identity, critical_rules, scope,
verbatim policy, refusal template] + L2[Layer 2: Runtime reminders
injected every turn +
long-conversation re-anchor] + M[Claude] + T{Tool use?} + L3[Layer 3: Tool-side enforcement
input validation +
protocol guard
eligibility before return] + L4[Layer 4: Output validation
regex grounding checks,
markdown / off-topic / ID / date] + OK[Reply to user] + BAD[Safe fallback,
bad reply dropped from history] + + U --> L1 + L1 --> L2 + L2 --> M + M --> T + T -- yes --> L3 + L3 --> M + T -- no --> L4 + L4 -- ok --> OK + L4 -- violations --> BAD + ``` + + Layer 1 is the system prompt itself. It tells the model what Bookly + is, what it can and cannot help with, what the return policy actually + says (quoted verbatim, not paraphrased), and exactly which template + to use when refusing. Layer 2 adds short reminder blocks on every + turn so the model re-reads the non-negotiable rules at the + highest-attention position right before the user turn. Layer 3 lives + in `tools.py`: the tool handlers refuse unsafe calls regardless of + what the model decides. Layer 4 lives at the end of the agent loop + and does a deterministic regex pass over the final reply looking + for things like fabricated order IDs, markdown leakage, and + off-topic engagement. + + # Request lifecycle + + A single user message travels this path: + + ```mermaid + sequenceDiagram + autonumber + participant B as Browser + participant N as nginx + participant S as FastAPI + participant A as agent.run_turn + participant C as Claude + participant TL as tools.dispatch_tool + + B->>N: POST /api/chat { message } + N->>S: proxy_pass + S->>S: security_headers middleware + S->>S: resolve_session (cookie) + S->>S: rate limit (ip + session) + S->>A: run_turn(session_id, message) + A->>A: SessionStore.get_or_create
+ per-session lock + A->>C: messages.create(tools, system, history) + loop tool_use + C-->>A: tool_use blocks + A->>TL: dispatch_tool(name, args, state) + TL-->>A: tool result + A->>C: messages.create(history+tool_result) + end + C-->>A: final text + A->>A: validate_reply (layer 4) + A-->>S: reply text + S-->>B: { reply } + ``` + + # Module layout + + Five Python files form the core. They depend on each other in one + direction only -- there are no cycles. + + ```mermaid + graph LR + MD[mock_data.py
ORDERS, POLICIES, RETURN_POLICY] + C[config.py
Settings] + T[tools.py
schemas, handlers, dispatch] + A[agent.py
SessionStore, run_turn, validate] + SV[server.py
FastAPI, middleware, routes] + + MD --> T + MD --> A + C --> T + C --> A + C --> SV + T --> A + A --> SV + ``` + + The rest of this document visits each module in dependency order: + configuration first, then the data fixtures they read, then tools, + then the agent loop, then the HTTP layer on top. + + # Configuration + + Every setting that might reasonably change between environments + lives in one place. The two required values -- the Anthropic API + key and the session-cookie signing secret -- are wrapped in + `SecretStr` so an accidental `print(settings)` cannot leak them to + a log. + + Everything else has a default that is safe for local development + and reasonable for a small production deployment. A few knobs are + worth noticing: + + - `max_tool_use_iterations` bounds the Layer-3 loop in `agent.py`. + A model that keeps asking for tools forever will not burn API + credit forever. + - `session_store_max_entries` and `session_idle_ttl_seconds` cap + the in-memory `SessionStore`, so a trivial script that opens + millions of sessions cannot OOM the process. + - `rate_limit_per_ip_per_minute` and + `rate_limit_per_session_per_minute` feed the sliding-window + limiter in `server.py`. + + """ + ) + + out += _chunk("python", "config-py", "config.py", config_py) + "\n\n" + + out += textwrap.dedent( + """\ + # Data fixtures + + Bookly does not talk to a real database. Four fixture orders are + enough to cover the interesting scenarios: a delivered order that + is still inside the 30-day return window, an in-flight order that + has not been delivered yet, a processing order that has not + shipped, and an old delivered order outside the return window. + Sarah Chen owns two of the four so the agent has to disambiguate + when she says "my order". + + The `RETURN_POLICY` dict is the single source of truth for policy + facts. Two things read it: the system prompt (via + `_format_return_policy_block` in `agent.py`, which renders it as + the `
Bookly
+Introduction
+Bookly is a customer-support chatbot for a bookstore. It handles +three things: looking up orders, processing returns, and answering a +small set of standard policy questions. Everything else it refuses, +using a verbatim template.
+The interesting engineering is not the feature set. It is the +guardrails. A chat agent wired to real tools can hallucinate order +details, leak private information, skip verification steps, or wander +off topic – and the consequences land on real customers. Bookly defends +against that with four independent layers, each of which assumes the +previous layers have failed.
+This document is both the prose walkthrough and the source code. The +code you see below is the code that runs. Tangling this file produces +the Python source tree byte-for-byte; weaving it produces the HTML you +are reading.
+The four guardrail layers
+Before anything else, it helps to see the layers laid out in one +picture. Each layer is a separate defence, and a malicious or confused +input has to defeat all of them to cause harm.
+
Layer 1 is the system prompt itself. It tells the model what Bookly
+is, what it can and cannot help with, what the return policy actually
+says (quoted verbatim, not paraphrased), and exactly which template to
+use when refusing. Layer 2 adds short reminder blocks on every turn so
+the model re-reads the non-negotiable rules at the highest-attention
+position right before the user turn. Layer 3 lives in
+tools.py: the tool handlers refuse unsafe calls regardless
+of what the model decides. Layer 4 lives at the end of the agent loop
+and does a deterministic regex pass over the final reply looking for
+things like fabricated order IDs, markdown leakage, and off-topic
+engagement.
Request lifecycle
+A single user message travels this path:
+
Module layout
+Five Python files form the core. They depend on each other in one +direction only – there are no cycles.
+
The rest of this document visits each module in dependency order: +configuration first, then the data fixtures they read, then tools, then +the agent loop, then the HTTP layer on top.
+Configuration
+Every setting that might reasonably change between environments lives
+in one place. The two required values – the Anthropic API key and the
+session-cookie signing secret – are wrapped in SecretStr so
+an accidental print(settings) cannot leak them to a
+log.
Everything else has a default that is safe for local development and +reasonable for a small production deployment. A few knobs are worth +noticing:
+-
+
max_tool_use_iterationsbounds the Layer-3 loop in +agent.py. A model that keeps asking for tools forever will +not burn API credit forever.
+session_store_max_entriesand +session_idle_ttl_secondscap the in-memory +SessionStore, so a trivial script that opens millions of +sessions cannot OOM the process.
+rate_limit_per_ip_per_minuteand +rate_limit_per_session_per_minutefeed the sliding-window +limiter inserver.py.
+
"""Application configuration loaded from environment variables.
+
+Settings are read from `.env` at process start. The Anthropic API key and
+the session-cookie signing secret are the only required values; everything
+else has a sensible default so the app can boot in dev without ceremony.
+"""
+
+from __future__ import annotations
+
+import secrets
+
+from pydantic import Field, SecretStr
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+
+class Settings(BaseSettings):
+ model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8", extra="ignore")
+
+ # Required secrets -- wrapped in SecretStr so accidental logging or repr
+ # does not leak them. Access the raw value with `.get_secret_value()`.
+ anthropic_api_key: SecretStr
+ # Signing key for the server-issued session cookie. A fresh random value is
+ # generated at import if none is configured -- this means sessions do not
+ # survive a process restart in dev, which is the desired behavior until a
+ # real secret is set in the environment.
+ session_secret: SecretStr = Field(default_factory=lambda: SecretStr(secrets.token_urlsafe(32)))
+
+ anthropic_model: str = "claude-sonnet-4-5"
+ max_tokens: int = 1024
+ # Upper bound on the Anthropic HTTP call. A stuck request must not hold a
+ # worker thread forever -- see the tool-use loop cap in agent.py for the
+ # paired total-work bound.
+ anthropic_timeout_seconds: float = 30.0
+
+ server_host: str = "127.0.0.1"
+ server_port: int = 8014
+
+ # Session store bounds. Protects against a trivial DoS that opens many
+ # sessions or drives a single session to unbounded history length.
+ session_store_max_entries: int = 10_000
+ session_idle_ttl_seconds: int = 1800 # 30 minutes
+ max_turns_per_session: int = 40
+
+ # Hard cap on iterations of the tool-use loop within a single turn. The
+ # model should never legitimately need this many tool calls for a support
+ # conversation -- the cap exists to stop a runaway loop.
+ max_tool_use_iterations: int = 8
+
+ # Per-minute sliding-window rate limits. Enforced by a tiny in-memory
+ # limiter in server.py; suitable for a single-process demo deployment.
+ rate_limit_per_ip_per_minute: int = 30
+ rate_limit_per_session_per_minute: int = 20
+
+ # Session cookie configuration.
+ session_cookie_name: str = "bookly_session"
+ session_cookie_secure: bool = False # Flip to True behind HTTPS.
