@design-ai/cli 4.64.0 → 4.65.0

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "design-ai",
-  "version": "4.64.0",
+  "version": "4.65.0",
   "description": "Senior product designer for any AI coding agent. 20 skills, 17 commands, 4 review agents covering UI/UX, website improvement, design systems, motion, illustration, print, video, game UI, conversational, and spatial design. Korean market depth.",
   "author": {
     "name": "Sungjin Park",

package/CHANGELOG.md

@@ -2,6 +2,32 @@
 
 User-facing release notes for design-ai. Versions follow semver.
 
+## v4.65.0 — MCP ↔ SDK parity: recall and opt-in learning-write tools (2026-07)
+
+Closes the parity gap the 2026-07-07 planning window fact-checked: the MCP server — the primary agent surface — had no recall tool and no learning-write tools, while the SDK shipped both. An MCP-connected agent can now run the same recall → author → check → capture loop the Agent SDK walkthrough demonstrates. The MCP server grows from 10 to **14 tools**; all additions are additive and the write tools are strictly opt-in.
+
+### Added
+- **`design_ai_recall`** (read-only) — combined corpus + learning-profile recall for a query, ranked by the deterministic lexical scorer (wraps `learn --recall`), with `limit` (1-20) and learning-`category` filters.
+- **`design_ai_learn_remember`**, **`design_ai_learn_feedback`** (outcome enum `keep`/`improve`/`avoid`, mirroring the CLI's real `--outcome` flag), **`design_ai_learn_capture`** (check an artifact, then capture its non-pass results — the only compound read+write tool). All three mirror the SDK `learn.*` boundary: they write ONLY the local learning profile (`DESIGN_AI_LEARNING_FILE` or its default), only when explicitly called, and never the network. The shared write-boundary language in the tool descriptions is guarded by a test, and the server's instructions name the write set explicitly so agents keep preferring read-only tools by default.
+
+### Changed
+- MCP input validation gains enum support: `category` and `outcome` are validated against the CLI's own value sets before any CLI invocation.
+
+### Verified
+- All 8 audits passed.
+- `npm run release:check`.
+- `npm run release:metadata`.
+- `git diff --check`.
+- Real-server stdio E2E: initialize → tools/list (14) → live recall query → remember/capture against a temp learning profile with the file verified after each write.
+- Main-branch GitHub Actions (`Design-AI audit`, `Deploy doc site`) passed for the constituent commits.
+
+### Versions
+- `package.json` + `.claude-plugin/plugin.json`: 4.64.0 → 4.65.0.
+- `vscode-extension/package.json`: remains 0.4.1.
+
+### What this enables
+- MCP consumers (Claude Code, Codex, any MCP client) get full parity with the Agent SDK: standalone combined recall, and an explicit, clearly-bounded way to record preferences, feedback, and check-derived learnings into the local profile — completing the learn loop on the surface where design-ai is most consumed.
+
 ## v4.64.0 — Route coverage sweep: marketing-page route and 18-class routing (2026-07)
 
 Generalizes the two dogfood rounds' converged finding — route-table coverage is the recurring gap class — into one systematic batch. Eighteen common task classes were probed through `route()`; thirteen were low/misrouted/zero-match despite the knowledge corpus already covering nearly every one of them (a pricing-page brief scored 0 with `pricing-page-design.md` sitting in the corpus). All fixed in one pass; evidence tables in `docs/ROUTE-COVERAGE-SWEEP.md`. Additive and backward-compatible.

package/README.ko.md

@@ -12,7 +12,7 @@
 
 > **모델이 아니에요. 파인튜닝도 아니에요.** 디자인 전문 지식을 구조화한 코퍼스 + 에이전트가 바로 실행할 수 있는 지시문이에요. 범용 LLM을 이번 세션에서만큼은 시니어 디자이너로 바꿔주는 셈이에요.
 
-> **배포 상태, 2026-07-07 확인:** 로컬 `npm run release:check`는 통과했고, GitHub Pages 문서는 live 상태이며, GitHub Release `v4.63.0`과 npm `@design-ai/cli@4.63.0`(`latest`) publish가 provenance와 함께 확인됐어요. 이번 릴리스는 도그푸드 루프 2회전이에요 — 데이터 밀집 화면 클래스를 위한 `dashboard-design` 라우트(금액 표기·테이블 접근성·밀도 check 요건)와 정산 워크드 예제, 그리고 `flow-design` 키워드 심화(한글 스텝 플로우 어휘 + 상태 설계 복합어, illustration 라우트와 양방향 검증). live `npm run registry:smoke`가 published `@design-ai/cli@4.63.0`에 대해 깨끗이 통과했어요(retrieval 표면과 route enrichment 포함). Homebrew formula는 `v4.63.0`에 pinning되어 있고, VS Code Marketplace에는 `sungjin.design-ai-vscode@0.4.1`이 공개되어 있어요. 자세한 내용은 [`docs/external-status.md`](docs/external-status.md)를 확인하세요.
+> **배포 상태, 2026-07-07 확인:** 로컬 `npm run release:check`는 통과했고, GitHub Pages 문서는 live 상태이며, GitHub Release `v4.64.0`과 npm `@design-ai/cli@4.64.0`(`latest`) publish가 provenance와 함께 확인됐어요. 이번 릴리스는 라우트 커버리지 스윕이에요 — 신규 `marketing-page` 라우트(랜딩·가격 페이지·이메일 템플릿, CTA/위계/이메일 클라이언트 check 요건)와 SaaS 랜딩 워크드 예제, `dashboard-design` 차트·권한 확장, `flow-design` 9클래스 키워드 배치로 프로브한 18개 클래스 전부가 medium+ 신뢰도로 라우팅돼요. live `npm run registry:smoke`가 published `@design-ai/cli@4.64.0`에 대해 깨끗이 통과했어요(retrieval 표면과 route enrichment 포함). Homebrew formula는 `v4.64.0`에 pinning되어 있고, VS Code Marketplace에는 `sungjin.design-ai-vscode@0.4.1`이 공개되어 있어요. 자세한 내용은 [`docs/external-status.md`](docs/external-status.md)를 확인하세요.
 
