WP-RN-05 record: Phase-1 package renames — the full §3.1/§3.2 matrix

Executed 2026-09-29 on branch phase1/rename-train of a5c-ai/fabric (base 531975433 = origin/staging). Spec: docs/specs/platform-gaps/15-private-isolation-rename.md §3.1 (43 publishables), §3.2 (8 renamable private workspaces + the 8-shell retirement row), §3.4/§3.5 (env vars and directories do NOT change in RN-1), §4.3 (guards G1–G6) and "WP-RN-05"; intent 00-intent.md D10 (major bump, no shims) and D17–D21; roadmap 50-roadmap-workpackages.md Phase 1 / M1. Builds on the Phase-0 records phase0/publish-guards.md (selector deferral), phase0/no-double-rename.md (baseline oracle) and phase0/scope-relock.md (@fabricstack).

The public repository a5c-ai/babysitter was never written to: gh api repos/a5c-ai/babysitter --jq .pushed_at → 2026-09-16T19:59:50Z at start and end (only GET calls were issued); git remote -v → upstream no_push (push). Nothing was published to any registry, no tag was created, no PR was opened.

What changed

Every change is driven by the machine-readable matrix scripts/lib/rename-spec.cjs (which parses spec 15), never by a hand-typed list. **Directories do not move** (RN-3, WP-RN-10): every package keeps its current directory; only names, publishConfig, dependency edges, import specifiers, scripts, workflow invocations, the lockfile and versions change.

Counts: node scripts/lib/publishable-packages.cjs | jq length → **43**, all @fabricstack/* (0 outside the scope); git diff --stat of the three commits: 65 + 3186 + tooling files.

Commands run (from `C:/work/fabric`)

node <scratch>/rename-train.cjs --apply --no-md          # phase A manifests (rename-spec.cjs matrix), phase B sweep
node scripts/sync-workspace-versions.mjs --version 7.0.0  # lockstep (after pointing it at the private scope)
npm install --legacy-peer-deps                            # lockfile regeneration (24 s: added 50, removed 51, changed 1)
node scripts/ci/no-double-rename.cjs --write-baseline     # 59 packages: 51 new names + 8 retired shells
npm run build:runtime && npm run build:observer && npm run build:cloud && npm run build:atlas && npm run build:kip
npm run build:adapters && npm run build:hooks-adapter && npm run build:kradle && npm run build:realtime-flow && npm run build:adapters-extensions && npm run build:kradle-web
npm run test:architecture && npm run test:sdk && npm run test:effect-baseline && npm run test:release-tooling
npm run test:publish-guards && npm run test:no-double-rename && npm run test:package-renames && npm run test:binary-renames
node scripts/release-version.cjs verify-manifests --version 7.0.0
<orchestrator gate, verbatim — see "Gates">

Tooling flips and the private-train flag decision

Spec 15 §4.3 G1 requires the inventory selector to change "in the same commit" as the manifests. WP-RN-04 had implemented the RN-1 selector behind FABRIC_PRIVATE_TRAIN=1. This WP flips it:

Tests adjusted, and why

• scripts/__tests__/publishable-packages.test.mjs: census is the 43 spec 15 names; a new test pins inventory == §3.1 new-name column, each in its pre-RN-3 directory; entries must pin AR, carry no access, be at 7.0.0; the git-tracked oracle uses the G1 predicate.
• scripts/ci/publish-guards.test.mjs: the two "selector deferral" tests became "selector (RN-1)" tests (43 @fabricstack entries, CLI byte-identical to the G1 oracle with the flag unset, =1 and =0; a legacy or public-access manifest is never selected); G1 rules: a publishable legacy manifest always trips legacy-scope-name, a private legacy shell does not; G2(d) test uses a genuinely legacy internal name (built with join() so a future sweep cannot rewrite it) and proves @a5c-ai/compendium is not flagged, plus the real tree has zero divergences; the eight workflows that now name @fabricstack/* workspaces are the expected touchingPrivate set and the three legacy npmjs workflows are asserted NOT swept; G5: tolerance keyed on new names, carry-over notices present, under the flag all four bin rules trip; G6: unconditional.
• scripts/ci/no-double-rename.test.mjs: the baseline oracle is now the §3.1/§3.2 **new-name** column plus the 8 retirement-listed names (59), @fabricstack/ops-console replaces a5c-commander; an old name in a spec directory now also trips legacy-scope-new-unit (it is no longer baselined); the "moved elsewhere" notice case uses the new name; the spec-10 allow-rule test cites an existing unit by a join()-built legacy name. Gate fix: the "baselined at another directory" notice was unreachable for private-scope names (it sat in the legacy-scope branch) — now independent of scope.
• scripts/ci/fixtures/no-double-rename/positive/baseline-unchanged: @fabricstack/catalog at packages/atlas, pinned to AR (renamed in place); publish-guards/g1/positive/pinned/packages/legacy-unit: @a5c-ai/legacy is now private: true (the retirement-list shape; a publishable legacy unit is no longer a positive); publish-guards/g5/negative/public-train-bin: @fabricstack/orchestrator declares atlas (another row's bin) instead of its own carry-over babysitter, so the negative still trips in default mode.
• scripts/__tests__/fixtures/fix011/* (9 verifier fixtures): renamed to @fabricstack/fix011-<case> and pinned to AR — G6 refuses any other scope before the steps those tests exercise; verify-release-artifacts.test.mjs reads reports as fabricstack__<id>.json.
• scripts/__tests__/{release-version,release-promotion,verify-published-release,publish-package-from-tag}.test.mjs: synthetic fixture repositories carry the root fabric block (the selector and G2(d) read it) and private-scope, AR-pinned fixture packages (@fabricstack/fixture-*) instead of @a5c-ai/fixture-* / @fixture/* with access: public (which the selector no longer selects). The npm-stub tests among them cannot run on Windows (pre-existing); CI verifies them.
• scripts/__tests__/release-matrix.test.mjs: EXPECTED_HOOKS_LEAVES sorted (harness-hooks-agent-runtime sorts first now); leaf prefix @fabricstack/harness-hooks-.
• packages/adapters/hooks/cli/src/__tests__/{cli,packaging}.test.ts: hooksAdapterPackageName.
• Four release-verification.test.ts fixtures (adapters/core, genty/tui, genty/ui, kradle/installer): publishConfig.registry instead of access: public.

Residuals of `grep -rn "@a5c-ai/" --include=package.json --include=*.ts --include=*.tsx --include=*.js --include=*.mjs --include=*.cjs --include=*.json -l packages apps scripts library | sort`

Verbatim (with a full install) the command also lists installed third-party copies under node_modules/ (e.g. the public @a5c-ai/* packages pulled in by @a5c-ai/compendium) and dist//dist-types//.next/ build output — 364 such files at the time of the full install; excluding those, the **64** tracked-relevant files and their one-line justification:

Deliberately outside the RN-1 sweep (not residual defects): .github/workflows/{publish-packages-from-tag,fix-npm-tags,live-stack-published}.yml (legacy npmjs train; naming @fabricstack/* there would be a G2(c) finding — WP-RN-07), packages/atlas/graph/** (105 YAML + 169 Markdown catalogue records describing the public ecosystem by its public npm names, which D17 keeps), all Markdown/.mdc (992 hits: READMEs, skills, process-library docs — WP-RN-13), plugins/ and blueprints/ (WP-RN-11), docs/ other than the three snippet-executed pages and the two tool inputs, docs-site/, .a5c/, .claude/, .codex/, .changeset/ (removed by WP-PKG-02), e2e-tests/**/*.md, video/.

