Versioned Skills: Unit B Exercises

Student edition · 90 minutes of dedicated work · 2026-09-09

Runs on Google Colab as it ships today, or on any local Python 3.12+ kernel.

This is one of two practical units for Chapter 7. Unit A constructs and connects the mechanism; Unit B investigates a controlled failure, repairs it and transfers the invariant. Each is a complete ninety-minute session, with its own setup and required conceptual introductions. Basic Python variables, conditions, loops, functions, lists and dictionaries are the starting knowledge. Libraries and specialized concepts used here are introduced below before the main task.

By the end you should be able to:

1. Explain the chapter's mechanism using a prediction and an observed intermediate result.
2. Repair the failure: An active procedure may require a tool the current worker cannot call. Stage and activate a local skill through the existing API, then inspect context with and without its required tool capability.
3. Solve admit only complete skill requirements using changed inputs and an independent expectation.
4. Retain your implementation, failed/corrected observations, causal explanation and limits.

Installation is preparation time. These are planning estimates, not measured completion times. Use the reference primers when a term is unfamiliar; in Unit B retrieve an explanation before re-reading it. Run All checks that the artifact executes. Unfinished student functions deliberately produce NEEDS_WORK. Keep your first attempt before opening answers.

This notebook belongs to the construction edition. Its supplied teaching runtime is embedded, so it can run without the textbook or another notebook. Where code uses REFERENCE_LESSON, that is the frozen runtime exercise identifier; the reader-facing chapter and saved unit identifiers use the current edition. Building against a supplied runtime is not proof that you have constructed all of its dependencies.

Run the self-contained setup

Open this notebook in Google Colab with the badge above, or use a local Python 3.12+ Jupyter kernel, with Pydantic 2. If needed, run %pip install "pydantic==2.13.4" once in a separate cell and restart the kernel. Package installation needs internet; the lesson itself needs no repository download, API key or prior notebook. The closing extension can use an OpenAI key from Colab's Secrets pane to put a live model behind the same tools; without a key it replays a recorded transcript and says so.

The collapsed cell contains 90 frozen teaching files. Base85 represents compressed bytes as text; zlib decompresses them; SHA-256 checks that the decoded files match this edition. These are supplied packaging operations, not learner algorithms. tempfile creates an isolated working copy; Path handles file locations; sys.path tells Python where the supplied modules live; a .pth file, the mechanism an editable install uses, tells reviewed local subprocesses the same. The code is available for inspection below and performs no package installation itself. The subsequent lesson teaches the libraries used by the mechanisms you will implement.

Run setup on every fresh kernel. It writes scratch runtime files separately from your retained practical-work/ch07-b folder. Rerunning setup restores the frozen support files and keeps your saved work. Restarting a kernel clears variables, not saved submission files. Source basis: Sovereign Agent 444c5f6. Some tasks use reviewed local subprocesses; they are not an OS sandbox.

Supplied offline setup and teaching files

The complete supplied setup cell is in the downloadable notebook. Run it there before attempting the cells below. This web edition omits only that compressed setup payload.

Commit to a prediction before the examples

prediction_notes = {
    "prediction": "Write the expected behavior before running the worked example.",
    "reason": "Name the input and rule behind that prediction.",
    "falsifier": "Name an observation that would prove the explanation wrong.",
    "revision": "After execution, explain what changed in your understanding.",
}

Reading the Python vocabulary used in this notebook

You need basic assignments, if, loops, functions, lists and dictionaries. The less familiar features used by the supplied code are introduced here. A library is reusable code that Python can import. The standard library ships with Python; Pydantic is an additional package. An import makes a name available, but does not mean that you have completed the exercise.

JSON is text for exchanging structured values. A Python dictionary is an in-memory object; the JSON representation is a string. Use json.dumps to encode and json.loads to decode. Decoding proves that text has valid JSON syntax, not that its fields match our business contract. Predict which of the following two decoded objects could describe a stock count.

import json

intro_data = {"sku": "MANGO", "count": 3}
intro_text = json.dumps(intro_data, sort_keys=True)
print(type(intro_data).__name__, type(intro_text).__name__, intro_text)
print(json.loads(intro_text))
print("Also valid JSON:", json.loads('["not", "a", "stock", "record"]'))
assert json.loads(intro_text) == intro_data

The first result is a dictionary; the second is a list. Before indexing a decoded object, check the shape that your function promises to accept. An exception interrupts the normal path. raise ValueError(...) refuses an invalid value; try/except lets a caller inspect that expected refusal. Catch the expected class, rather than turning every programming error into apparent success. finally runs cleanup even when an earlier operation raises.

An annotation, such as count: int, documents the expected type. It does not by itself enforce the type at runtime. A class defines a kind of object; an instance holds one object's data. @dataclass asks Python to generate routine construction and comparison methods from annotated fields. frozen=True prevents ordinary reassignment of the instance's fields; it does not make every object nested inside those fields immutable. A method is a function attached to a class; self refers to the instance receiving the call.

from dataclasses import dataclass


@dataclass(frozen=True)
class IntroObservation:
    operation: str
    count: int


intro_observation = IntroObservation("count-mango", 3)
print(intro_observation.operation, intro_observation.count)
assert intro_observation == IntroObservation("count-mango", 3)

A callback is a function passed to another function. This is how the classroom harness invokes your implementation. The argument candidate below is a function object; parentheses perform the call. Predict the two answers before execution, then trace the result to the callback.

def intro_apply(candidate, value):
    return {"input": value, "observed": candidate(value)}


def intro_double(value):
    return value * 2


print(intro_apply(intro_double, 3))
print(intro_apply(lambda value: value + 2, 3))
assert intro_apply(intro_double, 3)["observed"] == 6

lambda value: value + 2 is a small anonymous function. A closure is a function that retains access to values from its surrounding scope. It can bind a tool to a shop snapshot. A shallow copy duplicates only the outer container; copy.deepcopy also copies nested containers used in these fixtures. A set stores distinct values; required <= allowed asks whether every required item is allowed. frozenset is the corresponding immutable set. A tuple groups ordered values; (value,) is a one-item tuple, including the comma.

