Claude Code: The Reliability Playbook: Real Projects, Patterns, and Bounded Autonomy

Before 340 test files went to an agent, the suite was counted three ways and gave three answers. One gap traced to a line added to a test config in March 2024 to get a red build green. Seven test files had stopped running, and every build since had reported green.

A check that has stopped checking looks exactly like a check that passed.

This book names the habit behind lines like that: gate faith, trusting a check that has never been shown to catch anything. Under it sits the larger problem, the unwatched hour, the stretch of agent work nobody supervised, where correctness becomes something you were told.

Claude Code: The Reliability Playbook is the second rung of a three-rung ladder, the rung where the unit of work becomes a task you are not watching. It asks one question over and over: how do you know? The answers that count are a named artifact or a command that exits nonzero. Every gate in it has to catch a defect somebody planted on purpose before it earns the right to guard anything real.

It is a calibration manual, not a tour of hooks and skills. Chapter 8 gives itself an afternoon and five seeded defects, each aimed at a different rung. Rung four passed a contract break, because that rung had been wired to a mock for a year. Two rungs had nothing against them. Seven defects fill seven rows, and the seventh is stopped by a person, eleven minutes in.

Fourteen chapters, one named frame each, each built on a session with a failure in it:

• The delegation contract. Four fields written before you hand over a task: scope, done condition, forbidden zone, evidence to leave behind.
• The loaded set and the quiet policy. Which instruction files actually reached the model, and a policy where routine work never prompts and the dangerous path is unreachable.
• The hook fixture and the trigger line. A recorded payload and the exit code a gate must produce, testable with no agent in the room, and the one sentence of a skill that decides whether it fires.
• The clean-clone test and the cited plan. A job that clones the repository on a machine nobody configured, and a plan that lists the files it read.
• Gate calibration and the isolation assertion. Seed a known defect, run the ladder, publish which rung caught it. Then refuse to start when two sessions would share a database.
• The bounded-autonomy loop and the uncovered cell. Inventory, pilot, rules, gates, waves, journal, resume. And the matrix row with no test behind it, which is how a feature passes every check and is wrong.
• The init assertion, the authority budget, and the paired number. Prove what actually loaded. Count the power handed to connected tools. Never report a cost figure without the number that says the saving was real.

None of this needs a platform team, a new vendor, or anybody's approval. It assumes the rung below, plus git, a test runner, and jq at the terminal. The companion repository generates the 340-file suite in its pre-migration state and ships the seven seeded defects, so the calibration is a thing you can run.

Fifteen figures were cut rather than printed, each named on the page with its reason, and version-sensitive mechanics carry the date they were checked. For the developer who walked away for an hour, came back to a green suite and a branch nobody could sign, and threw it away. The Claude Code Ladder, Book 2: DELEGATE, the rung where you leave the room and the evidence has to hold.