Spec 15 — Private isolation & rename (WP-RN-*)

Date: 2026-09-27 · Status: Gate-B draft · Binding decisions: **D17–D21** (docs/specs/platform-gaps/00-intent.md:38-46), constrained by D10 (break freely), D14 (public-facing branding out of scope), D15 (attribution), and intent acceptance criterion 7 (00-intent.md:63).

Inputs consumed in full: R1 research/fabric-current-state.md, R3 research/packaging-state.md, R4 research/delta-map.md, Gate-A synthesis research/synthesis-adoption-proposal.md; pre-existing rename planning **absorbed, not re-derived**: docs/adapters/terminology-and-structure-gaps/{README,package-renames,binary-renames,env-var-renames,directory-moves,code-identifier-renames}.md and docs/blueprints-rename/README.md. Fresh evidence in this spec was re-verified at fabric HEAD 5e46b998e.

The name-mapping matrix in §3 is a **decision table**: it is proposed here and **locked verbatim at Gate B** (D18, D21). After Gate B no work package — in this spec or in specs 10/20/30/40/50 — may introduce a package, binary, env prefix, directory, or chart/CRD name that is not in (or derived by the stated rules from) this table.

---

1. The two-repo hard split (D17)

1.1 Current topology (evidence)

• The checked-out repo is the **private mirror a5c-ai/fabric**; its publish.yml header states: "Deploy pipeline of the private mirror a5c-ai/fabric … npm publishing, release tags and the docs site stay in the public a5c-ai/babysitter repository and are deliberately absent here" (.github/workflows/publish.yml:3-6; R3 §1.1).
• npm publication of the 43 public packages under @a5c-ai/* runs from the **public** repo's copy of the publish workflow (R3 inferred finding 4, research/packaging-state.md:365-367). Release tags follow babysitter/<branch>/v<version> (.github/workflows/release-tags.yml:1-14; R3 §1.1).
• The publishable inventory is derived at run time by scripts/lib/publishable-packages.cjs:122-161 (selector at :136-137: workspace member, private !== true, publishConfig.access === "public") — 43 packages, all @a5c-ai/*, lockstep 6.0.x (R3 §1.1, research/packaging-state.md:28-32).

1.2 Target topology (binding, D17)

Invariants (mirror intent acceptance criterion 7, 00-intent.md:63):

1. **No work package in any spec alters the public repo** — not its package names, not its publishing, not its tags, not its workflows. 2. Mirroring public→private **stops now**; fabric ceases to be a mirror and becomes a first-class diverged repository. 3. Upstream public fixes flow **only** by opportunistic cherry-pick (one-way, public→private). Private work never flows back. 4. Renamed packages are **never** published to public npmjs (D19; guards in §4.3).

1.3 Stop-mirror mechanics

• Terminate the repo-level sync that maintains fabric as a mirror of babysitter (whatever transport it uses — scheduled workflow, push mirror, or manual sync; the tree itself carries no sync workflow, consistent with mirroring being configured at the hosting level, .github/workflows/publish.yml:3-6 describing fabric as "mirror").
• Rewrite the publish.yml header block (publish.yml:3-6) to declare fabric the private product repo, not a mirror.
• Add upstream = https://github.com/a5c-ai/babysitter as a **fetch-only** remote convention; CI and branch protection must reject any push remote pointing at the public repo (§ WP-RN-02).
• Retag discipline: fabric's release tags move to fabric/<branch>/v<version> so cherry-picked commits can never collide with the public babysitter/<branch>/v* namespace (release-tags.yml:1-14; § WP-RN-07).

1.4 Cherry-pick-only intake workflow

• Intake is **pull-based and opportunistic**: a maintainer reviews git log upstream/main at their own cadence; there is no scheduled auto-merge and no shared history requirement going forward.
• Every intake lands as a PR labeled upstream-intake, carrying: upstream SHA(s), the path map applied (old @a5c-ai/directory names → §3 names), and a statement of what was deliberately dropped.
• **Path translation**: after rename phase RN-1 the package.json names differ; after phase RN-3 the directories differ. Intake before RN-3 is a plain git cherry-pick plus name fix-ups; after RN-3 it requires path rewriting (git format-patch + git apply -p… --directory=<new> or rename-aware cherry-pick -X find-renames). Consequence (binding sequencing rule): any **bulk** upstream intake still wanted is scheduled before WP-RN-10 executes; afterwards intake is expected to shrink to targeted fixes, matching D17's "opportunistic" wording.
• Conflict policy: on conflict, fabric's version wins by default — divergence is licensed (D17 "free to diverge in implementation"); the intake PR documents dropped hunks instead of contorting fabric to stay merge-clean.