Gates (from `C:/work/fabric`)

Orchestrator gate, verbatim, final run after every change: N=$(node scripts/lib/publishable-packages.cjs | jq -r 'length') && test "$N" = "43" && node scripts/lib/publishable-packages.cjs | jq -r '.[] | (.name // .)' | grep -vc "^@fabricstack/" | grep -qx 0 && node scripts/release-version.cjs verify-manifests --version 7.0.0 && node scripts/ci/publish-isolation-guard.cjs --self-test && node scripts/ci/no-double-rename.cjs --self-test && node scripts/ci/public-push-guard.cjs && npm run guard:packages && echo RN05-PASS →

All synchronized manifests are at release version 7.0.0.
publish-isolation-guard self-test: 14/14 assertions passed
publish-isolation-guard: self-test passed; repository scan clean (101 files, 0 findings, 8 legacy-train notice(s))
ok   repository scan — 59 manifest(s), 0 findings, 0 notice(s), 0 new unit(s)
no-double-rename self-test: 10/10 assertions passed (7 negative, 2 positive fixtures; retirement list: 8 units)
public-push-guard: scanned 215 files, 0 finding(s), 0 suppressed
✓ package-integrity guard passed (lockfile sync, native bindings, metadata, arch boundaries).
RN05-PASS

Suites (Node built-ins only, green after the last change): test:publish-guards 17/17; test:no-double-rename 16/16; scripts/__tests__/publishable-packages.test.mjs 8/8; gateway-engine-floor.test.mjs 6/6; test:package-renames → "51 renamed packages (43 publishable + 8 private) and 8 retirement-listed shells; 113 allowed legacy reference file(s)"; test:binary-renames → "single-bin policy (G5) holds over 32 workspace bin names"; test:effect-baseline 13/13 (with the full install; see the environment note); check-spec10-names --self-test passed; check:sdk-command-templates / check:plugin-commands synchronized; docs:snippets and docs:lint:style clean, docs:freshness fails only on the pre-existing stale-date list (identical on pristine staging).

Builds (full install, sequential): build:runtime (459 s), build:cloud, build:atlas, build:kip, build:adapters, build:hooks-adapter, build:kradle, build:realtime-flow, build:adapters-extensions, build:kradle-web — all PASS.

**Environment (Windows, this machine) — cannot pass here, CI verifies:** build:observer fails with Cannot find module '../lightningcss.win32-x64-msvc.node', and test:sdk / test:architecture (vitest) with Cannot find module './rolldown-binding.win32-x64-msvc.node'. The lockfile carries no Windows native bindings (only lightningcss-linux-x64-gnu; its non-link nodes are byte-identical to origin/staging, so this is not a rename effect); .github/workflows/ci.yml works around the same gap on Linux with npm install --no-save @rolldown/binding-linux-x64-gnu lightningcss-linux-x64-gnu …, and there is no Windows counterpart in the repository. An attempt to mirror it (npm install --no-save of the win32 bindings, 5 min) still left them absent, and the npm ci started to restore a faithful install was stopped by the session for low system memory, leaving node_modules partial (typescript, next, yaml, effect, @effect/vitest, patch-package missing). Consequently the vitest-based suites (test:sdk, test:architecture, the four package-level release-verification.test.ts, the hooks-CLI packaging/cli tests) and the five package-level verify:release gates could not be re-run on this machine after the last edits to them; the CI test / observer-dashboard / workspace-coverage jobs (ubuntu) verify them on the pushed commits. package-lock.json was never modified by any of this (md5 ab51e0832bee9404974780a1329a8898 before and after). The same partial install had removed eslint, which .githooks/pre-commit (scripts/validate-library-js-syntax.cjs --staged) requires; for the third commit the hook was run unchanged, with eslint@9.39.5 resolved from an isolated scratch install through NODE_PATH (nothing added to the repository tree, no hook bypassed; the commit stages no library/*.js, so the hook reports "No matching library JavaScript files to validate").

test:release-tooling (node --test scripts/__tests__/*.test.mjs): pristine staging on this machine → **111 pass / 37 fail of 148** (the FIX-010/FIX-011 workflow-shape tests and the npm-on-PATH stub tests, Windows-impossible; list saved with the WP output). After all changes, with the partial install → 103 pass / 46 fail of 149: the 37 pristine failures, 5 environmental ones (yaml / tsx missing: three verify-release-artifacts workflow-shape tests, and the two hooks-leaf packed-install tests, which additionally 404 against npmjs for @fabricstack/catalog@7.0.0 until the first AR publication — M1 exit criterion 1), and 4 caused by this WP that were then fixed (@a5c-ai/fixture-* regexes; legacyInternalNameSet() fallback when spec 15 is not reachable from a copied scripts/lib). Re-verified per file afterwards: release-version.test.mjs + verify-published-release.test.mjs → **zero failures that are not in the pristine list**, and four pristine failures now pass (FIX-001: the release version may be carried by the publication workspace release plan, FIX-001: the same tag re-tags the release version once the workspace is synchronized to it, FIX-010: channel promotion is a separate job …, FIX-010: the publish workflow calls the published-consumer workflow …). release-promotion.test.mjs cannot load in the partial install (yaml); with the full install its only non-pristine failure was one of the four fixed regexes. No test was deleted or skipped.

