Repository navigation
integrate(mdx-8je): Cloudflare-native kernel round 1 — prune to workerd-only, mdxdb root, @mdxe/cli-core + @mdxe/tui seam, sqlite pools + $init, mdxld 1.9.2 alias release - #14
Merged
Conversation
…mark harness (mdx-8je.13)
Kestrel ADR-0052 §4 doctrine: renderers are pure tree → bytes; a TUI is a shell
that displays bytes and handles input, never a fourth renderer. @mdxe/tui is the
seam @mdxe/ink (.14) and @mdxe/opentui (.15) implement.
- Viewer { mount(stream, input), unmount }, MountHandle, Terminal, KeyEvent /
PasteEvent / ResizeEvent, Action, Keymap, ViewerError (NOT_A_TTY, ALREADY_MOUNTED)
- input abstraction: decodeInput (CSI keys, ctrl/meta chords, bracketed paste),
defaultKeymap, mapActions, createInputSource over a live terminal
- tty guard: canAttach / assertAttachable — never attaches on a piped stdout or
stdin (raw mode fails open on pipes)
- conformance suite (@mdxe/tui/conformance): strip(paint(bytes)) === bytes,
identical action mapping, refuses non-TTYs, unmount releases raw mode and
paints nothing afterwards, quit unmounts; runner-agnostic runConformance plus
describeViewerConformance binding; createFakeTerminal + fixtures exported
- benchmark harness (@mdxe/tui/benchmark): startupMs, installBytes (recursive
dir size, lazily imports node:fs), repaint p50/p95/max per streamed frame;
formatReport emits JSON; `pnpm --filter @mdxe/tui bench` runs it on the stub
- reference stub viewer (@mdxe/tui/stub) the suite is proven green against
- root entry is dependency- and framework-free; heavier pieces are lazy subpaths
Verified: 29/29 tests, tsc, eslint, tsup build, bench emits valid JSON.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…yaml} their own tsconfig so tsc stops sweeping the monorepo (mdx-8je.4) These packages ran `tsc --noEmit` with no tsconfig.json, so tsc walked up to the root tsconfig.json, which has no `include`, and type-checked all ~1,300 files in the repo: 11,802 phantom errors per package that hid the packages' real errors. Each package now extends @mdxld/config/typescript/library.json with include src/** and test/**, and declares @mdxld/config as a devDependency (lockfile importer blocks updated to match). The issue names markdown, typescript and diff; csv, html, json, xlsx and yaml had the byte-identical defect and get the byte-identical fix, so the new repo-level guard can assert the invariant without an exclusion list. tests/tsconfig-includes.test.ts asserts, for every package under packages/: any package whose typecheck script runs tsc has its own tsconfig.json, and every tsconfig's include/files entries, effective include roots after extends, and resolved file list stay inside the package directory (208 assertions; goes red when a tsconfig is removed). Result for the three named packages: 35,406 -> 32 errors, all real, all in src/index.test.ts (tracked in mdx-qxi). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…pm collapse (mdx-8je.2) The primitives/ submodule and workspace glob were already removed in 7346df8; this finishes the job and restores the type ai-database needs. - mdxld: export `MDXLD<TData>` as a deprecated alias of `MDXLDDocument`. Published ai-database@2.4.0 (dist/types.d.ts) does `import type { MDXLD } from 'mdxld'` and `ThingExpanded extends MDXLD`; no current mdxld exported that name. Type-level test guards the alias (structural equality + interface extension) under `tsc --noEmit`. - mdxld: ambient `*?raw` module declaration so primitives.test.ts typechecks (it was the only thing failing `pnpm --filter mdxld typecheck`). - @mdxdb/fs tsconfig: drop the `paths` alias into the deleted primitives/packages/ai-database checkout; ai-database now resolves from npm. - vitest.mdx.config.ts: drop the alias into primitives/packages/ai-sandbox. - test/repo/primitives-collapse.test.ts (+ vitest.repo.config.ts, `pnpm test:repo`, chained into `test` / `test:unit`): asserts no primitives/ workspace entry, submodule, or directory; no ai-* dep via workspace:/file:/link:; no tsconfig or vitest alias into a local primitives checkout. - Docs: CLAUDE.md + mdxe/mdxld READMEs describe primitives as npm ^2.4.0. Verified: pnpm install --frozen-lockfile (exit 0); mdxld typecheck, test (165 passing), lint clean on touched files; @mdxdb/fs builds against npm ai-database/mdxld; pnpm test:repo (5 passing). Model gaps filed: mdx-8je.20 (published ai-database imports MDXLD that no published mdxld exports; pnpm resolves its mdxld from the registry, so a mdxld release is needed), mdx-8je.21 (ai-sandbox unpublished but imported by @mdxe/vitest and @mdxld/evaluate; ai-tests unused so not added). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…test-pool-workers (mdx-8je.5) The Durable Object imports cloudflare:workers, which only workerd provides, so the suite is now two pools: - vitest.config.ts (node): schema + MDXClient tests under tests/*.test.ts - vitest.workers.config.ts (workerd): MDXDatabase DO tests under tests/workers/, with the DO declared in wrangler.test.jsonc (new_sqlite_classes). `pnpm test` runs both. The pre-rewrite tests (compliance, miniflare, schema) targeted the API removed in 19b37ab (events/actions/artifacts/search, createInMemoryBinding, ai-database/tests) and could not pass under any pool, so they are replaced by tests against the current _data/_rels surface. See mdx-8je.19. Harness notes (mdx-8je.23): isolatedStorage is off because the pool cannot snapshot SQLite-backed DOs (each test uses its own DO name instead), and the optional @mdxe/isolate peer is aliased to a throwing stub so the suite does not require a sibling build. The DO-name -> $id contract does not hold in workerd (ctx.id.name is undefined); kept as it.fails, tracked in mdx-8je.24. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ke function-body output runnable @mdxld/evaluate still dynamically imported 'ai-sandbox', which was renamed to ai-evaluate on 2025-12-20, so `tsc` failed with TS2307 and the dts build broke. - @mdxld/evaluate: depend on ai-evaluate ^2.4.0 (already in the lockfile via mdxe); import it in evaluateInSandbox; externalize it in tsup; forward the already-declared `timeout` and `allowNetwork` SandboxOptions, which were documented (default false) but never sent - ai-evaluate allows network by default. Drop the unsafe `as SandboxResult` cast: EvaluateResult is structurally assignable. - Tests: evaluate() of a trivial MDX module returns its default export; run(); frontmatter; and evaluateInSandbox() must answer with ai-evaluate's own worker_loaders refusal rather than a module-resolution error, so the test actually witnesses the import resolving. - @mdxld/compile: the new evaluate() test exposed that `function-body` output (the default) was not a function body. transformJSX hard-coded esbuild `format: 'esm'`; a top-level `return` makes esbuild treat the input as CommonJS, wrap it in a __commonJS shim and emit `export default`, which `new Function` rejects. TransformJSXOptions gains `format` (default 'esm'; `null` leaves the code unwrapped) and compile() passes null for function-body. Tests pin both output shapes. Verified: pnpm --filter @mdxld/evaluate build && test (6/6) && typecheck; pnpm --filter @mdxld/compile build && test (6/6) && typecheck; pnpm install --lockfile-only --frozen-lockfile --offline. Refs: mdx-8je.3 Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ds (mdx-8je.1)
oauth.do's root entry ships a tsup-mangled .d.ts (`ensureLoggedIn as e`,
TS2460/TS2339 for importers) and its runtime ESM bundle does not export
ensureLoggedIn at all, so `import { ensureLoggedIn } from 'oauth.do'` was
both a type error and `undefined` at runtime. Every published 0.1.x and
0.2.x build has the same shape, so pinning cannot fix it. Only the
`oauth.do/node` subpath re-exports the symbol under its real name.
- mdxe: add src/auth.ts as the single import site (re-exports from
oauth.do/node); commands/deploy.ts, commands/db.ts and the managed-deploy
E2E import from it. src/auth.test.ts asserts ensureLoggedIn/getToken are
functions so future drift fails in one place.
- @mdxe/do, @mdxe/cloudflare, @mdxe/github, @mdxe/vercel: import from
oauth.do/node (they do not depend on mdxe); test mocks follow.
- mdxe tests/deploy-extended: mock the auth wrapper. deployViaManagedApi
authenticates before it builds, so the spawn-mocked build tests were
silently depending on a live auth.apis.do login.
- mdxe optional-deps.d.ts: ambient shims for @mdxe/workers/local and
@hono/node-server, which dev-server.ts imports dynamically but mdxe never
declares; their TS2307s were hiding behind the oauth.do error in the same
DTS build. Tracked as mdx-nzh.
Verified: tsup build green for mdxe, @mdxe/do, @mdxe/cloudflare,
@mdxe/github, @mdxe/vercel; unit suites green (mdxe 759/762, the 3 reds are
the live-network E2E cases while auth.apis.do returns 522 - mdx-d5s).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Both packages duplicated primitives that already ship in the workspace: @mdxe/rpc reimplemented RPCServer/RPCClient over fetch while advertising capnweb (ai-functions RPC / RPCPromise), and @mdxai/agentkit overlapped autonomous-agents, which mdxai already depends on. - git rm packages/@mdxe/rpc and packages/@mdxai/agentkit - drop the docs page for agentkit and its fumadocs .source entries - README / CLAUDE.md / ROADMAP / mdxe + @mdxe/mcp docs now point at ai-functions RPC and autonomous-agents instead - tests/removed-packages.test.ts asserts no package.json in the workspace names or depends on either package, the directories stay gone, and mdxe keeps re-exporting RPC / RPCPromise types from ai-functions directly - tests/readme-taxonomy.test.ts updated for the removed @mdxe/rpc row - pnpm-lock.yaml: only the two importer blocks removed - changeset (minor for mdxe and mdxai; mdxe never re-exported @mdxe/rpc) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…LI (mdx-8je.16)
New packages/mdxe/src/cli/{context,caller,errors,args,orient,tokens}.ts, ported from kestrel (MIT, attributed in each header) with mdxe env names (MDXE_AGENT, MDXE_COLOR) and one extra ladder rung for HTTP faces (Accept).