Paths and cleanup. Path represents a filesystem location. path / "file.json" constructs a child path; read_text and write_text read and write text. A context manager, used with with, manages entry and exit. A temporary-directory context removes its contents on exit. Save your submission outside temporary runtime directories. Reopening a file is different from reusing a Python variable: the former tests retained bytes, while the latter only tests this kernel.

from pathlib import Path
from tempfile import TemporaryDirectory

with TemporaryDirectory() as intro_folder:
    intro_path = Path(intro_folder) / "observation.json"
    intro_path.write_text(json.dumps(intro_data), encoding="utf-8")
    intro_reopened = json.loads(intro_path.read_text(encoding="utf-8"))
    assert intro_reopened == intro_data
    print("Read from a file:", intro_reopened)

Retrieval check: explain JSON versus a dictionary, annotation versus validation, class versus instance, and defining a callback versus invoking it. Change the callback above so an incorrect implementation visibly changes the observed output. This distinction will matter when grading your connected work. Reference: Python's JSON, dataclasses, and pathlib documentation.

Pydantic: turn an input dictionary into a checked object

Pydantic is an additional Python library for validating data. Its model is a class describing fields, not a neural network. Inherit from BaseModel, declare annotated fields, then call model_validate on incoming data. A field without a default is required. A field with a default can be omitted. The result is an instance whose values you read with dot notation.

from pydantic import BaseModel, ConfigDict, Field, ValidationError


class IntroCourseRequest(BaseModel):
    model_config = ConfigDict(strict=True, extra="forbid")
    name: str = Field(min_length=1)
    quantity: int = Field(gt=0, le=1000)
    note: str = ""


intro_request = IntroCourseRequest.model_validate({"name": "mango", "quantity": 4})
print(intro_request.name, intro_request.quantity, repr(intro_request.note))
assert intro_request.note == ""

The annotation says the field's type. Field supplies constraints: gt=0 means greater than zero, le=1000 means at most 1000, and min_length=1 excludes an empty name. ConfigDict sets model-wide behavior. strict=True rejects conversions for this integer field, including "4", 4.0 and True; extra="forbid" rejects undeclared keys. Pydantic can otherwise convert some compatible inputs, so choose this boundary deliberately rather than assuming every accepted input arrived in the expected type.

Predict which rule refuses each payload. ValidationError reports a failed contract. Its errors() entries contain loc, the field location, and type, the failure category. Catching that expected exception lets the notebook inspect the failure and continue.

intro_bad_requests = [
    {"name": "mango", "quantity": "4"},
    {"name": "mango", "quantity": True},
    {"name": "", "quantity": 4},
    {"name": "mango", "quantity": 0},
    {"name": "mango", "quantity": 4, "approved": True},
    {"quantity": 4},
]
for intro_bad_request in intro_bad_requests:
    try:
        IntroCourseRequest.model_validate(intro_bad_request)
    except ValidationError as intro_error:
        print([(item["loc"], item["type"]) for item in intro_error.errors(include_input=False)])
    else:
        raise AssertionError("An invalid input crossed the declared contract")

Use model_dump() for a Python dictionary, model_dump_json() for JSON text, and model_validate_json() to parse and validate JSON. model_json_schema() describes the contract; it is neither an instance's current values nor an invocation of the business handler.

intro_serialized = intro_request.model_dump_json()
intro_schema = IntroCourseRequest.model_json_schema()
assert IntroCourseRequest.model_validate_json(intro_serialized) == intro_request
print("Actual values:", intro_request.model_dump())
print("Quantity contract:", intro_schema["properties"]["quantity"])
assert intro_schema["properties"]["quantity"]["exclusiveMinimum"] == 0

Four is valid input to this schema even if the shop needs six. Pydantic checks the declared shape and constraints; the handler still needs authoritative stock, price and permission. Ordinary assignments to an existing instance are not automatically revalidated unless configured for assignment validation. This lesson validates new input at the boundary and uses the resulting values. Explain these limits before relying on a model object in a transaction or tool call.

Chapter 2's full introduction expands this pattern with a separate data-repair checkpoint. This notebook contains the required pattern here so prior Pydantic experience is not needed. References: models, fields, and strict mode.

SQLite from the first row to an atomic change

A dictionary disappears when its process ends. A database can retain records so a later process can resume from evidence. SQLite is an embedded database: Python's sqlite3 library opens a local database file without starting a separate database server. SQL is the language used to define, select and change its records. A table has named columns and rows. A primary key identifies a row; a query asks for rows satisfying a condition.

Start with a deliberately small preference table. Read the SQL as instructions: create the table, insert one named value, then select the value for one session. ? is a parameter placeholder; the values are passed separately so they are data, not SQL instructions. fetchone() returns one row or None; it does not guarantee that a matching row exists.

import sqlite3
from pathlib import Path
from tempfile import TemporaryDirectory

with TemporaryDirectory() as intro_sql_folder:
    intro_db_path = Path(intro_sql_folder) / "example.sqlite"
    intro_db = sqlite3.connect(intro_db_path, autocommit=True)
    intro_db.execute("CREATE TABLE preference (id INTEGER PRIMARY KEY, session TEXT, value TEXT)")
    intro_db.execute("INSERT INTO preference (session,value) VALUES (?,?)", ("lucy", "09:00"))
    intro_row = intro_db.execute(
        "SELECT value FROM preference WHERE session=?", ("lucy",)
    ).fetchone()
    print("Matching row:", intro_row)
    assert intro_row == ("09:00",)
    assert (
        intro_db.execute(
            "SELECT value FROM preference WHERE session=?", ("another-session",)
        ).fetchone()
        is None
    )
    intro_db.close()
    intro_reopen = sqlite3.connect(intro_db_path, autocommit=True)
    assert intro_reopen.execute("SELECT count(*) FROM preference").fetchone()[0] == 1
    intro_reopen.close()
print("The row survived closing and reopening its connection.")