+ session_cookie_max_age_seconds: int = 60 * 60 * 8 # 8 hours
+
+
+# The type ignore is needed because pydantic-settings reads `anthropic_api_key`
+# and `session_secret` from environment / .env at runtime, but mypy sees them as
+# required constructor arguments and has no way to know about that.
+settings = Settings() # type: ignore[call-arg]Data fixtures
+Bookly does not talk to a real database. Four fixture orders are +enough to cover the interesting scenarios: a delivered order that is +still inside the 30-day return window, an in-flight order that has not +been delivered yet, a processing order that has not shipped, and an old +delivered order outside the return window. Sarah Chen owns two of the +four so the agent has to disambiguate when she says “my order”.
+The RETURN_POLICY dict is the single source of truth for
+policy facts. Two things read it: the system prompt (via
+_format_return_policy_block in agent.py, which
+renders it as the <return_policy> section the model
+must quote) and the check_return_eligibility handler (which
+enforces the window in code). Having one copy prevents the two from
+drifting apart.
POLICIES is a tiny FAQ keyed by topic. The
+lookup_policy tool returns one of these entries verbatim
+and the system prompt instructs the model to quote the response without
+paraphrasing. This is a deliberate anti-hallucination pattern: the less
+the model has to generate, the less it can make up.
RETURNS is the only mutable state in this file.
+initiate_return writes a new RMA record to it on each
+successful return.
"""In-memory data fixtures for orders, returns, and FAQ policies.
+
+`ORDERS` and `RETURN_POLICY` are read by both the system prompt (so the prompt
+quotes policy verbatim instead of paraphrasing) and the tool handlers (so the
+two never drift apart). `RETURNS` is mutated by `initiate_return` at runtime.
+"""
+
+from datetime import date, timedelta
+
+# A frozen "today" so the four-order fixture stays deterministic across runs.
+TODAY = date(2026, 4, 14)
+
+
+def _days_ago(n: int) -> str:
+ return (TODAY - timedelta(days=n)).isoformat()
+
+
+RETURN_POLICY: dict = {
+ "window_days": 30,
+ "condition_requirements": "Items must be unread, undamaged, and in their original packaging.",
+ "refund_method": "Refunds are issued to the original payment method.",
+ "refund_timeline_days": 7,
+ "non_returnable_categories": ["ebooks", "audiobooks", "gift cards", "personalized items"],
+}
+
+
+# Four orders covering the interesting scenarios. Sarah Chen has two orders so
+# the agent must disambiguate when she says "my order".
+ORDERS: dict = {
+ "BK-10042": {
+ "order_id": "BK-10042",
+ "customer_name": "Sarah Chen",
+ "email": "sarah.chen@example.com",
+ "status": "delivered",
+ "order_date": _days_ago(20),
+ "delivered_date": _days_ago(15),
+ "tracking_number": "1Z999AA10123456784",
+ "items": [
+ {"title": "The Goldfinch", "author": "Donna Tartt", "price": 16.99, "category": "fiction"},
+ {"title": "Sapiens", "author": "Yuval Noah Harari", "price": 19.99, "category": "nonfiction"},
+ ],
+ "total": 36.98,
+ },
+ "BK-10089": {
+ "order_id": "BK-10089",
+ "customer_name": "James Murphy",
+ "email": "james.murphy@example.com",
+ "status": "shipped",
+ "order_date": _days_ago(4),
+ "delivered_date": None,
+ "tracking_number": "1Z999AA10987654321",
+ "items": [
+ {"title": "Project Hail Mary", "author": "Andy Weir", "price": 18.99, "category": "fiction"},
+ ],
+ "total": 18.99,
+ },
+ "BK-10103": {
+ "order_id": "BK-10103",
+ "customer_name": "Sarah Chen",
+ "email": "sarah.chen@example.com",
+ "status": "processing",
+ "order_date": _days_ago(1),
+ "delivered_date": None,
+ "tracking_number": None,
+ "items": [
+ {"title": "Tomorrow, and Tomorrow, and Tomorrow", "author": "Gabrielle Zevin", "price": 17.99, "category": "fiction"},
+ ],
+ "total": 17.99,
+ },
+ "BK-9871": {
+ "order_id": "BK-9871",
+ "customer_name": "Maria Gonzalez",
+ "email": "maria.gonzalez@example.com",
+ "status": "delivered",
+ "order_date": _days_ago(60),
+ "delivered_date": _days_ago(55),
+ "tracking_number": "1Z999AA10555555555",
+ "items": [
+ {"title": "The Midnight Library", "author": "Matt Haig", "price": 15.99, "category": "fiction"},
+ ],
+ "total": 15.99,
+ },
+}
+
+
+# Verbatim FAQ entries returned by `lookup_policy`. The agent quotes these
+# without paraphrasing.
+POLICIES: dict[str, str] = {
+ "shipping": (
+ "Standard shipping is free on orders over $25 and takes 3-5 business days. "
+ "Expedited shipping (1-2 business days) is $9.99. We ship to all 50 US states. "
+ "International shipping is not currently available."
+ ),
+ "password_reset": (
+ "To reset your password, go to bookly.com/account/login and click \"Forgot password.\" "
+ "Enter the email on your account and we will send you a reset link. "
+ "The link expires after 24 hours. If you do not receive the email, check your spam folder."
+ ),
+ "returns_overview": (
+ "You can return most items within 30 days of delivery for a full refund to your original "
+ "payment method. Items must be unread, undamaged, and in their original packaging. "
+ "Ebooks, audiobooks, gift cards, and personalized items are not returnable. "
+ "Refunds typically post within 7 business days of us receiving the return."
+ ),
+}
+
+
+# Mutated at runtime by `initiate_return`. Keyed by return_id.
+RETURNS: dict[str, dict] = {}Tools: Layer 3 enforcement
+Four tools back the agent: lookup_order,
+check_return_eligibility, initiate_return, and
+lookup_policy. Each has an Anthropic-format schema (used in
+the tools argument to messages.create) and a
+handler function that takes a validated arg dict plus the per-session
+guard state and returns a dict that becomes the tool_result
+content sent back to the model.
The most important guardrail in the entire system lives in this file.
+handle_initiate_return refuses unless
+check_return_eligibility has already succeeded for the same
+order in the same session. This is enforced in code, not in the prompt –
+if a model somehow decides to skip the eligibility check, the tool
+itself refuses. This is “Layer 3” in the stack: the model’s last line of
+defence against itself.
A second guardrail is the privacy boundary in
+handle_lookup_order. When a caller supplies a
+customer_email and it does not match the email on the
+order, the handler returns the same order_not_found error
+as a missing order. This mirror means an attacker cannot probe for which
+order IDs exist by watching response differences. The check uses
+hmac.compare_digest for constant-time comparison so
+response-time side channels cannot leak the correct email prefix
+either.
Input validation lives in _require_* helpers at the top
+of the file. Every string is control-character-stripped before length
+checks so a malicious \x00 byte injected into a tool arg
+cannot sneak into the tool result JSON and reappear in the next turn’s
+prompt. Order IDs, emails, and policy topics are validated with tight
+regexes; unexpected input becomes a structured
+invalid_arguments error that the model can recover from on
+its next turn.
TypedDict argument shapes make the schema-to-handler
+contract visible to the type checker without losing runtime validation –
+the model is an untrusted caller, so the runtime checks stay.
"""Tool schemas, dispatch, and Layer 3 (tool-side) guardrail enforcement.
+
+Each tool has an Anthropic-format schema (used in the `tools` argument to
+`messages.create`) and a handler. Handlers are typed with `TypedDict`s so the
+contract between schema and handler is visible to the type checker; inputs
+are still validated at runtime because the caller is ultimately the model.
+
+The most important guardrail in the whole system lives here:
+`handle_initiate_return` refuses unless `check_return_eligibility` has already
+succeeded for the same order in the same session. This protects against the
+agent skipping the protocol even if the system prompt is ignored entirely.
+"""
+
+from __future__ import annotations
+
+import hmac
+import re
+import uuid
+from dataclasses import dataclass, field
+from datetime import date
+from typing import Any, Callable, TypedDict
+
+try:
+ from typing import NotRequired # Python 3.11+
+except ImportError: # pragma: no cover -- Python 3.10 fallback
+ from typing_extensions import NotRequired # type: ignore[assignment]
+
+from mock_data import ORDERS, POLICIES, RETURN_POLICY, RETURNS, TODAY
+
+
+# ---------------------------------------------------------------------------
+# Validation helpers
+# ---------------------------------------------------------------------------
+
+# Validator limits. These are deliberately tight: tool arguments come from
+# model output, which in turn reflects user input, so anything that would not
+# plausibly appear in a real support conversation is rejected.
+ORDER_ID_RE = re.compile(r"^BK-\d{4,6}$")
+EMAIL_RE = re.compile(r"^[^@\s]{1,64}@[^@\s]{1,255}\.[^@\s]{1,10}$")
+TOPIC_RE = re.compile(r"^[a-z][a-z_]{0,39}$")
+ITEM_TITLE_MAX_LENGTH = 200
+REASON_MAX_LENGTH = 500
+ITEM_TITLES_MAX_COUNT = 50
+
+# Control characters are stripped from any free-form input. Keeping them out
+# of tool payloads means they cannot end up in prompts on later turns, which
+# closes one prompt-injection surface.