1.5 Divergence policy

• Fabric may freely diverge in implementation, dependencies, and architecture (D17). The t3code migration (spec 10) and **all** new work land fabric-only.
• The public repo's API/behavior is **not** a compatibility constraint on fabric (D10 break-freely reinforces this); the only shared obligation is the one-way fix intake of §1.4.
• Security fixes are the priority intake class: journal/replay/trust findings fixed publicly (the F-01…F-08 cluster of 20-functionality-coverage.md, per synthesis §3) must be evaluated for intake within the intake owner's next review pass, and vice-versa fabric-side fixes to that cluster are **not** pushed upstream (they may be re-implemented publicly by whoever maintains the public repo, outside this effort).

---

2. New npm scope (D18/D19 — proposed decision)

**Decision: the new scope is @fabricstack** (re-locked by owner decision on 2026-09-28, see the amendment note below; final lock at Gate B).

Rationale:

• **Neutral and descriptive** (D18: "no mythology"): "the fabric stack" names the product's platform layer; it carries no character/myth name and no a5c corporate marker.
• **Guard-friendly**: it shares no prefix with @a5c-ai, so publish-isolation checks (§4.3) and code greps can key on the scope string unambiguously; a stray @a5c-ai/ import after cutover is mechanically detectable.
• **Matches the repo** (fabric) without being the bare word: the bare @fabric scope is generic and collision-prone on public npm (the fabric package name is long-taken there), while fabricstack is specific enough to defensively register.
• **Free on public npmjs** (verified read-only on 2026-09-28, evidence in phase0/scope-relock.md): GET https://registry.npmjs.org/-/org/fabricstack/user → 404, GET https://registry.npmjs.org/fabricstack → 404, GET https://registry.npmjs.org/-/v1/search?text=%40fabricstack&size=20 → total: 0 (no @fabricstack/* package). Calibration on the same day: the same probes for agentfabric return 200 / 200 / 5 results.
• **Registry-independent**: on GCP Artifact Registry any scope works, but the scope must still be defensively registerable on npmjs to prevent squatting (§4.3 G4) — which the previous choice turned out not to be.

Alternatives considered:

• @agentfabric — the original proposal; **rejected on 2026-09-28: taken.** Third-party-owned on public npmjs (the scoped package sdk 1.0.1, maintainer defimessiah, created 2026-03-18; unscoped agentfabric 0.0.2, maintainer mdirshadengineer, created 2026-04-25; /-/org/agentfabric/user → 200), so the G4 premise ("defensively registerable") is false and any out-of-tree consumer without the AR routing line would resolve our names against a stranger's packages (dependency confusion).
• @fabric-platform — rejected: taken (/-/org/fabric-platform/user → 200 on 2026-09-28).
• @a5c-platform — rejected: keeps the corporate a5c marker in a surface D18 wants neutral, and its shared a5c prefix with the public scope weakens grep-based isolation guards.
• @fabric — rejected: generic; high squat/collision risk on public npm; weak as a guard token (matches prose and the html canvas library ecosystem).
• @sovereignty / @sovereign — rejected: leaks the company brand into the neutral private surface and contradicts the spirit of D14 (public-facing branding out of scope) by minting a brand-adjacent name.

**Amendment (2026-09-28).** This section originally proposed @agentfabric. While preparing the defensive npmjs registration (§4.3 G4, WP-RN-04) the scope was found to be third-party-owned on public npmjs (its scoped package sdk 1.0.1 by defimessiah, published 2026-03-18; unscoped agentfabric by mdirshadengineer). At breakpoint phase0.npmjs-org on 2026-09-28 the owner chose "Re-lock to @fabricstack" over keeping the colliding scope. The re-lock changed the single constant (fabric.privateScope in the root package.json), the root .npmrc routing line, the guard fixtures, the smoke package and every scope mention in specs 10/15/20/30/40/50 and the phase-0 records; the Artifact Registry repository itself is scope-agnostic and was not changed. Record: docs/specs/platform-gaps/phase0/scope-relock.md. The fabricstack org registration on npmjs remains an owner action (phase0/g4-npmjs-org-runbook.md).