Index [0] selects the first column of a returned tuple. sqlite3.Row is an alternative row factory that also permits named-column access. dict(row) then produces an ordinary dictionary. The book's Database wrapper supplies that configuration and its schema; the wrapper is course code, while sqlite3 is the standard library. You will use the public connection and transaction methods explained at the exercise boundary, rather than needing to reconstruct the wrapper.

Now consider a budget. Moving five cents from reserved to spent requires two values to change together. A transaction makes a group of local changes commit together or roll back together. The example explicitly controls SQL transactions with autocommit=True and SQL statements. BEGIN IMMEDIATE starts a write transaction; COMMIT keeps its changes; ROLLBACK discards them. Predict the row after the deliberately raised exception. Catching an error alone would not undo the first update; the rollback is the operation that restores the prior state.

intro_ledger = sqlite3.connect(":memory:", autocommit=True)
intro_ledger.execute(
    "CREATE TABLE budget (id INTEGER PRIMARY KEY, reserved INTEGER, spent INTEGER)"
)
intro_ledger.execute("INSERT INTO budget VALUES (1,5,0)")
intro_ledger.execute("BEGIN IMMEDIATE")
try:
    intro_ledger.execute("UPDATE budget SET reserved=0 WHERE id=1")
    raise ValueError("injected failure before the matching spend update")
except ValueError:
    intro_ledger.execute("ROLLBACK")
assert intro_ledger.execute("SELECT reserved,spent FROM budget").fetchone() == (5, 0)
intro_ledger.execute("BEGIN IMMEDIATE")
intro_ledger.execute("UPDATE budget SET reserved=0,spent=5 WHERE id=1")
intro_ledger.execute("COMMIT")
print(
    "After a complete change:", intro_ledger.execute("SELECT reserved,spent FROM budget").fetchone()
)
intro_ledger.close()

The literal :memory: creates a temporary database inside this connection; it is useful for the small experiment, but the earlier file example establishes persistence. Neither example proves that a remote supplier rolls back when the local transaction rolls back. An external operation has its own state and evidence.

SQL you will meet later. UPDATE ... SET ... WHERE ... changes selected rows. AND combines conditions. ORDER BY makes an ordering explicit; absent that clause, do not rely on row order. count(*) counts rows; sum(amount) totals a column and can be NULL on an empty input; coalesce(sum(amount),0) uses zero for that empty aggregate. A UNIQUE constraint rejects duplicate identities. GROUP BY status computes one aggregate per status.

An invariant is a condition that must remain true across operations, such as nonnegative reserved money. A snapshot is a consistent view at one point; two separate reads can describe different moments unless their transaction contract binds them. In the book, with db.immediate() groups related writes. It is a course-defined context manager with the commit/rollback purpose you just observed. Do not assume that an arbitrary with connection has identical behavior under every SQLite autocommit setting.

Your prediction: two workers both read ten remaining cents outside a transaction and each approve seven. Why can both believe the next order fits? Explain what must be checked together with the write. Then change the example's initial reserved amount and repeat the failure. Reference: Python's SQLite tutorial and transaction control.

A skill is versioned guidance with declared requirements

Lucy wants the same opening check every morning: inspect stock, review deliveries and prepare drafts for shortages. A skill packages a reusable procedure with a name, version and declared tool requirements. Its text guides a model; it does not create a new Python capability or grant permission to call a registered tool. Keep “what the procedure requests” separate from “what this worker is allowed to execute.”

The book stores readable skill definitions in TOML, a configuration format with named keys, strings, arrays and tables. Python's tomllib reads TOML into ordinary Python objects. The loads method reads a string; load reads a binary file. Parsing checks syntax, while your application still checks required fields and meaning. You do not need a separate TOML package on Python 3.11 or newer.

Read one procedure as data

Predict the Python type of requires, then compare the required capabilities with the worker's allowlist. A subset test means all requirements must be available; a nonempty intersection would prove only that at least one requirement is available.

import tomllib

intro_skill = tomllib.loads("""
name = "opening-review"
version = 2
requires = ["list_stock", "supplier"]
instructions = "Inspect stock and compare the supplier quote before drafting."
""")
intro_required = set(intro_skill["requires"])
for intro_allowed in ({"list_stock"}, {"list_stock", "supplier"}, set()):
    print(sorted(intro_allowed), "eligible:", intro_required <= intro_allowed)
assert not intro_required <= {"list_stock"}
assert intro_required <= {"list_stock", "supplier"}

A staged version is available for inspection and evaluation. An active version is the one admitted into current context. Staging alone should not change behavior. Keeping versions immutable lets an evaluation name the exact procedure it tested. If a file's contents change under the same version identity, the evidence no longer has a stable subject.

Build context within an explicit byte budget

Context is the input assembled for the next model request: selected skills, current preferences, relevant history and the new user request. More context is not automatically better. A limit forces an explicit admission policy and prevents an unbounded history from consuming the budget. This implementation measures UTF-8 encoded JSON bytes, which differ from characters and model tokens. Predict why the non-ASCII example below has a larger encoded length.

import json

for intro_text in ("cafe", "café", "🍦"):
    intro_bytes = intro_text.encode("utf-8")
    print(repr(intro_text), "characters:", len(intro_text), "UTF-8 bytes:", len(intro_bytes))
intro_items = [{"kind": "skill", "text": "Inspect stock."}, {"kind": "memory", "text": "10:00"}]
intro_context = []
intro_limit = 60
for intro_item in intro_items:
    intro_candidate = [*intro_context, intro_item]
    if len(json.dumps(intro_candidate, ensure_ascii=False).encode("utf-8")) <= intro_limit:
        intro_context = intro_candidate
print("Admitted context:", intro_context)
assert len(json.dumps(intro_context, ensure_ascii=False).encode("utf-8")) <= intro_limit

Notice that the example measures the complete candidate representation, including JSON brackets, keys and separators. Adding the lengths of text values alone would undercount it. The actual exercise preserves provenance and the declared ordering while admitting only bounded items. It does not silently truncate an instruction halfway through a sentence and call that the same procedure.