+_CONTROL_CHAR_RE = re.compile(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f]")
+
+
+class ToolValidationError(ValueError):
+ """Raised when a tool argument fails validation.
+
+ The dispatcher catches this and converts it into a tool-result error so
+ the model can recover on its next turn instead of crashing the request.
+ """
+
+
+def _require_string(value: Any, field_name: str, *, max_length: int) -> str:
+ if not isinstance(value, str):
+ raise ToolValidationError(f"{field_name} must be a string")
+ cleaned = _CONTROL_CHAR_RE.sub("", value).strip()
+ if not cleaned:
+ raise ToolValidationError(f"{field_name} is required")
+ if len(cleaned) > max_length:
+ raise ToolValidationError(f"{field_name} must be at most {max_length} characters")
+ return cleaned
+
+
+def _require_order_id(value: Any) -> str:
+ order_id = _require_string(value, "order_id", max_length=16)
+ if not ORDER_ID_RE.match(order_id):
+ raise ToolValidationError("order_id must match the format BK-NNNN")
+ return order_id
+
+
+def _require_email(value: Any, *, field_name: str = "customer_email") -> str:
+ email = _require_string(value, field_name, max_length=320)
+ if not EMAIL_RE.match(email):
+ raise ToolValidationError(f"{field_name} is not a valid email address")
+ return email
+
+
+def _optional_email(value: Any, *, field_name: str = "customer_email") -> str | None:
+ if value is None:
+ return None
+ return _require_email(value, field_name=field_name)
+
+
+def _require_topic(value: Any) -> str:
+ topic = _require_string(value, "topic", max_length=40)
+ topic = topic.lower()
+ if not TOPIC_RE.match(topic):
+ raise ToolValidationError("topic must be lowercase letters and underscores only")
+ return topic
+
+
+def _optional_item_titles(value: Any) -> list[str] | None:
+ if value is None:
+ return None
+ if not isinstance(value, list):
+ raise ToolValidationError("item_titles must be a list of strings")
+ if len(value) > ITEM_TITLES_MAX_COUNT:
+ raise ToolValidationError(f"item_titles may contain at most {ITEM_TITLES_MAX_COUNT} entries")
+ cleaned: list[str] = []
+ for index, entry in enumerate(value):
+ cleaned.append(_require_string(entry, f"item_titles[{index}]", max_length=ITEM_TITLE_MAX_LENGTH))
+ return cleaned
+
+
+def _emails_match(supplied: str | None, stored: str | None) -> bool:
+ """Constant-time email comparison with normalization.
+
+ Returns False if either side is missing. Uses `hmac.compare_digest` to
+ close the timing side-channel that would otherwise leak the correct
+ prefix of a stored email.
+ """
+ if supplied is None or stored is None:
+ return False
+ supplied_norm = supplied.strip().lower().encode("utf-8")
+ stored_norm = stored.strip().lower().encode("utf-8")
+ return hmac.compare_digest(supplied_norm, stored_norm)
+
+
+def _is_within_return_window(delivered_date: str | None) -> tuple[bool, int | None]:
+ """Return (within_window, days_since_delivery)."""
+ if delivered_date is None:
+ return (False, None)
+ delivered = date.fromisoformat(delivered_date)
+ days_since = (TODAY - delivered).days
+ return (days_since <= RETURN_POLICY["window_days"], days_since)
+
+
+# ---------------------------------------------------------------------------
+# TypedDict argument shapes
+# ---------------------------------------------------------------------------
+
+
+class LookupOrderArgs(TypedDict, total=False):
+ order_id: str
+ customer_email: NotRequired[str]
+
+
+class CheckReturnEligibilityArgs(TypedDict):
+ order_id: str
+ customer_email: str
+
+
+class InitiateReturnArgs(TypedDict, total=False):
+ order_id: str
+ customer_email: str
+ reason: str
+ item_titles: NotRequired[list[str]]
+
+
+class LookupPolicyArgs(TypedDict):
+ topic: str
+
+
+@dataclass
+class SessionGuardState:
+ """Per-session protocol state used to enforce tool ordering rules.
+
+ Sessions are short-lived chats, so plain in-memory sets are fine. A
+ production deployment would back this with a session store.
+ """
+
+ eligibility_checks_passed: set[str] = field(default_factory=set)
+ returns_initiated: set[str] = field(default_factory=set)
+
+
+# ---------------------------------------------------------------------------
+# Tool schemas (Anthropic format)
+# ---------------------------------------------------------------------------
+
+LOOKUP_ORDER_SCHEMA: dict[str, Any] = {
+ "name": "lookup_order",
+ "description": (
+ "Look up the status and details of a Bookly order by order ID. "
+ "Optionally pass the customer email to verify ownership before returning details. "
+ "Use this whenever the customer asks about an order."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "order_id": {
+ "type": "string",
+ "description": "The order ID, formatted as 'BK-' followed by digits.",
+ },
+ "customer_email": {
+ "type": "string",
+ "description": "Optional email used to verify the customer owns the order.",
+ },
+ },
+ "required": ["order_id"],
+ },
+}
+
+CHECK_RETURN_ELIGIBILITY_SCHEMA: dict[str, Any] = {
+ "name": "check_return_eligibility",
+ "description": (
+ "Check whether an order is eligible for return. Requires both order ID and the email "
+ "on the order. Must be called and succeed before initiate_return."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "order_id": {"type": "string"},
+ "customer_email": {"type": "string"},
+ },
+ "required": ["order_id", "customer_email"],
+ },
+}
+
+INITIATE_RETURN_SCHEMA: dict[str, Any] = {
+ "name": "initiate_return",
+ "description": (
+ "Start a return for an order. Only call this after check_return_eligibility has "
+ "succeeded for the same order in this conversation, and after the customer has "
+ "confirmed they want to proceed."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "order_id": {"type": "string"},
+ "customer_email": {"type": "string"},
+ "reason": {
+ "type": "string",
+ "description": "The customer's stated reason for the return.",
+ },
+ "item_titles": {
+ "type": "array",
+ "items": {"type": "string"},
+ "description": "Optional list of specific item titles to return. Defaults to all items.",
+ },
+ },
+ "required": ["order_id", "customer_email", "reason"],
+ },
+}
+
+LOOKUP_POLICY_SCHEMA: dict[str, Any] = {
+ "name": "lookup_policy",
+ "description": (
+ "Look up a Bookly customer policy by topic. Use this whenever the customer asks "
+ "about shipping, password reset, returns overview, or similar standard policies. "
+ "Returns the verbatim policy text or topic_not_supported."
+ ),
+ "input_schema": {
+ "type": "object",
+ "properties": {
+ "topic": {
+ "type": "string",
+ "description": "Policy topic, e.g. 'shipping', 'password_reset', 'returns_overview'.",
+ },
+ },
+ "required": ["topic"],
+ },
+ # Cache breakpoint: marking the last tool with `cache_control` extends the
+ # prompt cache over the whole tools block so schemas are not re-tokenized
+ # on every turn. The big system prompt already has its own breakpoint.
+ "cache_control": {"type": "ephemeral"},
+}
+
+TOOL_SCHEMAS: list[dict[str, Any]] = [
+ LOOKUP_ORDER_SCHEMA,
+ CHECK_RETURN_ELIGIBILITY_SCHEMA,
+ INITIATE_RETURN_SCHEMA,
+ LOOKUP_POLICY_SCHEMA,
+]
+
+
+# ---------------------------------------------------------------------------
+# Handlers
+# ---------------------------------------------------------------------------
+
+
+def handle_lookup_order(args: LookupOrderArgs, state: SessionGuardState) -> dict[str, Any]:
+ order_id = _require_order_id(args.get("order_id"))
+ customer_email = _optional_email(args.get("customer_email"))
+
+ order = ORDERS.get(order_id)
+ if order is None:
+ return {"error": "order_not_found", "message": f"No order found with ID {order_id}."}
+
+ # Privacy: when an email is supplied and does not match, return the same
+ # error as a missing order so callers cannot enumerate which IDs exist.
+ if customer_email is not None and not _emails_match(customer_email, order["email"]):
+ return {"error": "order_not_found", "message": f"No order found with ID {order_id}."}
+
+ return {"order": order}
+
+
+def handle_check_return_eligibility(
+ args: CheckReturnEligibilityArgs, state: SessionGuardState
+) -> dict[str, Any]:
+ order_id = _require_order_id(args.get("order_id"))
+ customer_email = _require_email(args.get("customer_email"))
+
+ order = ORDERS.get(order_id)
+ if order is None or not _emails_match(customer_email, order["email"]):
+ return {
+ "error": "auth_failed",
+ "message": "Could not verify that order ID and email together. Please double-check both.",
+ }
+
+ if order["status"] != "delivered":
+ return {
+ "eligible": False,
+ "reason": (
+ f"This order has status '{order['status']}', not 'delivered'. "
+ "Returns can only be started after an order has been delivered."
+ ),
+ "policy": RETURN_POLICY,
+ }
+
+ within_window, days_since = _is_within_return_window(order.get("delivered_date"))
+ if not within_window:
+ return {
+ "eligible": False,
+ "reason": (
+ f"This order was delivered {days_since} days ago, which is outside the "
+ f"{RETURN_POLICY['window_days']}-day return window."
