Maintenance Tiers — 33-Package Classification

Last refreshed: 2026-06-25. Direction 4 of the 2026-06 optimization brief. The strategy memo (2026-06-competitiveness.md) made the case in prose; this file makes it actionable.

Why this file exists

Single maintainer + 37 packages + an 80-kLOC flagship demo (bscode) is a load-bearing problem disguised as an inventory problem. Every package implies a release cadence, an issue queue, a CHANGELOG row, and a slot on the 1.0-freeze checklist (api-stability.md).

This page assigns every package to one of three tiers so adopters, co-maintainer candidates, and the bot that drafts the bi-weekly release note all read the same prioritization.

Tier definitions

• Core (★). The "embedded runtime" thesis lives here. Continuous investment, full SemVer protection at 1.0 (where applicable), graduated experimental surfaces by 2026-12-15. An issue tagged core is on the maintainer's first daily pass.
• Narrative (◆). Carries the story to the leaders we want to be embedded in. Tracks upstream contracts; experimental status is acceptable as long as the upstream is moving. Best contributor pipeline for the co-maintainer ask.
• Maintenance-mode (▽). Useful, working, but not receiving proactive feature work. Bug fixes only; security patches honored. Documented dominance by GenericOpenAICompatModel + recipe (for model presets) or by a core primitive (for the others). Candidates for archival in a future 1.x cycle once download data confirms low usage.

A package's tier is set by what carries the embedded-runtime thesis, not by code size or test count.

Classification table

Package	Tier	Rationale (one line)
@wasmagent/core	★	The runtime. Every other package depends on it.
@wasmagent/kernel-quickjs	★	Cross-platform JS WASM kernel — the entry the Cloudflare codemode "third-party executor" hook expects.
@wasmagent/kernel-pyodide	★	Edge-safe Python; no competitor kernel ships this. Fills the "Python execution" diff vs Cloudflare.
@wasmagent/kernel-wasmtime	★	True WASM (Javy → QuickJS-in-WASM); paired with kernel-quickjs covers the WASM tier.
@wasmagent/kernel-remote	★	Third (microVM) tier — E2B / Cloudflare Sandbox. Required to honor "three-tier" claim.
@wasmagent/mcp-server	★	The two-tool code-mode MCP shape. Direct competitor to Cloudflare Code Mode MCP.
@wasmagent/evals-runner	★	Public-leaderboard play (Direction 2). Statistics axis is the differentiator vs Mastra.
@wasmagent/cli	★	Single binary surface for WasmAgent run / devtools / evals / model. Discovery entry for newcomers.
@wasmagent/compliance	★	Runtime compliance + ComplianceEvalRecord emitter. Core of the verifiable training loop.
@wasmagent/mcp-firewall	★	Runtime MCP policy + consent ledger. Core of the governance / security narrative.
@wasmagent/capability-compiler	◆	Compile CapabilityManifest → runtime policy rules. Companion to mcp-firewall.
@wasmagent/aisdk	◆	Vercel AI SDK adapter. Primary contributor pipeline; in flight upstream (issue #16063).
@wasmagent/mastra-sandbox	◆	Mastra sandbox provider; in flight upstream (issue #17884).
@wasmagent/claude-agent-sdk	◆	Anthropic Claude Agent SDK adapter. Tracks v0/v1 SDK type evolution.
@wasmagent/openai-agents	◆	OpenAI Agents JS adapter. Tracks @openai/agents Tool<T> shape.
@wasmagent/devtools	◆	Cross-framework local Studio (D5 GenAI semconv ingest). Direction 5 promotes this further.
@wasmagent/otel-exporter	◆	OTLP/HTTP exporter. Companion to devtools' framework-agnostic story.
@wasmagent/agent-prompts	◆	Reusable system prompts. Small surface, but feeds adapter packages.
@wasmagent/react	◆	useAgentRun() for Next.js. Pairs with the AI SDK adapter in chat-UI scenarios.
@wasmagent/cloudflare-worker	◆	Reference Worker entry — proves edge story. Pairs with kernel-quickjs.
@wasmagent/model-anthropic	◆	Anthropic adapter — needed for prompt-cache discipline (a core differentiator).
@wasmagent/model-openai	◆	OpenAI adapter — also the parent class of GenericOpenAICompatModel recipes.
@wasmagent/model-local	◆	Embedded local LLM (node-llama-cpp + grammar). Unique vs every competitor; offline-only stories.
@wasmagent/tools-browser	◆	Browser automation. Pairs with bscode's verifier and is a natural code-mode tool surface.
@wasmagent/tools-rag	◆	Embedding + vector store tools. Common scenario; small surface.
@wasmagent/tools-web	◆	Web search adapters (Tavily / Brave / Perplexity). Common scenario; small surface.
@wasmagent/model-deepseek	▽	OpenAI-compatible endpoint — covered by GenericOpenAICompatModel + recipe.
@wasmagent/model-doubao	▽	OpenAI-compatible endpoint — covered by GenericOpenAICompatModel + recipe.
@wasmagent/model-minimax	▽	OpenAI-compatible endpoint — covered by GenericOpenAICompatModel + recipe.
@wasmagent/model-moonshot	▽	OpenAI-compatible endpoint (Kimi) — covered by GenericOpenAICompatModel + recipe.
@wasmagent/model-qwen	▽	OpenAI-compatible endpoint (DashScope) — covered by GenericOpenAICompatModel + recipe.
@wasmagent/model-zhipu	▽	OpenAI-compatible endpoint (GLM) — covered by GenericOpenAICompatModel + recipe.
@wasmagent/a2a	▽	A2A (Agent2Agent) adapter. Cross-framework protocol; low organic demand so far. Re-tier on signal.
@wasmagent/ag-ui	▽	AG-UI protocol adapter. Same shape as a2a — re-tier when the upstream protocol stabilizes.
@wasmagent/ui-cards	▽	Card-block parser. Useful, but not on the embedded-runtime path.
@wasmagent/ui-cards-react	▽	React renderer for ui-cards. Will move with ui-cards — re-tier if either gets external traction.
@wasmagent/eliza-rollout-plugin	▽	elizaOS community plugin — emit rollout-wire/v1 records from elizaOS runs. Re-tier on community demand.