History has a configuration boundary

A completed old conversation may have used another preference or skill revision. Replaying it as current guidance can reintroduce an obsolete decision. The context builder therefore needs to select revision-matching completed history. “Completed” concerns that prior episode's state; it is not proof that every sentence it produced was factually correct.

In the core exercise, the context function assembles active eligible skills, explicit preferences and bounded history. Its caller is the model-request path. Unit B removes the requirement check and demonstrates that an active skill requiring an unavailable tool enters context. The repair belongs in eligibility. Strengthening the skill's prose cannot replace the missing set condition.

Explain before constructing: what differs between a malformed TOML file, a staged but inactive skill, an active ineligible skill and an active eligible skill that does not fit the byte budget? Give one observation for each. Then write a new procedure with no tool requirements and explain why an empty required set is eligible even for an empty allowlist. Eligibility still grants no new capabilities. Reference: Python's tomllib documentation.

Choose an explicit starting point for this independent notebook

This Unit B runs without Unit A. By default it prepares a supplied reference starting point and labels its provenance. It is not evidence that you built Unit A. To investigate your own successful implementation, set LEARNER_HANDOFF to its saved path before running the cell. An invalid selected file refuses; it is never silently replaced with the reference.

SourceTask supplies copied-source execution and handoff validation; RuntimeLab supplies the controlled failure experiment. Their public operations are introduced beside the main exercise. The artifact stores identity and observations; no variables from another kernel are required.

LEARNER_HANDOFF = None

Prepare and validate the supplied starting artifact

import json
import runpy
import shutil
import textwrap
from pathlib import Path

COURSE_INPUT = COURSE_WORK / "ch07-unit-a-handoff-v1.json"
if LEARNER_HANDOFF is not None:
    learner_input = Path(LEARNER_HANDOFF).expanduser().resolve()
    if not learner_input.is_file():
        raise FileNotFoundError("The selected learner handoff does not exist")
    if learner_input != COURSE_INPUT.resolve():
        shutil.copy2(learner_input, COURSE_INPUT)
    HANDOFF_ORIGIN = "LEARNER_SELECTED"
else:
    source_task_class = runpy.run_path(
        str(COURSE_ROOT / "book/always_on/exercises/source_tasks_v1.py")
    )["SourceTask"]
    reference_task = source_task_class(COURSE_ROOT, 5)
    try:
        reference_task.install(textwrap.dedent(reference_task.fragment))
        reference_observation = reference_task.visible("SUPPLIED_REFERENCE_START")
        if reference_observation["status"] != "PASS":
            raise RuntimeError("The supplied starting point did not pass its connection check")
        reference_task.save(COURSE_INPUT, reference_observation)
    finally:
        reference_task.close()
    HANDOFF_ORIGIN = "SUPPLIED_REFERENCE"
print("Starting evidence:", HANDOFF_ORIGIN)
print("The core task below validates the selected artifact before using it.")

Understand the supplied execution interface

The course runtime is provided so your implementation can be connected to real callers and storage. SourceTask(ROOT, chapter) makes a private copy. install(source) replaces only the declared function; visible() invokes the real chapter probe; save(path, result) retains a successful implementation and its evidence. load(path) checks the saved identities and hashes. inject_failure() changes the declared boundary; repair(fragment) replaces that broken fragment. close() removes the scratch copy after you retain evidence. These methods are supplied harness operations, not additional packages you must discover or install.

RuntimeLab provides the same copied-source failure experiment without the complete-function construction layer. Its run method records exit status, observations and the compared expectation. A subprocess log from an unfinished learner implementation is feedback about that implementation; it is not a successful connection. A syntax error in the notebook cell itself is a separate issue to fix. The task below names which interface it uses.

For direct-function units, the visible driver calls your callback without installing a source string. In either case, trace where your code is invoked. Supplied fixtures, database wrappers and replay models are labeled infrastructure; your own implementation and changed-case explanation are the evidence of learning.

Main practical: construct, connect and challenge

An active procedure may require a tool the current worker cannot call. This time you begin with your Unit A implementation and its saved evidence. Lucy gets a procedure that assumes unavailable capabilities and may repeatedly request impossible work.

Verify the handoff

The starting-point cell has selected the Unit A artifact explicitly. A selected learner handoff must validate; the default reference start is labeled separately. Run the setup and keep the runtime and implementation hashes in your submission.

import json
import os
import runpy
from pathlib import Path

ROOT = COURSE_ROOT

SourceTask = runpy.run_path(str(ROOT / "book/always_on/exercises/source_tasks_v1.py"))["SourceTask"]
REFERENCE_LESSON = 5
HANDOFF = Path("ch07-unit-a-handoff-v1.json")
handoff_status = "MISSING"
if HANDOFF.is_file():
    task = SourceTask(ROOT, REFERENCE_LESSON)
    try:
        handoff = task.load(HANDOFF)
        handoff_status = "VERIFIED"
        print("IMPLEMENTATION", handoff["implementation_sha256"])
    finally:
        task.close()
print("UNIT_A_HANDOFF", handoff_status)

Reproduce and diagnose

Predict the consequence of this injected boundary before executing it:

if True:

The controlled mutation changes the same implementation you submitted. It refuses if the declared mutation boundary no longer occurs exactly once; inspect an alternative implementation with the instructor before adapting the experiment.

baseline = broken = None
if handoff_status == "VERIFIED":
    task = SourceTask(ROOT, REFERENCE_LESSON)
    try:
        task.load(HANDOFF)
        baseline = task.visible("YOUR_BASELINE")
        if baseline["status"] != "PASS":
            raise ValueError("Saved Unit A code no longer satisfies the visible contract")
        task.inject_failure()
        broken = task.run("INJECTED_FAILURE", expected=task.spec["expected_broken"])
        print("BEFORE", baseline["observation"])
        print("AFTER", broken["observation"])
    finally:
        task.close()
else:
    print("HANDOFF_REQUIRED: complete Unit A before performing Unit B")

State a diagnosis using those two observations. Name a test that would prove your diagnosis wrong. Stage and activate a local skill through the existing API, then inspect context with and without its required tool capability.

Repair the boundary

Return the complete replacement for the injected fragment. Do not edit the oracle or print a desired observation. Repair the actual source. The starter keeps the defect so the learner outcome remains incomplete.

def repair_fragment():
    return "if True:"

Hint 1: the consequence

Lucy gets a procedure that assumes unavailable capabilities and may repeatedly request impossible work.

Hint 2: the evidence

Compare the two observations, then trace the changed field to context in src/sovereign_agent/assistant_context.py. Distinguish a schema refusal from a business-rule or authority refusal.

Hint 3: the design

Read active rows; require the full requires set; append preferences and bounded matching history; admit items only within the JSON byte budget.

def connect_repair(fragment):
    task = SourceTask(ROOT, REFERENCE_LESSON)
    try:
        task.load(HANDOFF)
        task.inject_failure()
        task.repair(fragment)
        return task.visible("YOUR_REPAIR")
    finally:
        task.close()


repair_result = None
if handoff_status == "VERIFIED":
    repair_result = connect_repair(repair_fragment())
    print("REPAIR", repair_result["status"], repair_result["observation"])
else:
    print("REPAIR_NOT_ATTEMPTED: missing Unit A evidence")

Transfer under a changed constraint

Compare staged-only, active-empty-requirements, partially permitted and fully permitted skills. A prompt must never expand the dispatcher allowlist.

Create a fresh task, load your handoff, inject the defect and apply your repair. Then change only the copied probe to exercise the new condition. Keep the actual observation and a prediction written beforehand. Explain why a visible-case lookup or a blanket refusal could pass the original example but fail this transfer.

The instructor's holdout applies your repair to a new copied runtime and checks both the positive case and the missing protection. An exact exception or changed state must cause a failure; no broad error is accepted as successful refusal.

Exit ticket

Submit the original handoff, baseline and broken observations, repair, transfer probe and results. State what Lucy would experience before and after the fix. Identify the guarantee that still requires separate evidence: Eligibility and fixture activation checks do not prove live model quality.

passed = repair_result is not None and repair_result["status"] == "PASS"
exercise_report = {
    "unit": "ch07-b",
    "attempted": int(repair_result is not None),
    "completed": int(passed),
    "failed": int(repair_result is not None and not passed),
    "skipped": int(repair_result is None),
    "connection": "PASS" if passed else "NOT_READY",
    "handoff": handoff_status,
}
print("EXERCISE_REPORT=" + json.dumps(exercise_report, sort_keys=True))

Changed-constraint construction: Admit only complete skill requirements

Allow twenty minutes. Spend three minutes predicting, ten implementing and tracing, five on a new case of your own, and two explaining the surviving limitation. This is dedicated work, not an invitation to run a supplied answer. Both units revisit the same invariant after different core experiences; in Unit B, attempt this task from memory before consulting Unit A.

Implement transfer_check(skills, allowed). Return names, preserving input order, for skills with active=True whose complete requires list is a subset of allowed. Inputs have unique names. An empty requirement list is eligible even when allowed is empty. Never change allowed based on instruction text.

Write your expected values before running the table. Keep one accepted case and one refusal. Your function is passed directly into the driver below. The driver copies inputs and checks they remain unchanged; it does not replace your implementation with the reference answer.

Hint 1: identify the authoritative inputs Name the source field for each output value. Which input changes while the rule remains the same?

Hint 2: choose the boundary cases Start with exact empty, exact equality and one value on each side of the boundary where valid. Do not add a special case for a visible product name or operation identity.

def transfer_check(skills, allowed):
    raise NotImplementedError("Separate active state from capability eligibility")

import copy
import json

TRANSFER_CASES = [
    (
        "partial capabilities",
        [[{"name": "opening", "active": True, "requires": ["stock", "price"]}], ["stock"]],
        [],
    ),
    (
        "all capabilities",
        [[{"name": "opening", "active": True, "requires": ["stock", "price"]}], ["stock", "price"]],
        ["opening"],
    ),
    (
        "empty requirements",
        [[{"name": "greeting", "active": True, "requires": []}], []],
        ["greeting"],
    ),
    ("staged only", [[{"name": "greeting", "active": False, "requires": []}], []], []),
]


def same_transfer_value(actual, expected):
    if type(actual) is not type(expected):
        return False
    if isinstance(expected, dict):
        return actual.keys() == expected.keys() and all(
            same_transfer_value(actual[key], value) for key, value in expected.items()
        )
    if isinstance(expected, list):
        return len(actual) == len(expected) and all(
            same_transfer_value(a, e) for a, e in zip(actual, expected, strict=True)
        )
    return actual == expected


