Build Log
Repo Cleanup, Stranded-Work Recovery, and the SmithV2 Toggle
September 25, 2026
A 3-phase repo cleanup surfaced 5 stale worktrees (one holding real unmerged security fixes) and a live URL rename bug, then built a session-level toggle so Zeb can trial a revised instruction set without a risky one-way cutover.
Timeline
What shipped
Removed orphaned products/medinaCHF/public/MCHFLogo_old.avif (zero code references)
Deleted 73MB of local scratch cruft (.tmp/, openai-tunnel/) — both gitignored/untracked
Found and merged 4 real, 3-week-stranded commits from the owner-loop-s3 git worktree (a security fix for /api/health leaking raw NeonDbError text to unauthenticated callers, plus health-check path corrections for all 4 canvas apps) into master — resolved 5 real merge conflicts by hand, verified clean via tsc/node --check. NOT YET VERIFIED LIVE: a post-build check found canvas-1's /api/health and root both 404 in production right now, meaning the merge is durable in git history but has not reached production — see nextItems
Removed all 5 stale git worktrees and their branches after confirming each was either fully merged or, for owner-loop-s3, successfully recovered first
Fixed CLAUDE.md's Apps/Experiments tables — added 14 real, actively-committed apps that were completely undocumented (oldest gap: over 2 months, medina-square)
Corrected forgekit-core and forgekit-ui's package-registry status from a stale single-consumer CANONICAL claim to COMPOUNDING with grep-verified evidence (5 and 3 real consumers respectively); added 5 packages that existed with zero table entry
Fixed a live public-site bug: renamed forgekit-website's /forgehome route to /hearth (the product was renamed in June; the URL never followed), updated all 6 real internal references, added a 308 redirect, verified via a real production build + live curl test
Investigated leashline.forgekits.build's live 404 via the real Vercel project list — found the Hearth Vercel project was also never renamed from 'forgehome', and a Deploy Gate transcript-layer check (documented as Forge-only) was incorrectly blocking a leashline redeploy; stopped rather than fabricate gate evidence to force it through
Reviewed Zeb's SmithV2 experimental instruction-set revision (CLAUDE.md, PROJECT_INDEX.md, HOW_TO_WORK_WITH_ZEB.md, concurrency.md, ignition-core.md, ignition-harvest.md) against the live v1 files in detail
Built the ignite v1/v2 session-level toggle (Smith's own memory: feedback_ignite_v1_v2_toggle.md) — a behavioral convention, not a file-swap mechanism, since CLAUDE.md/.claude/rules are auto-injected before Smith sees a session's first message
Found and fixed a real regression in SmithV2's revision: the Neon/GRANT KNOWN TRAPS row had lost its actual technical finding (bare sql.unsafe() silently fails to commit GRANT, the interpolated-template-literal fix, the separate grantor/grantee bug) in favor of a 'go recover this from history' pointer — restored the full finding
Updated the toggle memory after Zeb restructured SmithV2/ to mirror the real repo layout (root files at top level, 3 rules files nested under SmithV2/.claude/rules/)
Committed SmithV2/ and the Neon/GRANT fix to master