+ ),
+ "policy": RETURN_POLICY,
+ }
+
+ state.eligibility_checks_passed.add(order_id)
+ return {
+ "eligible": True,
+ "reason": (
+ f"Order delivered {days_since} days ago, within the "
+ f"{RETURN_POLICY['window_days']}-day window."
+ ),
+ "items": order["items"],
+ "policy": RETURN_POLICY,
+ }
+
+
+def handle_initiate_return(args: InitiateReturnArgs, state: SessionGuardState) -> dict[str, Any]:
+ order_id = _require_order_id(args.get("order_id"))
+ customer_email = _require_email(args.get("customer_email"))
+ reason = _require_string(args.get("reason"), "reason", max_length=REASON_MAX_LENGTH)
+ item_titles = _optional_item_titles(args.get("item_titles"))
+
+ # Layer 3 protocol guard: the agent must have called check_return_eligibility
+ # for this exact order in this session, and it must have passed.
+ if order_id not in state.eligibility_checks_passed:
+ return {
+ "error": "eligibility_not_verified",
+ "message": (
+ "Cannot initiate a return without a successful eligibility check for this "
+ "order in the current session. Call check_return_eligibility first."
+ ),
+ }
+
+ if order_id in state.returns_initiated:
+ return {
+ "error": "already_initiated",
+ "message": "A return has already been initiated for this order in this session.",
+ }
+
+ order = ORDERS.get(order_id)
+ # Paired assertion: we already checked eligibility against the same order,
+ # but re-verify here so a future edit that makes ORDERS mutable cannot
+ # silently break the email-binding guarantee.
+ if order is None or not _emails_match(customer_email, order["email"]):
+ return {"error": "auth_failed", "message": "Order/email mismatch."}
+
+ # Explicit: an empty list means "no items selected" (a caller error we
+ # reject) while `None` means "default to all items on the order".
+ if item_titles is not None and not item_titles:
+ return {"error": "no_items_selected", "message": "item_titles cannot be an empty list."}
+ titles = item_titles if item_titles is not None else [item["title"] for item in order["items"]]
+
+ return_id = f"RMA-{uuid.uuid4().hex[:8].upper()}"
+ record = {
+ "return_id": return_id,
+ "order_id": order_id,
+ "customer_email": order["email"],
+ "items": titles,
+ "reason": reason,
+ "refund_method": RETURN_POLICY["refund_method"],
+ "refund_timeline_days": RETURN_POLICY["refund_timeline_days"],
+ "next_steps": (
+ "We've emailed a prepaid shipping label to the address on file. Drop the package at "
+ "any carrier location within 14 days. Your refund will post within "
+ f"{RETURN_POLICY['refund_timeline_days']} business days of us receiving the return."
+ ),
+ }
+ RETURNS[return_id] = record
+ state.returns_initiated.add(order_id)
+ return record
+
+
+def handle_lookup_policy(args: LookupPolicyArgs, state: SessionGuardState) -> dict[str, Any]:
+ topic = _require_topic(args.get("topic"))
+
+ text = POLICIES.get(topic)
+ if text is None:
+ return {
+ "error": "topic_not_supported",
+ # Echo the normalized topic, not the raw input, so nothing the
+ # caller injected is ever reflected back into model context.
+ "message": f"No policy entry for topic '{topic}'.",
+ "available_topics": sorted(POLICIES.keys()),
+ }
+ return {"topic": topic, "text": text}
+
+
+# ---------------------------------------------------------------------------
+# Dispatch
+# ---------------------------------------------------------------------------
+
+
+_HANDLERS: dict[str, Callable[[Any, SessionGuardState], dict[str, Any]]] = {
+ "lookup_order": handle_lookup_order,
+ "check_return_eligibility": handle_check_return_eligibility,
+ "initiate_return": handle_initiate_return,
+ "lookup_policy": handle_lookup_policy,
+}
+
+
+def dispatch_tool(name: str, args: dict[str, Any], state: SessionGuardState) -> dict[str, Any]:
+ handler = _HANDLERS.get(name)
+ if handler is None:
+ return {"error": "unknown_tool", "message": f"No tool named {name}."}
+ if not isinstance(args, dict):
+ return {"error": "invalid_arguments", "message": "Tool arguments must be an object."}
+ try:
+ return handler(args, state)
+ except ToolValidationError as exc:
+ # Return validation errors as structured tool errors so the model can
+ # recover. Never surface the message verbatim from untrusted input --
+ # `_require_string` already stripped control characters, and the error
+ # messages themselves are constructed from field names, not user data.
+ return {"error": "invalid_arguments", "message": str(exc)}Agent loop
+This is the biggest file. It wires everything together: the system +prompt, runtime reminders, output validation (Layer 4), the in-memory +session store with per-session locking, the cached Anthropic client, and +the actual tool-use loop that drives a turn end to end.
+System prompt
+The prompt is structured with XML-style tags
+(<identity>, <critical_rules>,
+<scope>, <return_policy>,
+<tool_rules>, <tone>,
+<examples>, <reminders>). The
+critical rules are stated up front and repeated at the bottom (primacy
+plus recency). The return policy section interpolates the
+RETURN_POLICY dict verbatim via
+_format_return_policy_block, so the prompt and the
+enforcement in tools.py cannot disagree.
Four few-shot examples are embedded directly in the prompt. Each one +demonstrates a case that is easy to get wrong: missing order ID, quoting +a policy verbatim, refusing an off-topic request, disambiguating between +two orders.
+Runtime reminders
+On every turn, build_system_content appends a short
+CRITICAL_REMINDER block to the system content. Once the
+turn count crosses LONG_CONVERSATION_TURN_THRESHOLD, a
+second LONG_CONVERSATION_REMINDER is added. The big
+SYSTEM_PROMPT block is the only one marked
+cache_control: ephemeral – the reminders vary per turn and
+we want them at the highest-attention position, not in the cached
+prefix.
Layer 4 output validation
+After the model produces its final reply, validate_reply
+runs four cheap deterministic checks: every BK-NNNN string
+in the reply must also appear in a tool result from this turn, every ISO
+date in the reply must appear in a tool result, the reply must not
+contain markdown, and if the reply contains off-topic engagement phrases
+it must also contain the refusal template. Violations are collected and
+returned as a frozen ValidationResult.
The off-topic patterns used to be loose substring matches on a +keyword set. That false-positived on plenty of legitimate support +replies (“I’d recommend contacting…”). The current patterns use word +boundaries so only the intended phrases trip them.
+Session store
+SessionStore is a bounded in-memory LRU with an idle
+TTL. It stores Session objects (history, guard state, turn
+count) keyed by opaque server-issued session IDs. It also owns the
+per-session locks used to serialize concurrent turns for the same
+session, since FastAPI runs the sync chat handler in a
+threadpool and two simultaneous requests for the same session would
+otherwise corrupt the conversation history.
The locks-dict is itself protected by a class-level lock so two +threads trying to create the first lock for a session cannot race into +two different lock instances.
+Under the “single-process demo deployment” constraint this is enough. +For multi-worker, the whole class would get swapped for a Redis-backed +equivalent.
+The tool-use loop
+_run_tool_use_loop drives the model until it stops
+asking for tools. It is bounded by
+settings.max_tool_use_iterations so a runaway model cannot
+burn credit in an infinite loop. Each iteration serializes the
+assistant’s content blocks into history, dispatches every requested
+tool, packs the results into a single tool_result user-role
+message, and calls Claude again. Before each call,
+_with_last_message_cache_breakpoint stamps the last message
+with cache_control: ephemeral so prior turns do not need to
+be re-tokenized on every call. This turns the per-turn input-token cost
+from O(turns^2) into O(turns) across a
+session.
run_turn
+run_turn is the top-level entry point the server calls.
+It validates its inputs, acquires the per-session lock, appends the user
+message, runs the loop, and then either persists the final reply to
+history or – on validation failure – drops the bad reply and returns a
+safe fallback. Dropping a bad reply from history is important: it
+prevents a hallucinated claim from poisoning subsequent turns.
Warning logs never include the reply body. Session IDs and reply +contents are logged only as short SHA-256 hashes for correlation, which +keeps PII out of the log pipeline even under active incident +response.
+"""Bookly agent: system prompt, guardrails, session store, and the agentic loop.
+
+This module wires four guardrail layers together:
+
+1. The system prompt itself (XML-tagged, primacy+recency duplication, verbatim
+ policy block, refusal template, few-shot examples for edge cases).
+2. Runtime reminder injection: a short "non-negotiable rules" block appended
+ to the system content on every turn, plus a stronger reminder once the
+ conversation gets long enough that the original prompt has decayed in
+ effective attention.
+3. Tool-side enforcement (lives in `tools.py`): handlers refuse unsafe calls
+ regardless of what the model decides.
+4. Output validation: deterministic regex checks on the final reply for
+ ungrounded order IDs/dates, markdown leakage, and off-topic engagement
+ without the refusal template. On failure, the bad reply is dropped and the
+ user gets a safe canned message -- and the bad reply is never appended to
+ history, so it cannot poison subsequent turns.