def run_transfer(candidate, cases):
    observations = []
    for label, arguments, expected in cases:
        supplied = copy.deepcopy(arguments)
        before = copy.deepcopy(supplied)
        raised = None
        try:
            actual = candidate(*supplied)
        except NotImplementedError:
            raised = "NotImplementedError"
            actual = {"unfinished": True}
        except Exception as error:
            raised = type(error).__name__
            actual = {"raises": raised}
        expects_error = isinstance(expected, dict) and set(expected) == {"raises"}
        correct = (
            raised == expected["raises"]
            if expects_error
            else (raised is None and same_transfer_value(actual, expected))
        )
        passed = correct and same_transfer_value(supplied, before)
        observations.append(
            {"case": label, "expected": expected, "observed": actual, "passed": passed}
        )
        print("PASS" if passed else "NEEDS_WORK", label, "expected", expected, "observed", actual)
    return observations


transfer_observations = run_transfer(transfer_check, TRANSFER_CASES)
TRANSFER_PASSED = all(row["passed"] for row in transfer_observations)
print("TRANSFER_STATUS", "PASS" if TRANSFER_PASSED else "NEEDS_WORK")

Design a counterexample and retrieve the mechanism

Add one new case with an independently calculated expected outcome to TRANSFER_CASES and rerun the driver. Change one condition at a time. Then deliberately replace your candidate with a constant answer in a temporary copy and show a case that rejects it. Restore your implementation. Explain why that counterexample is stronger than repeating the original example with a new name.

Without viewing the worked example, write the invariant in words and trace one observed value back to its input. Identify which part is a local fixture result and which claim would need a live provider, host or external-system observation. Keep a first attempt even if you used a hint.

Save your evidence and explain the result

Fill the prediction notes and your explanation before saving. Include the exact observed value, the input or retained row that caused it, your code's invocation point, one failed hypothesis, and the strongest claim the evidence still cannot support. A completed code cell alone does not earn explanation credit. Do not label reference-start behavior as your own Unit A construction.

Keep this edited notebook, the Markdown if used for notes, saved handoff files, and the JSON record below. Your work folder survives scratch cleanup and can be reopened in a new kernel. An instructor can ask for an unseen case after the visible checks; keep your implementation general.

explanation_notes = {
    "causal_trace": "Explain the input, learner invocation and observed result.",
    "failed_hypothesis": "Describe a prediction the evidence changed.",
    "remaining_limit": "Name the guarantee not established by this experiment.",
}
course_submission = {
    "unit": "ch07-b",
    "planned_minutes": 90,
    "starting_evidence": globals().get("HANDOFF_ORIGIN", "INDEPENDENT_UNIT_A"),
    "prediction": prediction_notes,
    "explanation": explanation_notes,
    "core_report": exercise_report,
    "transfer": transfer_observations,
    "explanation_review": "HUMAN_REVIEW_REQUIRED",
}
submission_path = COURSE_WORK / "ch07-b-submission-v1.json"
submission_path.write_text(
    json.dumps(course_submission, indent=2, sort_keys=True), encoding="utf-8"
)
print("Saved evidence:", submission_path)
print(
    "COURSE_REPORT="
    + json.dumps(
        {
            "unit": "ch07-b",
            "transfer_passed": TRANSFER_PASSED,
            "starting_evidence": course_submission["starting_evidence"],
            "edition": "student",
        },
        sort_keys=True,
    )
)

Extension: put a language model behind the assembled context

Everything above assembled context and then inspected it. In production that output is the first message a model reads, and the model then asks for tools. This closing section puts a real model behind the function you constructed and shows what the eligibility rule does to a run, not only to a printed list.

The boundary is the pair context builds and Dispatcher enforces. context admits a skill only when its complete requires list is inside the worker's allowed set; the dispatcher refuses any call outside that same set with tool_not_allowed. The model sees guidance only when it can act on it, and cannot act on guidance it was not given. Preferences are operator-owned: remember is never a tool. Completed work returns as past_work only while the memory revision still matches.

The loop is agent_loop.run_loop from Chapter 3. llm_lab.py supplies two models for it: ScriptedModel replays turns written in this notebook, OpenAIModel sends the same messages to a hosted model when a key is present. The record always says which one ran.

Prediction: the next cell activates opening_check through the real activation gate, records one preference, and builds two workers over the same shop: one with all three shop tools, one without draft_order. Write down which item kinds each worker's context will hold before you run it.

import json
from pathlib import Path
from typing import Any

from pydantic import BaseModel, ConfigDict, Field

from reference_organizations.store.agent import OfflineShopModel, seed_lucy, shop_dispatcher
from reference_organizations.store.evaluation import CASES, candidate_checks, evaluate
from sovereign_agent.agent_loop import Limits, run_loop
from sovereign_agent.assistant_context import (
    activate_skill,
    context,
    preferences,
    remember,
    skill_snapshot,
    stage_skill,
)
from sovereign_agent.assistant_work import claim, enqueue, finish
from sovereign_agent.database import Database
from sovereign_agent.llm_lab import ScriptedModel, describe, scripted_turn, tool_trace
from sovereign_agent.tool_dispatch import Dispatcher, ExecutableTool

LLM_SKILLS_DB = COURSE_WORK / "ch07-llm-skills.sqlite"
SESSION = "lucy:morning"
PROMPT = "Prepare replenishment drafts from current stock. State USD amounts."


def fresh_shop_database(path):
    """Start from an empty ledger so reruns and replays observe the same rows."""
    for stale in (path, Path(f"{path}-wal"), Path(f"{path}-shm"), path.with_suffix(".authority")):
        stale.unlink(missing_ok=True)
    db = Database(path)
    seed_lucy(db)
    return db


class PreferenceQuery(BaseModel):
    model_config = ConfigDict(extra="forbid", strict=True)
    query: str = Field(min_length=1, max_length=200)


def build_worker(db, allowed):
    """A worker is the shop's real tools behind one explicit allowlist."""

