Session handoff — 2026-08-11 — the B-8/truthfulness/cache arc closes

For the next session. Written at the Operator's direction at session close.

Repo heads and tags (verify, don't trust — git log --oneline -1 each)

• loomworks-engine main at 64610b4 (CR-2026-202). Tags this session, in order: cr-2026-197-b8-companion-drafts (46ab350), cr-2026-198-b84-polling-determinism (4ae5509), cr-2026-199-truthful-replies (5e6cd17), cr-2026-200-b81-grant-pathway (469fc22), cr-2026-201-b87-all-claim-replies (f9372d6), cr-2026-202-b77-prompt-cache (64610b4), plus untagged d172da3+f338f3f (B-83) and 722592a (B-85) and a7e26bc (B-87 denial). Eight consecutive watched-green push runs through CR-201; CR-202's run verdict is in its completion note.
• loomworks (live frontend) main at 35a49c9. loomworks-ui is STALE (June 24) — never read it as the surface.
• loomworks-record main — build list at v0.64; completion notes through CR-2026-202.

The open queue

• B-78 — the Manifestation write via conversation (split from B-8; stated-facts consent named as its shape; needs a new capability — none exists for derive).
• B-79 — what "confirm" means (the shape-confirmed dispatch hook is registered only by tests; production renders happen only via explicit POST).
• B-80 — the three specific approval-card components dead against wire pending_approval (generic card fixed in CR-197; 891-rows/zero-cards was why nobody saw it).
• B-86 — remedy RULED, CR undrafted: persist the operation outcome on the turn; the 32-intent classifier change DEFERRED unless a live recurrence demands it. The steering finding (composed replies route the grant sentence to the wired path; fallback prose routes to a re-ask — 18 trials, zero misroutes) stands as its own result.
• The five splits (B-87's second half) — their own CR after B-86 is settled, weighed WITH the steering finding; the mechanism decision (prefix-composition vs structured side-channel — setting_changes is the working precedent) and commit_project_draft's empty-slot prerequisite as its own named step. Scoping: inspection-briefs/loomworks-b86-b87-scoping-report-v0_1.
• B-77's second Kind B entry — the Manifestation preview freezes seed-derived ordering; untouched by CR-202.
• Operator items: B-63 (model-authored answers, seed-level), B-65 (the foundation document's four dangling topics), B-69 (seed mutability — the standing instruction below remains IN FORCE).
• Awaiting nothing: B-81–B-85, B-87's four remedies, B-84, B-77's cache half — all closed.

Standing disciplines a fresh session would otherwise rediscover

• The gate string, exact form only: work halts at a Checkpoint until the Operator writes GATE <X> APPROVED — PROCEED. Never draft approval text in advance. Mirror every gate-review file to ~/Downloads at STOP-AND-REPORT, and name the copies.
• Observe the failure before trusting the check — five modes seen this arc: (1) deliberately break the invariant and watch the named message fire; (2) prove the fixture non-vacuous first (the B-51 shape); (3) distinguish your break from a known flake by message, rate, and mechanism (B-84); (4) write tests before the fix and run them against pre-fix code; (5) watch the push run rather than assuming it (two assumed-vs-observed divergences were caught only by watching: red-for-two-was-six; sonnet-default-was-tier-routed-Opus-4.7).
• B-69's standing instruction, verbatim heading seed-mutability impact on every CR: record every contact point, Kind A (signal computed, stops at a boundary — closeable), Kind B (code assumes the seed cannot move — not closeable, waits on B-69), Kind C (checked, no contact — a null finding IS an entry; say what was checked). Seed-adjacent is not seed-dependent (ShapeEvent.seed_version_at_production is the worked example).
• Descriptions of state are not state: stale artefacts get acted on as state; close descriptions in the same act; correct by superseding (next version file), never in place. A broken gate makes even the records assert falsehoods — audit claims against observed runs (the CR-194/196 "push-verified" corrections).
• A failure of a kind suppresses detection of its kind — the first fix should expect to find more (red-on-red; race-under-race). Ask "what was this failure standing in front of?" as part of the fix.
• Split the sentence, not the turn: claims whose truth only the server holds are server-composed; conversation stays model prose. The persona rule is necessary and never sufficient — structure is what holds. validate_intent_templates makes formatter-discard unconstructible; additions to format_operation_result must also go in OPERATION_RESULT_INTENTS.
• CR-2026-169 D-1 is PARTIALLY SURRENDERED (its record v0.2): temperature pinned where the API honours it; the responder falls back visibly (llm_compat.fallen_pin_seams); the same refusal message in any other seam's failure path means a second seam has fallen — wire the helper there then.
• Eye-tests: dedicated throwaway engagement under Marvin-Test2, torn down to zero rows; capture every turn's classified_intent BEFORE teardown; credit-ledger rows are append-only — never delete flows; Test2's current_engagement_id gets captured by visiting — reset it. The dev engine runs LOOMWORKS_ENV=production, no --reload, under a supervisor (logs at ~/Library/Logs/loomworks-engine.err.log) — kill by PID and verify the listener's start time, or eye-tests run stale code. The frontend dev script serves port 3001. The composer drops the first click+type after navigation — click, wait 2s, type, zoom-verify before Return.

Where the detail lives

Memory: loomworks_b81_scoping_findings (the whole arc), loomworks_cr_2026_197_b8_companion_drafts, loomworks_failure_kind_suppression_property, loomworks_b8_slice_two_scoping_read. Record: the build list v0.64 changes-from chain reads as the arc's narrative; checkpoints and completion notes under their CR numbers.

DUNIN7 — Done In Seven LLC — Miami, Florida Session handoff — 2026-08-11 — v0.1

Addendum — serving-state recovery (recorded before the 2026-08-11 redeploy)

• Engine (port 8000): supervised by ~/Library/LaunchAgents/com.dunin7.loomworks-engine.plist — the "supervisor" that respawns on kill all session. Bounce = kill the listener PID (lsof -nP -iTCP:8000 -sTCP:LISTEN -t); it respawns on the current working tree. Logs: ~/Library/Logs/loomworks-engine.{out,err}.log.
• Production surface (port 3002, app.dunin7.com): ~/Library/LaunchAgents/com.dunin7.loomworks-operator-layer.plist, serving the build at /Users/dunin7/loomworks/.next. Build before this redeploy: BUILD_ID oNW-ze-GPnN0rqm7UpE8N, mtime Aug 10 11:24 (pre-35a49c9, so no generic-card fix). Redeploy = npm run build in /Users/dunin7/loomworks, then launchctl kickstart -k gui/$(id -u)/com.dunin7.loomworks-operator-layer (or kill the next-server PID and let launchd respawn).
• Revert if a rebuild fails: git checkout <prior surface commit> in /Users/dunin7/loomworks (head at redeploy time: 35a49c9), rebuild, kickstart — the build is derived state; the repo is the source of truth. A failed build leaves the OLD .next intact until the next successful build overwrites it; do not delete .next by hand.
• Dev surface (3001, loomworks-dev.dunin7.com): not a service; npm run dev in /Users/dunin7/loomworks when needed; 502 on the dev hostname simply means it isn't running.