+
+Anthropic prompt caching is enabled on the large system-prompt block AND on
+the last tool schema and the last message in history, so per-turn input cost
+scales linearly in the number of turns instead of quadratically.
+"""
+
+from __future__ import annotations
+
+import functools
+import hashlib
+import json
+import logging
+import re
+import threading
+import time
+from collections import OrderedDict
+from dataclasses import dataclass, field
+from typing import Any
+
+from anthropic import Anthropic
+
+from config import settings
+from tools import SessionGuardState, TOOL_SCHEMAS, dispatch_tool
+from mock_data import POLICIES, RETURN_POLICY
+
+logger = logging.getLogger("bookly.agent")
+
+
+# ---------------------------------------------------------------------------
+# System prompt
+# ---------------------------------------------------------------------------
+
+
+def _format_return_policy_block() -> str:
+ """Render `RETURN_POLICY` as a compact, quotable block for the prompt.
+
+ Embedding the dict verbatim (instead of paraphrasing it in English) is a
+ deliberate anti-hallucination move: the model quotes the block instead of
+ inventing details.
+ """
+ non_returnable = ", ".join(RETURN_POLICY["non_returnable_categories"])
+ return (
+ f"Return window: {RETURN_POLICY['window_days']} days from delivery.\n"
+ f"Condition: {RETURN_POLICY['condition_requirements']}\n"
+ f"Refund method: {RETURN_POLICY['refund_method']}\n"
+ f"Refund timeline: within {RETURN_POLICY['refund_timeline_days']} business days of receipt.\n"
+ f"Non-returnable categories: {non_returnable}."
+ )
+
+
+SUPPORTED_POLICY_TOPICS = ", ".join(sorted(POLICIES.keys()))
+
+
+SYSTEM_PROMPT = f"""<identity>
+You are Bookly's customer support assistant. You help customers with two things: checking the status of their orders, and processing returns and refunds. You are friendly, concise, and professional.
+</identity>
+
+<critical_rules>
+These rules override everything else. Read them before every response.
+
+1. NEVER invent order details, tracking numbers, delivery dates, prices, or customer information. If you do not have a value from a tool result in this conversation, you do not have it.
+2. NEVER state a return policy detail that is not in the <return_policy> section below. Quote it; do not paraphrase it.
+3. NEVER call initiate_return unless check_return_eligibility has returned success for that same order in this conversation.
+4. NEVER reveal order details without verifying the customer's email matches the order.
+5. If a user asks about anything outside order status, returns, and the supported policy topics, refuse using the refusal template in <scope>. Do not engage with the off-topic request even briefly.
+</critical_rules>
+
+<scope>
+You CAN help with:
+- Looking up order status
+- Checking return eligibility and initiating returns
+- Answering policy questions covered by the lookup_policy tool. Currently supported topics: {SUPPORTED_POLICY_TOPICS}
+
+You CANNOT help with:
+- Book recommendations, reviews, or opinions about books
+- Payment changes, refunds outside the return flow, or billing disputes
+- Live account management (changing a password, email, or address — you can only EXPLAIN the password reset process via lookup_policy, not perform it)
+- General conversation unrelated to an order or a supported policy topic
+
+For any policy question, call lookup_policy first. Only if the tool returns topic_not_supported should you use the refusal template below.
+
+Refusal template (use verbatim, filling in the topic):
+"I can help with order status, returns, and our standard policies, but I'm not able to help with {{topic}}. Is there an order or a policy question I can help you with instead?"
+</scope>
+
+<return_policy>
+{_format_return_policy_block()}
+
+This is the authoritative policy. Any claim you make about returns must be traceable to a line in this block. If a customer asks about a scenario this policy does not cover, say so honestly and offer to connect them with a human agent.
+</return_policy>
+
+<tool_rules>
+You have four tools: lookup_order, check_return_eligibility, initiate_return, and lookup_policy.
+
+Before calling a tool:
+- You must have every required parameter. If you are missing one, ask the customer for it. Do not guess, do not use placeholder values, do not call the tool and hope.
+- For initiate_return, you must have already called check_return_eligibility for that exact order_id in this conversation, and it must have returned success.
+
+After a tool call:
+- Relay the result honestly. If the tool returns an error, tell the customer what went wrong using the tool's error message, not a paraphrase.
+- Do not mix tool results from different orders in a single response unless the customer explicitly asked about multiple.
+- For lookup_policy, quote the returned policy text; do not summarize or embellish. If lookup_policy returns topic_not_supported, fall through to the refusal template in <scope>.
+</tool_rules>
+
+<clarifying_rules>
+Ask one clarifying question at a time, not a list. Common cases:
+
+- Customer mentions "my order" without an order ID: ask for the order ID. Tell them it starts with "BK-" and is in their confirmation email.
+- Customer gives an order ID but no email, and wants a return: ask for the email on the order.
+- A customer has multiple orders and was ambiguous: ask which order they mean, listing them by ID and status only.
+- Customer wants to initiate a return: after eligibility is confirmed, summarize what will happen (which items, refund method, timeline) and ask for explicit confirmation before calling initiate_return.
+</clarifying_rules>
+
+<tone>
+- Friendly and warm, but not chatty. One or two sentences per turn is usually right.
+- Use the customer's first name once you know it, but not in every message.
+- Plain text only. No markdown, no bullet points, no headers, no asterisks for emphasis. The chat UI does not render markdown.
+- Never apologize more than once for the same issue.
+</tone>
+
+<examples>
+Example 1 — missing order ID:
+User: "Where's my order?"
+Assistant: "Happy to check on that for you. Could you share your order ID? It starts with 'BK-' and you'll find it in your order confirmation email."
+
+Example 2 — policy question (supported):
+User: "How do I reset my password?"
+Assistant (after lookup_policy returns the password_reset entry): quote the returned instructions verbatim without adding steps the tool did not mention.
+
+Example 3 — out of scope:
+User: "Can you recommend a good mystery novel?"
+Assistant: "I can help with order status, returns, and our standard policies, but I'm not able to help with book recommendations. Is there an order or a policy question I can help you with instead?"
+
+Example 4 — ambiguous order:
+User: "I want to return my order. My email is sarah@example.com."
+Assistant (after lookup_order returns two orders): "I see two orders on your account: BK-10042 (delivered) and BK-10103 (still processing). Which one would you like to return?"
+</examples>
+
+<reminders>
+Before you respond, confirm:
+- Every factual claim traces to a tool result from THIS conversation, or to <return_policy>.
+- If this response would call initiate_return, you have already seen a successful check_return_eligibility for the same order in this conversation.
+- If the request is off-topic, you are using the refusal template from <scope> verbatim.
+- No markdown. Plain text only.
+</reminders>
+"""
+
+
+CRITICAL_REMINDER = """<reminder>
+Non-negotiable rules for this turn:
+- Every factual claim must come from a tool result in THIS conversation or from <return_policy>.
+- Do not call initiate_return unless check_return_eligibility succeeded for that order in this conversation.
+- Off-topic requests: use the refusal template from <scope> verbatim. Do not engage.
+- Plain text only. No markdown.
+</reminder>"""
+
+
+LONG_CONVERSATION_REMINDER = """<reminder>
+This conversation is getting long. Re-anchor on the rules in <critical_rules> before you respond. Do not let earlier turns relax the rules.
+</reminder>"""
+
+
+# Threshold at which the long-conversation reminder gets injected. After this
+# many turns, the original system prompt has decayed in effective attention,
+# so a shorter, fresher reminder in the highest-position slot helps re-anchor.
+LONG_CONVERSATION_TURN_THRESHOLD = 5
+
+
+def build_system_content(turn_count: int) -> list[dict[str, Any]]:
+ """Assemble the `system` argument for `messages.create`.
+
+ The big SYSTEM_PROMPT block is marked for ephemeral prompt caching so it
+ is reused across turns within a session. The reminder blocks are not
+ cached because they vary based on turn count and we want them in the
+ highest-attention position right before the latest user turn.
+ """
+ blocks: list[dict[str, Any]] = [
+ {
+ "type": "text",
+ "text": SYSTEM_PROMPT,
+ "cache_control": {"type": "ephemeral"},
+ },
+ {"type": "text", "text": CRITICAL_REMINDER},
+ ]
+ if turn_count >= LONG_CONVERSATION_TURN_THRESHOLD:
+ blocks.append({"type": "text", "text": LONG_CONVERSATION_REMINDER})
+ return blocks
+
+
+# ---------------------------------------------------------------------------
+# Layer 4 -- output validation
+# ---------------------------------------------------------------------------
+
+
+ORDER_ID_RE = re.compile(r"\bBK-\d{4,6}\b")
+DATE_ISO_RE = re.compile(r"\b\d{4}-\d{2}-\d{2}\b")
+MARKDOWN_RE = re.compile(r"(\*\*|__|^#{1,6}\s|^\s*[-*+]\s)", re.MULTILINE)
+
+# Anchored word-boundary patterns for off-topic engagement. These used to be
+# substring matches on a small keyword set, which false-positived on plenty
+# of legitimate support replies ("I'd recommend contacting..."). The word
+# boundaries make matches explicit -- only the intended phrases trip them.