 ## 한눈에 보는 커버리지
 
@@ -172,7 +172,7 @@ design-ai는 한국 시장을 1순위로 만들어졌고, 글로벌 시장 패
 
 ## 상태
 
-전체 단계 로그는 [`docs/ROADMAP.md`](docs/ROADMAP.md), 현재 완료 범위는 [`docs/PRODUCT-READINESS.md`](docs/PRODUCT-READINESS.md)에서 확인하세요. 현재 **v4.63.0**: 도그푸드 루프 2회전(`flow-design`+`dashboard-design` 라우트, trust & safety 코퍼스, 디자인 지식 전용 recall, `pack` 바이트 예산 보장, SDK 워크스루), ranked-search eval 체크포인트(`search --eval`), Agent SDK(`@design-ai/cli/sdk`) — TypeScript 선언 + opt-in `learn.*` 로컬 쓰기 네임스페이스, public npm publish, provenance-backed GitHub Actions release, public registry smoke, Website Console MCP readiness, workspace learning restore/eval coverage, handoff bundle verification, 90%+ component coverage가 완료됐어요.
+전체 단계 로그는 [`docs/ROADMAP.md`](docs/ROADMAP.md), 현재 완료 범위는 [`docs/PRODUCT-READINESS.md`](docs/PRODUCT-READINESS.md)에서 확인하세요. 현재 **v4.64.0**: 18개 태스크 클래스 전면 라우트 커버리지(`flow-design`·`dashboard-design`·`marketing-page` 라우트, trust & safety 코퍼스, 디자인 지식 전용 recall, `pack` 바이트 예산 보장, SDK 워크스루), ranked-search eval 체크포인트(`search --eval`), Agent SDK(`@design-ai/cli/sdk`) — TypeScript 선언 + opt-in `learn.*` 로컬 쓰기 네임스페이스, public npm publish, provenance-backed GitHub Actions release, public registry smoke, Website Console MCP readiness, workspace learning restore/eval coverage, handoff bundle verification, 90%+ component coverage가 완료됐어요.
 
 핵심 디자인 컨설팅 워크플로우는 로컬 릴리스 기준으로 준비되어 있어요. 웹사이트 개선 컨트롤 타워는 zero-dependency static Web App과 `website-improvement` route/skill/command로 제공되고, Site Profile, audit checklist, MCP readiness, refactor prompt, handoff evidence tracking, bundle export/verify/repair를 한 번에 다뤄요. 로컬 학습 선호도는 `design-ai learn`으로 관리해요 — profile bootstrap, feedback 캡처, 읽기 전용 signal registry, 반복 QA 신호에서 만드는 skill 제안, 그리고 backup/restore/curate/audit까지 전부 로컬에서만 동작하는 opt-in 기능이에요. AI 모델 학습이나 fine-tuning은 여전히 현재 배포 범위 밖이에요.
 

package/README.md

@@ -12,7 +12,7 @@ A model-agnostic design knowledge base + skill system. Drop it in front of any A
 
 > **Not a model. Not a fine-tune.** A structured corpus of design expertise + agent-ready instructions that turn a general-purpose LLM into an expert.
 
-> **Distribution status, checked 2026-07-07:** local `npm run release:check` passes, GitHub Pages docs are live, GitHub Release `v4.63.0` is published, `@design-ai/cli@4.63.0` is public on npm (`latest`) with provenance, and the live `npm run registry:smoke` passes cleanly against `@design-ai/cli@4.63.0`. The Homebrew formula is pinned to `v4.63.0`, and `sungjin.design-ai-vscode@0.4.1` is public on the VS Code Marketplace. See [`docs/external-status.md`](docs/external-status.md).
+> **Distribution status, checked 2026-07-07:** local `npm run release:check` passes, GitHub Pages docs are live, GitHub Release `v4.64.0` is published, `@design-ai/cli@4.64.0` is public on npm (`latest`) with provenance, and the live `npm run registry:smoke` passes cleanly against `@design-ai/cli@4.64.0`. The Homebrew formula is pinned to `v4.64.0`, and `sungjin.design-ai-vscode@0.4.1` is public on the VS Code Marketplace. See [`docs/external-status.md`](docs/external-status.md).
 
 ## Coverage at a glance
 
@@ -214,7 +214,7 @@ Refresh refs/ on demand: `./tools/extractors/run-all.sh`.
 
