WP-PKG-01 record: exports maps for all 43 publishables + exactness verifier

Executed 2026-09-29 on branch phase1/rename-train of a5c-ai/fabric (base 3beee69a6, the WP-PKG-02 head; draft PR #4 open, CI on every push). Spec: docs/specs/platform-gaps/40-packaging.md §2.1 (exports-map program), §2.3 (publics documentation) and "WP-PKG-01" (objective, affected paths, acceptance criteria, test requirements); spec 15 §1.2 invariants and §5 rule 6 (exports maps ride the RN-1 train); intent 00-intent.md D10 (break freely, major bump — in-repo consumers migrated in the same effort). Builds on rn-05-rename.md (names), rn-06-bins.md (collision-free bin set, consumed here) and pkg-02-versioning.md (7.0.0 lockstep train).

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.

1. Policy (encoded once, in `scripts/lib/exports-map.cjs`)

Shared by the packed-artifact verifier (scripts/verify-release-artifacts.mjs, step exactness), the in-repo resolver check (scripts/ci/exports-exactness.cjs) and the publics page generator, so the PR gate and the release gate read the same map the same way:

Node's resolution semantics are followed where they matter (exact subpath keys, condition objects in declaration order, arrays as fallbacks, null as a block); patterns are refused, never matched.

2. Surface table: package → exported subpaths → probe that must fail

Derived from evidence, never guessed: every @fabricstack/<pkg>/<subpath> string in an import context across packages/ apps/ library/ scripts/ e2e-tests/ plugins/ blueprints/ (3,303 workspace imports resolved by the scanner), the bins' own entry points, documented subpaths (docs/, package READMEs, JSDoc @examples) and the pre-existing typesVersions/main/module/types fields. The full per-package table (subpaths, the data-driven negative probe and the local matrix verdict) is generated from the verifier reports in §5; the decisions are:

harness-sdk (the spec's "exports only ." example): checked against deep-import usage — the only subpath consumer was @fabricstack/harness-sdk/package.json (a test), now satisfied by the added ./package.json; no other subpath is used, so the map stays "." + ./package.json.

Storage/journal and breakpoint seams the spec mentions as examples (§2.1) are reachable through the SDK root (src/index.ts re-exports ./storage, ./breakpoints); no consumer or sibling spec imports them as subpaths, so they were **not** minted — adding a subpath is a one-line, guarded change when a WP cites one.

3. Deep imports of now-private paths: every site and its decision

Zero remaining findings: node scripts/ci/exports-exactness.cjs → 43 publishable manifest(s), 6379 source file(s), 18979 specifier(s), 3303 workspace import(s) resolved, 0 finding(s).

4. Verifier exactness (AC1, AC2) — `scripts/verify-release-artifacts.mjs`

New step exactness between typecheck and bins (so a platform bin-shim failure cannot hide it), **on by default**, --exactness explicit in the workflow, --no-exactness for triage only:

1. exportsMapProblems(manifest, { tarballFiles }) — the §1 policy over the *packed* manifest, every target must be in the tarball. 2. nonExportedProbes(manifest, tarballFiles) — a **data-driven** probe manifest generated from the tarball inventory: up to 3 real shipped .js/.cjs/.mjs/.json files that no subpath resolves to (dist/ first, shortest first — e.g. harness-hooks-core/dist/api.js, harness-core/dist/pty.js, fix011-good-package/cli.js, bin-only-metapackage/bin/fix011-bin-only.js), plus the spec's synthetic <pkg>/dist/internal-anything. Written to fix011-exactness-probes.json in the clean consumer; fix011-exactness-probe.mjs tries each with require() and import(); both must fail with ERR_PACKAGE_PATH_NOT_EXPORTED (RESOLVED, MODULE_NOT_FOUND or any other code is a failure). A map-less package therefore fails on its leaking file *and* on the synthetic probe. 3. Together with steps 6–7 (every declared runtime subpath imports; every typed subpath typechecks in the fresh consumer) the importable surface equals the map. The report records declaredSubpaths, tarballFileProbes and each probe's two outcomes.

Fixtures (scripts/__tests__/fixtures/fix011/): new exports-wildcard (healthy artifact with "./*" → fails [wildcard-export]), new exports-leaky (no map, internal.js shipped → require -> RESOLVED named, synthetic probe reported); bin-only-metapackage, failing-verify-release, bundler-runtime gained closed/complete maps because exactness is on by default. Tests added to scripts/__tests__/verify-release-artifacts.test.mjs (probe record on the healthy fixture, wildcard red, leaky red, --no-exactness skipped, workflow passes --exactness, bin-only closed map keeps bin/ unimportable).

