Build Log

Gate Enforcement Gap

July 16, 2026

Gates were checklists enforced by Claude's own memory — built a real PreToolUse hook that mechanically blocks commits/pushes without fresh gate evidence.

OSEnforcementHooksCrucible

PreToolUse hook shipped

live commit attempts proving deny/deny/allow

Council personas run (architect, QE, steward)

Timeline

Start

Zeb names the gap

Gates are self-graded — no mechanism stops Claude from skipping triggers.md

Early

Crucible attempt fails

npm run crucible:architect crashed under non-interactive Bash — misdiagnosed as inconsistency

Mid

Root cause found

runPre() uses readline, blocks on stdin — deterministic failure under any non-interactive caller, not flaky

Mid

Council ran clean via crucible:council

One API call, three personas (architect/QE/steward) + synthesis, converged on process-vs-substance ceiling

Late

Hook built, tested, wired

.claude/hooks/gate-check.mjs proven against 3 live git commit --dry-run attempts

What shipped

A PreToolUse hook (.claude/hooks/gate-check.mjs) that mechanically blocks git commit/git push unless a fresh, substantive gate-evidence marker exists — wired into .claude/settings.json

Marker schema: .claude/gate-evidence.json (gitignored, session-scratch only — not the durable record)

Fixed crucible.mjs docs to mark interactive persona modes (pre/architect/customer/etc.) as terminal-only, pointing to crucible:council for programmatic use

Anvil + Strike entries capturing the process-vs-substance enforcement ceiling

“you have the ability to run this via the root level env that gives you API access to chat. last time you had no problem invoking this. why is this so inconsistent?”
— Zeb, after the crucible:architect crash — the push that led to actually reading the script instead of guessing