Roadmap & status
Weft is built in phase order. Each phase has explicit acceptance criteria, and a later phase’s packages are not started before the current phase’s acceptance passes.
Status: released as weft 1.1.0 on npm (@michaelfromyeg/weft-*), renamed from “loom”.
Phases 0-4 plus the post-v1 work in Phase 5 are done; everything under “Beyond v1” is the
honest not-yet-built boundary.
Phase 0: Prove the compile (Claude Code only), DONE
Packages: @michaelfromyeg/weft-schema, @michaelfromyeg/weft-core, @michaelfromyeg/weft-adapter-kit, @michaelfromyeg/weft-adapter-claude,
@michaelfromyeg/weft-cli.
Acceptance (all met, exercised by pnpm test):
weft buildemits a valid.claude-plugin/marketplace.json+plugin.json+ placed component files, with zero hand-written JSON.claude plugin validate <generated-marketplace>exits 0 (and--strict).weft installplaces files under the scope and writes aweft.lock.- Re-running
build/installproduces identical hashes (deterministic). - Unit tests for schema parsing (incl. the YAML 1.2 “Norway” cases) and an e2e fixture test pass under Vitest.
@michaelfromyeg/weft-adapter-kitwas created in Phase 0 (not Phase 1 as the spec lists it) because@michaelfromyeg/weft-adapter-claudeimplements itsHarnessAdapterinterface. Only the interfaces + helpers exist so far; theHarnessDriverimplementations land in Phase 1.
Brought forward from later phases (already implemented)
- Multi-plugin marketplace build.
weft buildon amarketplace.yamlresolves and compiles many plugins into one native catalog (the company-marketplace workflow). An entryversionoverride flows into the compiledplugin.json. - Piecemeal install.
weft install <plugin> --only <component>installs a single skill/MCP; the manifest reflects only the selection. - Generated CLI map.
weft docsemits a full Markdown CLI reference from the command tree (see cli.md); never hand-maintained.
Vocabulary note: the canonical authoring unit is a plugin (cross-harness), and a marketplace packages many plugins. There is no separate “bundle” concept. See concepts.md.
Phase 1: Breadth + drivers + deterministic evals, DONE
Packages: the four remaining @michaelfromyeg/weft-adapter-* (codex, cursor, copilot, opencode),
@michaelfromyeg/weft-eval (runner + five HarnessDrivers via execa).
- All five adapters implemented;
weft buildemits native manifests for every harness. - A plugin installs to every harness present on the machine (
weft installdetects via the drivers and skips/reports the rest;--alloverrides). weft evalruns trace + output assertions via the real headless drivers (Claude / Codex / Cursor / OpenCode NDJSON/JSONL parsing; Copilot has no trace, sotracedegrades tooutput) and reports per-harness coverage honestly, incl. UNTESTED.- Piecemeal install (
--only) and namespacing/alias. - A community adapter can be authored against
@michaelfromyeg/weft-adapter-kitwithout importing core internals (verified: every adapter imports only adapter-kit + schema). - Remote resolver.
github:/git/file://sources are git-cloned into~/.weft/cacheand pinned to a SHA;dependsvendors selected components into a merged tree (piecemeal + drift-aware copy of shared assets) with cycle detection; deps recorded inweft.lock. - Secrets (
ConfigVardeclare-not-store) resolved from env/default to a gitignored.weft/secrets.local.json, never to the lockfile/plugin/index. weft updatere-resolves, recompiles, and re-places ONLY artifacts whose content hash changed (content-addressed; an unchanged artifact is never rewritten).
Many adapter facts beyond Claude are marked
// TODO(verify):in-code where upstream docs are thin (e.g. Copilot marketplace shape, Cursor remote MCP header serialization). They are isolated behind each adapter’stargetSchemaper spec §2. (Codex’s plugin + marketplace format is now CONFIRMED as of the v0.121 adapter rewrite – see harness-research.md.)
Key verified facts for the drivers (see harness-research.md):
Claude needs --output-format stream-json --verbose for a trace (plain json is a single
result object); Codex exec --json; Cursor --output-format stream-json; Copilot has no
structured trace (degrade to output); OpenCode run --format json or the SDK /event.