 ## Status
 
-See [`docs/ROADMAP.md`](docs/ROADMAP.md) for the full phase log and [`docs/PRODUCT-READINESS.md`](docs/PRODUCT-READINESS.md) for the current completion boundary. Currently at **v4.63.0**: two dogfood-loop turns (`flow-design` + `dashboard-design` routes, trust & safety corpus, design-knowledge-only recall, `pack` byte-budget guarantee, SDK walkthrough), the ranked-search eval checkpoint (`search --eval`), the Agent SDK (`@design-ai/cli/sdk`) with shipped TypeScript types and the opt-in `learn.*` local-write namespace, public npm publish, provenance-backed GitHub Actions release, public registry smoke, Website Console MCP readiness, workspace learning restore/eval coverage, handoff bundle verification, and 90%+ component coverage are complete.
+See [`docs/ROADMAP.md`](docs/ROADMAP.md) for the full phase log and [`docs/PRODUCT-READINESS.md`](docs/PRODUCT-READINESS.md) for the current completion boundary. Currently at **v4.64.0**: full route coverage across 18 probed task classes (`flow-design`, `dashboard-design`, `marketing-page` routes, trust & safety corpus, design-knowledge-only recall, `pack` byte-budget guarantee, SDK walkthrough), the ranked-search eval checkpoint (`search --eval`), the Agent SDK (`@design-ai/cli/sdk`) with shipped TypeScript types and the opt-in `learn.*` local-write namespace, public npm publish, provenance-backed GitHub Actions release, public registry smoke, Website Console MCP readiness, workspace learning restore/eval coverage, handoff bundle verification, and 90%+ component coverage are complete.
 
 Core design consulting workflows are locally release-ready. The website improvement control tower ships as a zero-dependency static Web App plus a `website-improvement` route/skill/command, covering Site Profiles, audit checklists, MCP readiness, refactor prompts, handoff evidence tracking, and bundle export/verify/repair. Local learning preferences are available through `design-ai learn` — profile bootstrap, feedback capture, a read-only signals registry, skill-proposal generation from repeated QA signals, and full backup/restore/curate/audit tooling, all local-only and opt-in. AI model training or fine-tuning remains outside the shipped scope.
 

package/cli/lib/mcp-server.mjs

@@ -6,8 +6,10 @@ import { readFileSync } from "node:fs";
 import path from "node:path";
 
 import { DESIGN_AI_HOME, PACKAGE_ROOT } from "./paths.mjs";
+import { LEARNING_CATEGORIES, LEARNING_FEEDBACK_OUTCOMES } from "./learn-args.mjs";
 
 const PROTOCOL_VERSION = "2025-11-25";
+const LEARNING_WRITE_BOUNDARY = "Writes ONLY the local learning profile (DESIGN_AI_LEARNING_FILE or its default), and only when explicitly called.";
 const MAX_TOOL_OUTPUT_BYTES = 220_000;
 const DESIGN_AI_BIN = path.join(PACKAGE_ROOT, "cli", "bin", "design-ai.mjs");
 const MCP_RESPONSE_METHODS = new Set([
@@ -23,6 +25,10 @@ function optionalString(description = "") {
   return { type: "string", description };
 }
 
+function optionalEnumString(values, description = "") {
+  return { type: "string", enum: values, description };
+}
+
 function optionalBoolean(description = "") {
   return { type: "boolean", description };
 }
@@ -133,6 +139,21 @@ export const MCP_TOOLS = [
       additionalProperties: false,
     },
   },
+  {
+    name: "design_ai_recall",
+    title: "Recall corpus and learning context",
+    description: "Recall brief-relevant shipped corpus knowledge plus local learning-profile entries for a query, ranked by the deterministic lexical scorer. Read-only.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        query: { type: "string", minLength: 1, description: "Recall query text." },
+        limit: optionalInteger({ description: "Maximum results per list (corpus and learning), 1-20.", minimum: 1, maximum: 20 }),
+        category: optionalEnumString(LEARNING_CATEGORIES, "Optional learning category filter, scopes only the learning list."),
+      },
+      required: ["query"],
+      additionalProperties: false,
+    },
+  },
   {
     name: "design_ai_check",
     title: "Check Markdown artifact quality",
@@ -182,6 +203,49 @@ export const MCP_TOOLS = [
       additionalProperties: false,
     },
   },
+  {
+    name: "design_ai_learn_remember",
+    title: "Remember a learning preference",
+    description: `Record a local learning-profile preference for prompt personalization. ${LEARNING_WRITE_BOUNDARY}`,
+    inputSchema: {
+      type: "object",
+      properties: {
+        text: { type: "string", minLength: 1, description: "Preference text to remember." },
+        category: optionalEnumString(LEARNING_CATEGORIES, "Optional learning category. Defaults to preference."),
+      },
+      required: ["text"],
+      additionalProperties: false,
+    },
+  },
+  {
+    name: "design_ai_learn_feedback",
+    title: "Record learning feedback",
+    description: `Record keep/improve/avoid feedback as a local learning-profile entry. ${LEARNING_WRITE_BOUNDARY}`,
+    inputSchema: {
+      type: "object",
+      properties: {
+        text: { type: "string", minLength: 1, description: "Feedback text." },
+        outcome: optionalEnumString(LEARNING_FEEDBACK_OUTCOMES, "Optional feedback outcome: keep, improve, or avoid. Defaults to improve."),
+        category: optionalEnumString(LEARNING_CATEGORIES, "Optional learning category. Defaults to workflow."),
+      },
+      required: ["text"],
+      additionalProperties: false,
+    },
+  },
+  {
+    name: "design_ai_learn_capture",
+    title: "Check artifact and capture learning",
+    description: `Check a Markdown artifact, then capture its non-pass results as local learning-profile entries. The only compound read+write tool. ${LEARNING_WRITE_BOUNDARY}`,
+    inputSchema: {
+      type: "object",
+      properties: {
+        content: { type: "string", minLength: 1, description: "Markdown artifact content to check via stdin, then capture." },
+        route: optionalString("Optional route id, such as component-spec."),
+      },
+      required: ["content"],
+      additionalProperties: false,
+    },
+  },
   {
     name: "design_ai_version",
     title: "Show design-ai version",
@@ -218,6 +282,9 @@ function assertMcpInputValue(name, value, schema) {
     if (schema.minLength && value.length < schema.minLength) {
       throw new Error(`${name} must be at least ${schema.minLength} characters`);
     }
+    if (schema.enum && !schema.enum.includes(value)) {
+      throw new Error(`${name} must be one of: ${schema.enum.join(", ")}`);
+    }
     return;
   }
 