Env-var master prefix and data-dir token derived from the same decision: prefix families per component (§3.4) with FABRIC_ reserved for genuinely platform-wide variables, and the runtime data directory token .fabric/ (§3.5). Binary names are per-component words, not scope-prefixed (§3.3) — a CLI named fabricstack-… for every tool would be hostile; t3code itself ships t3 (R3 §2.1).

---

3. The complete neutral-name mapping matrix (D18) — decision table, locked at Gate B

Conventions used below:

• Every current name/dir/bin is CONFIRMED from the live inventory (scripts/lib/publishable-packages.cjs executed at HEAD: 43 packages; bins read from each manifestPath; private workspaces read from root workspaces globs). The 9 packages/ subsystems named by the intent (00-intent.md:50) plus apps/commander are all explicitly present.
• **Single-bin policy**: each package ships at most the bins listed — every legacy alias bin is dropped, not aliased. The public repo keeps the old bins for old users (D17), so the private train needs no deprecation shims (D10). This resolves the babysitter bin collision (both @a5c-ai/babysitter and @a5c-ai/babysitter-sdk declare bin babysitter — packages/babysitter/package.json:7-9 vs packages/babysitter-sdk/package.json:9-14; R3 §1.6, research/packaging-state.md:176-178) **and** the two further collisions confirmed at HEAD: @a5c-ai/babysitter-sdk also declares bins adapters and adapters-hooks, colliding with @a5c-ai/adapters (bin adapters) and @a5c-ai/hooks-adapter-cli (bin adapters-hooks).
• D18's example vocabulary is fully covered: orchestrator(-sdk)=babysitter(-sdk), control-plane=kradle, agent-runtime=genty, memory=kip-sdk, catalog=atlas, harness-gateway=adapters-gateway, model-gateway=transport-adapter, trust=trust-core, console=client/UI family, operator=kradle-installer, process-library=library/, packages=the packages/ root (unchanged).
• The pre-existing plan package-renames.md/directory-moves.md targeted @a5c-ai/adapters-{feature} under packages/adapters/ — the **structural** part of that plan (everything under one umbrella; hooks under adapters/hooks/) is already landed at HEAD (the live inventory shows all adapter dirs under packages/adapters/…), and this matrix keeps that structure while replacing the umbrella word per D18. The binary-renames.md backward-compat alias guidance is superseded by the single-bin policy above (it predates D17's two-repo model). The env-var-renames.md ADAPTER_* → AGENT_MUX_* mapping is adopted verbatim (§3.4). code-identifier-renames.md and docs/blueprints-rename/README.md are absorbed into phases RN-4 (§5, WP-RN-11).

3.1 Publishable packages (all 43)

Grouping note: rows 4–18 and 20–33 form the **harness family** (harness-* under packages/harness/), rows 34–37 the **agent-runtime family**, rows 38–39 (+incoming console apps, §3.3) the **console family**, rows 42–43 (+private siblings) the **control-plane family** — the grouping matches the four D2 surfaces and keeps the landed umbrella structure of directory-moves.md (its moves are already executed at HEAD; only the umbrella word changes).

3.2 Private (non-publishable) workspaces

library/ (process library, in-repo but non-workspace, R1 §3.4) → process-library/ (D18's example name), phase RN-3. blueprints/ and plugins/ follow docs/blueprints-rename/README.md (§5, WP-RN-11) — that rename is partially landed (a top-level blueprints/ tree already exists, R1 §1).

3.3 t3code-incoming units (D21 — land DIRECTLY under new names; no double rename)

Migrated units from spec 10 never carry an @a5c-ai/* name or an old-style directory, even for a day. Spec 10's WPs must reference these names:

3.4 Env-var prefix families (phase RN-2 surface; decision table)

Counted at fabric HEAD 5e46b998e over **git-tracked files under packages/ apps/ scripts/** (git grep only sees tracked files, so node_modules/, dist/ and other build output are excluded by construction). "Occurrences" = total token matches; "unique" = distinct matched variable names. Reproduce with, per <PREFIX> (run at repo root):

git grep -ohE '\b<PREFIX>_[A-Z0-9_]*[A-Z0-9]' -- packages apps scripts | wc -l              # occurrences
git grep -ohE '\b<PREFIX>_[A-Z0-9_]*[A-Z0-9]' -- packages apps scripts | sort -u | wc -l    # unique names

These are **sizing figures**, not the rename inventory: the token pattern also matches same-shaped code constants and doc mentions, all of which the rename sweeps anyway. Completeness is gated by WP-RN-09 AC1's zero-hit grep, not by these counts.

