Debug Fetcher

name: debug-fetcher description: > Automated URL fetch failure handling with strategy exhaustion, memory learning, and human-in-the-loop recovery. Use when fetches fail and you need intelligent retry, pattern learning, and human collaboration. allowed-tools: Bash, Read, Write, Task triggers:

• debug fetch
• debug fetcher
• why did fetch fail
• fetch failure
• analyze fetch failure
• retry fetch
• resilient fetch
• extraction failure
• empty extracted text
• result.text contains HTML
• content extraction bug metadata: short-description: Failure-to-recovery automation for URL fetching

provides: - debug-fetcher composes: [, task-monitor]

Debug-Fetcher Skill

Automated fetch failure handling that:

1. Queries /memory first - applies learned strategies before trying defaults
2. Exhausts all strategies - direct, playwright, wayback, brave, jina, proxy, UA rotation
3. Stores successes - saves working strategies to /memory for future runs
4. Collaborates with humans - uses /interview when all automated strategies fail

Quick Start

# Fetch single URL with failure handling
./run.sh fetch https://example.com

# Fetch batch with failure handling
./run.sh fetch-batch urls.txt

# Check what was learned about a domain
./run.sh recall example.com

# Export all learned strategies
./run.sh export-learnings

How It Works

URL Request
    │
    ▼
┌──────────────────────────┐
│  1. Query /memory        │
│  "What works for this    │
│   domain?"               │
└──────────────────────────┘
    │
    ▼
┌──────────────────────────┐
│  2. Try learned strategy │
│     (if exists)          │
└──────────────────────────┘
    │
    ▼ (fail or no learned strategy)
┌──────────────────────────┐
│  3. Exhaust strategies:  │
│  - direct fetch          │
│  - playwright            │
│  - wayback machine       │
│  - brave alternates      │
│  - jina reader           │
│  - proxy rotation        │
│  - user-agent rotation   │
└──────────────────────────┘
    │
    ▼ (all fail)
┌──────────────────────────┐
│  4. Launch /interview    │
│  Ask human for help:     │
│  - Credentials?          │
│  - Mirror URL?           │
│  - Manual download?      │
│  - Skip this URL?        │
└──────────────────────────┘
    │
    ▼
┌──────────────────────────┐
│  5. Store to /memory     │
│  - Successful strategy   │
│  - Domain patterns       │
│  - Human-provided info   │
└──────────────────────────┘

Memory Schema

Each learned strategy stores:

Commands

fetch-batch Options

Task-Monitor Integration

debug-fetcher integrates with the centralized task-monitor for live progress tracking:

# Run batch fetch with task-monitor (enabled by default)
./run.sh fetch-batch urls.txt

# View progress in task-monitor TUI
cd ~/.pi/skills/task-monitor
uv run python monitor.py tui --filter debug-fetcher

# Or check state file directly
cat /path/to/debug-fetcher/debug_fetcher_task_state.json | jq

State file schema:

{
  "completed": 50,
  "total": 100,
  "progress_pct": 50.0,
  "strategies": {
    "direct": {"attempts": 30, "successes": 25, "rate": 0.833},
    "playwright": {"attempts": 20, "successes": 18, "rate": 0.9}
  },
  "stats": {
    "success_count": 45,
    "failure_count": 5,
    "success_rate": 0.9,
    "learned_count": 12
  },
  "failures": [...],
  "status": "running"
}

NDJSON Streaming Output

For long-running batch jobs, use --json-stream to output one JSON object per line:

./run.sh fetch-batch urls.txt --json-stream | tee results.jsonl

# Each line:
# {"url": "https://...", "success": true, "winning_strategy": "playwright", "attempts": 2, "timing_ms": 1234}

This enables:

• Real-time progress monitoring via tail -f results.jsonl | jq
• Integration with streaming parsers
• Resume from partial runs

Environment Variables

Integration with Fetcher

Debug-fetcher wraps the standard fetcher skill and adds failure handling capabilities. All fetcher environment variables (BRAVE_API_KEY, FETCHER_EMIT_MARKDOWN, etc.) are respected.