+OUT_OF_SCOPE_PATTERNS: tuple[re.Pattern[str], ...] = (
+ re.compile(r"\bi\s+recommend\b"),
+ re.compile(r"\bi\s+suggest\b"),
+ re.compile(r"\byou\s+should\s+read\b"),
+ re.compile(r"\bgreat\s+book\b"),
+ re.compile(r"\bfavorite\s+book\b"),
+ re.compile(r"\bwhat\s+should\s+i\s+read\b"),
+ re.compile(r"\breview\s+of\b"),
+)
+
+REFUSAL_PHRASE = "i'm not able to help with"
+
+
+@dataclass(frozen=True)
+class ValidationResult:
+ ok: bool
+ violations: tuple[str, ...] = ()
+
+
+def _collect_grounded_values(
+ tool_results: list[dict[str, Any]], pattern: re.Pattern[str]
+) -> set[str]:
+ """Pull every substring matching `pattern` out of the tool result JSON."""
+ grounded: set[str] = set()
+ for entry in tool_results:
+ text = json.dumps(entry.get("result", {}))
+ grounded.update(pattern.findall(text))
+ return grounded
+
+
+def validate_reply(reply: str, tool_results_this_turn: list[dict[str, Any]]) -> ValidationResult:
+ """Run deterministic checks on the final assistant reply.
+
+ Heuristic, not exhaustive. Catches the cheap wins -- fabricated order IDs,
+ made-up dates, markdown leakage, and obvious off-topic engagement. For
+ anything subtler we rely on layers 1-3.
+ """
+ if not isinstance(reply, str):
+ raise TypeError("reply must be a string")
+ if not isinstance(tool_results_this_turn, list):
+ raise TypeError("tool_results_this_turn must be a list")
+
+ violations: list[str] = []
+
+ grounded_ids = _collect_grounded_values(tool_results_this_turn, ORDER_ID_RE)
+ for match in ORDER_ID_RE.findall(reply):
+ if match not in grounded_ids:
+ violations.append(f"ungrounded_order_id:{match}")
+
+ grounded_dates = _collect_grounded_values(tool_results_this_turn, DATE_ISO_RE)
+ for match in DATE_ISO_RE.findall(reply):
+ if match not in grounded_dates:
+ violations.append(f"ungrounded_date:{match}")
+
+ if MARKDOWN_RE.search(reply):
+ violations.append("markdown_leaked")
+
+ lowered = reply.lower()
+ engaged_off_topic = any(pattern.search(lowered) for pattern in OUT_OF_SCOPE_PATTERNS)
+ if engaged_off_topic and REFUSAL_PHRASE not in lowered:
+ violations.append("off_topic_engagement")
+
+ return ValidationResult(ok=not violations, violations=tuple(violations))
+
+
+# ---------------------------------------------------------------------------
+# Session store
+# ---------------------------------------------------------------------------
+
+
+SAFE_FALLBACK = (
+ "I hit a problem generating a response. Could you rephrase your question, "
+ "or share an order ID so I can try again?"
+)
+
+CONVERSATION_TOO_LONG_MESSAGE = (
+ "This conversation has gone long enough that I need to reset before I keep "
+ "making mistakes. Please start a new chat and I'll be happy to help from there."
+)
+
+TOOL_LOOP_EXCEEDED_MESSAGE = (
+ "I got stuck working on that request. Could you try rephrasing it, or share "
+ "an order ID so I can try a fresh approach?"
+)
+
+
+@dataclass
+class Session:
+ history: list[dict[str, Any]] = field(default_factory=list)
+ guard_state: SessionGuardState = field(default_factory=SessionGuardState)
+ turn_count: int = 0
+
+
+class SessionStore:
+ """Bounded in-memory session store with LRU eviction and idle TTL.
+
+ Also owns the per-session locks used to serialize turns for the same
+ session_id when FastAPI runs the sync handler in its threadpool. The
+ creation of a per-session lock is itself guarded by a class-level lock to
+ avoid a double-create race.
+
+ Designed for a single-process demo deployment. For multi-worker, swap
+ this class out for a Redis-backed equivalent.
+ """
+
+ def __init__(self, *, max_entries: int, idle_ttl_seconds: int) -> None:
+ if max_entries <= 0:
+ raise ValueError("max_entries must be positive")
+ if idle_ttl_seconds <= 0:
+ raise ValueError("idle_ttl_seconds must be positive")
+ self._max_entries = max_entries
+ self._idle_ttl_seconds = idle_ttl_seconds
+ self._entries: OrderedDict[str, tuple[Session, float]] = OrderedDict()
+ self._store_lock = threading.Lock()
+ self._session_locks: dict[str, threading.Lock] = {}
+ self._locks_lock = threading.Lock()
+
+ def get_or_create(self, session_id: str) -> Session:
+ if not isinstance(session_id, str) or not session_id:
+ raise ValueError("session_id is required")
+ now = time.monotonic()
+ with self._store_lock:
+ self._evict_expired_locked(now)
+ entry = self._entries.get(session_id)
+ if entry is None:
+ session = Session()
+ self._entries[session_id] = (session, now)
+ self._enforce_size_cap_locked()
+ return session
+ session, _ = entry
+ self._entries[session_id] = (session, now)
+ self._entries.move_to_end(session_id)
+ return session
+
+ def lock_for(self, session_id: str) -> threading.Lock:
+ """Return the lock guarding turns for `session_id`, creating if new."""
+ if not isinstance(session_id, str) or not session_id:
+ raise ValueError("session_id is required")
+ with self._locks_lock:
+ lock = self._session_locks.get(session_id)
+ if lock is None:
+ lock = threading.Lock()
+ self._session_locks[session_id] = lock
+ return lock
+
+ def clear(self) -> None:
+ """Drop all sessions. Intended for tests and admin operations only."""
+ with self._store_lock:
+ self._entries.clear()
+ with self._locks_lock:
+ self._session_locks.clear()
+
+ def __len__(self) -> int:
+ with self._store_lock:
+ return len(self._entries)
+
+ def __contains__(self, session_id: object) -> bool:
+ if not isinstance(session_id, str):
+ return False
+ with self._store_lock:
+ return session_id in self._entries
+
+ def __getitem__(self, session_id: str) -> Session:
+ with self._store_lock:
+ entry = self._entries.get(session_id)
+ if entry is None:
+ raise KeyError(session_id)
+ return entry[0]
+
+ def _evict_expired_locked(self, now: float) -> None:
+ expired = [
+ sid for sid, (_, last) in self._entries.items() if now - last > self._idle_ttl_seconds
+ ]
+ for sid in expired:
+ del self._entries[sid]
+ with self._locks_lock:
+ self._session_locks.pop(sid, None)
+
+ def _enforce_size_cap_locked(self) -> None:
+ while len(self._entries) > self._max_entries:
+ sid, _ = self._entries.popitem(last=False)
+ with self._locks_lock:
+ self._session_locks.pop(sid, None)
+
+
+SESSIONS = SessionStore(
+ max_entries=settings.session_store_max_entries,
+ idle_ttl_seconds=settings.session_idle_ttl_seconds,
+)
+
+
+# ---------------------------------------------------------------------------
+# Anthropic client
+# ---------------------------------------------------------------------------
+
+
+@functools.lru_cache(maxsize=1)
+def _get_client() -> Anthropic:
+ """Return the shared Anthropic client.
+
+ Cached so every turn reuses the same HTTP connection pool. Tests swap
+ this out with `monkeypatch.setattr(agent, "_get_client", ...)`.
+ """
+ return Anthropic(
+ api_key=settings.anthropic_api_key.get_secret_value(),
+ timeout=settings.anthropic_timeout_seconds,
+ )
+
+
+# ---------------------------------------------------------------------------
+# Content serialization helpers
+# ---------------------------------------------------------------------------
+
+
+def _extract_text(content_blocks: list[Any]) -> str:
+ parts: list[str] = []
+ for block in content_blocks:
+ if getattr(block, "type", None) == "text":
+ parts.append(getattr(block, "text", ""))
+ return "".join(parts).strip()
+
+
+def _serialize_assistant_content(content_blocks: list[Any]) -> list[dict[str, Any]]:
+ """Convert SDK content blocks back into JSON-serializable dicts for history."""
+ serialized: list[dict[str, Any]] = []
+ for block in content_blocks:
+ block_type = getattr(block, "type", None)
+ if block_type == "text":
+ serialized.append({"type": "text", "text": getattr(block, "text", "")})
+ elif block_type == "tool_use":
+ serialized.append(
+ {
+ "type": "tool_use",
+ "id": block.id,
+ "name": block.name,
+ "input": getattr(block, "input", None) or {},
+ }
+ )
+ return serialized
+
+
+def _with_last_message_cache_breakpoint(
+ messages: list[dict[str, Any]],
+) -> list[dict[str, Any]]:
+ """Return a shallow-copied message list with a cache breakpoint on the last block.
+
+ Marking the last content block with `cache_control: ephemeral` extends the
+ prompt cache through the full conversation history so prior turns do not
+ need to be re-tokenized on every call. We avoid mutating the stored history
+ because the stored form should stay canonical.
