CC10: Claude Features — Thinking, Caching, Citations, Files

resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=4096,
    thinking={"type": "enabled", "budget_tokens": 8000},
    messages=[{"role": "user", "content":
        "We have 200 microservices, 80% Java. Should we migrate auth from JWT "
        "to mTLS service-to-service? Walk through the trade-offs."
    }],
)

for block in resp.content:
    if block.type == "thinking":
        # Don't show users; do log for debugging
        print("[thinking]", block.thinking[:200], "...")
    elif block.type == "text":
        print("[answer]", block.text)

long_claude_md = open("CLAUDE.md").read()  # imagine 5K tokens
schema_docs = open("schema.md").read()      # imagine 3K tokens

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": long_claude_md + "\n\n" + schema_docs,
            "cache_control": {"type": "ephemeral"},   # ← cache this prefix
        },
    ],
    messages=[{"role": "user", "content": "How should I model a refund?"}],
)
print(resp.usage)
# usage shows: cache_creation_input_tokens (first call), cache_read_input_tokens (later calls)

{
  "type": "text",
  "text": "...",
  "cache_control": {"type": "ephemeral", "ttl": "1h"}   # 1-hour TTL
}

import base64
with open("ui-mockup.png", "rb") as f:
    img_b64 = base64.standard_b64encode(f.read()).decode()

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=1024,
    messages=[{"role": "user", "content": [
        {"type": "image", "source": {
            "type": "base64", "media_type": "image/png", "data": img_b64,
        }},
        {"type": "text", "text": "What React component structure would you propose for this UI?"},
    ]}],
)

with open("design-doc.pdf", "rb") as f:
    pdf_b64 = base64.standard_b64encode(f.read()).decode()

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=2048,
    messages=[{"role": "user", "content": [
        {"type": "document", "source": {
            "type": "base64", "media_type": "application/pdf", "data": pdf_b64,
        }},
        {"type": "text", "text": "Summarize the proposed migration plan in 5 bullets."},
    ]}],
)

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=1024,
    messages=[{"role": "user", "content": [
        {
            "type": "document",
            "source": {"type": "text", "data": open("policy.md").read()},
            "title": "Refund Policy v3",
            "context": "Internal policy doc, last updated 2025-08-12",
            "citations": {"enabled": True},   # ← opt in
        },
        {"type": "text", "text": "Can a customer get a refund 35 days after purchase?"},
    ]}],
)

# resp.content has text blocks; each text block can have a `citations` array
# pointing to the document and character ranges that grounded the sentence.
for block in resp.content:
    if block.type == "text":
        print(block.text)
        for cite in (block.citations or []):
            print(f"  → {cite.cited_text!r} (chars {cite.start_char_index}-{cite.end_char_index})")

tools = [{"type": "code_execution_20250522", "name": "code_execution"}]

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=4096,
    tools=tools,
    extra_headers={"anthropic-beta": "code-execution-2025-05-22"},
    messages=[{"role": "user", "content":
        "Generate a histogram of these numbers: 1,1,2,3,3,3,4,4,5,5,5,5,7."
    }],
)
# Claude writes Python that uses matplotlib, the sandbox runs it,
# returns the figure as a tool_result, Claude embeds it in the answer.

file = client.beta.files.upload(
    file=open("sales.csv", "rb"),
)   # one upload, reusable
print(file.id)   # "file_abc123"

resp = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=4096,
    tools=[{"type": "code_execution_20250522", "name": "code_execution"}],
    extra_headers={"anthropic-beta": "code-execution-2025-05-22,files-api-2025-04-14"},
    messages=[{"role": "user", "content": [
        {"type": "container_upload", "file_id": file.id},   # mount file into sandbox
        {"type": "text", "text": "Compute total sales by region. Write the result to result.csv."},
    ]}],
)
# Claude's code_execution writes /mnt/result.csv; you can fetch it via the Files API.

import json
from anthropic import Anthropic

# Pad with realistic CLAUDE.md content. Aim for >1024 tokens (Sonnet minimum).
SYSTEM = ("# Project: PublicRecords\n\n" +
          ("- All amounts in cents.\n- Use Drizzle ORM.\n- Tests must pass.\n" * 200))

client = Anthropic()

def call(question: str):
    return client.messages.create(
        model="claude-sonnet-4-6", max_tokens=128,
        system=[{
            "type": "text", "text": SYSTEM,
            "cache_control": {"type": "ephemeral"},
        }],
        messages=[{"role": "user", "content": question}],
    )

r1 = call("How should money be represented?")
print(json.dumps(r1.usage.model_dump(), indent=2))
# Expect: cache_creation_input_tokens > 0, cache_read_input_tokens == 0

for q in ["Which ORM?", "Do tests need to pass before commit?",
          "What's the unit for amounts?"]:
    r = call(q)
    print(q, "→", r.usage.model_dump())
    # Expect: cache_creation_input_tokens == 0, cache_read_input_tokens > 0

# Add even ONE extra space at the start of SYSTEM
SYSTEM_BROKEN = " " + SYSTEM
r = client.messages.create(
    model="claude-sonnet-4-6", max_tokens=128,
    system=[{"type": "text", "text": SYSTEM_BROKEN,
             "cache_control": {"type": "ephemeral"}}],
    messages=[{"role": "user", "content": "Which ORM?"}],
)
print(r.usage.model_dump())
# Cache miss again; cache_creation_input_tokens > 0

from anthropic.types import Usage
def cost(u: Usage, in_per_m=3.0, cache_in_per_m=3.75, cache_read_per_m=0.30,
         out_per_m=15.0):
    return (u.input_tokens * in_per_m
            + (u.cache_creation_input_tokens or 0) * cache_in_per_m
            + (u.cache_read_input_tokens     or 0) * cache_read_per_m
            + u.output_tokens * out_per_m) / 1_000_000

# Compare cold first call vs a warm one. Expect ~10x savings on input portion.