@@ -328,6 +395,13 @@ export function buildCliInvocation(toolName, input = {}) {
     return { args, stdin };
   }
 
+  if (toolName === "design_ai_recall") {
+    args.push("learn", "--recall", assertString(input.query, "query"), "--json");
+    maybePush(args, "--limit", input.limit);
+    maybePush(args, "--category", input.category);
+    return { args, stdin };
+  }
+
   if (toolName === "design_ai_check") {
     args.push("check", "--stdin");
     stdin = assertString(input.artifact, "artifact");
@@ -356,6 +430,26 @@ export function buildCliInvocation(toolName, input = {}) {
     return { args, stdin };
   }
 
+  if (toolName === "design_ai_learn_remember") {
+    args.push("learn", "--remember", assertString(input.text, "text"), "--json");
+    maybePush(args, "--category", input.category);
+    return { args, stdin };
+  }
+
+  if (toolName === "design_ai_learn_feedback") {
+    args.push("learn", "--feedback", assertString(input.text, "text"), "--json");
+    maybePush(args, "--outcome", input.outcome);
+    maybePush(args, "--category", input.category);
+    return { args, stdin };
+  }
+
+  if (toolName === "design_ai_learn_capture") {
+    args.push("check", "--stdin", "--learn", "--yes", "--json");
+    stdin = assertString(input.content, "content");
+    maybePush(args, "--route", input.route);
+    return { args, stdin };
+  }
+
   if (toolName === "design_ai_version") {
     args.push("version", "--json");
     return { args, stdin };
@@ -547,7 +641,7 @@ export async function handleMcpRequest(message, { runCli = runDesignAiCli } = {}
         name: "design-ai",
         version: readPackageVersion(),
       },
-      instructions: "Use design-ai MCP tools for local, deterministic design expertise: route briefs, generate prompts/packs, search/show the design corpus, check Markdown artifacts, and validate Website Improvement MCP readiness. Prefer read-only tools unless the user explicitly asks to record local learning usage.",
+      instructions: "Use design-ai MCP tools for local, deterministic design expertise: route briefs, generate prompts/packs, search/show the design corpus, recall combined corpus+learning context, check Markdown artifacts, and validate Website Improvement MCP readiness. Prefer read-only tools unless the user explicitly asks to record local learning usage. The opt-in write tool set is design_ai_learn_remember, design_ai_learn_feedback, and design_ai_learn_capture — each writes only the local learning profile, and only when explicitly called.",
     });
   }
 

package/docs/NEXT-SURFACE-DECISION.md

@@ -1,7 +1,7 @@
 # Next surface decision
 
-> Status: proposed — awaiting maintainer sign-off
-> Date: 2026-07-03
+> Status: living decision record — original recommendation accepted and executed (v4.57–v4.61); re-evaluated 2026-07-06 (composed slice, executed v4.62–v4.64) and 2026-07-07 (MCP ↔ SDK parity, accepted). See the dated sections below for each decision's status.
+> Date: 2026-07-03 (original) / 2026-07-07 (latest re-evaluation)
 
 This record answers option 3 of the "Recommended next decision" in [`PRODUCT-READINESS.md`](PRODUCT-READINESS.md): if the goal is "best design tool" as a broader product, which surface do we deepen next — CLI, VS Code extension, web UI, Figma plugin, or agent SDK?
 
@@ -157,3 +157,28 @@ Scored with the same seven criteria, in the post-execution landscape: CLI deepen
 **Run the two leaders as one composed slice: a real SDK consumer example that drives a design task end-to-end (`route` → `pack` → `check` → `learn.captureFromCheck`), documented as the SDK walkthrough.** The example is simultaneously the SDK adoption artifact (D) and a dogfood pass (G): building it exercises the shipped surfaces against a genuine task, and any friction or corpus gaps it surfaces become the evidence-backed next backlog — the same loop that produced the v4.58 corpus additions.
 