Incident during execution (recorded for honesty)

A read-only comparison against pristine staging was done through a temporary git worktree at /tmp/fabric-staging with a node_modules link back into this checkout; git worktree remove --force followed that link into the node_modules/@a5c-ai/genty-desktop-app workspace junction and emptied packages/genty/desktop-app/ (14 tracked files) at 11:34. It was restored from 531975433, re-swept (5 .tsx files) and re-pinned before anything was pushed; git ls-files -d is empty and no other path was affected (checked against git ls-files -d, the lockfile and node_modules link health). No worktree or link is left behind.

Gate-B clarifications for the owner (review round 1)

• **@fabricstack/ops-console (apps/commander)**: spec 15 §3.2 (line 137) says the commander rename "rides RN-3", while the binding WP-RN-05 objective (line 293) is "the 8 renamable private workspaces carry their @fabricstack/* names" and §3.2 lists exactly eight renamable rows including the commander. This record reads "rides RN-3" as the *directory* move (apps/commander did not move) and executed the *name* rename here; the adversarial review judged this the only reading consistent with the WP text, not a deviation. If the owner intended the name to wait for RN-3 as well, the revert is the one manifest line plus its two dependency edges.
• **scripts/check-package-metadata.cjs RELEASE_MATRIX_SURFACES = []**: the release-coverage gate checks nothing until WP-RN-07 registers the AR publish workflow there (the inherited publish-packages-from-tag.yml publishes the public train and by D19 may never name a @fabricstack/* package). Carried on the WP-RN-07 checklist together with the tag prefix (babysitter/ → fabric/) and the retargeting of the three legacy npmjs workflows.

Review round 1 (adversarial review, score 74 → fixes)

Findings and what changed are recorded per work package: B1/M2/M3/M6 in rn-06-bins.md ("Review round 1"), M1 and the CI job split in this file's commits (f8224cae7 release-tooling job split; 7fdb9becd publish.yml hooks-leaf names + templated-legacy-name guard in publish-isolation-guard.cjs and check-package-renames.cjs, the latter now wired into the publish-isolation job), M4 in pkg-01-exports.md, M5 in rn-06-bins.md. Two further split-token rename casualties of the RN-1 sweep (which matched whole @a5c-ai/<name> tokens and therefore could not see path.join("@a5c-ai", "babysitter-sdk")) were fixed in ed01994df: packages/babysitter-sdk/src/cli/main/runSupport.ts and packages/genty/platform/.../planProcess/validation.ts linked the process-local SDK fallback under the legacy name, where require("@fabricstack/orchestrator-sdk") never looked. The ADR docs/v6-spec-and-roadmap/decisions/ADR-001-agent-platform-seam-contract.md (asserted by the architecture gate) was swept in b76f636dd.

Not done here (by design)

• Bins keep their legacy names (WP-RN-06 renames them, retires the G5 carry-over tolerance and the four pinned collisions, and flips spec-bins-missing on by default).
• The release machinery (publish-packages-from-tag.yml, fix-npm-tags.yml, live-stack-published.yml, publish.yml publication steps, scripts/publish-package-from-tag.mjs, tag prefix fabric/) is WP-RN-07's; it must also register the AR publish workflow in check-package-metadata.cjs RELEASE_MATRIX_SURFACES and retire the G2(c)/G3 flag.
• Docs, READMEs, skills, blueprints/plugins, Atlas catalogue data: WP-RN-13 / WP-RN-11.
• The first AR publication of the 43 renamed packages (M1 exit criterion 1): until it happens, clean-consumer verifications that install a renamed dependency from a registry cannot pass (see "Gates").