+ """
+ if not messages:
+ return messages
+ head = messages[:-1]
+ last = dict(messages[-1])
+ content = last.get("content")
+ if isinstance(content, str):
+ last["content"] = [
+ {"type": "text", "text": content, "cache_control": {"type": "ephemeral"}}
+ ]
+ elif isinstance(content, list) and content:
+ new_content = [dict(block) for block in content]
+ new_content[-1] = {**new_content[-1], "cache_control": {"type": "ephemeral"}}
+ last["content"] = new_content
+ return head + [last]
+
+
+def _hash_for_logging(value: str) -> str:
+ """Short stable hash for log correlation without leaking content."""
+ return hashlib.sha256(value.encode("utf-8")).hexdigest()[:16]
+
+
+# ---------------------------------------------------------------------------
+# Agent loop
+# ---------------------------------------------------------------------------
+
+
+class ToolLoopLimitExceeded(RuntimeError):
+ """Raised when the tool-use loop hits `settings.max_tool_use_iterations`."""
+
+
+def _run_tool_use_loop(
+ session: Session,
+ system_content: list[dict[str, Any]],
+ client: Anthropic,
+) -> tuple[Any, list[dict[str, Any]]]:
+ """Drive the model until it stops asking for tools.
+
+ Returns the final Anthropic response object plus the cumulative list of
+ tool results produced during the turn (used by Layer 4 validation to
+ check for ungrounded claims in the final reply).
+
+ Raises `ToolLoopLimitExceeded` if the model keeps asking for tools past
+ `settings.max_tool_use_iterations`. This is the structural guard that
+ prevents one bad request from burning API credit in an infinite loop.
+ """
+ tool_results_this_turn: list[dict[str, Any]] = []
+
+ response = client.messages.create(
+ model=settings.anthropic_model,
+ max_tokens=settings.max_tokens,
+ system=system_content,
+ tools=TOOL_SCHEMAS,
+ messages=_with_last_message_cache_breakpoint(session.history),
+ )
+
+ for _ in range(settings.max_tool_use_iterations):
+ if getattr(response, "stop_reason", None) != "tool_use":
+ return response, tool_results_this_turn
+
+ assistant_blocks = _serialize_assistant_content(response.content)
+ session.history.append({"role": "assistant", "content": assistant_blocks})
+
+ tool_result_blocks: list[dict[str, Any]] = []
+ for block in response.content:
+ if getattr(block, "type", None) != "tool_use":
+ continue
+ name = block.name
+ args = getattr(block, "input", None) or {}
+ tool_id = block.id
+ result = dispatch_tool(name, args, session.guard_state)
+ tool_results_this_turn.append({"name": name, "result": result})
+ tool_result_blocks.append(
+ {
+ "type": "tool_result",
+ "tool_use_id": tool_id,
+ "content": json.dumps(result, ensure_ascii=False),
+ }
+ )
+
+ session.history.append({"role": "user", "content": tool_result_blocks})
+
+ response = client.messages.create(
+ model=settings.anthropic_model,
+ max_tokens=settings.max_tokens,
+ system=system_content,
+ tools=TOOL_SCHEMAS,
+ messages=_with_last_message_cache_breakpoint(session.history),
+ )
+
+ # Fell out of the for loop without hitting `end_turn` -- the model is
+ # still asking for tools. Refuse the request rather than loop forever.
+ raise ToolLoopLimitExceeded(
+ f"Tool-use loop did not terminate within {settings.max_tool_use_iterations} iterations"
+ )
+
+
+def run_turn(session_id: str, user_message: str) -> str:
+ """Run one user turn end-to-end and return the assistant's reply text.
+
+ Wires together: session lookup with locking, history append, system
+ content with reminders, the tool-use loop, output validation, and the
+ safe-fallback path on validation failure.
+ """
+ if not isinstance(session_id, str) or not session_id:
+ raise ValueError("session_id is required")
+ if not isinstance(user_message, str) or not user_message.strip():
+ raise ValueError("user_message is required")
+
+ session_lock = SESSIONS.lock_for(session_id)
+ with session_lock:
+ session = SESSIONS.get_or_create(session_id)
+
+ # Bounded conversations. Past the limit we refuse rather than let
+ # history grow unbounded, which keeps memory usage predictable and
+ # avoids the model losing coherence late in a chat.
+ if session.turn_count >= settings.max_turns_per_session:
+ return CONVERSATION_TOO_LONG_MESSAGE
+
+ session.history.append({"role": "user", "content": user_message})
+ system_content = build_system_content(session.turn_count)
+ client = _get_client()
+
+ try:
+ response, tool_results_this_turn = _run_tool_use_loop(session, system_content, client)
+ except ToolLoopLimitExceeded:
+ logger.warning(
+ "tool_loop_exceeded session=%s turn=%s",
+ _hash_for_logging(session_id),
+ session.turn_count,
+ )
+ session.turn_count += 1
+ return TOOL_LOOP_EXCEEDED_MESSAGE
+
+ reply_text = _extract_text(response.content)
+ validation = validate_reply(reply_text, tool_results_this_turn)
+ if not validation.ok:
+ # Redact PII: log only violation codes plus hashes of session and
+ # reply. Never log the reply body -- it may contain customer name,
+ # email, order ID, or tracking number.
+ logger.warning(
+ "validation_failed session=%s turn=%s violations=%s reply_sha=%s",
+ _hash_for_logging(session_id),
+ session.turn_count,
+ list(validation.violations),
+ _hash_for_logging(reply_text),
+ )
+ # Do NOT append the bad reply to history -- that would poison
+ # future turns.
+ session.turn_count += 1
+ return SAFE_FALLBACK
+
+ session.history.append(
+ {"role": "assistant", "content": _serialize_assistant_content(response.content)}
+ )
+ session.turn_count += 1
+ return reply_textHTTP surface
+The FastAPI app exposes four routes: GET /health,
+GET / (redirects to /static/index.html),
+POST /api/chat, and GET /architecture (this
+very document). Everything else is deliberately missing – the OpenAPI
+docs pages and the redoc pages are disabled so the public surface is as
+small as possible.
Security headers
+A middleware injects a strict Content-Security-Policy and friends on
+every response. CSP is defense in depth: the chat UI in
+static/chat.js already renders model replies with
+textContent rather than innerHTML, so XSS is
+structurally impossible today. The CSP exists to catch any future
+regression that accidentally switches to innerHTML.
The /architecture route overrides the middleware CSP
+with a more permissive one because pandoc’s standalone HTML has inline
+styles.
Sliding-window rate limiter
+SlidingWindowRateLimiter keeps a deque of timestamps per
+key and evicts anything older than the window. The
+/api/chat handler checks twice per call – once with an
+ip: prefix, once with a session: prefix – so a
+single attacker cannot exhaust the per-session budget by rotating
+cookies, and a legitimate user does not get locked out by a noisy
+neighbour on the same IP.
Suitable for a single-process demo deployment. A multi-worker +deployment would externalize this to Redis.
+Session cookies
+The client never chooses its own session ID. On the first request a
+new random ID is minted, HMAC-signed with
+settings.session_secret, and set in an HttpOnly,
+SameSite=Lax cookie. Subsequent requests carry the cookie; the server
+verifies the signature in constant time
+(hmac.compare_digest) and trusts nothing else. A leaked or
+guessed request body cannot hijack another user’s conversation because
+the session ID is not in the body at all.
/api/chat
+The handler resolves the session, checks both rate limits, then calls
+into agent.run_turn. The Anthropic exception hierarchy is
+caught explicitly so a rate-limit incident and a code bug cannot look
+identical to operators: anthropic.RateLimitError becomes
+503, APIConnectionError becomes 503,
+APIStatusError becomes 502, ValueError from
+the agent becomes 400, anything else becomes 500.
/architecture
+This is where the woven literate program is served. The handler reads
+static/architecture.html (produced by pandoc from this
+file) and returns it with a relaxed CSP. If the file does not exist yet,
+the route 404s with a clear message rather than raising a 500.
"""FastAPI app for Bookly. Hosts /api/chat, /health, and the static chat UI.
+
+Security posture notes:
+
+- Sessions are server-issued and HMAC-signed. The client never chooses its
+ own session ID, so a leaked or guessed body cannot hijack someone else's
+ chat history. See `_resolve_session`.
+- Every response carries a strict Content-Security-Policy and related
+ headers (see `security_headers`). The chat UI already uses `textContent`
+ for model replies, so XSS is structurally impossible; CSP is defense in
+ depth for future edits.
+- In-memory sliding-window rate limiting is applied per IP and per session.
+ Suitable for a single-process demo deployment; swap to a shared store for
+ multi-worker.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import hmac
+import logging
+import secrets
+import threading
+import time
+from collections import defaultdict, deque
+from pathlib import Path
+
+import anthropic
+from fastapi import FastAPI, HTTPException, Request, Response
+from fastapi.responses import HTMLResponse, RedirectResponse
+from fastapi.staticfiles import StaticFiles
+from pydantic import BaseModel, Field
+
+import agent
+from config import settings
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s %(message)s")
+logger = logging.getLogger("bookly.server")
+
+app = FastAPI(title="Bookly", docs_url=None, redoc_url=None)
+
+
+# ---------------------------------------------------------------------------
+# Security headers
+# ---------------------------------------------------------------------------
+
+
+_SECURITY_HEADERS: dict[str, str] = {
+ # Tight CSP: only same-origin assets, no inline scripts, no embedding.