Examples

Learning from Failures

After fetching a batch of URLs, debug-fetcher stores successful strategies:

# Fetch a batch
./run.sh fetch-batch urls.txt --output results.jsonl

# View what was learned
./run.sh recall attack.mitre.org
# Output:
# Domain: attack.mitre.org
# Strategy: playwright
# Success rate: 95%
# Last used: 2025-01-30

# Next time, playwright will be tried first for attack.mitre.org
./run.sh fetch https://attack.mitre.org/techniques/T1059

Human-in-the-Loop Interview

When all strategies fail, an interview is generated:

# Fetch batch with failures
./run.sh fetch-batch difficult_urls.txt

# Interview generated at: /tmp/interview_abc123.json
# Run: ./agents/skills/interview/run.sh /tmp/interview_abc123.json

# Example interview questions:
# - "Failed 5 URLs from nytimes.com. Do you have credentials?"
# - "archive.org not working. Try a mirror URL?"

YouTube URL Handling

YouTube URLs are automatically detected and handled via the /ingest-youtube skill:

# YouTube URLs use transcript extraction
./run.sh fetch https://www.youtube.com/watch?v=abc123
# Uses: /ingest-youtube skill for transcript extraction
# Falls back to other strategies if transcript unavailable

Batch Analysis

After a batch run, analyze patterns:

from debug_fetcher.batch_analyzer import analyze_batch, get_failure_summary

# Get summary
summary = get_failure_summary(results)
# {
#   "total": 1000,
#   "success": 850,
#   "failed": 150,
#   "success_rate": "85.0%",
#   "top_failing_domains": [
#     {"domain": "nytimes.com", "count": 45},
#     {"domain": "wsj.com", "count": 30}
#   ],
#   "patterns": [
#     "All 45 URLs from nytimes.com returned HTTP 403",
#     "High failure rate: 50% of failures are paywalled sites"
#   ]
# }

Fetch Failure vs Extraction Failure

Critical distinction: "Empty content" can mean two different things:

Diagnosing the Difference

# Step 1: Check HTTP status
jq '.items[0].status' consumer_summary.json
# 200 = fetch succeeded, issue is extraction
# 403/401/etc = fetch failed, try other strategies

# Step 2: Check raw download
head -c 500 downloads/*.html
# If contains real content → extraction bug
# If contains JS shell only → needs Playwright
# If contains error page → fetch failed

# Step 3: Check extracted text
head -c 500 extracted_text/*.txt
# If contains "<!DOCTYPE html>" → extraction returned HTML (BUG)
# If contains clean text → working correctly
# If empty → extraction failed

# Step 4: Verify method used
jq '.items[0].method' consumer_summary.json
# "aiohttp" = direct fetch
# "playwright" = JS rendering used

Real-World Example (Fixed 2026-02-04)

Bug: result.text contained raw HTML instead of extracted text.

Evidence:

downloads/abc.html     → <!DOCTYPE html>...(full HTML)
extracted_text/abc.txt → <!DOCTYPE html>...(same HTML - BUG!)

Fix: evaluate_result_content() now replaces result.text with assessment.text (trafilatura output).

Lesson: Always compare downloads/ vs extracted_text/ - they should be different formats.

Recovery Actions

When human provides help via interview:

Files

.pi/skills/debug-fetcher/
├── SKILL.md           # This file
├── run.sh             # Entry point
├── pyproject.toml     # Dependencies
└── debug_fetcher/     # Python package
    ├── __init__.py
    ├── cli.py                 # CLI commands
    ├── memory_schema.py       # FetchStrategy dataclass
    ├── memory_bridge.py       # Recall/learn from /memory
    ├── strategy_engine.py     # Strategy exhaustion loop
    ├── batch_analyzer.py      # Analyze batch failures
    ├── interview_generator.py # Generate /interview JSON
    ├── interview_processor.py # Process interview responses
    ├── recovery_executor.py   # Execute recovery actions
    └── pdf_bridge.py          # Cross-skill integration with debug-pdf

Companion Skill: debug-pdf

debug-fetcher and debug-pdf work together in the pipeline:

URL → debug-fetcher → /fetcher → /extractor → debug-pdf
         ↓                           ↓
      fetch fail               extraction fail
         ↓                           ↓
    retry/recover            analyze PDF issues
         ↓                           ↓
      /memory                     /memory

Shared failure patterns:

Cross-skill notifications:

• When debug-fetcher successfully fetches a PDF but detects issues (password protected, scanned), it notifies debug-pdf via agent-inbox
• When debug-fetcher fails to fetch a PDF URL, it notifies debug-pdf for tracking

Related Skills

• /memory - Stores learned fetch strategies
• /interview - Human collaboration for unrecoverable URLs
• /ingest-youtube - YouTube transcript extraction
• /fetcher - Core URL fetching functionality
• /extractor - Content extraction from fetched documents
• /debug-pdf - Companion skill for PDF extraction failures