Anth Advanced Troubleshooting

Anthropic Advanced Troubleshooting

Issue: Context Window Overflow

# Symptom: invalid_request_error about token count
# Diagnosis: pre-check with Token Counting API
import anthropic

client = anthropic.Anthropic()

count = client.messages.count_tokens(
    model="claude-sonnet-4-20250514",
    messages=conversation_history,
    system=system_prompt
)
print(f"Input tokens: {count.input_tokens}")
# Claude Sonnet: 200K context, Claude Opus: 200K context

# Fix: truncate oldest messages or summarize
def trim_conversation(messages: list, max_tokens: int = 180_000) -> list:
    """Keep recent messages within token budget."""
    # Always keep first (system context) and last 5 messages
    if len(messages) <= 5:
        return messages
    return messages[:1] + messages[-5:]  # Crude but effective

Issue: Tool Use Not Triggering

# Symptom: Claude responds with text instead of calling tools
# Diagnosis checklist:
# 1. Tool description must clearly state WHEN to use the tool
# 2. User message must match the tool's trigger condition

# BAD description (too vague):
{"name": "search", "description": "Search for things"}

# GOOD description (clear trigger):
{"name": "search_products", "description": "Search the product catalog by name, category, or price range. Use whenever the user asks about products, pricing, or availability."}

# Force tool use if needed:
message = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    tools=tools,
    tool_choice={"type": "any"},  # Must call at least one tool
    messages=[{"role": "user", "content": "Find products under $50"}]
)

Issue: Streaming Drops or Corruption

# Symptom: stream ends prematurely or text is garbled
# Cause: network interruption, proxy timeout, or large response

# Fix: implement reconnection with content tracking
def resilient_stream(client, **kwargs):
    """Stream with reconnection on failure."""
    collected_text = ""
    max_retries = 3

    for attempt in range(max_retries):
        try:
            with client.messages.stream(**kwargs) as stream:
                for text in stream.text_stream:
                    collected_text += text
                    yield text
                return  # Success
        except Exception as e:
            if attempt == max_retries - 1:
                raise
            # Note: Claude streams are NOT resumable
            # Must restart from beginning
            collected_text = ""
            print(f"Stream interrupted, retrying ({attempt + 1}/{max_retries})")

Issue: Unexpected Stop Reason

# Debug unexpected truncation
msg = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=4096,  # Was it too low?
    messages=[{"role": "user", "content": long_prompt}]
)
print(f"Stop reason: {msg.stop_reason}")
print(f"Output tokens: {msg.usage.output_tokens}")
print(f"Max tokens: 4096")
# If output_tokens == max_tokens, response was truncated

Issue: Response Quality Degradation

# Checklist for quality issues:
# 1. System prompt too long or contradictory?
# 2. Conversation history too noisy (too many turns)?
# 3. Wrong model for task complexity?
# 4. Temperature too high for deterministic tasks?

# Debug: log the full request for review
import json
request_params = {
    "model": model,
    "max_tokens": max_tokens,
    "system": system[:200] + "...",  # Truncated for logging
    "message_count": len(messages),
    "temperature": temperature,
}
print(f"Request config: {json.dumps(request_params, indent=2)}")

Diagnostic Curl Commands

# Test specific model availability
for model in claude-haiku-4-20250514 claude-sonnet-4-20250514 claude-opus-4-20250514; do
  echo -n "$model: "
  curl -s -o /dev/null -w "%{http_code}" https://api.anthropic.com/v1/messages \
    -H "x-api-key: $ANTHROPIC_API_KEY" \
    -H "anthropic-version: 2023-06-01" \
    -H "content-type: application/json" \
    -d "{\"model\":\"$model\",\"max_tokens\":8,\"messages\":[{\"role\":\"user\",\"content\":\"hi\"}]}"
  echo
done

Overview

This guide isolates difficult Claude API failures by testing one variable at a time: token budget, tool schema and choice, stream transport, stop reason, or prompt configuration. It complements the common status-code guide and should produce evidence that is safe to share with an operator.

Prerequisites

• Use an approved sandbox workspace, a pinned model ID, and synthetic messages/tools that cannot access production data or perform side effects.
• Have a bounded request timeout, retry cap, token-counting access where enabled, and a known-good baseline request for comparison.
• Configure telemetry to retain request ID, model, token counts, stop reason, event counts, and latency only; redact prompts, completions, tool arguments, headers, and secrets.

Instructions

1. Reproduce the smallest failing case in the sandbox and record a correlation ID. Change one input at a time, starting with token count and request shape.
2. For context failures, count tokens before sending and trim or summarize using an explicit policy that keeps required system context. For tool failures, validate the schema and use a no-op tool before enabling any real action.
3. For streaming failures, count received events and restart the complete non-resumable request with a bounded retry; deduplicate downstream presentation by correlation ID.
4. Compare stop reason, usage, latency, and output-shape assertions against the known-good baseline. Run one canary against the approved environment before promotion.
5. If the canary changes scope, output policy, retention, or error rate, halt and roll back the prompt/model/configuration change. Remove synthetic fixtures after the receipt is written.

Output

Return a troubleshooting receipt with correlation_id, hypothesis, changed variable, model, input/output token counts, stop reason, stream event counts, retry attempts, baseline comparison, canary status, rollback reference, and cleanup status. Keep all prompt, completion, tool-input, and credential fields redacted.

Error Handling

• A token-counting call that fails is not evidence that the message call is safe; stop at the preflight gate and report the provider error without sending the full request.
• Never treat a partial stream as a complete answer. Mark it incomplete, discard or quarantine it, and restart only when the operation is safe to repeat.
• A tool-use response is untrusted input to the tool executor. Validate name and arguments against an allowlist, require approval for side effects, and reject unknown or malformed calls.
• If a quality regression cannot be isolated, freeze promotion, preserve the redacted baseline comparison, and revert to the last known-good model/prompt pair.

Examples

Use a synthetic tool lookup_fixture whose only permitted input is fixture_id=demo-001; run the same prompt with and without tool_choice, and record tool_call_count, schema result, and side_effects=0. For a dropped stream, record events_received=17, complete=false, restart once with the same correlation policy, and expose only the final redacted result.

Resources

• Error Reference
• Token Counting
• Tool Use Guide

Next Steps

For load testing, see anth-load-scale.