+ # The UI is plain HTML+JS under /static, all same-origin.
+ "Content-Security-Policy": (
+ "default-src 'self'; "
+ "script-src 'self'; "
+ "style-src 'self'; "
+ "img-src 'self' data:; "
+ "connect-src 'self'; "
+ "object-src 'none'; "
+ "base-uri 'none'; "
+ "frame-ancestors 'none'; "
+ "form-action 'self'"
+ ),
+ "X-Content-Type-Options": "nosniff",
+ "X-Frame-Options": "DENY",
+ "Referrer-Policy": "no-referrer",
+ "Permissions-Policy": "geolocation=(), microphone=(), camera=()",
+}
+
+
+@app.middleware("http")
+async def security_headers(request: Request, call_next):
+ response = await call_next(request)
+ for header_name, header_value in _SECURITY_HEADERS.items():
+ response.headers.setdefault(header_name, header_value)
+ return response
+
+
+# ---------------------------------------------------------------------------
+# Sliding-window rate limiter (in-memory)
+# ---------------------------------------------------------------------------
+
+
+class SlidingWindowRateLimiter:
+ """Per-key request counter over a fixed trailing window.
+
+ Not meant to be bulletproof -- this is a small demo deployment, not an
+ edge-network WAF. Enforces a ceiling per IP and per session so a single
+ caller cannot burn the Anthropic budget or exhaust memory by spamming
+ `/api/chat`.
+ """
+
+ def __init__(self, *, window_seconds: int = 60) -> None:
+ if window_seconds <= 0:
+ raise ValueError("window_seconds must be positive")
+ self._window = window_seconds
+ self._hits: defaultdict[str, deque[float]] = defaultdict(deque)
+ self._lock = threading.Lock()
+
+ def check(self, key: str, max_hits: int) -> bool:
+ """Record a hit on `key`. Returns True if under the limit, False otherwise."""
+ if max_hits <= 0:
+ return False
+ now = time.monotonic()
+ cutoff = now - self._window
+ with self._lock:
+ bucket = self._hits[key]
+ while bucket and bucket[0] < cutoff:
+ bucket.popleft()
+ if len(bucket) >= max_hits:
+ return False
+ bucket.append(now)
+ return True
+
+
+_rate_limiter = SlidingWindowRateLimiter(window_seconds=60)
+
+
+def _client_ip(request: Request) -> str:
+ """Best-effort client IP for rate limiting.
+
+ If the app is deployed behind a reverse proxy, set the proxy to add
+ `X-Forwarded-For` and trust the first hop. Otherwise fall back to the
+ direct client address.
+ """
+ forwarded = request.headers.get("x-forwarded-for", "")
+ if forwarded:
+ first = forwarded.split(",", 1)[0].strip()
+ if first:
+ return first
+ if request.client is not None:
+ return request.client.host
+ return "unknown"
+
+
+# ---------------------------------------------------------------------------
+# Session cookies (server-issued, HMAC-signed)
+# ---------------------------------------------------------------------------
+
+
+_SESSION_COOKIE_SEPARATOR = "."
+
+
+def _sign_session_id(session_id: str) -> str:
+ secret = settings.session_secret.get_secret_value().encode("utf-8")
+ signature = hmac.new(secret, session_id.encode("utf-8"), hashlib.sha256).hexdigest()
+ return f"{session_id}{_SESSION_COOKIE_SEPARATOR}{signature}"
+
+
+def _verify_signed_session(signed_value: str) -> str | None:
+ if not signed_value or _SESSION_COOKIE_SEPARATOR not in signed_value:
+ return None
+ session_id, _, signature = signed_value.partition(_SESSION_COOKIE_SEPARATOR)
+ if not session_id or not signature:
+ return None
+ expected = _sign_session_id(session_id)
+ # Compare the full signed form in constant time to avoid timing leaks on
+ # the signature bytes.
+ if not hmac.compare_digest(expected, signed_value):
+ return None
+ return session_id
+
+
+def _issue_new_session_id() -> str:
+ return secrets.token_urlsafe(24)
+
+
+def _resolve_session(request: Request, response: Response) -> str:
+ """Return the session_id for this request, issuing a fresh cookie if needed.
+
+ The client never chooses the session_id. Anything in the request body
+ that claims to be one is ignored. If the cookie is missing or tampered
+ with, we mint a new session_id and set the cookie on the response.
+ """
+ signed_cookie = request.cookies.get(settings.session_cookie_name, "")
+ session_id = _verify_signed_session(signed_cookie)
+ if session_id is not None:
+ return session_id
+
+ session_id = _issue_new_session_id()
+ response.set_cookie(
+ key=settings.session_cookie_name,
+ value=_sign_session_id(session_id),
+ max_age=settings.session_cookie_max_age_seconds,
+ httponly=True,
+ secure=settings.session_cookie_secure,
+ samesite="lax",
+ path="/",
+ )
+ return session_id
+
+
+# ---------------------------------------------------------------------------
+# Request/response models
+# ---------------------------------------------------------------------------
+
+
+class ChatRequest(BaseModel):
+ # `session_id` is intentionally NOT accepted from clients. Sessions are
+ # tracked server-side via the signed cookie.
+ message: str = Field(..., min_length=1, max_length=4000)
+
+
+class ChatResponse(BaseModel):
+ reply: str
+
+
+# ---------------------------------------------------------------------------
+# Routes
+# ---------------------------------------------------------------------------
+
+
+@app.get("/health")
+def health() -> dict:
+ return {"status": "ok"}
+
+
+@app.get("/")
+def root() -> RedirectResponse:
+ return RedirectResponse(url="/static/index.html")
+
+
+@app.post("/api/chat", response_model=ChatResponse)
+def chat(body: ChatRequest, http_request: Request, http_response: Response) -> ChatResponse:
+ session_id = _resolve_session(http_request, http_response)
+
+ ip = _client_ip(http_request)
+ if not _rate_limiter.check(f"ip:{ip}", settings.rate_limit_per_ip_per_minute):
+ logger.info("rate_limited scope=ip")
+ raise HTTPException(status_code=429, detail="Too many requests. Please slow down.")
+ if not _rate_limiter.check(f"session:{session_id}", settings.rate_limit_per_session_per_minute):
+ logger.info("rate_limited scope=session")
+ raise HTTPException(status_code=429, detail="Too many requests. Please slow down.")
+
+ try:
+ reply = agent.run_turn(session_id, body.message)
+ except anthropic.RateLimitError:
+ logger.warning("anthropic_rate_limited")
+ raise HTTPException(
+ status_code=503,
+ detail="Our AI provider is busy right now. Please try again in a moment.",
+ )
+ except anthropic.APIConnectionError:
+ logger.warning("anthropic_connection_error")
+ raise HTTPException(
+ status_code=503,
+ detail="We couldn't reach our AI provider. Please try again in a moment.",
+ )
+ except anthropic.APIStatusError as exc:
+ logger.error("anthropic_api_error status=%s", exc.status_code)
+ raise HTTPException(
+ status_code=502,
+ detail="Our AI provider returned an error. Please try again.",
+ )
+ except ValueError:
+ # Programmer-visible input errors (e.g., blank message). Surface a
+ # 400 rather than a 500 so clients can distinguish.
+ raise HTTPException(status_code=400, detail="Invalid request.")
+ except Exception:
+ logger.exception("chat_failed")
+ raise HTTPException(status_code=500, detail="Something went wrong handling that message.")
+
+ return ChatResponse(reply=reply)
+
+
+# Absolute path so the mount works regardless of the process working directory.
+_STATIC_DIR = Path(__file__).resolve().parent / "static"
+_ARCHITECTURE_HTML_PATH = _STATIC_DIR / "architecture.html"
+
+
+# Pandoc-generated literate program. The HTML comes from weaving Bookly.lit.md
+# and contains inline styles (and inline SVG from mermaid-filter), so the
+# default strict CSP must be relaxed for this one route.
+_ARCHITECTURE_CSP = (
+ "default-src 'self'; "
+ "style-src 'self' 'unsafe-inline'; "
+ "script-src 'none'; "
+ "img-src 'self' data:; "
+ "object-src 'none'; "
+ "base-uri 'none'; "
+ "frame-ancestors 'none'"
+)
+
+
+@app.get("/architecture", response_class=HTMLResponse)
+def architecture() -> HTMLResponse:
+ """Serve the woven literate program for the Bookly codebase."""
+ try:
+ html = _ARCHITECTURE_HTML_PATH.read_text(encoding="utf-8")
+ except FileNotFoundError:
+ raise HTTPException(
+ status_code=404,
+ detail="Architecture document has not been built yet.",
+ )
+ response = HTMLResponse(content=html)
+ response.headers["Content-Security-Policy"] = _ARCHITECTURE_CSP
+ return response
+
+
+app.mount("/static", StaticFiles(directory=str(_STATIC_DIR)), name="static")