) -- Use `resolve-library-id` then `query-docs` for up-to-date library documentation. This is the most reliable source because it pulls directly from official package documentation.
2. **Official documentation** -- Framework and library official docs via WebFetch if Context7 does not cover the library.
3. **Web search** -- Use WebSearch as a fallback when Context7 and official docs are insufficient. Cross-reference multiple results before citing.
### Rules
- **Never guess framework conventions** -- If you cannot confirm a convention from the source hierarchy, state "could not confirm" rather than assuming.
- **Cite the source tier** -- When reporting externally-sourced information, note where it came from: "(confirmed via Context7)", "(from official docs)", or "(from web search -- verify independently)".
- **Source code is ground truth** -- External documentation describes how things SHOULD work. The codebase shows how things ACTUALLY work. When they conflict, report both and flag the discrepancy.
- **Web search results degrade fastest** -- Blog posts and Stack Overflow answers may reference outdated APIs. Always note the date if visible and prefer recent results.
---
## References
This skill follows the **Pipeline Architecture** pattern with artifact-based reporting and gate enforcement. Context is ephemeral; all findings must be saved to files at each phase to survive context compression and enable reuse across sessions.
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 |