[shafty023]

Purpose

One release (vibe 1.0.66) for the open vibe lessons ledger, combining what was #313, #314, #317 and the unopened orchestration change, per the operator's rule that lessons go into one PR (each PR costs a full CI run). Every gate runs inside a step vibe already has, and the prose rules a gate enforces are deleted.

Lessons carried

Review gates (handoff-inventory diffRules): symphony-alpha #8958 withDb.tx partial commit, migration lock strategy and timestamp, raw writes bypassing activity capture (#8958, #8960), Desktop outbox lanes without synchronizing-susan and the shared give-up (#8960, #8965), ArtifactLink deletes beyond the replaced edge (#8960), flag-off output on shared paths (#8960, #8963, #8962 phone width, #8963 widths), dropped ticketed why-comments (#8948, #8962), Storybook placement and frozen v1.0 closure (#8947, #8962, #8963), story outcome flips as releases (#8959), lost story IDs, GridTableV2 parity and endpoint e2e specs as required runs (#8962, #8967), migration order after every main sync (#8965). One-line owner rules for #8956 to #8967 review findings (cache/precondition, id case, scope gates, composite ports, identity keys, focus from promises, void handlers, comparators, query shape, cursor bounds, throw-to-Result callers, skew tests, persisted snapshot shapes, in-row clipping, contrast, disclosure propagation, narrow selection and target size, sibling story removals, flag-on states per claimed surface).

PR body facts (inventory prFacts, blocking.prBody / blocking.prEvidence): overlap from open PR file and import intersections (#8947, #8956, #8962, #8963, #8965), flipped and deleted assertions and changed contracts with file:line (#8947, #8948, #8953, #8956, #8957, #8959), new refusals under Breaking changes (#8963), ticket "say in the PR" asks (#8956), cited rulings resolve to approved artifacts (#8962), stacked bases state no CI ran (#8967), server-only sessions skip UI ceremony (#8948), pending device and play runs block creation, PostHog re-read before "(flag-pending)" (#8987), Roadmap dependencies and open PRs resolved before designing a write path (#8958, #8963).

Claimed children and provenance (live-ticket-check.mjs --child-evidence): children reach In Review only with per-item evidence from the ticket worker's map, never the writer's audit (#8947, #8953, #8958, #8959, #8960, #8963, #8967, PRD-729); "active elsewhere" needs real activity (#8959); PRD flag and boundary clauses are criteria (#8957, #8959, #8962, #8965); no PRD target narrowing (#8953); operator clauses met per surface and live fallbacks (#8962); the session's own text is never a source, criteria never strengthened, precedent after ticket scope, human threads and sibling requirements (#8956, #8957, #8959, orchestrator ruling); every new user-reachable capability, refusal rule, visual element and numeric limit needs a human source (#8953 status write built to pass its own e2e, #8963, #8967).

Orchestration, resume and PR follow-up: readiness ceremony cut (the pre-capture main-sync-inputs turn and its extra bootstrapped checkout are removed; a failing producer reports command, exit code, error and log; restated main-sync prose replaced by pointers), Codex-to-Claude harnessSwitch migration, review threads resolved only through review-threads.mjs resolve with an independent verified verdict (#8987, ISS-9488), human threads re-read on every later look, process and state rules (no unscoped kill, no temp-dir records, no Codex computer use, visual checks after push), dependencies never idle a session (PRD-728, #8960).

Operator rulings 2026-10-20: tests after any change, fix or main merge are exactly the code_tests_for set plus the inventory's required tests; fleet-orchestrator gains criteria fixed at plan time with a two-way criteria-to-diff QA trace, a specialist role table with fallbacks, and one batched lessons PR per repo.

Cost

• Instruction length (vibe agents/*.md and skills/**/*.md, evals excluded): 16,195 lines / 156,483 words on main, 16,214 / 156,919 here (+19 lines, +436 words). About 14 of those lines are the requested fleet-orchestrator role table and QA rules; the four lesson sets were each flat or negative before combining.
• Whole PR: 75 files, +2844 / -437, most of it gate code (diff-rules.mjs, pr-facts.mjs, review-threads.mjs, the children gate) and their tests.
• Session wall time: publication runs input readiness once instead of twice (one pristine bootstrapped checkout fewer); server-only sessions skip Storybook, Desktop and visual QA; test runs are targeted. The inventory adds about 2 to 6 seconds on large diffs.

Commands run (combined head)

• uv run pytest plugins/vibe: 361 passed
• node --test handoff, fleet-orchestrator and precedent-check suites: 113 passed
• uv run ruff check . clean; uv run pyright clean; tools/vibe-workers typecheck clean and npm run build reproduces the committed dist bundles
• Full tools/vibe-workers vitest was not run locally (CPU-bound laptop); CI confirms it.

Version

vibe 1.0.66 in both manifests, the only bump in this PR, on top of main's 1.0.65.