**Verifier bug found and fixed on the way**: since the CI-triage fix 1117e6993 each package's clean consumer resolves internal dependencies from tarballs packed in this run, but verifyPackage deleted its working directory (including pack/) after every package — so in any multi-package run every dependent failed at install with ENOENT on the file: override (first seen as @fabricstack/orchestrator-sdk opening the already-removed harness-hooks-core tarball). The tarball is now copied to the shared internal-deps directory before cleanup; regression test "a selected dependency stays installable by a dependent verified later in the same run" (trust → harness-policy, both bin-less, real packages, built by the verifier itself) pins it. CI triage round 1 retargeted the test: it originally ran hooks-core → hooks-claude with --skip-build, which passed only on a checkout that had already been built — the Release tooling tests CI step runs before any workspace build, so the surfaces check found an empty dist/. The verifier now builds the pair like the other real-package suites do, and the trust/policy closure (trust, harness-tasks, harness-policy) does not touch the catalog build that the two hooks suites already run concurrently.

5. Local matrix (43 packages, `--skip-build --exactness`)

Command: node scripts/verify-release-artifacts.mjs --skip-build --exactness --report-dir <scratch>/matrix (one process, dependency order, checkout-packed internal deps). Windows limits of this machine (known from RN-05/RN-06, unrelated to exports): the bins step spawns npm's .cmd shim directly (spawnSync … ENOENT) for every bin-bearing package, and harness-core's own verify:release gate spawns bare npm (same ENOENT class). Both happen **after** the exactness step, so exactness is proven locally wherever the row says so; CI (ubuntu, release-artifact-verifier.yml) verifies the whole chain.

Outcome: the run was **stopped by the Claude Code session at 27/43 for critically low system memory** (not a verifier failure; per the session notice it was not restarted). Reports exist for 27 packages; the 16 rows marked NOT RUN were never reached and are verified by CI only. Counts over the 27 reports: exactness passed 27/27 (never skipped, never failed); 20 fully green (every step incl. bins where present); 6 bin-bearing packages stopped only at the Windows .cmd-shim bin smoke AFTER exactness passed; 1 (harness-core) stopped only at its own verify:release gate (bare npm spawn) AFTER exactness passed. Earlier single-package runs (harness-hooks-core, the verifier fixtures, the hooks-core→hooks-claude regression pair) are consistent. Raw counters: .

{"pass":20,"envBins":6,"envVerify":1,"other":16,"exactnessPassed":27,"exactnessNotReached":0}

6. In-repo check (AC3) — `scripts/ci/exports-exactness.cjs`

• --self-test: 8 fixtures under scripts/ci/fixtures/exports-exactness/ (@fabricstack/exx-* names, AR registry, outside the workspace globs — pass every other guard's scan): clean (0 findings, publics page in sync), wildcard-export, private-deep-import (packages/consumer/src/index.ts:12 @fabricstack/exx-lib/dist/internal.js), missing-subpath-file (./missing → dist/missing.js absent), no-exports-map (+ the consumer's /util import becomes unresolvable), bin-only-root-import, publics-drift, publics-missing → 8/8 fixtures behaved (7 negative, 1 positive).
• Repository scan (above): 0 findings.
• scripts/ci/exports-exactness.test.mjs behind npm run test:exports-exactness → 8/8: self-test, fixture matrix with diagnostics, repository scan, all 43 manifests policy-clean with no * anywhere (exports or typesVersions) and the synthetic probe unexported, the SDK's exact subpath set, specifier extraction over 11 import shapes (+ 4 non-imports ignored), publics generation deterministic/idempotent, CI wiring present.
• Wired into ci.yml job publish-isolation (dependency-free, before npm ci) as step "Exports exactness (WP-PKG-01)": self-test → repository scan → npm run test:exports-exactness. release-artifact-verifier.yml: new step "Exports-map exactness self-test and repository scan", --exactness passed to the matrix, pull_request paths extended with the new library and check.

7. Docs-site publics (AC5)

The repo has no generator of SDK reference pages from manifests (docs:* scripts generate process-library/CLI-example artifacts and run lint/snippets/links/freshness). Added docs/reference/package-publics.md, generated by npm run docs:publics (node scripts/ci/exports-exactness.cjs --write-publics): front matter the freshness gate requires (title, description, last_updated — restamped only when the body changes), one section per package with directory, bins and a table subpath | import specifier | runtime target | types; 43 sections. Drift is a finding of the CI check and of the node:test suite. Picked up by the docs-site automatically (sidebars.js autogenerates reference/). Gates: docs-style-check 7 rules / 0 issues, docs-link-check clean, markdownlint-cli2 0 errors over the page; docs-freshness-report fails only on four pre-existing stale generated docs (adapters/reference/04-agent-events.md, generated/cli-examples-verification.md, GAP-PROMPT-011.md, reference-repos/README.md, 71–114 days), untouched here.

