Repository navigation
v0.336.0 - Checks say what they actually look at, and the theory pages match their sources
2026-11-12. The second and third releases' worth from the triage of the maintainer's upstream-candidates list,
shipped together: checks and refusals that said more or less than they did, and skill and theory text that drifted
from its sources.
What a check reports
- A Monitor command is checked when the turn ends. The shell-state guard compares the files that switch a gate
off before and after each shell command; a command the Monitor tool starts gets no after-check from Claude Code,
so its changes were never compared. A Stop handler now compares every snapshot left open, and files the person
approved meanwhile are taken in so Stop does not undo them. The contract sentence that said "every shell command"
is corrected. - The gate's own preflight run no longer spends the background checks.
gate.shrenews a missing stamp by
runningpreflight.shwith its output discarded, and that run delivered the session's background checks to
nowhere and removed their marker. The gate's run now writes the stamp and stops. - The closing path reads a door-opened diamond. A diamond opened through a catalogue door carries
object_ref, which the closing path ignored, so it printed "NOTHING ON RECORD". It now reads it, and matches a
diamond id as a whole id: a short id such asl2no longer matches any link that contains it. The pending
evidence row now says the gate is ruled by hand, instead of naming a trigger the script never checked. - The purpose-stance gate scoped to one diamond blocks only on that diamond's solutions (its front runner or
the solution itsobject_refnames); others print a WARN. - The open item offers what the ledger accepts. "run, rule, snooze, drop" is now one table mapped to the
ledger's values, so the offer andadvisory_ledger.pycannot drift. The item says "was last ruled on" where it
said "assessed", which it never measured. - The undersized-test ask is a yes/no question, with the reason drafted for the owner.
- The exposure-change advisory leaves a row in its fires log, like the other advisory hooks.
- Release counting on an untagged repo reads "ok, 0 minor releases" when every version subject is a patch,
not "cannot measure". - Citations marked retracted, recanted or misquoted count as annotated.
- The upstream-candidates check treats every status but the closed ones as open, for LANDED and for the
different-builder rule; gate parity reads everyscripts/*.pynamed in CI, not onlycheck_*;
retirement candidates print under "named by nothing that runs (unwired, not necessarily dead)". - The key-shape advisory warns when a write leaves more evidence sources than source classes in a block.
- Refusals name their ways out. A refused delivery names the prototype-folder exit first, a refused release
says the person may run the deploy themselves, and the shell-safety footer no longer says "in this project
before" in projects where it never was. - Tests for 45 mutation survivors in six checks; the seeded sample now kills 130 of 132 (was 84%).
What the skills and theory pages say
-
interview files "will users pay" under value, not viability, and states the shape of
who; skills that record
or contact other people carry a consent step; launch-tier asks leavers why; dora-check's rework rate says when
its keyword method lost its input (under 2% of commits matching); define-done asks whether the opposite of the
outcome is a choice someone could make; the stale-diamond text says what runs; the cycle-owed lines no longer
carry the framework project's cycle class to every adopter; devils-advocate's Technique 7 evidence is current. -
Theory citations checked against their sources: Lauchengco's four fundamentals named; Wardley's survivorship
caveat on the x-axis and the cycle line; Team Topologies' cognitive-load claim, evidence base and second-edition
platform grouping; Accelerate's construct, which leaves change-fail rate out; Sinek's five just-cause criteria
(advisory); Allen's litmus test and need taxonomy in user-needs-map, and the rule that an attribution fix is
checked against the source, not only the error; G-P7 no longer sources corrections to Hoskins's friction log;
the theory-fidelity ranking is no longer called "correct" on unsourced numbers; theory rows follow the move from
phases to decisions. -
Alternatives considered.
- Monitor commands: comparing at the next shell command instead of at Stop (it broke the parallel-commands
case); keeping a Monitor snapshot open past Stop (it would undo the person's own terminal edits). - Gate preflight: making preflight's deliveries idempotent instead (each of three deliveries would need its
own once-only record; the gate's run needs only the stamp). - Closing-path evidence row: building a real check (there is no reader to reuse: the gate is a hand-set status
with no date, so a check needs two new definitions; listed for the maintainer). - Open item: offering the ledger's own verbs to the person (it changes what every user sees, and "run" is not
a ruling). - Gate parity: adding
validate_canvas.pyto the local gate set (pre-push already runs it first with a
checked interpreter; waived with that reason). - Consent step: a
consent:canvas key (a new field under the four-step rule; written as a step instead). - Theory text: rewriting earlier errata rows (appended to them, as errata's own convention says).
- Not built, each needing a choice and listed for the maintainer: a per-diamond record for BVSSH and bias input
rows, a marker for hooks that record nothing on purpose, known debt for the different-builder rule, and the
platformvalue in the team-shape schema.
- Monitor commands: comparing at the next shell command instead of at Stop (it broke the parallel-commands