- resolveOutputCtx: --json/--format > Accept > --agent/agent env > CI > !isTTY > TTY; NO_COLOR presence beats --color always; bad --format/--color is USAGE even alongside --json; frozen ctx, width=Infinity off-TTY.
- resolveCaller: env beats TTY (CLAUDECODE, CLAUDE_CODE_CHILD_SESSION, CURSOR_AGENT, GEMINI_CLI, OPENCODE, AI_AGENT, AGENT, MDXE_AGENT, CODEX_*), then CI, then TTY probe; interactive only for a confident human.
- errors: CliError with stable codes; EXIT {0,1,2,3,4,5,6,130}; REFUSED(6) distinguishes fail-closed-working from a crash; fail() renders to stderr only.
- args: extractGlobals peels the render/meta flags before the router; unknown flags rejected before value consumption.
- orient: bare invocation renders a task-grouped orientation (text/json/human) and exits 0 without reading stdin — it never starts the dev server.
- tokens: labelled TokenCount (tiktoken lazy-optional / chars-approx / fable-native offline cache that fails closed on a miss or base drift).
cli.ts: main(argv) resolves ctx+caller once, fast-paths help/version before any dynamic import, threads the frozen ctx via CliOptions.ctx, returns the exit code (process.exit only at the entrypoint, guarded so importing the module never runs main). parseArgs fails closed: unknown command, unknown flag, malformed --env, bad enumerations are USAGE (exit 2) instead of silent fallthrough to dev; requireValue/validatePort throw instead of process.exit(1); --path aliases --dir; tail receives --json/--no-color from the resolved ctx.
Tests: ported context/caller matrices plus args/errors/orient/tokens units, a source-level light-import guard, and a spawned-bin e2e (esbuild code-split build) proving CLAUDECODE=1 + open stdin => orientation, exit 0, no hang; USAGE exits 2 with stderr-only errors.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…e/mdx-8je # Conflicts: # CLAUDE.md # packages/mdxe/README.md
…te/mdx-8je # Conflicts: # CLAUDE.md
… worktree-wf_3d4080bf-321-17
… worktree-wf_3d4080bf-321-17 # Conflicts: # CLAUDE.md # packages/mdxe/README.md
… worktree-wf_3d4080bf-321-17 # Conflicts: # CLAUDE.md
ReviewReviewed the merge diff (113 files, ~5.4k additions) integrating the 7 FindingsMajor — stale/broken guard test contract ( Minor — CLAUDE.md dependency tree missed the Minor / worth a second look — unsandboxed eval as the default API ( Verified clean / positives worth calling out
Test coverageNo package with meaningful source changes was found lacking a corresponding test file; the newest packages ( Nice work keeping this integration tight — the two real gaps above are documentation/consistency issues rather than runtime bugs, and the RPC re-export was already flagged by the author as a known follow-up. 🤖 Generated with Claude Code |
…-8je.30) Repo-level invariant tests were split between tests/ (on no script's include path; turbo only runs per-package scripts) and test/repo/ (pnpm test:repo, branch 3), so removed-packages, tsconfig-includes and readme-taxonomy sat on no CI path at all. - test/repo/ is the single canonical home; tests/ is .mdx fixtures only - vitest.repo.config.ts includes test/repo/**/*.test.ts, and sweeps tests/**/*.test.ts so a misplaced guard still executes instead of silently going dark - test/repo/guard-test-home.test.ts fails the build with a git-mv pointer if a *.test.ts appears under tests/, and asserts the package.json + ci.yml wiring cannot regress - tests/readme-taxonomy.test.ts moved to test/repo/ - package.json: test:repo script, chained into test and test:unit (byte-identical to branch 3's lines so the merge is clean) - ci.yml: "Run repo guard tests" step (pnpm test:repo) in the test job - CLAUDE.md + test/repo/README.md document the convention Post-merge note: branches 5 and 7 add removed-packages.test.ts and tsconfig-includes.test.ts under tests/; the guard will name them and they need `git mv tests/<f> test/repo/`. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… ai-sandbox (mdx-8je.21)
@mdxe/vitest's generateTestCode() pushed `import { evaluate } from 'ai-sandbox'`
into every generated test file that needs the sandbox (JSX, hooks, Hono app).
ai-sandbox was renamed ai-evaluate on 2025-12-20; npm ai-sandbox is a 0.0.0
placeholder and the primitives/ submodule alias is gone, so those files could
not resolve in any install.