No dual-read compatibility shims: the old-and-new-env fallback pattern proposed in env-var-renames.md ("read both, prefer new") targeted a public in-place migration; under D17/D10 the private train breaks cleanly in one release, and only the **public-facing plugin surfaces that run inside third-party harnesses** (hook adapters installed into user machines from the public repo) keep whatever the public repo ships — which this effort does not touch.

3.5 Runtime data directories (phase RN-3 surface)

A one-shot orchestrator migrate-data-dir command moves existing .a5c/ state; no dual-path reading is retained after it runs (D10).

3.6 Chart / CRD / K8s names (phase RN-5 surface)

CRD group renames are the most invasive rename in the program (stored CRs must be migrated cluster-side; API group changes are breaking for every client) — hence last phase, with an explicit CR migration job (WP-RN-12).

---

4. Private distribution: GCP Artifact Registry npm (D19)

4.1 Registry provisioning

- @fabricstack:registry=https://us-central1-npm.pkg.dev/boot-464019/fabric-npm/ - //us-central1-npm.pkg.dev/boot-464019/fabric-npm/:always-auth=true

• **Where**: the existing GCP project boot-464019, region us-central1 — the same project/region already hosting the Docker Artifact Registry used by the GKE pipeline (us-central1-docker.pkg.dev, .github/workflows/publish.yml:3-6,880-893; R3 §1.3). New npm-format repository fabric-npm; registry URL https://us-central1-npm.pkg.dev/boot-464019/fabric-npm/.
• **Auth**: CI reuses the existing Workload Identity Federation step (google-github-actions/auth@v2, already wired in deploy_kradle, R3 §1.3) with the CI service account granted roles/artifactregistry.writer on fabric-npm; publish steps mint a short-lived access token (gcloud auth print-access-token) injected as //us-central1-npm.pkg.dev/boot-464019/fabric-npm/:_authToken=${TOKEN} via environment — never written to a committed file. Developers use gcloud ADC plus the google-artifactregistry-auth npm helper to refresh tokens into their user-level ~/.npmrc; humans get roles/artifactregistry.reader by default, writer only for release owners.
• **Committed root .npmrc** (scope routing only, no credentials):
• **Dist-tag/channel scheme carries over unchanged**: main→latest, staging→staging, develop→develop with commit-pinned prereleases and the single resolver (scripts/lib/release-version.cjs:44-46,146-151; R3 §1.1) — Artifact Registry npm repos support dist-tags, so the resolver logic ports as-is (WP-RN-07).

4.2 Publish pipeline relocation

Because npm publishing today lives **only** in the public repo (publish.yml:3-6), the private train needs its own publish workflow in fabric: the tag/recovery/verifier trio already present in this tree (release-tags.yml, publish-packages-from-tag.yml, release-artifact-verifier.yml; R3 §1.1) is retargeted at the AR registry and the renamed 43-package matrix, with tags fabric/<branch>/v<version> (§1.3). The public repo's pipeline is not modified (invariant §1.2.1). This also settles P-13's artifact-home question for npm bytes: private npm artifacts live in AR under boot-464019; spec 40 inherits that decision.

4.3 Publish-isolation guards (binding; intent acceptance criterion 7)

---

5. Phasing (D20) and interaction rules with the t3code migration (D21)

Phase order is fixed by D20: **packages + binaries first; env vars, directories, code identifiers, chart/CRD names in later phases.**

**Interaction rules (binding on specs 10/20/30/40/50, per D21):**

1. **No double rename.** Any unit migrated from t3code lands directly under its §3.3 name, scope, directory, and bin. A spec-10 WP that lands a unit under a t3code or @a5c-ai name violates this spec. 2. **RN-0/RN-1 precede T-01.** Fork intake (T-01) requires the AR registry, guards, and the scope to exist, otherwise the first landed fork would need renaming later. T-02…T-33 depend transitively. 3. **Later phases do not block migration.** Spec-10 work must not wait for RN-2…RN-5; migrated code may temporarily *reference* old env vars/dirs of not-yet-renamed fabric units, and RN-2/RN-3 sweeps include the migrated units in their rename scope. 4. **Retired units are never renamed** (§3.2): T-27's retirement list is the single exemption list for all rename completeness checks. 5. **Bulk upstream intake before RN-3** (§1.4): directory moves multiply cherry-pick friction; the intake owner is warned in the WP-RN-10 acceptance gate. 6. **Spec 40 consistency**: packaging WPs (P-01 exports maps, P-07 npm launcher, P-09 chart, P-13 artifact home) operate on the renamed matrix and the AR registry; P-01's exports-map program and the D10 major bump land in the same train as RN-1 so publics are defined once, under the new names (R3 spec input 8).