 - **Status:** accepted — maintainer signed off 2026-07-06; the composed slice is the next work item
+
+## Re-evaluation (2026-07-07)
+
+The 2026-07-06 composed slice executed and compounded: two dogfood rounds plus a systematic 18-class route-coverage sweep shipped as v4.62.0–v4.64.0. Route coverage has converged (18/18 probed classes at medium+; 22 routes; example-qa 22/22), so further sweeps are diminishing returns. Scored against the same seven criteria in the new landscape:
+
+| Candidate | Leverage | Reach | Build cost | Maintenance | Distribution | Risk | Synergy | Total |
+|---|---|---|---|---|---|---|---|---|
+| MCP ↔ SDK surface parity | 5 | 5 | 5 | 4 | 5 | 4 | 5 | **33** |
+| Dogfood round 3 (browse/IA class) | 4 | 3 | 4 | 5 | 3 | 5 | 4 | 28 |
+| Corpus automation (spec extractor v2) | 3 | 3 | 3 | 4 | 3 | 4 | 2 | 22 |
+| VS Code deepening | 4 | 3 | 3 | 3 | 4 | 3 | 3 | 23 |
+| Web UI expansion | 4 | 3 | 3 | 4 | 3 | 4 | 3 | 24 |
+| Skill-proposal apply path | 3 | 3 | 3 | 3 | 4 | 2 | 4 | 22 |
+
+- **MCP ↔ SDK surface parity** — fact-checked gap: the MCP server exposes 10 tools with **no `design_ai_recall`** and **no learning-write tools**, while the SDK ships `recall` plus the `learn.*` namespace (remember/feedback/captureFromCheck). MCP is the primary agent surface, yet an MCP-connected agent today cannot recall combined corpus+learning context as a standalone verb, nor record what it learned — the exact loop the SDK walkthrough demonstrates. The fix is the proven playbook: thin wrappers over the same `cli/lib` functions, with the write tools opt-in and clearly marked (the server instructions already say "prefer read-only tools unless the user explicitly asks to record local learning"). Highest leverage (serves every MCP consumer), lowest cost (adapter-only), direct synergy with everything shipped this month.
+- Dogfood round 3 stays the standing loop — cheap to rerun, but its route-coverage yield has converged; next run should target curation quality (browse/IA) rather than routing.
+- VS Code / Web UI / Figma: unchanged triggers, still no demand signal — deferred.
+
+### Recommendation (2026-07-07)
+
+**Close the MCP ↔ SDK parity gap: add `design_ai_recall` (read-only) and an opt-in learning-write tool set (`design_ai_learn_remember`, `design_ai_learn_feedback`, `design_ai_learn_capture`) mirroring the SDK's `learn.*` boundary — writes only the local learning profile, clearly marked in tool descriptions, never enabled implicitly.** Same verification pattern as every surface addition: MCP protocol tests, smoke assertions, docs (MCP server guide + SDK/MCP parity note).
+
+- **Status:** accepted — maintainer signed off 2026-07-07; MCP ↔ SDK parity is the next work item
+
+_Done (2026-07-07): added all four tools to `cli/lib/mcp-server.mjs` (14 tools total, up from 10). `design_ai_recall` maps to `learn --recall <query> --json [--limit N] [--category kind]`; `design_ai_learn_remember` maps to `learn --remember <text> --json [--category kind]`; `design_ai_learn_feedback` maps to `learn --feedback <text> --json [--outcome keep|improve|avoid] [--category kind]` (the CLI does expose `--outcome`, verified in `cli/lib/learn-args.mjs`'s `normalizeFeedbackOutcome`/`LEARNING_FEEDBACK_OUTCOMES`, so the MCP schema mirrors it rather than omitting it); `design_ai_learn_capture` maps to `check --stdin --learn --yes --json [--route id]` (verified the full flag combination runs cleanly, matching the SDK's `learn.captureFromCheck` path). All three write tools carry the shared "Writes ONLY the local learning profile ... and only when explicitly called" marker in their descriptions, and the serverInfo `instructions` string now names the write set explicitly. Added `category`/`outcome` enum validation to the MCP schema layer (`assertMcpInputValue` gained an `enum` branch) so invalid values are rejected before CLI invocation, mirroring `LEARNING_CATEGORIES`/`LEARNING_FEEDBACK_OUTCOMES`. Added mapping tests for all four tools (all param combinations), a 14-tool tools/list assertion, a write-boundary-marker description test, and a tools/call happy-path test for `design_ai_recall` with an injected runner, in `cli/lib/mcp-server.test.mjs`. `tools/audit/smoke_assertions.py`'s MCP assertion checks tool presence (`design_ai_route`, `design_ai_search`), not an exact count or full name list, so no smoke-assertion change was required; verified by grepping `package-smoke.py`/`registry-smoke.py` for the same presence-only pattern. Documented the new tools and the write-boundary note in `docs/integrations/design-ai-mcp-server.md` (no `.ko` variant exists) and added a one-line SDK/MCP parity cross-reference to `docs/SDK.md`. Full verification: `npm test` (593/593 pass), `npm run audit` (8/8), `npm run release:metadata` (pass), `npm run package:smoke`, `npm run registry:smoke:self-test`, and `npm run ci:local` all green. Live stdio E2E: initialize → tools/list (14 tools confirmed) → `design_ai_recall` with a real Korean query ("접근성 버튼") → `design_ai_learn_remember` against a temp `DESIGN_AI_LEARNING_FILE` (verified the file gained the entry) → `design_ai_learn_capture` against the same temp profile (verified entry count went from 1 to 9, one remembered plus eight captured from the failing check)._

