Product readiness

Current status: v5.1.0 is published. npm latest, GitHub Release, and the Homebrew formula identify the same additive P6-P13 release. The public contract has 29 MCP tools and 20 SDK exports. Model training is not part of the shipped product.

Current source status: P14 public distribution, the internal documentation homepage pilot, deployed Pages QA, and the three-slot external pilot launch framework are complete. P15 recruitment is live through the default-branch Issue Form and public issue #56. P16 marketing-site pilot operations, anonymous status tracking, separate consent, fixed problem classification, and distinct-owner decision rules are prepared; direct outreach has not started, and real external participation remains consent-gated and unverified. P17A now validates the metadata, activation, progressive disclosure, local source authority, completion criteria, line budgets, and inventory parity of all 21 skills, while removing guessed MCP operations and dated operational claims from durable playbooks. This is current-source internal hardening, not a new public capability or release. CLI design-ai start, SDK start(), MCP design_ai_start, and Website Console share one canonical read-only start payload. CLI design-ai inspect, SDK inspectHtml(), and MCP design_ai_inspect_html share the canonical quality report. The inspector confirms only supported static HTML evidence and leaves runtime behavior unverified. CLI design-ai review, SDK reviewHtml(), MCP design_ai_review_html, and Website Console now compose those two contracts into one linked review session with exact source identity, ordered stages, and a pending human decision. CLI design-ai review-handoff, SDK reviewHandoff(), MCP design_ai_review_handoff, and Website Console preserve that session as a self-validating but undelivered transfer. Optional browser evidence must match the exact quality-report source and declared viewport set; receiving-agent validation and implementation remain pending. CLI design-ai review-handoff-verify, SDK verifyReviewHandoff(), MCP design_ai_verify_review_handoff, and Website Console now let the named consumer revalidate the exact handoff and emit a separate receipt. The receipt carries forward evidence and approvals while keeping identity, transport, acceptance, target-repository intake, and implementation unverified. CLI design-ai review-intake, MCP design_ai_review_intake, and Website Console ground the handoff in one target checkout. They inspect only supported root project metadata and local Git state, report path and remote drift plus pre-existing changes, and stop before application-source review or implementation. CLI review-scope / review-scope-approve, SDK proposeImplementationScope() / approveImplementationScope(), MCP design_ai_review_scope / design_ai_approve_review_scope, and Website Console then preserve one immutable proposal and a separate approval. Approval covers listed selectors only; release and external-write gates remain pending. CLI review-evidence, MCP design_ai_review_evidence, and Website Console now compare one exact approval with the current local Git state and only the evidence files declared by the implementation record. They do not run reported commands, read application source, write files, or authorize commit, push, deployment, or external writes. Missing verification and runtime observations remain visible as warnings instead of being converted into a successful claim. CLI review-pilot, SDK recordPilotEvidence(), MCP design_ai_review_pilot, and Website Console then join the exact P11 evidence, original P6 workflow, and one consented pilot record. The result derives bounded time, finding, approval, implementation, and risk metrics while keeping real, synthetic, inferred, and unverified claims separate. This contract can prove one internal dogfood run; it cannot independently establish identity, feedback, external adoption, production quality, or business outcomes. CLI review-compare, SDK compareReviews(), MCP design_ai_compare_reviews, and Website Console then compare two exact canonical quality reports for the same subject and context. They preserve source references, bytes, SHA-256 identities, all eight lens transitions, and resolved, persistent, introduced, or uncertain finding decisions. A missing finding is not promoted to resolved unless its candidate lens passes. The comparison remains read-only and keeps target mutation, commit, push, deployment, external writes, production quality, and adoption outside its claim boundary. The separate CLI-only design-ai verify-browser runner can collect that runtime evidence from a user-supplied adapter after explicit approval, without adding a browser dependency or changing the P2 report. Five Korean product review packs and Website Console evidence review complete P4. The CLI-only design-ai benchmark runner completes P5 with four packaged synthetic cases for new design, existing-product refactor, Korean product UX, and multi-agent handoff. It validates exact contracts and finding-id sets, records SHA-256 evidence, and publishes case studies without an aggregate quality score or a real customer adoption claim.

The P16 readiness record links the merged operating guide, lifecycle labels, main CI, Pages deployment, desktop/mobile checks, and the unchanged zero-participation claim boundary.

This document separates shipped product scope from future product ideas so the roadmap does not imply that every possible AI feature is already complete.