8. Gates

8a. Review round 1 (M4): a live target-missing gate on a built tree

The review found that no job proved the declared targets on a built tree: the publish-isolation job scans an unbuilt checkout (221 targets in 40 packages *deferred*, by design) and the packed-artifact gate (release-artifact-verifier.yml) is disabled_manually, so a target that went missing after build was caught by nothing on a PR. Fixed:

• scripts/ci/exports-exactness.cjs --require-built (alias --strict): allowUnbuilt is off, nothing is deferred, every declared target of every publishable must exist on disk or it is a target-missing finding (exit 1). The report line says exports-exactness (strict, built tree: no unbuilt deferral): …. scripts/ci/exports-exactness.test.mjs gained the test "--require-built: an unbuilt package is refused (target-missing), never deferred; the built real tree passes strictly" (the unbuilt-package fixture: 4 findings, 0 deferred, CLI exit 1 with no deferral line; the tolerant scan of the same tree still exits 0; on the real tree strict findings == tolerant deferrals), and the CI-wiring test now also asserts the strict step sits after the last build step of the test job. Suite: 11/11.
• .github/workflows/ci.yml test job: new step **"Build agent-runtime-cli and operator (strict exports gate inputs)"** (the two publishables no earlier step of that job builds; tsc --build directly for agent-runtime-cli because its own build script re-runs build:runtime) followed by **"Exports exactness on the built tree (strict, WP-PKG-01)"** = node scripts/ci/exports-exactness.cjs --require-built, placed after "Build babysitter-harness" and before the agent-runtime suites. The tolerant scan stays in the dependency-free publish-isolation job (unbuilt checkout).
• Local proof on this (fully built) checkout: exports-exactness (strict, built tree: no unbuilt deferral): 43 publishable manifest(s), 6385 source file(s), 19013 specifier(s), 3305 workspace import(s) resolved, 0 finding(s) (exit 0). CI proof: see "CI evidence" below once the test job reaches the step.

**Still pending (owner decision)**: the full 43-package *packed* matrix (AC1 negative probes + AC2 fresh-consumer typecheck inside each tarball) runs only in release-artifact-verifier.yml, which is owner-disabled (workflow id 352401988); the local run reached 27/43 (§5). Re-enabling that workflow — or a dispatched run of scripts/verify-release-artifacts.mjs --exactness over the matrix with its report attached — is required before 7.0.0 and is outside what this task may do (no workflow enabling). The strict on-disk gate above proves the *source* of every tarball target on every PR; it does not replace the packed proof.

9. Honesty / deferred

1. **Matrix coverage on this machine**: exactness is proven locally for the 27 packages the run reached (all 27 passed the exactness step); the 16 listed as NOT RUN (agent-runtime-cli/core/engine/platform, console-tui/ui, control-plane, harness-channels/extensions/hooks-agent-runtime/mock/tools/triggers, observer, operator, orchestrator) were never reached because the Claude Code session stopped the run at 27/43 for critically low system memory and instructed not to restart it. Their maps pass the same static policy (exports-exactness over the checked-in tree, target-missing against disk) and the same probe generator; CI's release-artifact-verifier.yml (with --exactness) proves them on the pushed commits. No package is claimed verified that was not. 2. **Windows limits (pre-existing, unrelated to exports)**: the bins step spawns npm's .cmd shim directly (spawnSync … ENOENT) and harness-core's own verify:release gate spawns bare npm — both after exactness. test:babysitter-metapackage cannot run here (its before hook feeds GNU tar an absolute C:\… path); the stub/consumer fix in that test was made by reading and mirrors the real map shape (./cli resolved through exports; verified by hand in a temp consumer: require('@fabricstack/orchestrator-sdk/cli') → MODULE_NOT_FOUND naming the absolute dist\cli\main.js when the file is missing, ERR_PACKAGE_PATH_NOT_EXPORTED for the old deep path). vitest suites (rolldown win32 binding absent) and observer build:app (lightningcss) still cannot run here; the observer was typechecked with tsc instead. 3. **Verifier defect fixed in scope**: the multi-package run of verify-release-artifacts.mjs was broken since 1117e6993 (dependents installed from a tarball already deleted with the dependency's working directory); fixed by retaining the tarball in internal-deps, with a regression test. Without this fix the full matrix could not run at all. 4. **Scope calls**: ./cli and ./runtime on the SDK are deliberate publics (a spawnable CLI entry and a small commit-path graph) rather than consumer rewrites to the root, because the root cannot provide either property; ./config/./testing are public because they are documented. No wildcard was added anywhere, no test was deleted or skipped, no map was widened to make a probe pass. The private atlas/webui tsconfig paths (@fabricstack/catalog/* → source) is compile-time only and out of exports scope. 5. **Private workspace members without a map** (e.g. @fabricstack/memory) are legacy-resolved and only counted by the in-repo check — extending the policy to private members is a follow-up, not an AC of this WP. 6. artifacts/docs-qa/freshness-report.json (tracked, generated) was rewritten by the local freshness run and restored; it is not part of this change.