---

6. Work packages

WP-RN-01: Execute the two-repo hard split

• **Objective**: Fabric stops being a mirror of public a5c-ai/babysitter; divergence policy (§1.5) is documented in-repo; the public repo is untouched.
• **Affected paths**: hosting-level mirror configuration (outside the tree); .github/workflows/publish.yml:3-6 (header rewrite); new docs/DIVERGENCE-POLICY.md; README.md repo-role statement.
• **Acceptance criteria**: (1) no automated public→private sync occurs after cutover, verified **strictly read-only** (invariant §1.2.1 stays absolute — this WP creates no branch, tag, commit, or any other ref on the public repo, not even a scratch one): (a) the hosting-level mirror transport is confirmed disabled, with the settings/API evidence captured in the WP closing note, **and** (b) passive propagation watch — at cutover, record the newest upstream/main commit SHA; over an observation window of ≥7 days that must contain at least one upstream commit post-dating cutover (extend the window until one exists — upstream is actively maintained), git fetch upstream plus a before/after git for-each-ref audit of fabric shows that no upstream commit entered any fabric ref except via an explicitly labeled upstream-intake PR (§1.4); (2) publish.yml no longer self-describes as a mirror; (3) divergence policy doc states invariants §1.2.1–4 verbatim; (4) zero commits to the public repo originate from this WP (its reflog/PR list audited).
• **Test requirements**: CI check from WP-RN-02 (public-remote guard) green; docs-quality gate passes on the new policy doc.
• **Dependencies**: none (first mover).
• **Size**: S

WP-RN-02: Cherry-pick-only intake workflow and public-push guard

• **Objective**: Codify the §1.4 intake process; make pushing fabric content to the public repo mechanically impossible from CI.
• **Affected paths**: new docs/UPSTREAM-INTAKE.md; new CI job in .github/workflows/ci.yml (guard); repository branch-protection settings; PR label + template upstream-intake.
• **Acceptance criteria**: (1) intake doc specifies remote setup (fetch-only upstream), PR labeling, path-map documentation and the conflict policy (fabric wins); (2) a CI guard fails any workflow/config referencing a5c-ai/babysitter as a push target or checkout-with-write; (3) a rehearsal cherry-pick of one real upstream commit lands via the documented flow with its path map recorded.
• **Test requirements**: guard job has a self-test fixture (a deliberately-bad workflow snippet in a test asserting the guard trips); rehearsal PR merged and referenced.
• **Dependencies**: WP-RN-01.
• **Size**: S

WP-RN-03: Scope decision + Artifact Registry npm provisioning

• **Objective**: @fabricstack scope adopted (§2, locked at Gate B); AR npm repo fabric-npm live in boot-464019/us-central1 with WIF-based CI auth and documented developer auth.
• **Affected paths**: GCP project boot-464019 (new AR repo + IAM bindings); root .npmrc (new, scope routing only); docs/user-guide/… private-install doc (new page); CI auth steps in the fabric publish workflow (built in WP-RN-07).
• **Acceptance criteria**: (1) npm ping --registry https://us-central1-npm.pkg.dev/boot-464019/fabric-npm/ succeeds from CI via WIF token; (2) a throwaway @fabricstack/smoke package publishes to AR from CI and installs on a developer machine following only the written auth doc; (3) committed .npmrc contains scope routing and zero credentials; (4) IAM: CI SA writer, humans reader-by-default (binding list reviewed).
• **Test requirements**: publish/install smoke automated as a workflow-dispatch job; secret-scan on .npmrc history.
• **Dependencies**: WP-RN-01 (fabric owns its own pipeline identity).
• **Size**: M

WP-RN-04: Publish-isolation guards G1–G6

• **Objective**: Renamed packages can never publish to public npmjs (D19); bin collisions can never recur.
• **Affected paths**: scripts/lib/publishable-packages.cjs:136-137 (selection criterion change); scripts/release-version.cjs:24-27 (verify-manifests extension); new CI job in ci.yml; .github/workflows/release-artifact-verifier.yml (G6 assertion); ci.yml:99-102 binary-rename guardrail extension (G5); npmjs org fabricstack (G4, external).
• **Acceptance criteria**: (1)–(6) = guards G1–G6 of §4.3 each implemented and red-until-green demonstrated: a fixture package with publishConfig.access: "public"+npmjs registry must fail G1/G2; a fixture duplicate bin must fail G5; a mis-scoped tarball must fail G6; (7) repo secret audit shows no npmjs credential (G3) recorded in the WP's closing note.
• **Test requirements**: each guard ships with a negative-fixture test in the guard job itself; test:binary-renames extended suite green.
• **Dependencies**: WP-RN-03 (registry URL exists to pin).
• **Size**: M