- Emit `ai-evaluate/node` (exported as SANDBOX_MODULE). Generated files run
under vitest in Node, so the /node entry (Miniflare fallback) is the right
one; the root entry requires a Cloudflare worker_loaders binding.
- Declare ai-evaluate >=2.4.0 as an optional peerDependency: the generated
import resolves from the consumer project via the vitest transform plugin,
the same way `vitest` itself is a peer. Add it as a devDependency so the
test suite can witness resolution. Lockfile: only the @mdxe/vitest importer
entry is added (ai-evaluate@2.4.0 was already resolved); the unrelated
--no-frozen-lockfile churn (mdx-33k) was reverted by hand.
- Tests: no sandbox import for plain tests; sandbox tests import from
ai-evaluate/node and never ai-sandbox; the emitted specifier actually
imports to a real evaluate() (ai-evaluate/node lazy-loads miniflare, so
this has no side effects).
- README: document sandboxed tests and the optional peer. CLAUDE.md: the
`mdxe test` line still named ai-sandbox.
The other call site named by the issue, @mdxld/evaluate evaluateInSandbox,
was already repointed to ai-evaluate in 8cbbba9 (mdx-8je.3); nothing in the
repo references ai-sandbox as a module specifier any more (remaining hits
are comments and a wrangler service name).
ai-tests: not added. ai-evaluate's `tests` mode already provides
describe/it/expect inside the sandbox; ai-tests is only the optional TEST
service binding for Workers deployments and nothing here consumes it.
Verified: pnpm --filter @mdxe/vitest test (132/132, was 129) and build green;
typecheck (32x TS2532) and lint (10) fail identically before and after this
change on pre-existing lines - filed as mdx-b7m.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ot ship (mdx-8je.27)
ai-functions@2.4.0 exports no RPC or RPCPromise (dist/index.d.ts has zero
occurrences; no ./rpc subpath). The type re-export in src/index.ts was TS2305
and failed tsup's DTS step, so mdxe built no .d.ts. Remove the re-export and
the README/INTEGRATION docs that advertised 'import type { RPC, RPCPromise }
from mdxe', plus the no-op 'should have RPC types available' test.
Witnessed: tsc --noEmit 40 -> 38 errors, 0 in src/index.ts; tsup DTS exits 0
and ships dist/index.d.ts when combined with the oauth.do fix from mdx-8je.1
(b8e95c7), which is the only remaining DTS error.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ate/mdx-8je .31 reconciled .2 -> .6 -> .13 in parallel with round 1's integration; net delta on this branch is the @mdxe/tui entry in the dependency tree, the root README's stale ./primitives/ link -> npm link, and one CLAUDE.md line. Conflict: CLAUDE.md ai-functions line (two wordings of the same claim) -> kept round 1's wording.
…ate/mdx-8je Conflict: vitest.repo.config.ts add/add (round-1 .2's include of test/repo only vs .30's include of test/repo plus a sweep of tests/ so a stray guard still executes) -> took .30's, a superset.
…ds from tests/ to test/repo/ The post-merge step .30 named: round-1's .6 and .4 added these under tests/, which no script ran; test/repo/guard-test-home.test.ts now flags them. Moved with git mv and repointed the repo root from __dirname/.. to __dirname/../..
nathanclevenger
changed the title
…ot ship (mdx-8je.27)
ai-functions@2.4.0 exports no RPC or RPCPromise (dist/index.d.ts has zero
occurrences; no ./rpc subpath). The type re-export in src/index.ts was TS2305
and failed tsup's DTS step, so mdxe built no .d.ts. Remove the re-export and
the README/INTEGRATION docs that advertised 'import type { RPC, RPCPromise }
from mdxe', plus the no-op 'should have RPC types available' test.
Witnessed: tsc --noEmit 40 -> 38 errors, 0 in src/index.ts; tsup DTS exits 0
and ships dist/index.d.ts when combined with the oauth.do fix from mdx-8je.1
(b8e95c7), which is the only remaining DTS error.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ing root scripts (mdx-8je.32) The Test job ran 'pnpm test:integration' and 'pnpm test:e2e' but root package.json defines neither (ERR_PNPM_NO_SCRIPT), so the job went red after the unit-test step on every push/PR to main. The package-level test:integration / test:e2e scripts (@mdxdb/github, @mdxdb/do, @mdxdb/studio, @mdxui/editor, mdxld-vscode) all need resources CI does not provision (GITHUB_TOKEN + test repo, ClickHouse, wrangler/workerd, Playwright browsers, VS Code download), so wiring them as turbo tasks would keep the job red. They remain runnable per package locally; a CI lane for them is tracked separately. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…mdx-8je.34)
mdx-8je.27 dropped 'export type { RPC, RPCPromise } from ai-functions' from
packages/mdxe/src/index.ts (TS2305; ai-functions@2.4.0 ships neither symbol),
but the mdx-8je.6 guard pinned that exact text and the mdx-8je.31 docs
re-asserted it, so integrate + .27 turned test/repo/removed-packages.test.ts
red on pnpm test:repo / CI (mdx-8je.30 sweep).