package/docs/ROADMAP.md

@@ -1,5 +1,41 @@
 # Roadmap
 
+## Phase 771 — MCP Parity Release (v4.65.0) ✓ ready
+
+Ships the Phase 770 MCP ↔ SDK parity work as an npm version: `design_ai_recall` plus the opt-in learning-write tool set (`design_ai_learn_remember`/`feedback`/`capture`), growing the MCP server from 10 to 14 tools with the SDK `learn.*` write boundary mirrored and test-guarded. Additive; write tools are strictly opt-in.
+
+### Verified
+- All 8 audits passed.
+- `npm run release:check` (unit tests incl. MCP mapping/boundary tests, strict audits, whitespace, package contents, release metadata, release self-tests, packed-tarball smoke).
+- `npm run release:metadata`.
+- `git diff --check`.
+- Real-server stdio E2E (14 tools; live recall; temp-profile writes verified).
+- Main-branch GitHub Actions (`Design-AI audit`, `Deploy doc site`) passed for the constituent commits.
+
+### Versions
+- `package.json` + `.claude-plugin/plugin.json`: 4.64.0 → 4.65.0.
+- `vscode-extension/package.json`: remains 0.4.1.
+
+### What this enables
+- The learn loop closes on the primary agent surface: MCP clients recall combined corpus+learning context and record learnings with the same explicit write boundary the SDK ships — CLI, SDK, and MCP are now at full surface parity.
+
+### What's still ahead
+- Next planning window per [NEXT-SURFACE-DECISION.md](NEXT-SURFACE-DECISION.md); dogfood loop remains cheap to rerun on curation-quality questions (browse/IA).
+
+## Phase 770 — MCP ↔ SDK parity (implemented, unreleased)
+
+Closes the parity gap the 2026-07-07 planning window fact-checked ([NEXT-SURFACE-DECISION.md](NEXT-SURFACE-DECISION.md)): the MCP server had no recall tool and no learning-write tools while the SDK shipped both, so an MCP-connected agent could not run the recall→learn loop the SDK walkthrough demonstrates.
+
+### Delivered
+- [x] `design_ai_recall` (read-only): combined corpus+learning recall over `learn --recall`, with limit and category-enum validation.
+- [x] Opt-in write set mirroring the SDK `learn.*` boundary: `design_ai_learn_remember`, `design_ai_learn_feedback` (outcome enum keep/improve/avoid — mirrored only after verifying the CLI's `--outcome` flag exists), `design_ai_learn_capture` (check + capture, the only compound read+write tool). All three carry a shared write-boundary description (writes ONLY the local learning profile, only when explicitly called), guarded by a test that fails if the safety language is dropped; serverInfo instructions name the write set explicitly.
+- [x] `assertMcpInputValue` gains enum validation (category/outcome checked against `LEARNING_CATEGORIES`/`LEARNING_FEEDBACK_OUTCOMES` before CLI invocation).
+- [x] Real-server stdio E2E: initialize → tools/list (10 → **14** tools) → live `design_ai_recall` query → `learn_remember` and `learn_capture` against a temp `DESIGN_AI_LEARNING_FILE` with the profile file verified after each write. 593/593 tests; smoke assertions are presence-based (no drift risk, verified in both smoke files).
+- [x] Docs: MCP server guide tool table + write-boundary note; SDK ↔ MCP parity cross-reference in `docs/SDK.md`.
+
+### Remaining
+- [ ] Release: bundle into the next npm version (external-status "10 tools" evidence line updates at release handoff).
+
 ## Phase 769 — Route Coverage Sweep Release (v4.64.0) ✓ ready
 
 Ships the Phase 768 route-coverage sweep as an npm version: the `marketing-page` route with its worked example, the `dashboard-design` chart/permissions enrichment, and the `flow-design` nine-class keyword batch — 18/18 probed task classes now land on target routes at medium+ confidence. Additive and backward-compatible.

package/docs/SDK.md

@@ -6,6 +6,8 @@
 
 See [`AGENT-SDK.md`](AGENT-SDK.md) for the full design rationale, phased plan, and open questions. This page is the public reference for the shipped Phase A (read-only) and Phase B (local-write) surface.
 
+MCP parity: the SDK's `recall` and `learn.*` (`remember`, `feedback`, `captureFromCheck`) map 1:1 to the MCP tools `design_ai_recall` and `design_ai_learn_*` (`design_ai_learn_remember`, `design_ai_learn_feedback`, `design_ai_learn_capture`) — see [`integrations/design-ai-mcp-server.md`](integrations/design-ai-mcp-server.md).
+
 ## Install and import
 