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

• **Objective**: All 43 publishables and the 8 renamable private workspaces carry their @fabricstack/* names; every internal dependency edge updated; retirement-listed shells excluded per §3.2.
• **Affected paths**: every package.json named in §3.1/§3.2 (name, publishConfig, internal dependencies); root package.json workspaces (unchanged globs — dirs move only in RN-3); package-lock.json regeneration; source import specifiers of workspace packages (@a5c-ai/… → @fabricstack/…) across packages/ apps/ scripts/ library/.
• **Acceptance criteria**: (1) publishable-packages.cjs prints exactly 43 @fabricstack/* entries; (2) repo-wide grep for @a5c-ai/ in package manifests and source import statements returns only the retirement list and historical docs; (3) full workspace build + test:sdk + architecture gate pass (ci.yml:114-132); (4) verify-manifests passes with the G2(d) extension active.
• **Test requirements**: all PR-gating CI jobs (ci.yml jobs at :15,:54,:319,:372,:459 — R1 §4.1) green on the rename commit; packed-artifact verifier dry-run over the renamed matrix.
• **Dependencies**: WP-RN-03, WP-RN-04.
• **Size**: L

WP-RN-06: Phase-1 binary renames + bin-collision resolution

• **Objective**: Every package ships exactly the §3.1 bin column under the single-bin policy; the babysitter bin collision and the adapters/adapters-hooks duplicate-bin collisions are permanently resolved.
• **Affected paths**: bin maps in packages/babysitter/package.json:7-9, packages/babysitter-sdk/package.json:9-14, and the 13 other bin-bearing manifests enumerated in §3.1; the bin entry scripts they point at; docs referencing old bin names.
• **Acceptance criteria**: (1) union of all bin keys across publishables is duplicate-free and equals the §3.1 bin set exactly; (2) npm i -g of @fabricstack/orchestrator + @fabricstack/harness-cli from AR on a clean machine installs without EEXIST bin conflicts; (3) each renamed bin passes a --version/--help smoke.
• **Test requirements**: extended test:binary-renames (G5) green; release-artifact-verifier bin smoke tests (release-artifact-verifier.yml:5-9) pass over renamed bins.
• **Dependencies**: WP-RN-05 (same train).
• **Size**: M

WP-RN-07: Release machinery retarget (versions, tags, waves, verifier → AR)

• **Objective**: The incident-hardened release train (resolver, immutable tags, waves, packed-artifact gate, recovery publish — R3 §1.1) publishes the renamed matrix to Artifact Registry under fabric/<branch>/v<version> tags, keeping channel semantics main→latest / staging→staging / develop→develop.
• **Affected paths**: scripts/release-version.cjs + scripts/lib/release-version.cjs:44-46,146-151 (registry + tag-prefix parameters); scripts/release-matrix.cjs; .github/workflows/release-tags.yml, publish-packages-from-tag.yml, release-artifact-verifier.yml, fix-npm-tags.yml; a new publish-npm-private.yml in fabric (§4.2).
• **Acceptance criteria**: (1) an end-to-end staging release publishes all 43 renamed packages to AR in dependency waves, moves the staging dist-tag, and creates one fabric/staging/v… tag; (2) release-version.cjs verify passes against AR; (3) recovery path publish-packages-from-tag.yml proven by a rehearsal from-tag publish under a candidate dist-tag; (4) the public repo's tag namespace shows no new tags from this train.
• **Test requirements**: published-consumer validation job (per publish-packages-from-tag.yml:20-27 pattern) runs against AR: fresh project, .npmrc scope routing, install + import + bin smoke of @fabricstack/orchestrator-sdk.
• **Dependencies**: WP-RN-04, WP-RN-05, WP-RN-06.
• **Size**: L