Phase 2: Index, federation, badges, CI, DONE
Package: @michaelfromyeg/weft-index (build + client + MCP-Registry federation + badges + publish gate).
weft index <dirs...>builds aweft.index/1(metadata only) from a set of plugins.--federateingests the MCP RegistryGET /v0.1/servers(injectable fetch; offline-tested) intofederated[]+ MCP-only entries.valid/testedbadges compute from validation + eval results (testedonly when a real harness passes;harnessCoverageis the passing set, never UNTESTED).- Opt-in aggregate telemetry (
installscount; no per-user data). weft publishruns the deterministic gate (static valid + trace/output evals) and exits 1 on failure;.github/actions/weft-publish+publish.ymlblock a failing publish in CI.
Phase 3: Trust & subjective evals, DONE
- Judge evals (injectable model, advisory unless
gate:true) plus differential evals that compare a case’s deterministic score to a committed baseline; a regression below the threshold blocks.evals/.baselines/snapshotting viaweft publish --snapshot. - Security scan (
scannedbadge): a built-in heuristic scanner over executable/hook/ passthrough artifacts (garak / AI-Infra-Guard would plug in for production). - Signing (
signedbadge): ed25519 over the lockfile’s artifact-hash digest;weft sign/weft verifydetect both a bad signature and tampered on-disk artifacts. (sigstore/cosign keyless signing is the intended production backend.) - Managed-mode install gating:
weft install --managed <namespaces>blocks a non-allowlisted namespace (and supports required-badge policies).
Out of scope for v1 (noted in the spec): a hosted CI eval tier (BYO-keys local is the default) and an index UI. The deterministic gate already runs the same driver invocations a hosted runner would.
Phase 4: Lifecycle & federation extras
Beyond the original spec, two lifecycle commands round out the loop:
weft importreverse-compiles an existing native plugin or marketplace into the Weft model, so you can cross-compile assets you already maintain (federate, don’t wall off). All five harnesses are supported through each adapter’simportNative, so import is any-to-any (weft import --from <harness>); a Weft -> Claude -> Weft -> Claude round-trip passesclaude plugin validate --strict.weft uninstallremoves everythinginstallplaced, using the paths recorded inweft.lock, then deletes the lockfile.
Phase 5: Rename, target-side lifecycle, and breadth, DONE
Renamed from loom to weft (no backwards compat) and released 1.0.0 then 1.1.0 to
npm; the repo is github.com/michaelfromyeg/weft. Tag-triggered CI publishes every
@michaelfromyeg/weft-* package and cuts the matching GitHub release.
- Rebrand with no compat:
weft.yaml/weft.lock/weft.index/1/weft_min_version, theweftCLI,~/.weft/cache,.weft/secrets. A weft install will not read old loom artifacts. - The lockfile is the install target’s ledger:
weft.locklives with the project you installed into (not the source), holds every installed plugin in one file, and accumulates across installs;weft uninstallruns from the project (one plugin or all). This is what makes remote and marketplace installs behave sensibly. weft install <marketplace>installs every plugin in a marketplace across the detected/specified harnesses in one command (the same primitive at marketplace scale).- Remote + npm install: a
github:/git/npm:/owner/repotarget is fetched into~/.weft/cache(git clone ornpm pack), with a trailing//subdir. (Also under Resolver depth below.) - Evals work for every component kind. The runner is component-agnostic; the Claude driver
allows
Skill(skills) +Task(sub-agents) underacceptEdits(which keeps the model answering rather than executing); passthrough gained anevalsfield; hook/passthrough are evaluated via asetup+verifyshell check.fixtures/all-kinds+ a stub-driver test cover all six kinds. MCP/command real activation needs harness-specific tool names the static driver can’t enumerate, so those are reported UNTESTED there, never faked. - Output assertions honor
minPassRateacrosssamples, so a flaky one-shot LLM answer passes on a majority instead of needing all N (mirrors trace assertions). - Generic skills-only adapter for the
.agents/skills/family (zed/gemini/amp/aider). (Detailed under More harnesses below.) weft eval --compare <ref>: the “vibes” A/B. Run each case against a prior git ref (or a dir) and the working tree, and print the two transcripts side by side for a human (or the pairwise judge) to pick.weft build --bare: write a single target straight to--outwith no<target>/subdir, so a repo root becomes a harness-native marketplace.weft build . --target claude --out . --baremakesclaude plugin marketplace add github:owner/repowork;--target codexwrites.agents/plugins/marketplace.jsonforcodex plugin marketplace add owner/repo.