- test/repo/removed-packages.test.ts: replace the source-text pin with its
negation - mdxe imports nothing from ai-functions, and no source file under
packages/ imports RPC / RPCPromise from ai-functions. Witnessed red when the
re-export is appended back, green without it.
- packages/mdxe/package.json + pnpm-lock.yaml: drop ai-functions, which
nothing in mdxe imports after .27 (frozen-lockfile clean).
- Docs (CLAUDE.md, README.md, packages/mdxe/{INTEGRATION,ECOSYSTEM}.md,
packages/@mdxe/mcp/README.md, .changeset/remove-rpc-and-agentkit.md):
stop claiming mdxe re-exports RPC types or that ai-functions ships RPC;
point at rpc.do (RPC<T>(url), RPCPromise; capnweb via @dotdo/capnweb),
which @mdxdb/rpc, @mdxdb/server and @mdxdb/do already depend on.
Verified: pnpm test:repo 241/241; mdxe tsc 30 errors, 0 in src/index.ts,
0 mentioning ai-functions; mdxe vitest 866 passed / 7 failed, the same 7
failing at baseline (oauth.do drift mdx-8je.1, unbuilt @mdxe/fumadocs dist).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… into integrate/mdx-8je
nathanclevenger
changed the title
… (mdx-8je.45)
`changeset status` (and therefore scripts/release.sh) failed at config validation:
- `ignore: ["@mdxld/config"]` errored for all 61 publishable packages that
devDepend on it. @mdxld/config is private, so it is never published anyway; drop
the ignore entry and set privatePackages {version:false, tag:false} instead.
- linked group `["mdxdb", "@mdxdb/*"]` errored because no `mdxdb` root package
exists yet (mdx-8je.11 creates it). Use one brace glob so the group is valid now
and picks up the root package once it lands.
Verified: `changeset status` and `changeset version` now pass config validation
(remaining non-fatal dependency warnings tracked in mdx-8je.54).
Same change as 6b28125 on worktree-wf_3d4080bf-321-16 (filed there under mdx-8je.20).
Closes mdx-8je.45
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ds (mdx-8je.1)
oauth.do's root entry ships a tsup-mangled .d.ts (`ensureLoggedIn as e`,
TS2460/TS2339 for importers) and its runtime ESM bundle does not export
ensureLoggedIn at all, so `import { ensureLoggedIn } from 'oauth.do'` was
both a type error and `undefined` at runtime. Every published 0.1.x and
0.2.x build has the same shape, so pinning cannot fix it. Only the
`oauth.do/node` subpath re-exports the symbol under its real name.
- mdxe: add src/auth.ts as the single import site (re-exports from
oauth.do/node); commands/deploy.ts, commands/db.ts and the managed-deploy
E2E import from it. src/auth.test.ts asserts ensureLoggedIn/getToken are
functions so future drift fails in one place.
- @mdxe/do, @mdxe/cloudflare, @mdxe/github, @mdxe/vercel: import from
oauth.do/node (they do not depend on mdxe); test mocks follow.
- mdxe tests/deploy-extended: mock the auth wrapper. deployViaManagedApi
authenticates before it builds, so the spawn-mocked build tests were
silently depending on a live auth.apis.do login.
- mdxe optional-deps.d.ts: ambient shims for @mdxe/workers/local and
@hono/node-server, which dev-server.ts imports dynamically but mdxe never
declares; their TS2307s were hiding behind the oauth.do error in the same
DTS build. Tracked as mdx-nzh.