WP-RN-08: t3code intake naming rules — no-double-rename gate

• **Objective**: Spec 10's migrated units land only under §3.3 names; a mechanical gate enforces it for the life of the migration.
• **Affected paths**: docs/specs/platform-gaps/10-t3code-migration.md (consumes §3.3 as its naming source of truth); the WP-RN-04 CI guard job (extra assertion: no package.json name matching @t3code/*, t3-*, or @a5c-ai/* for new units); intake PR template.
• **Acceptance criteria**: (1) spec 10's final text references @fabricstack/{contracts,acp,codex-protocol,ssh,tailscale,environment,client-runtime,console-web,console-desktop,relay} and no interim names (Gate-B cross-check); (2) the CI assertion trips on a fixture @t3code/x manifest; (3) the first real landed fork (T-01/T-02 output) passes the gate un-waived.
• **Test requirements**: negative fixture in the guard job; Gate-B compliance matrix row (R-03) cites this WP.
• **Dependencies**: WP-RN-04; blocks T-01.
• **Size**: S

WP-RN-09: Phase-2 env-var renames

• **Objective**: The §3.4 prefix table is fully applied: ADAPTER_*→AGENT_MUX_* per env-var-renames.md (adopted verbatim, incl. its workflow/action.yml/pipeline-template sweep), BABYSITTER_*→ORCHESTRATOR_* (663 occurrences, §3.4 count), KRADLE_*→CONTROL_PLANE_* (1193), GENTY_*→AGENT_RUNTIME_* (67), ATLAS_*→CATALOG_* (133), KIP_*→MEMORY_* (386), A5C_*→FABRIC_* (35 in the counted set + .github/ workflow bodies); no dual-read shims retained.
• **Affected paths**: all source/config/workflow files carrying the counted prefixes (source sweep across packages/ apps/ scripts/); .github/workflows/*.yml env blocks; packages/adapters/triggers pipeline templates (per env-var-renames.md CI/CD section); GKE Secret Manager prefixes and repo variables (A5C_CLOUD_*, publish.yml:1285-1357) coordinated with the ops owner.
• **Acceptance criteria**: (1) repo-wide grep for \b(ADAPTER|ADAPTERS|BABYSITTER|KRADLE|GENTY|ATLAS|KIP|A5C)_[A-Z] in source/workflows returns zero hits outside historical docs and the T-27 retirement list; (2) live-stack matrix (R1 §4.3) green post-rename — it is the strongest env-sensitive suite; (3) deployed staging environment comes up with renamed Secret Manager prefixes.
• **Test requirements**: full PR CI + live-stack.yml push-default lanes + kradle deployed-stack e2e (publish.yml:423-424) green on the rename commit.
• **Dependencies**: WP-RN-05 (names exist), WP-RN-07 (train can ship it); does not block spec-10 WPs (interaction rule 3).
• **Size**: L

WP-RN-10: Phase-3 directory moves + data-dir migration

• **Objective**: The tree matches the §3.1/§3.2 directory column (packages/{orchestrator,harness,agent-runtime,console,control-plane,catalog,trust,memory,observer,model-gateway}, apps/ops-console, process-library/); runtime data dirs move .a5c/→.fabric/, ~/.a5c/→~/.fabric/ with a one-shot migration command; the landed structure of directory-moves.md is preserved under the new umbrella words.
• **Affected paths**: git mv of every directory in the matrix; root package.json workspaces globs; tsconfig references; import paths; .github/workflows/*.yml build paths; packages/atlas/graph/ source refs (all five reference classes enumerated in directory-moves.md Phase 3); kanban run discovery (packages/adapters/gateway/src/kanban/lib/source-discovery.ts:12-34) and every .a5c/ literal; new orchestrator migrate-data-dir command in the orchestrator SDK CLI.
• **Acceptance criteria**: (1) zero directories named for old components remain under packages//apps/ except the T-27 retirement list; (2) full build + all PR CI jobs green; (3) migrate-data-dir moves a populated .a5c/ fixture (runs + journal) and a subsequent run replays from .fabric/ byte-identically (journal checksums compared); (4) bulk-intake warning: WP owner confirms in the closing note that the intake owner (§1.4) signed off before execution.
• **Test requirements**: journal replay integration test over migrated fixture; kanban source-discovery unit tests updated and green; live-stack smoke on one harness lane.
• **Dependencies**: WP-RN-05, WP-RN-09.
• **Size**: L

