Plan Fix

Investigate bugs and create TDD fix plans in PLANS.md.

Purpose

• Investigate bugs found after processing files (extraction errors, wrong data, missing matches)
• Debug deployment failures using Railway MCP
• Test and iterate Gemini prompts when extraction issues are suspected
• Create investigation report documenting findings and root cause
• Generate TDD-based fix plan in PLANS.md
• Does NOT implement fixes (integrates with plan-implement)

Pre-flight Check

Before doing anything, read PLANS.md and check if it contains incomplete work:

• If PLANS.md has content but NO "Status: COMPLETE" at the end → STOP
• Tell the user: "PLANS.md has incomplete work. Please review and clear it before planning new items."
• Do not proceed.

If PLANS.md is empty or has "Status: COMPLETE" → proceed with investigation.

Arguments

$ARGUMENTS should contain the bug description with context:

• What happened vs what was expected
• File IDs if relevant (Google Drive file IDs)
• Error messages or unexpected values
• Deployment ID if it's a deployment issue
• Any other context that helps investigation

Context Gathering

IMPORTANT: Do NOT hardcode MCP names or folder paths. Always read CLAUDE.md to discover:

1. Available MCP servers - Look for the "MCP SERVERS" section to find:

   • Google Drive MCP for file access (gdrive_search, gdrive_read_file, gsheets_read, etc.)
   • Railway MCP for deployment debugging (get-logs, list-deployments, list-services, list-variables)
   • Gemini MCP for prompt testing (gemini_analyze_pdf)

2. Folder structure - Look for "FOLDER STRUCTURE" section to understand:

   • Where documents are stored
   • Naming conventions for bank folders, card folders, broker folders

3. Document types - Look for "DOCUMENT CLASSIFICATION" section to understand:

   • Document type → destination mapping
   • ADVA role in each document type

4. Spreadsheet schemas - Look for "SPREADSHEETS" section or read SPREADSHEET_FORMAT.md

Investigation Workflow

Step 1: Classify the Bug Type

Based on $ARGUMENTS, determine the bug category:

Step 2: Gather Evidence

For Codebase Issues:

• Use Grep/Glob for searching the codebase
• Use Task tool with subagent_type=Explore for broader exploration
• Read relevant source files and tests

For Deployment Issues:

1. check-railway-status - Verify CLI is working
2. list-services - Find the affected service
3. list-deployments - Get recent deployment IDs and statuses
4. get-logs with logType: "deploy" - Check deployment logs
5. get-logs with logType: "build" - Check build logs if deployment failed
6. list-variables - Verify environment configuration

For Document/Drive Issues:

• gdrive_search - Find the problematic file
• gdrive_read_file or gdrive_get_pdf - Read file contents
• gsheets_read - Check spreadsheet data

For Prompt Issues (when extraction is consistently wrong):

1. Get the source PDF using gdrive_get_pdf
2. Use gemini_analyze_pdf to test alternative prompts
3. Compare current prompt output vs expected output
4. Iterate on prompt wording until extraction improves
5. Document the improved prompt for implementation

Step 3: Document Findings

Write PLANS.md with this structure:

# Bug Fix Plan

**Created:** YYYY-MM-DD
**Bug Report:** [Summary from $ARGUMENTS]
**Category:** [Extraction | Deployment | Matching | Storage | Prompt]

## Investigation

### Context Gathered
- **MCPs used:** [list which MCPs were used and why]
- **Files examined:** [list files checked - Drive files, spreadsheets, code files, logs]

### Evidence
[Detailed findings from investigation with specific data points]

### Root Cause
[Clear explanation of why the bug occurs]

## Fix Plan

### Fix 1: [Title matching the issue]
1. Write test in [file].test.ts for [scenario that reproduces the bug]
2. Implement fix in [file].ts

### Fix 2: [Title]
...

## Post-Implementation Checklist
1. Run `bug-hunter` agent - Review changes for bugs
2. Run `test-runner` agent - Verify all tests pass
3. Run `builder` agent - Verify zero warnings

Gemini Prompt Testing Guidelines

When investigating extraction issues:

1. Get the problematic PDF using gdrive_get_pdf
2. Read current prompt from src/gemini/prompts.ts
3. Test with gemini_analyze_pdf using variations of the prompt
4. Document what works - Include the improved prompt in the fix plan
5. Note: gemini_analyze_pdf is for testing ONLY, not production analysis

Example prompt testing workflow:

1. Current prompt extracts "fechaCierre" as null
2. Test prompt variation A: "Look for 'Fecha de Cierre' or 'CIERRE'"
3. Test prompt variation B: "The closing date appears near the top of the statement"
4. Variation B correctly extracts the date
5. Add to fix plan: Update prompts.ts with variation B

Railway Debugging Guidelines

When investigating deployment issues:

1. Check status first - check-railway-status confirms CLI access
2. List recent deployments - list-deployments with json: true for structured data
3. Get targeted logs - Use filter parameter to search for errors:
   • filter: "@level:error" - Find error-level logs
   • filter: "TypeError" - Search for specific errors
4. Check environment - list-variables to verify config

Rules

• Refuse to proceed if PLANS.md has incomplete work
• Do NOT hardcode MCP names or folder paths - always read from CLAUDE.md
• Investigation only - do not modify source code
• All fixes must follow TDD (test first)
• Include enough detail for another model to implement without context
• Always include post-implementation checklist
• For prompt issues, test multiple variations before recommending changes

CRITICAL: Scope Boundaries

This skill creates plans. It does NOT implement them.

1. NEVER ask to "exit plan mode" - This skill doesn't use Claude Code's plan mode feature
2. NEVER implement code - Your job ends when PLANS.md is written
3. NEVER ask ambiguous questions like "should I proceed?" or "ready to continue?"
4. NEVER start implementing after writing the plan, even if user says "yes" to something

Termination

When you finish writing PLANS.md, output this exact message and STOP:

✓ Plan created in PLANS.md

Next step: Run `plan-implement` to execute this plan.

Do not ask follow-up questions. Do not offer to implement. Just output the message and stop.