Verified: tsup build green for mdxe, @mdxe/do, @mdxe/cloudflare,
@mdxe/github, @mdxe/vercel; unit suites green (mdxe 759/762, the 3 reds are
the live-network E2E cases while auth.apis.do returns 522 - mdx-d5s).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…mdx-8je.1) Second break in the same mdxe DTS build, hidden behind the oauth.do error and invisible to the earlier verification: commit 7346df8 moved mdxe to ai-functions ^2.4.0, and the 2.4.0 primitives train no longer exports RPC or RPCPromise from ai-functions (they live in rpc.do now). src/index.ts still re-exported both, so `tsup` failed with TS2305 on the committed lockfile. Nothing in packages/, apps/ or examples/ imported those types via mdxe, and mdxe does not depend on rpc.do, so the re-export is removed rather than re-pointed (a new dependency would need a lockfile change). @mdxe/rpc has the same stale re-export; tracked as mdx-8je.55 (model-gap). Verified against the committed lockfile in a fresh worktree (not symlinked node_modules): turbo build green for mdxe, @mdxe/do, @mdxe/cloudflare, @mdxe/github, @mdxe/vercel; unit suites do 57/57, cloudflare 85/85, github 89/89, vercel 158/158, mdxe 759/762 (3 reds = managed-deploy E2E against auth.apis.do, still 522 - mdx-d5s). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… (mdx-8je.55) Items 1 and 2 of mdx-8je.55 are already true on integrate/mdx-8je, which this branch is based on: packages/@mdxe/rpc is deleted (mdx-8je.6), no source file imports RPC / RPCPromise from ai-functions (removed-packages guard, mdx-8je.34) and CLAUDE.md points RPC users at rpc.do. Item 3 - verification-by-symlink resolving a different lockfile than the committed one - gets a mechanism instead of a checkout policy: a repo guard that reads the importers section of pnpm-lock.yaml and asserts every primitive / .do dependency is installed under its importer at exactly the resolved version, and that the ai-* train is one version workspace-wide. Witnessed red with ai-functions faked at 0.4.0 under packages/mdxai, green after pnpm install --frozen-lockfile. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…base with @mdxdb/do as primary backend (mdx-8je.11)
packages/mdxdb@2.0.0 (the npm name was never published; 1.x never existed):
- DB(schema, { url?, env?, provider? }) — synchronous facade over ai-database's
DB(); the backend resolves lazily from DATABASE_URL (options.url, then
env.DATABASE_URL, then process.env.DATABASE_URL, then ./content).
- parseDatabaseUrl(): do://<name>[?binding=&vectorize=&r2=] (primary, @mdxdb/do),
sqlite://<name> (alias), ./content|file:// (@mdxdb/fs),
clickhouse[s]://user:pass@host:port/db (@mdxdb/clickhouse HTTP),
http(s):// (@mdxdb/api), :memory:. Unknown schemes, libsql:// and chdb://
fail closed — never a silent memory fallback.
- DOProvider: ai-database DBProvider over an MDXDurableObject Workers RPC stub
(typed structurally, no @mdxdb/do import). Flattens Things to
{ ...data, $id, $type }, $init(name) handshake with $id() fallback, maps DO
errors onto EntityAlreadyExistsError/EntityNotFoundError, writes the schema's
reverse predicate on edges so relatedBy works natively.
- ApiProvider over @mdxdb/api DBApiClient.
- withSchemaRelations(): 'Author.posts' reverse fields resolve on every backend
(single relation fields become edges on create/update).
- mdxdb/tests: shared provider contract suite (things, listing, relationships,
search, and the CLAUDE.md example through DB()).
Tests: 109 node (url, resolve, DO provider on an in-process fake, contract on
memory/@mdxdb/fs/fake DO, DB() facade) + 26 under @cloudflare/vitest-pool-workers
against the real @mdxdb/do MDXDurableObject in workerd, including the CLAUDE.md
"Database Interface" example verbatim. typecheck/lint/build green.
CLAUDE.md: Database Interface imports DB from 'mdxdb'; DATABASE_URL list now
matches what ships.
Model gaps filed: mdx-8je.56 (ai-database DI ignored by entity ops),
mdx-8je.57 (backref parsed without direction, DBPromise untyped getters),
mdx-8je.58 (chdb://, sqlite://path, libsql:// documented but unimplemented),
mdx-8je.59 (@mdxdb/do ships no .d.ts), mdx-8je.60 (stale mdxdb CLI references).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…(mdx-8je.7)
mdx.org.ai is Cloudflare-native only: code executes through Dynamic Worker
Loaders (workerd) in production and under Miniflare locally; Node and Bun are
thin CLI shells, never evaluation runtimes. Delete the packages that existed to
serve other deployment models and pin the surviving surface with an allowlist.
Removed (history stays in git):
- @mdxe/{node,bun,next,honox,electron,expo,remotion,slidev,vercel,github,payload}
- @mdxdb/{desktop,mobile,mongo,postgres,payload,studio,git}
- apps/desktop, apps/mobile (workspace:* consumers of electron/expo)
'?' entries decided: keep @mdxe/fumadocs (OpenNext on Workers), @mdxdb/fumadocs,
@mdxdb/github (Octokit/fetch), @mdxdb/rpc (rpc.do), @mdxdb/server (Hono),
@mdxdb/sources (until mdx-8je.17); drop @mdxdb/git (node:fs + isomorphic-git
http/node). @mdxdb/sqlite, @mdxe/cloudflare, cli-core and tui were missing from
the issue text and are kept (mdx-8je.62).
Tests first: test/repo/package-allowlist.test.ts enumerates packages/@mdxe and
packages/@mdxdb against the allowlist, asserts the removed dirs and app shells
are gone, and that no package.json or source file references a removed package.
Consumers:
- @mdxe/deploy: Platform is 'do' | 'cloudflare'; Vercel/GitHub providers,
types, helpers and vercel.json detection removed; deps now workspace:*.