WP-RN-11: Phase-4 code-identifier renames + blueprints-rename completion

• **Objective**: Apply code-identifier-renames.md (3,500+ adapter*→agentMux* identifier refs, file renames in the agent-platform harness layer, mobile namespace items reduced to whatever survives T-27) and complete docs/blueprints-rename/README.md (SDK src/plugins/→src/blueprints/, remaining CLI/docs/skill references; marketplace dir move if not yet landed) — both absorbed as written, with targets re-spelled to §3 names where those docs predate D18 (e.g. their agentMux* camelCase family is kept: it is neutral and matches the adopted AGENT_MUX_* env family, so re-deriving a new casing would be a double rename).
• **Affected paths**: per code-identifier-renames.md tables (packages/genty/platform/src/harness/adapters/* → renamed under packages/agent-runtime/platform/... post-RN-10); packages/babysitter-sdk/src/plugins/ (post-RN-10: packages/orchestrator/sdk/src/blueprints/); test files/mocks/fixtures listed in both docs.
• **Acceptance criteria**: (1) identifier greps from code-identifier-renames.md §"Source Code Identifiers" return zero old-pattern hits; (2) blueprint CLI surface is orchestrator blueprints:* with no plugin-named alias; (3) all touched packages' suites green; (4) no behavior change: the rename commits contain no logic edits (reviewed as mechanical).
• **Test requirements**: full PR CI; genty(-platform→agent-runtime-platform) 135-test package green (R1 §3.3); blueprint install/list e2e against a fixture marketplace.
• **Dependencies**: WP-RN-10.
• **Size**: L

WP-RN-12: Phase-5 chart / CRD / K8s renames

• **Objective**: Chart kradle→control-plane; CRD group kradle.a5c.ai→control-plane.fabric.internal across all 8 CRD files with a stored-CR migration; namespaces/release names per §3.6; deployed to all three GKE environments.
• **Affected paths**: packages/kradle/charts/ (post-RN-10: packages/control-plane/core/charts/ or the chart's P-09 home) — Chart.yaml:1-7, all crds/*.yaml, 23 templates, values.yaml; publish.yml:852-877,952-955,1080 (helm/namespace/CRD-apply steps); kradle-core reconcilers' group/kind references (packages/kradle/core/src/*.js CR clients); a CR migration job (export old-group CRs → transform → apply new group → verify → delete old CRDs).
• **Acceptance criteria**: (1) helm upgrade --install control-plane … deploys cleanly to develop, staging, then prod, in that order; (2) CR migration rehearsed on develop with zero orphaned old-group CRs after cutover (audited via kubectl get crd | grep kradle.a5c.ai = empty); (3) reconcile loop adopts migrated CRs (regression guard for the HEAD-commit bug class: "repository reconcile loop must only adopt its own environment's CRs", HEAD 5e46b998e message); (4) e2e smoke (publish.yml:1131) green in each environment post-cutover.
• **Test requirements**: kradle-core node:test suite (publish.yml:565-568) updated and green; deployed-stack e2e + smoke (publish.yml:423-424); chart render golden tests for renamed values keys.
• **Dependencies**: WP-RN-09, WP-RN-10; coordinates with P-09 (chart productization) so re-versioning and renaming land as one chart-consumer break, not two.
• **Size**: L

WP-RN-13: Docs/skills surface sweep + Gate-B name-lock record

• **Objective**: All in-repo docs, docs-site pages, skills, process-library references and install guides speak the §3 names; the locked matrix is recorded as the program's naming decision record.
• **Affected paths**: docs/ and docs-site/ (incl. docs/user-guide/getting-started/installation.md:126-131 rewritten for the private AR install path); library/ (→process-library/) skill/process texts; MATURITY-MATRIX.md row names; a new docs/NAMING-DECISION-RECORD.md embedding §2+§3 verbatim with Gate-B approval reference; deprecation banner notes in docs/adapters/terminology-and-structure-gaps/README.md and docs/blueprints-rename/README.md marking them absorbed-by-this-spec.
• **Acceptance criteria**: (1) docs-quality CI gate green; (2) grep of docs/ docs-site/ for old package/bin names returns only historical/changelog contexts and the absorbed planning docs' own tables; (3) the decision record exists and matches Gate-B-approved §3 byte-for-byte; (4) both absorbed planning-doc trees carry a pointer banner to this spec.
• **Test requirements**: docs QA scripts (package.json:116-123 docs gates, R3 §1.6) green; link-check over renamed doc paths.
• **Dependencies**: WP-RN-05 (names final); updated incrementally as RN-2…RN-5 land.
• **Size**: M

---

7. Consistency obligations on the other specs (D21 checklist for Gate B)

Work-package count in this spec: **13** (WP-RN-01 … WP-RN-13).