error-recovery

name	error-recovery
description	Intelligent error diagnosis, automatic recovery strategies, and prevention of recurring failures
author	AuthorAgent
version	1.0.0
triggers	["error","fix error","something broke","failed","not working","debug","troubleshoot","diagnose","recovery"]
permissions	["file:read","file:write"]

Error Recovery — Core Skill

When things go wrong, AuthorAgent doesn't just report the error — it diagnoses the root cause, attempts automatic recovery, and remembers the fix for next time.

Error Categories

1. AI Provider Errors

Symptoms: API timeouts, rate limits, auth failures, model unavailability

Error Detected: Gemini API returned 429 (rate limited)

Diagnosis: Too many requests in short period (>60/min)
Recovery Strategy:
  1. ✅ Wait 30 seconds and retry (attempt 1 of 3)
  2. ⏳ Switch to DeepSeek as fallback provider
  3. ⏳ If all providers fail, queue the task for later

Lesson Stored: "Gemini rate limits at ~60 req/min.
Add 500ms delay between batch requests."

Recovery Playbook:

• 429 Rate Limited → Wait + retry → Switch provider → Queue
• 401 Unauthorized → Check API key in vault → Alert user if missing
• 503 Service Unavailable → Retry with backoff → Switch provider
• Timeout → Reduce prompt length → Retry → Switch provider
• Content filtered → Rephrase prompt → Retry without triggering content filter
• Context too long → Truncate prior results → Summarize context → Retry

2. Goal Execution Errors

Symptoms: Step failures, invalid AI output, parsing failures

Error Detected: Goal step "Create character profiles" failed
AI returned unstructured text instead of character profiles

Diagnosis: Prompt was too open-ended for structured output
Recovery Strategy:
  1. ✅ Reformulate prompt with explicit structure requirements
  2. ✅ Add output format example to prompt
  3. ✅ Retry with reformulated prompt

New prompt: "Create character profiles in this exact format:
NAME: [name]
AGE: [age]
ROLE: [protagonist/antagonist/supporting]
..."

Result: ✅ Step succeeded on retry

Recovery Playbook:

• Unstructured output → Add format constraints → Retry
• JSON parse error → Strip markdown fences → Extract JSON → Retry with stricter prompt
• Output too short → Request longer output → Increase max tokens → Retry
• Repeated failure → Skip step → Flag for user review → Continue goal
• Nonsensical output → Switch provider → Reduce temperature → Retry

3. File System Errors

Symptoms: Write failures, permission denied, path not found

Error Detected: Cannot write to workspace/projects/my-novel/chapter-1.md
Permission denied

Diagnosis: Directory doesn't exist or wrong permissions
Recovery Strategy:
  1. ✅ Create directory with mkdir -p
  2. ✅ Retry write
  3. If still failing → Check sandbox permissions → Alert user

Result: ✅ Created directory and wrote file successfully

Recovery Playbook:

• ENOENT (file not found) → Create parent directories → Retry
• EACCES (permission denied) → Check sandbox rules → Alert user
• ENOSPC (disk full) → Alert user with disk usage info
• EISDIR (is a directory) → Correct the path → Retry

4. Network/Research Errors

Symptoms: Fetch timeouts, blocked domains, invalid responses

Recovery Playbook:

• Timeout → Retry with shorter timeout → Try different URL → Skip
• Blocked domain → Inform user → Suggest alternative source
• Empty response → Retry → Try cached version → Use AI knowledge instead
• SSL error → Alert user (don't bypass)

5. Memory/Context Errors

Symptoms: Context too long, corrupted memory files, missing context

Recovery Playbook:

• Context overflow → Summarize old context → Trim to fit → Retry
• Corrupted JSONL → Skip bad lines → Rebuild from valid entries → Alert user
• Missing memory file → Create empty file → Continue with fresh context

Automatic Recovery System

The Recovery Loop

1. Error detected
2. Classify error (provider/goal/file/network/memory)
3. Look up recovery playbook
4. Check improvement log for past fixes for this error type
5. Execute recovery strategy (up to 3 attempts)
6. If recovered:
   → Log the successful fix
   → Continue with the task
   → Store lesson for future prevention
7. If not recovered:
   → Pause the goal (don't fail silently)
   → Send clear error report to user
   → Suggest manual intervention steps
   → Store the failure for analysis

Error Prevention

The real power is preventing errors before they happen:

1. Pre-flight checks — Before each step, verify:

   • AI provider is responsive (quick ping)
   • Required files exist and are readable
   • Sufficient disk space for output
   • API key is still valid (not expired)

2. Pattern recognition — If an error occurred 3+ times:

   • Add automatic mitigation before the risky step
   • Example: Always add 500ms delay before Gemini calls (learned from rate limits)

3. Graceful degradation — If the best approach fails:

   • Fall back to simpler approach
   • Use a different provider
   • Reduce scope rather than fail completely

Error Log

All errors and recoveries are logged to workspace/.audit/error-log.jsonl:

{
  "timestamp": "2026-02-24T15:30:00Z",
  "category": "provider",
  "error": "Gemini API 429 rate limited",
  "goalId": "goal-5",
  "stepId": "goal-5-step-3",
  "recoveryAttempts": [
    { "strategy": "wait_retry", "success": false },
    { "strategy": "switch_provider", "provider": "deepseek", "success": true }
  ],
  "resolved": true,
  "lesson": "Rate limit hit during batch execution. Added 500ms delay.",
  "preventionRule": "delay_between_gemini_calls_500ms"
}

User-Facing Error Reports

When errors can't be auto-resolved, the user gets a clear report:

⚠️ I hit a problem and couldn't fix it automatically.

What happened: Gemini and DeepSeek both returned errors
while trying to write Chapter 5.

What I tried:
  1. Retry with Gemini (3x) — failed (rate limited)
  2. Switch to DeepSeek — failed (API key expired)
  3. Switch to Ollama — not configured

What you can do:
  • Check your DeepSeek API key (may need renewal)
  • Wait 5 minutes and say "continue" to retry
  • Add an Ollama endpoint as a backup provider

Your goal is paused at step 5/12. No work was lost.

Commands

• diagnose [error] — Analyze a specific error
• show error log — View recent errors and recoveries
• error stats — Error frequency and recovery success rates
• test recovery — Simulate an error to test recovery
• prevention rules — Show active error prevention rules