    def recall_preferences(args: PreferenceQuery) -> list[dict[str, Any]]:
        return preferences(db, SESSION, args.query, maximum=5)

    tools = [
        *shop_dispatcher(db).tools.values(),
        ExecutableTool(
            "recall_preferences",
            "Search Lucy's explicit preferences for this session. Read only.",
            PreferenceQuery,
            recall_preferences,
        ),
    ]
    return Dispatcher(tools, allowed=frozenset(allowed))


def context_kinds(db, allowed, **options):
    system_text = context(db, SESSION, PROMPT, allowed=allowed, **options)[0]["content"]
    return [item["kind"] for item in json.loads(system_text.split("provenance:\n", 1)[1])]


skills_db = fresh_shop_database(LLM_SKILLS_DB)
candidate = stage_skill(skills_db, COURSE_ROOT / "book/always_on/skills/opening-check-v1.toml")
activation = activate_skill(
    skills_db,
    candidate.name,
    candidate.version,
    evaluate=lambda skill: candidate_checks(
        evaluate(OfflineShopModel, skill=skill, cases=CASES[:3])
    ),
    required_cases=frozenset(f"{case.name}:0" for case in CASES[:3]),
)
print("Activation cases:", activation)
print("Active skills:", [(skill.name, skill.version) for skill in skill_snapshot(skills_db)[1]])
print("Skill requires:", candidate.requires)
remember(skills_db, SESSION, "format", "three bullets", "lucy/explicit-message")

phone_worker = build_worker(
    skills_db, ["list_stock", "supplier", "draft_order", "recall_preferences"]
)
readonly_worker = build_worker(skills_db, ["list_stock", "supplier", "recall_preferences"])
for label, worker in (("phone", phone_worker), ("read-only", readonly_worker)):
    names = [schema["function"]["name"] for schema in worker.schemas()]
    print(f"{label} worker sees tools {names}")
    print(f"{label} worker context kinds {context_kinds(skills_db, worker.allowed)}")
assert context_kinds(skills_db, phone_worker.allowed) == ["skill_guidance", "preference"]
assert context_kinds(skills_db, readonly_worker.allowed) == ["preference"]

The activation gate ran the candidate through the real evaluate against three authored cases with the offline model, and activated the version only because every named case passed. The preference is in both contexts because it does not depend on tools. The skill is only in the phone worker's context: the read-only worker cannot call draft_order, so the procedure that says "call draft_order" would be guidance it cannot follow.

Run the loop against a recorded transcript

The turns below are what a model replied, written down, and this run uses the read-only worker. Read them as a conversation. Turn one asks for stock. Turn two searches the preferences, asks for a supplier price and, having seen the shortage, tries to draft. Turn three answers in the three bullets Lucy asked for. The run is real work: it is queued with enqueue, claimed with claim, and its answer is recorded with finish, so the next context call for this session can return it as past_work.

Prediction: which call is refused, with which error, and what does the phone worker's context hold once this work is DONE? Recall that the dispatcher checks allowed before it even looks at the arguments, so the quantity 6 is never validated.

RECORDED_TURNS = [
    scripted_turn(calls=[{"name": "list_stock", "arguments": {}}]),
    scripted_turn(
        calls=[
            {"name": "recall_preferences", "arguments": {"query": "format of the brief"}},
            {"name": "supplier", "arguments": {"sku": "SKU-VANILLA"}},
            {"name": "draft_order", "arguments": {"sku": "SKU-VANILLA", "quantity": 6}},
        ]
    ),
    scripted_turn(
        "- Vanilla is 6 tubs below its reorder point; the supplier price is 250 cents, so a draft "
        "would total $15.00 USD.\n"
        "- Strawberry is 4 tubs short at 275 cents, $11.00 USD.\n"
        "- No draft was created: drafting is not available to this worker, so this is a "
        "recommendation, not a draft."
    ),
]
LOOP_LIMITS = Limits(model_calls=8, tool_calls=16, seconds=120, output_tokens=4096)

scripted_work = claim(
    skills_db, "read-only-worker", identifier=enqueue(skills_db, "ch07:scripted", SESSION, PROMPT)
)
assert scripted_work is not None
scripted_messages = context(skills_db, SESSION, PROMPT, allowed=readonly_worker.allowed)
scripted_result = run_loop(
    ScriptedModel(RECORDED_TURNS), readonly_worker, scripted_messages, limits=LOOP_LIMITS
)
print(describe(scripted_result, "scripted"))
scripted_trace = tool_trace(scripted_result)
assert scripted_result.status == "COMPLETED"
assert [step["ok"] for step in scripted_trace] == [True, True, True, False]
assert scripted_trace[3]["tool"] == "draft_order"
assert scripted_trace[3]["result"] == "tool_not_allowed"
assert scripted_trace[1]["result"][0]["value"] == "three bullets"
assert all(
    message.get("tool_call_id") for message in scripted_result.messages if message["role"] == "tool"
)
finish(skills_db, scripted_work, "DONE", scripted_result.answer)
print("Phone worker context now:", context_kinds(skills_db, phone_worker.allowed))
assert context_kinds(skills_db, phone_worker.allowed) == [
    "skill_guidance",
    "preference",
    "past_work",
]
assert skills_db.connection.execute("SELECT count(*) FROM assistant_orders").fetchone()[0] == 0

The refused draft is an ordinary failed observation and the loop carries on. Nothing about the arguments was inspected: tool_not_allowed came before Pydantic, because authorization precedes validation in Dispatcher.invoke. The finished answer is now a row in assistant_work, and context returns it as past_work with the work identifier as its source, because the session's memory revision has not moved since.

Run the same loop against a live model

Nothing changes except the model. build_model looks for an OPENAI_API_KEY in Colab's Secrets pane (the key icon in the left sidebar: add a secret with that name and switch on notebook access for it), then in the process environment for a local kernel. Without a key it returns the scripted model again and the record says so. A missing key is a normal condition, never a silent substitution.