 The SDK ships inside the `@design-ai/cli` package (no separate install):

package/docs/external-status.md

@@ -5,25 +5,25 @@
 
 ## Summary
 
-npm is publicly published at `@design-ai/cli@4.63.0` (npm dist-tag `latest` = `4.63.0`); publish run from tag `v4.63.0` succeeded with provenance and all pre-publish gates green. This release ships the dogfood loop's second turn: the `dashboard-design` route (data-dense screen class with money-format/table-a11y/density check requirements) with its settlement worked example, and `flow-design` keyword depth (Korean step-flow vocabulary + states-design compounds, verified in both directions against the illustration route). The live `npm run registry:smoke` passes cleanly against published `@design-ai/cli@4.63.0` ("Registry smoke passed"), covering the retrieval surfaces (index/ranked/embeddings/recall) and route enrichment. GitHub Release `v4.63.0` is published, and the Homebrew tap formula points at the `v4.63.0` release source tarball with a verified SHA-256. The VS Code extension `sungjin.design-ai-vscode` remains published at `0.4.1`. GitHub Pages docs are publicly reachable.
+npm is publicly published at `@design-ai/cli@4.64.0` (npm dist-tag `latest` = `4.64.0`); publish run from tag `v4.64.0` succeeded with provenance and all pre-publish gates green. This release ships the route coverage sweep: the new `marketing-page` route (landing/pricing/email-template classes with conversion/hierarchy/email-client check requirements) with its SaaS-landing worked example, plus `dashboard-design` chart/permissions enrichment and the `flow-design` nine-class keyword batch — 18/18 probed task classes now route at medium+ confidence. The live `npm run registry:smoke` passes cleanly against published `@design-ai/cli@4.64.0` ("Registry smoke passed"), covering the retrieval surfaces (index/ranked/embeddings/recall) and route enrichment. GitHub Release `v4.64.0` is published, and the Homebrew tap formula points at the `v4.64.0` release source tarball with a verified SHA-256. The VS Code extension `sungjin.design-ai-vscode` remains published at `0.4.1`. GitHub Pages docs are publicly reachable.
 
 ## Results
 
 | Surface | Checked target | Result | Evidence |
 |---|---|---|---|
-| npm registry | `@design-ai/cli` | Published latest is `4.63.0` (tag `v4.63.0`, provenance). Pre-publish packed-tarball smoke (incl. SDK import + `learn.remember` write) and live `npm run registry:smoke` both pass for `@design-ai/cli@4.63.0`. | `npm view @design-ai/cli version` → `4.63.0`; live registry smoke "Registry smoke passed" |
+| npm registry | `@design-ai/cli` | Published latest is `4.64.0` (tag `v4.64.0`, provenance). Pre-publish packed-tarball smoke (incl. SDK import + `learn.remember` write) and live `npm run registry:smoke` both pass for `@design-ai/cli@4.64.0`. | `npm view @design-ai/cli version` → `4.64.0`; live registry smoke "Registry smoke passed" |
 | GitHub Pages | `https://sungjin9288.github.io/design-ai/` | Published and reachable: HTTP `200`, design-ai MkDocs page rendered | `evidence/cli-logs/github-pages-status.log` |
-| Homebrew tap | `Formula/design-ai.rb` | Formula pinned to `v4.63.0` release source tarball with SHA-256 `024a720cfe1a18a1b77751232818f01d9a2826fac4772865b889bd3d91ce1ae8` (recomputed from the published tag tarball) | `Formula/design-ai.rb` |
+| Homebrew tap | `Formula/design-ai.rb` | Formula pinned to `v4.64.0` release source tarball with SHA-256 `96396304643902a55a8bae605e6efcf169c6131892236f7fb58a0a0ced152bf4` (recomputed from the published tag tarball) | `Formula/design-ai.rb` |
 | VS Code Marketplace | `sungjin.design-ai-vscode` | Published: run `28431571256` published `v0.4.1`, and the Marketplace Gallery API returned visible version `0.4.1` on 2026-07-02. | `evidence/cli-logs/vscode-marketplace-status.log`, `evidence/cli-logs/vscode-marketplace-secret-status.log`, `evidence/cli-logs/vscode-extension-vsce-package.log`, `evidence/cli-logs/vscode-publish-workflow-status.log` |
-| GitHub Release | `v4.63.0` | Published for tag `v4.63.0` at commit `d6ce680` | `gh release view v4.63.0` |
-| MCP server | `@design-ai/cli@4.63.0` / local clone | Public npm `design-ai-mcp` responds to initialize and tools/list with 10 tools; local Codex and Claude Code both report `design-ai` MCP as configured and connected. | `evidence/cli-logs/npm-registry-smoke.log`, `evidence/cli-logs/design-ai-mcp-client-status.log` |
+| GitHub Release | `v4.64.0` | Published for tag `v4.64.0` at commit `d0c5283` | `gh release view v4.64.0` |
+| MCP server | `@design-ai/cli@4.64.0` / local clone | Public npm `design-ai-mcp` responds to initialize and tools/list with 10 tools; local Codex and Claude Code both report `design-ai` MCP as configured and connected. | `evidence/cli-logs/npm-registry-smoke.log`, `evidence/cli-logs/design-ai-mcp-client-status.log` |
 
 ## Interpretation
 
 - The repository is locally release-ready based on `npm run release:check`.
 - GitHub Pages docs can now be described as publicly deployed.
-- Public npm install can now be described as published and smoke-tested at `4.63.0`.
-- Homebrew tap install can now be described as pinned to the `v4.63.0` release tarball; full tap audit/install/test remains a maintainer-side verification step because this Homebrew version rejects path-based `brew audit Formula/...` calls.
+- Public npm install can now be described as published and smoke-tested at `4.64.0`.
+- Homebrew tap install can now be described as pinned to the `v4.64.0` release tarball; full tap audit/install/test remains a maintainer-side verification step because this Homebrew version rejects path-based `brew audit Formula/...` calls.
 - VS Code Marketplace can now be described as published and publicly reachable at version `0.4.1`.
 - Claude Code and Codex can now be described as locally connected to the clone-backed `design-ai` MCP server, while public npm users can verify the published `design-ai-mcp` entrypoint from a clean working directory.
 