TL;DR

Area	Status	Evidence	Remaining work
Design consulting skills	Complete for v5.0.0	21 skills across design systems, interface craft, website improvement, UX audit, critique, handoff, motion, illustration, print, video, game UI, conversational UI, and spatial design	Keep knowledge fresh through normal stability review
Skill contract reliability	Implemented in current source; unreleased	Zero-dependency metadata and playbook validator, isolated negative fixtures, 21/21 skill check, manifest parity, and release-self-test integration; npm run release:check passes with 832 tests, 8 strict audits, 794 packaged files, a 0/0 documentation warning policy, SDK smoke, installed-bin smoke, and one-shot package smoke; no CLI, SDK, MCP, package version, publication, or P16 state change	Modularize the oversized smoke harness under P17B while preserving entry points and exact contract behavior
Smoke harness maintainability	P17B.1 implemented in current source; unreleased	One Website Console contract owner for 42 values, one validator module for eight shared validators, and one isolated snapshot/negative-fixture runner; exact baseline parity; stable shared/local/registry entry points; all new files under 400 lines and largest extracted function 85 lines	Move remaining Website Console scenario fixtures, assertion groups, and dual execution runners, then extract learning, review, and lifecycle domains without reducing coverage
Design agent workflows	Complete for v5.0.0	16 public slash commands, 4 review agents, route/prompt/artifact/pack/check/search/show/examples CLI workflows; shared implementation-plan, critique-loop, and DESIGN.md contracts across CLI, SDK, MCP, and Website Console	Collect adoption feedback and keep release evidence current
One start flow	Published in v5.1.0	Shared start operation, CLI/SDK/MCP exact JSON parity, Website Console import/review/export, closed input contracts, and installed-bin plus one-shot smoke coverage	Collect real workflow evidence without weakening the read-only boundary
Read-only quality engine	Published in v5.1.0	Shared static HTML operation across CLI/SDK/MCP, canonical eight-lens report validation, benchmark reproduction, immutable-source assertions, and npm run release:check evidence with 694 tests, 8 strict audits, 676 packaged files, warning-free docs, and installed-bin plus one-shot smoke coverage	Keep static confirmation separate from optional browser evidence
Browser verification runner	Published in v5.1.0	Separate versioned sidecar contract, explicit --yes plus approval reference, loopback-only preview boundary, post-run source-report digest guard, normalized responsive/keyboard/accessibility/reduced-motion/loading/error/repeated-action probes, artifact-linked findings, exact network-policy attestation, restored intermediate source and adapter write uncertainty, complete artifact checks, run-interval timestamps, POSIX process-group termination, and npm run release:check evidence with 706 tests, 8 strict audits, 682 packaged files, warning-free docs, and functional installed-bin plus one-shot smoke coverage	Add concrete adapters only through separate dependency and security review
Product specialization benchmarks	Published in v5.1.0	Read-only design-ai benchmark, four versioned synthetic journey fixtures, strict start/artifact/quality validators, SHA-256 source and contract evidence, exact finding-id false-negative/false-positive/uncertainty regression lists, transfer envelopes with consumer validation, four case studies, and npm run release:check evidence with 736 tests, 8 strict audits, 715 packaged files, a 0/0 documentation warning policy, and installed-bin plus one-shot package smoke	Collect separately consented real-user feedback before making any adoption or outcome claim
Canonical review workflow	Published in v5.1.0	Shared design-ai-review-workflow v1 across CLI review, SDK reviewHtml(), MCP design_ai_review_html, and Website Console; exact source bytes, nested start and quality contracts, SHA-256 linkage, ordered review stages, immutable Console import/export, package smoke drift rejection, and npm run release:check evidence with 749 tests, 8 strict audits, 722 packaged files, a 0/0 documentation warning policy, SDK import smoke, and installed-bin plus one-shot review smoke	Keep browser verification and implementation as separate approval-gated stages; design a general evidence bundle only after this chain is stable
Review evidence handoff	Published in v5.1.0	Shared design-ai-review-handoff v1 across CLI review-handoff, SDK reviewHandoff(), MCP design_ai_review_handoff, and Website Console; exact source-byte envelopes, workflow and optional browser-evidence linkage, immutable Console export, named but undelivered recipient state, pending consumer validation, and npm run release:check evidence with 760 tests, 8 strict audits, 728 packaged files, SDK import smoke, installed-bin plus one-shot handoff smoke, and desktop/mobile Console verification	Collect real multi-agent usage evidence without adding transport or implementation claims to this read-only contract