The live model is gpt-5.1 through the openai library, which Colab ships preinstalled; on a local kernel run %pip install openai once. One run costs a few thousand tokens. It runs as the same read-only worker, with a context that now includes the recorded result. It may read stock again, trust the past work, or try to draft anyway; what it cannot do is create a draft, because draft_order is outside this worker's allowlist.

Prediction: should a careful assistant call list_stock again when the context already holds a recorded result? Write down why, then run the cell and compare.

from sovereign_agent.llm_lab import DEFAULT_MODEL, build_model, resolve_api_key

chosen_model = build_model(RECORDED_TURNS, model=DEFAULT_MODEL)
print("Model source:", chosen_model.source, "| key found:", resolve_api_key() is not None)
live_work = claim(
    skills_db, "read-only-worker", identifier=enqueue(skills_db, "ch07:live", SESSION, PROMPT)
)
assert live_work is not None
live_kinds = context_kinds(skills_db, readonly_worker.allowed)
print("Live context kinds:", live_kinds)
assert live_kinds == ["preference", "past_work"]
live_messages = context(skills_db, SESSION, PROMPT, allowed=readonly_worker.allowed)
live_result = run_loop(chosen_model, readonly_worker, live_messages, limits=LOOP_LIMITS)
print(describe(live_result, chosen_model.source))
live_trace = tool_trace(live_result)
finish(
    skills_db,
    live_work,
    "DONE" if live_result.status == "COMPLETED" else "BLOCKED",
    live_result.answer,
)
assert live_result.status in {"COMPLETED", "MODEL_CALL_LIMIT", "TOOL_LIMIT", "MODEL_FAILED"}
# Whatever the model did, guidance it was not given cannot become a capability it has.
for step in live_trace:
    if step["tool"] == "draft_order":
        assert step["ok"] is False and step["result"] == "tool_not_allowed", step
assert "skill_guidance" not in live_messages[0]["content"]
assert skills_db.connection.execute("SELECT count(*) FROM assistant_orders").fetchone()[0] == 0

Challenge: shrink the budget, then move the revision

Both rules in context are data. First the byte budget: rerun context for the phone worker with byte_budget=256, the smallest the function accepts, and predict which kinds survive. Items are admitted in order, skill first, and an item is dropped when the JSON of everything selected so far plus that item exceeds the budget; a dropped item does not stop later, smaller items from being admitted. The skill's instructions alone are longer than 256 bytes.

Then the revision: forget this session's format preference. It deletes the preference rows and increments assistant_memory_revisions, so every completed work item recorded under the old revision stops matching. Predict the phone worker's kinds after that, then replay the recorded transcript against the phone worker. The transcript did not change, the worker did: the draft now succeeds, and the recorded final answer, which says no draft was created, is wrong. The observation with the matching tool_call_id is the evidence, not the prose.

In your notes, separate what the scripted run proves (eligibility, the allowlist, the budget, the revision, the transcript accounting) from what only a live run can show (the model's judgment about when to read stock again). The saved record keeps the two apart.

from sovereign_agent.assistant_context import forget

tight_kinds = context_kinds(skills_db, phone_worker.allowed, byte_budget=256)
print("Kinds under a 256-byte budget:", tight_kinds)
assert tight_kinds == ["preference"]

forget(skills_db, SESSION, "format")
forgotten_kinds = context_kinds(skills_db, phone_worker.allowed)
print("Phone worker kinds after forget:", forgotten_kinds)
assert forgotten_kinds == ["skill_guidance"]

phone_result = run_loop(
    ScriptedModel(RECORDED_TURNS),
    phone_worker,
    context(skills_db, SESSION, PROMPT, allowed=phone_worker.allowed),
    limits=LOOP_LIMITS,
)
phone_trace = tool_trace(phone_result)
print("Same transcript, phone worker:", [step["ok"] for step in phone_trace])
assert phone_result.status == "COMPLETED"
assert [step["ok"] for step in phone_trace] == [True, True, True, True]
assert phone_trace[3]["result"]["total_cents"] == 6 * 250
assert phone_trace[1]["result"] == []
assert "No draft was created" in phone_result.answer

llm_lab_report = {
    "unit": "ch07-b",
    "scripted": {
        "status": scripted_result.status,
        "tool_calls": scripted_result.tool_calls,
        "refused": [step["tool"] for step in scripted_trace if not step["ok"]],
        "context_kinds_after": ["skill_guidance", "preference", "past_work"],
    },
    "session": {
        "source": chosen_model.source,
        "model": DEFAULT_MODEL if chosen_model.source == "live" else None,
        "status": live_result.status,
        "model_calls": live_result.model_calls,
        "tool_calls": live_result.tool_calls,
        "context_kinds": live_kinds,
        "refused": [step["tool"] for step in live_trace if not step["ok"]],
        "answer": live_result.answer[:400],
    },
    "tight_budget": {"byte_budget": 256, "kinds": tight_kinds},
    "after_forget": {
        "kinds": forgotten_kinds,
        "phone_worker_draft_cents": phone_trace[3]["result"]["total_cents"],
    },
}
llm_report_path = COURSE_WORK / "ch07-b-llm-lab-report-v1.json"
llm_report_path.write_text(json.dumps(llm_lab_report, indent=2, sort_keys=True), encoding="utf-8")
print("Saved evidence:", llm_report_path)
print("LLM_LAB_REPORT=" + json.dumps(llm_lab_report["session"], sort_keys=True))
skills_db.close()

Keep building with Prof Rod

Found this material through a colleague, classroom or shared download? Get the complete book at profrod.ai/book and join the Prof Rod learner community. Bring one result, one question or one failure you learned from. Share this resource with another learner and keep its source links with it so they can find the full course and future updates.