@@ -32,12 +32,12 @@ npm is publicly published at `@design-ai/cli@4.63.0` (npm dist-tag `latest` = `4
 ```bash
 npm view @design-ai/cli version name time.modified dist-tags --json
 curl -sS -L -o /tmp/design-ai-pages.html -w 'http_code=%{http_code}\nurl_effective=%{url_effective}\n' https://sungjin9288.github.io/design-ai/
-curl -sL https://github.com/sungjin9288/design-ai/archive/refs/tags/v4.63.0.tar.gz | shasum -a 256
+curl -sL https://github.com/sungjin9288/design-ai/archive/refs/tags/v4.64.0.tar.gz | shasum -a 256
 ruby -c Formula/design-ai.rb
 brew style Formula/design-ai.rb
-gh release view v4.63.0 --repo sungjin9288/design-ai --json tagName,isDraft,isPrerelease,publishedAt,name,url,assets
+gh release view v4.64.0 --repo sungjin9288/design-ai --json tagName,isDraft,isPrerelease,publishedAt,name,url,assets
 gh run view 28569283984 --repo sungjin9288/design-ai --json status,conclusion,name,url,createdAt,updatedAt
-npm exec --yes --package=@design-ai/cli@4.63.0 -- design-ai-mcp
+npm exec --yes --package=@design-ai/cli@4.64.0 -- design-ai-mcp
 codex mcp get design-ai
 claude mcp list
 curl -sS -H 'Content-Type: application/json' \

package/docs/integrations/design-ai-mcp-server.md

@@ -14,11 +14,19 @@ The server is local, stdio-based, and deterministic. It wraps existing `design-a
 | `design_ai_search` | Search `knowledge/`, `examples/`, `skills/`, `docs/`, `agents/`, and `commands/`. Optional `ranked` opts into deterministic BM25-style results (`design-ai search --ranked`); requires no index and never builds one. | Read-only |
 | `design_ai_show` | Read a corpus file or line range. | Read-only |
 | `design_ai_examples` | Find worked examples by query or route. | Read-only |
+| `design_ai_recall` | Recall brief-relevant shipped corpus knowledge plus local learning-profile entries for a query, ranked by the deterministic lexical scorer (`design-ai learn --recall`). | Read-only |
 | `design_ai_check` | Check generated Markdown artifacts for grounding, accessibility, responsive notes, and unresolved markers. | Read-only |
 | `design_ai_site_mcp_check` | Validate Website Improvement MCP readiness from workspace JSON. | Read-only |
 | `design_ai_site_mcp_plan` | Generate a Website Improvement MCP action plan. | Read-only |
+| `design_ai_learn_remember` | Record a local learning-profile preference for prompt personalization. | Writes only the local learning profile, opt-in |
+| `design_ai_learn_feedback` | Record keep/improve/avoid feedback as a local learning-profile entry. | Writes only the local learning profile, opt-in |
+| `design_ai_learn_capture` | Check a Markdown artifact, then capture its non-pass results as local learning-profile entries. The only compound read+write tool. | Writes only the local learning profile, opt-in |
 | `design_ai_version` | Return CLI and corpus version metadata. | Read-only |
 
+### Opt-in write tools
+
+`design_ai_learn_remember`, `design_ai_learn_feedback`, and `design_ai_learn_capture` are the MCP mirror of the SDK's `learn.*` namespace (see [`../SDK.md`](../SDK.md)). Each writes ONLY the local learning profile (`DESIGN_AI_LEARNING_FILE` or its default path), and only when explicitly called by name — no other tool call implicitly triggers a write. `design_ai_learn_capture` maps to `design-ai check --stdin --learn --yes --json`, the same path as the CLI's `--learn --yes` capture flow.
+
 ## Start the server manually
 
 From a local clone:
@@ -147,6 +155,7 @@ Then generate a design_ai_site_mcp_plan and summarize blocking MCP gaps.
 - The default tools do not call external MCP servers.
 - Website Improvement MCP readiness tools inspect local workspace JSON only.
 - `design_ai_prompt` and `design_ai_pack` are read-only unless `withLearning` is set.
+- `design_ai_learn_remember`, `design_ai_learn_feedback`, and `design_ai_learn_capture` write only the local learning profile, never the network, and only when called explicitly.
 - The server does not mutate target repositories.
 - Do not pass secrets in briefs, workspace JSON, or artifacts.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@design-ai/cli",
-  "version": "4.64.0",
+  "version": "4.65.0",
   "description": "Senior product designer for any AI coding agent. Installs design-ai (20 skills, 17 commands, 4 agents) into Claude Code globally. Korean market depth plus website improvement control tower.",
   "bin": {
     "design-ai": "cli/bin/design-ai.mjs",