- mdxe: 'admin' command and 'db studio' gone; --platform accepts do|cloudflare;
DeployOptions.platform / SourceTypeInfo.adapter narrowed; @mdxe/deploy is
workspace:*; optional-deps and tsup externals trimmed.
- @mdxe/cloudflare: adapter 'do' replaces 'postgres' | 'mongo'.
- @mdxai/claude: mdxe_deploy platform enum do|cloudflare|custom.
- @mdxe/fumadocs, @mdxui/notebook: "test": "vitest run" so root pnpm test can
terminate (mdx-8je.61).
- Docs, READMEs, CLAUDE.md, ROADMAP, ECOSYSTEM, publish scripts, apps/docs
.source, readme-taxonomy guard updated; pnpm-lock trimmed (-14k lines).
Changeset: major for mdxe, @mdxe/deploy, @mdxe/cloudflare; minor @mdxai/claude.
npm deprecations for the 4 published packages are tracked in mdx-8je.64 (the
local npm token is 401).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…te/mdx-8je Adds test/repo/lockfile-drift.test.ts: every primitive / .do dependency in pnpm-lock.yaml importers must be installed under its importer at exactly the resolved version, and the ai-* train must be one version workspace-wide. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
nathanclevenger
changed the title
…ient (mdx-8je.60)
mdxdb 2.0 is a DB() facade library; it has no src/cli.ts, no DBClient, and no
Database/NotFoundError/createDB/createApiClient exports. Bring the automation
and docs in line with what actually ships:
- .github/workflows/mdxdb-release.yml: deleted. It ran `bun build src/cli.ts
--compile` in packages/mdxdb and published GitHub-release binaries; tagging
mdxdb-v* would run red. Nothing links to it.
- .github/workflows/mdxe-release.yml: release notes no longer claim a bundled
"mdxdb CLI".
- package.json `studio`: retargeted from the non-existent
packages/mdxdb/dist/cli.js to @mdxdb/studio's real bin (dist/cli.js --path).
- docs/(Structured Data)/mdxdb/{getting-started,adapters/index}.mdx: the
document-level Database interface and its option/result types come from
the adapters (@mdxdb/fs); `mdxdb` is documented as the DB() facade. Error
handling now matches reality: get() returns null on both surfaces, and the
facade throws EntityNotFoundError/DatabaseError from writes.
- examples/docs.buildercss.dev/views/editor.mdx, packages/mdxe/ECOSYSTEM.md,
packages/@mdxdb/fumadocs/README.md: samples use DB() / @mdxdb/api.
vitest.setup.ts (MemoryDBClient/DBClient imports) is intentionally left to
mdx-8je.39, which owns the root MDX test harness.
Verified: package.json parses, mdxe-release.yml parses as YAML, the three
changed .mdx files compile with @mdx-js/mdx 3.1.1, and every remaining
`from 'mdxdb'` import names an export in packages/mdxdb/src/index.ts (877abc3).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…built (mdx-8je.62) The mdx-8je.7 issue text listed @mdxdb as "do, vectorize, parquet, clickhouse, fs, git?, github?, api, rpc?, server?, sources" and @mdxe without cloudflare, cli-core or tui, while mdx-8je.5/.19/.22/.23/.24 built @mdxdb/sqlite, mdx-8je.26 built @mdxe/cli-core, mdx-8je.13 built @mdxe/tui and @mdxe/deploy needs @mdxe/cloudflare. The allowlist file is now the source of truth: every entry is tagged ISSUE / '?' / SIBLING with the reason it stays, the four '?' entries are decided (git DROP; github, rpc, server, fumadocs KEEP; sources KEEP until mdx-8je.17), and the "rpc-via-ai-functions" line is recorded as deliberately not a directory (rpc.do; test/repo/rpc-home.test.ts). Guard the reverse direction too: an allowlisted entry with no package.json on disk fails unless it is in PENDING with the issue that creates it (@mdxe/opentui -> mdx-8je.15); PENDING must itself be a subset of ALLOWED and not yet exist; ALLOWED and REMOVED_PACKAGES must be disjoint. Witnessed red on a stale entry, green on the tree. Prettier-formatted (the file was already non-conforming). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…o integrate/mdx-8je # Conflicts: # packages/mdxe/src/index.ts
… into integrate/mdx-8je # Conflicts: # packages/mdxe/README.md # packages/mdxe/src/index.ts
…te/mdx-8je # Conflicts: # pnpm-lock.yaml
… into integrate/mdx-8je # Conflicts: # package.json
…e.11 and mdx-8je.7 pnpm install --frozen-lockfile failed after the merges because the mdxdb root package's importer was missing (its branch conflicted with the prune's lockfile and HEAD's was kept). Regenerated with pnpm install; the only change is the added importer, no resolved-package churn. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The merge of worktree-wf_3d4080bf-321-2 kept a comment naming the removed RPC promise type verbatim, which test/repo/removed-packages.test.ts forbids in packages/mdxe/src/index.ts. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The brace glob was a workaround from before packages/mdxdb existed (mdx-8je.45 predates mdx-8je.11). Same match set; changeset status unchanged. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
mdx-8je.7 narrowed DeployOptions.platform to 'do' | 'cloudflare' | 'custom' and the CLI to 'do' | 'cloudflare', but src/schemas.ts still accepted vercel/github/netlify (and DeployOptionsSchema rejected 'do'). Align both enums, pin the rejection in validation.test.ts, drop the cli.test.ts case that cast 'vercel' to 'do', and remove the dead 'ai-functions' tsup external. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ate/mdx-8je # Conflicts: # packages/@mdxdb/sqlite/src/client.ts
…h report field - mdxdb resolve.ts mapped any error whose message contained 'not found' to AdapterNotInstalledError, which would mislabel an installed adapter's own init failure. Match only module-resolution errors (code or message). - @mdxe/tui BenchmarkReport gains an optional installPackages so @mdxe/ink's bench no longer smuggles it through a cast. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
nathanclevenger
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Lands epic mdx-8je (Cloudflare-native kernel, round 1) on main: rounds 1-4 from the integrate branch plus the round-5 work that had never been integrated.