Beyond v1 (planned)
Not built yet; tracked so the boundary is honest. Grouped by theme, roughly in priority order within each group. Nothing here blocks using Weft today; most items swap a production backend in behind an interface that already exists.
Trust & supply chain
The one real gap in the trust story is ownership. Everything else here is a backend swap.
- The
verifiedbadge: proveowner.namespaceownership via a GitHub / DNS / HTTP challenge, reusing the MCP Registry’s scheme rather than reinventing it. Today any metadata can claim any namespace. - sigstore / cosign keyless signing as the production
signedbackend (today: local ed25519), tied to npm provenance / SLSA attestation (releases already publish with provenance on). - garak / AI-Infra-Guard scanner integration for the
scannedbadge (today: a heuristic scan).
Authoring experience
weft dev(watch mode): recompile on file change for a tight authoring loop.weft diff: a dry run that shows exactly which artifacts an install/update would add, rewrite, or remove in a harness scope before touching disk, reading the prior set fromweft.lock(the content-hash plan already exists internally). Removal/orphan detection is the partweft build --checkdeliberately leaves to this command.
Shipped: weft build --check compares --out against a fresh compile (via diffPlanned
over the build plan’s content hashes) and exits 1 on drift, writing nothing. It is the
build-output side of the same content-hash machinery: the CI guard that keeps a committed
weft build . --target claude --out . --bare marketplace from silently drifting from source.
- Interactive A/B for
weft eval --compare: pick the better transcript in the terminal, and auto-run the pairwise judge over the pair (today it prints both sides for a human to read). - Publish the JSON Schemas to a CDN for
$schema-driven editor autocomplete, then an LSP / VS Code extension that validatesweft.yamland previews compiled output live. - More scaffolds:
weft init --template <mcp-wrapper|skill-pack|...>.
More harnesses
The adapter-kit seam is designed for this; each is a new package, no core change.
Shipped: a generic skills-only adapter (genericSkillsAdapter) for the directory-convention
harnesses that load SKILL.md from a <root>/skills/ layout (the .agents/skills/ family).
Zed, Gemini CLI, Amp, and Aider are wired up as targets today; adding another is one row in
the CLI registry. These get skills only; the five deep harnesses keep full plugin support.
- Full (non-skills) adapters for more harnesses where it’s worth the depth.
- An adapter conformance suite (golden fixtures) so a community adapter self-certifies against
the
HarnessAdaptercontract. - Richer OpenCode driver via
@opencode-ai/sdk/opencode serveSSE (full pending->running-> completed tool states) and the ACP transport.
Resolver depth
- Transitive (multi-level)
dependsand semver ranges on dependencies (today: one level, exact SHA pins).
Shipped: weft install/build now resolve a remote target, not just depends. A
github:/git/npm:/owner/repo ref is fetched into ~/.weft/cache (git clone or
npm pack) and installed from there, and a trailing //subdir selects a plugin or
marketplace nested in the source (e.g. weft install github:owner/repo//marketplace).
Discovery & operations
- A public index UI plus a hosted registry backend, and
weft searchover the federated index. - A hosted CI eval tier (BYO-keys local stays the default) and an eval compatibility-matrix badge across harnesses.
weft doctor: report installed harnesses and versions, and any drift betweenweft.lockand on-disk state.- Opt-in auto-update with channels, and a real telemetry transport (the data model and the deterministic gate already exist).
Distribution
- Standalone single-file binaries (
bun build --compile) per platform, attached to GitHub releases, soweftinstalls without a Node toolchain.
Invariants held throughout
- Compiler, not platform; the library is the product.
- Single source of truth; generated files are never hand-edited.
- Standards as inputs, adapters as the seam; upstream churn touches one adapter.
- Federate, don’t wall off.
- Same primitives at every scale; scope + policy differ, mechanism does not.
- Honest coverage; untested harnesses are reported, never faked.