Tier counts (2026-06-25): 10 ★ · 15 ◆ · 11 ▽.

What "maintenance-mode" means in practice

Each ▽ package carries a top-of-README.md banner with this exact shape (added in the same PR that publishes this file):

> **Maintenance-mode** (▽). This package is functional and security-patched, but is **not** receiving proactive feature work — it is dominated by `GenericOpenAICompatModel` + a recipe in [`docs/guides/openai-compat-recipes.md`](https://github.com/WasmAgent/wasmagent-js/blob/main/docs/guides/openai-compat-recipes.md). See [maintenance tiers](https://github.com/WasmAgent/wasmagent-js/blob/main/docs/strategy/maintenance-tiers.md) for the rationale. If you actively use this package and want it promoted to ◆ Narrative, open an issue tagged `tier:promote-request` with your use case.

For non-model ▽ packages (a2a, ag-ui, ui-cards, ui-cards-react), the banner replaces the "GenericOpenAICompatModel + recipe" sentence with the relevant "dominated by ..." reason from the table above.

The banner is the only behavior change — no package.json deprecation flag, no npm dist-tag move, no signal that would appear as a yank. Users who depend on these packages today should not see breakage; they should see honest signal about the maintenance posture.

How a package moves between tiers

• ▽ → ◆: open an issue tagged tier:promote-request. The case is: an upstream we want to be embedded in (per ROADMAP.md S1) actively recommends this package, OR organic npm downloads cross 1k/week sustained.
• ◆ → ▽: if a ◆ package's upstream contract has been stable for two minor versions and download data shows fewer than 100/week sustained for two consecutive quarters. Demotion lands with the same banner above.
• ★ → other: unanimous maintainer agreement and a strategy memo entry. ★ exists to mark "the embedded-runtime thesis"; demoting one is a strategy change, not a maintenance change.

Relation to the 1.0 freeze

The 2026-12-15 1.0 freeze (per api-stability.md) applies only to @wasmagent/core. Every other package follows independent SemVer. ◆ packages that track an upstream still in flux stay 0.x past 1.0; ▽ packages stay 0.x indefinitely until demand signals warrant promotion.

bscode funnel-cost reduction (2026-06-13)

The brief's Direction 4 also calls out bscode itself: "considere把 Web IDE 部分降为可选 (使漏斗维护成本与其转化价值匹配)." First-round progress 2026-06-13 (bscode commit linked from WasmAgent CHANGELOG.md):

• The new /recipes reverse-funnel route is architecturally separate from the IDE — it imports only useState from React; every IDE component (Editor / Terminal / FileTree / FileBrowser / WebTerminal / WebContainer hooks / JSZip) is excluded from the /recipes chunk.
• The 535-line FrameworkApiMap modal on the IDE home (/) is now next/dynamic lazy-loaded with {ssr: false} and only mounts on open. Removes the modal markup from the / first-paint chunk.
• JobsPanel (577 LOC) lives on its own /jobs route already, so it's a separate chunk by virtue of Next's app-dir routing.

Funnel-cost takeaway: a visitor arriving via ?source=bscode-<framework>-recipe UTM (the reverse-funnel audience) downloads roughly the /recipes/page.tsx chunk (~6 KB of React + the inline recipe data) — orders of magnitude smaller than the IDE shell. The IDE itself remains useful as a CodeAgent demo for visitors who arrive at /; the brief's "降为可选" is now delivered by routing, not by ripping the IDE out.

Further reductions tracked under governance:bscode-shrink issues (target: total LOC ≤ WasmAgent framework LOC by 2026-Q4).

Falsifiability

If after two quarters no ▽ package has either (a) been promoted or (b) showed sustained > 100 downloads/week, the right answer is to move to "deprecated → archive in 2027-Q1" on those packages rather than carry them forward. The signal is npm download data

• issue activity; the decision lands as a separate PR with the data attached.