Included issues
packages/mdxe/src/auth.ts; RPC re-export dropped), .2 (primitives from npm^2.4.0, MDXLD alias), .3, .4, .14, .16, .18, .21, .25, .27, .28, .29, .31, .33, .34, .35, .39, .40, .43, .44, .54, .55, .57, .58, .59, .61test/repo/package-allowlist.test.ts), .6 (@mdxe/rpc,@mdxai/agentkitremoved; RPC home isrpc.do), .62 (allowlist reconciled: keep sqlite, cloudflare, cli-core, tui), .60 (never-shipped mdxdb CLI references dropped;mdxdb-release.ymldeleted)packages/mdxdb2.0.0 —DB()facade over ai-database,DATABASE_URLresolution,@mdxdb/doprimary backend)$init(name)handshake — was missing from integrate; merged fromworktree-wf_3d4080bf-321-15)@mdxe/cli-coreleaf: .26, .30, .32 (ci.yml:test:integration/test:e2esteps removed — those root scripts never existed on main, so no coverage is lost; lanes tracked in .36), .45 (changesets config repaired)@mdxe/tuiseam + Ink 7 viewer: .13, .15 (opentui allowlisted ahead), .17@mdxe/viteston ai-evaluate: .37 (h/Fragment bound for JSX blocks).changeset/mdxld-restore-mdxld-alias.md→ mdxld 1.9.2;test/repo/mdxld-surface-for-primitives.test.ts)Not included: mdx-8je.9 (blocked on primitives aip-263g; no commits exist), mdx-8je.8 (mdxui deletion deferred — nothing under
packages/@mdxuiorpackages/mdxuiis removed here), mdx-8je.64 (npm deprecations run by the publisher after this merge).Integration fixes made on this branch
packages/mdxdbimporter was added.packages/mdxe/src/schemas.tsplatform enums aligned with the pruned targets (do/cloudflare); cast-hidden vercel test removed.mdxdb/src/resolve.tsmaps only module-resolution errors toAdapterNotInstalledError.@mdxe/tuiBenchmarkReport.installPackagestyped (no cast in@mdxe/inkbench).Verification (worktree, fresh
pnpm install --frozen-lockfile)pnpm test:repo: 10 files / 251 tests, typecheck lane clean.typecheck build testfor mdxld, mdxdb, mdxe, @mdxe/tui, @mdxe/ink, @mdxe/cli-core, @mdxld/evaluate, @mdxld/compile, @mdxdb/sqlite, @mdxdb/do: all builds and tests green (mdxe 800 passed / 3 skipped; mdxdb 109 + 26 workers; sqlite 56 + 38 workers; tui 30; ink 20; cli-core 75; mdxld 165). Typecheck: mdxe 7 pre-existing errors (37 on main), @mdxdb/do 62 pre-existing (66 on main) — filed as mdx-w89.1.turbo run build --continue: 87/91. Rootturbo run test --continue: 156/171. The failing set is exactly the pre-existing red recorded in mdx-8je.63 — builds: @mdxai/claude, mdxai, @mdxui/plan, storybook; tests: @mdxdb/{api,rpc,clickhouse,parquet,fumadocs}, @mdxui/{headless,jsx,radix,slides,video}, mdxai. Nothing new.Follow-ups filed under mdx-w89
mdx-w89.1 (@mdxdb/do typecheck + dts), mdx-w89.2 (root
ai-evaluateentry used outside workerd in mdxerunTestand @mdxld/evaluate), mdx-w89.3 (release versioning: linked groups push mdxe/@mdxe//@mdxdb/ to 3.0.0; ink/tui/mdxdb have no changesets), mdx-w89.4 (review nits).🤖 Generated with Claude Code