Measure what you claim, gate what you cannot undo
The quota warning never fired on macOS. quota-check read only
~/.claude/.credentials.json and returned silently when it was absent — but
Claude Code keeps OAuth in the login Keychain on macOS and writes no such file.
So the hook exited at its first line for every macOS OAuth user: exactly the
population it protects. It now tries the file first, then the Keychain on darwin,
with refreshed tokens written back only to the file store (a Keychain entry is
Claude Code's to manage). This is why three parallel benchmark runs exhausted a
session window earlier with no warning at all. 7 tests.
Estimates are now checked against reality. scripts/lib/cost-drift.mjs
compares the hardcoded ROUGH_COST_USD predictions against measured spend in
cost-history.log. Deterministic, zero tokens, no new tracking. It refuses to
pronounce below 30 samples, reports medians (one outlier decides nothing), and
surfaces agents that were measured but never estimated. First run: architect
2.67× over its estimate (n=27), median 1.385× → "estimates-adequate, do not
calibrate" — a question answered with data instead of opinion. 14 tests.
Gates follow reversibility, not just position (ADR-009). Gates here are
positional; /crystallize approve sat outside every stage boundary and shipped
ungated while activating a pattern injected into every future run of every
project. An audit found the remaining irreversible operations already covered
(infra-provisioner stops for approval before creating paid resources, publish
refuses on a dirty tree, devops refuses prod domains) — so no gate was added or
moved. What was missing is the rule that produced that coverage, now stated in
CLAUDE.md: an operation needs a human decision when undoing it is expensive —
escapes the machine, crosses a project boundary, costs money, or destroys
evidence. A refusal, a plan-and-stop, or an evidence requirement all satisfy it;
silence does not.
qa-engineer: Gherkin scenarios + mutation testing. Critical paths are stated
as Given/When/Then before test code is written — a path that cannot be phrased as
observable precondition → action → outcome is one you do not yet understand well
enough to test. Mutation testing then asks whether the suite would actually catch
a fault, since a green run only proves the tests executed. Both feed the report
and the substantiveness rubric. Neither invents a number: no mutation framework
configured reports exactly mutation: not configured, and no BDD runner is added
to a project that never asked for one. 11 eval cases weighted toward that honesty
boundary.
Also: a broken timeout binary (x86 on arm64, exit 126) no longer blocks a
push — the hook treats 126/127 as "the wrapper failed" and falls through; a
package.json "files" entry that pointed at a directory which never existed is
removed, with a test asserting every entry resolves; and docs/decisions/ DISCOVERY-NO-BUILD.md records four evaluated-and-rejected ideas with the
evidence that decided each, so they are not re-litigated.