Consumer validation receipt	Published in v5.1.0	Shared design-ai-review-handoff-receipt v1 across CLI review-handoff-verify, SDK verifyReviewHandoff(), MCP design_ai_verify_review_handoff, and Website Console; exact nested handoff bytes, matching self-declared consumer, evidence summary, unchanged approvals, original-byte export, npm run release:check evidence with 767 tests, 8 strict audits, 734 packaged files, SDK import smoke, installed-bin plus one-shot receipt smoke, desktop/mobile Console verification, a fresh Codex MCP call, and PR #42 required CI run 29408587642	Collect real multi-agent adoption evidence without expanding the receipt into identity, transport, acceptance, or implementation proof. Fresh Claude invocation is separately blocked by organization-level Claude Code subscription policy, so no Claude receipt claim is made
Target repository intake	Published in v5.1.0	Shared design-ai-target-repo-intake v1 across CLI review-intake, MCP design_ai_review_intake, and Website Console; exact receipt linkage, declared-path and consumer guards, absolute-path MCP file intake, root metadata inspection, unfiltered local Git evidence, symbolic-link refusal, remote-drift blocking, npm run release:check evidence with 778 tests, 8 strict audits, 739 packaged files, installed-bin plus one-shot smoke, desktop/mobile Console proof, a fresh Codex MCP call, and PR #43 required CI run 29413686416	Keep implementation scope as a separate approval and collect a real P10 proposal before any source read or target mutation
Implementation scope approval	Published in v5.1.0	Exact-source request, proposal, and approval v1 contracts across CLI, typed SDK, in-process MCP, and Website Console; derived baseline, ownership, selector, risk, verification, and gate validation; npm run release:check with 795 tests, 8 strict audits, 749 packaged files, installed-bin plus one-shot smoke, fresh MCP proposal and approval, desktop/mobile Console proof with exact original-byte export and approval -> proposal -> intake restore, and PR #44 required CI run 29418999657	Run P11 only against an unchanged approved baseline; keep commit, push, deployment, migration execution, and external writes separately gated
Implementation evidence	Published in v5.1.0	Exact-source request and result v1 contracts across CLI, in-process MCP, and Website Console; unchanged-baseline and selector enforcement; complete Git change accounting; regular non-symlink evidence hashing; failed, missing, and uncertain proof preservation; npm run release:check with 805 tests, 8 strict audits, 757 packaged files, a 0/0 documentation warning policy, and installed-bin plus one-shot smoke, a fresh MCP call, and desktop/mobile Console proof with exact original-byte export and evidence -> approval -> proposal -> intake restore, and PR #45 required CI run 29424169441	Resolve or explicitly accept evidence warnings before release approval; keep commit, push, deployment, migration execution, network calls, and external writes separately gated
Real pilot evidence	Merged in PR #46; internal dogfood complete	Exact P11, P6, and pilot-record source envelopes across CLI, typed SDK, in-process MCP, and Website Console; a consented Website Console dogfood recorded a 14-second first useful artifact, one accepted finding, complete implementation evidence, zero unresolved implementation risks, desktop/mobile browser proof, and an exact 416,114-byte Console import/export round trip; a fresh stdio MCP process kept full-output failure explicit at 452,923 bytes and returned a 4,021-byte compact summary with matching source hashes and measures	Keep external adoption, feedback authenticity, production quality, and business outcomes unverified until separate evidence exists
Verified design iteration	Published in v5.1.0	Shared design-ai-review-comparison v1 across CLI review-compare, SDK compareReviews(), MCP design_ai_compare_reviews, and Website Console; exact baseline/candidate source envelopes, matching subject and context guard, eight lens transitions, four finding decisions, compact identity-preserving output, desktop/mobile Console proof with exact export SHA-256, 44px navigation targets, keyboard skip-link success, no horizontal overflow, and no console warnings or errors; npm run release:check passes with 832 tests, 8 strict audits, 774 packaged files, warning-free docs, SDK import smoke, and installed-bin plus one-shot comparison smoke; a fresh stdio MCP process lists all 29 tools and returns a 3,934-byte compact comparison with matching source hashes; PR #47 and the resulting main-branch audit, real VS Code e2e, docs build, and Pages deployment passed	Collect real before-and-after product evidence before making external quality or outcome claims
Website improvement control tower	Complete for v5.0.0	Zero-dependency static Web App under docs/website-console/, static Workflow Graph rendering with graph JSON copy/export, browser-local handoff evidence tracking for executed work / verification results / remaining risks / next actions, read-only linked-code root metadata inspection through --linked-preview and design_ai_site_linked_preview, Console import/render/export of process-not-started, URL-not-probed, evidence-not-recorded preview readiness, CLI/bundle evidence export through design-ai site --json, --tasks, --report, and --bundle, verified bundle evidence metadata, MCP probe evidence, bundle-check JSON/human and bundle-handoff JSON/prompt boundary metadata for deterministic-local, no-external-call, and no-target-repo-mutation handoff validation, MCP probe count telemetry and package/shared smoke self-test coverage for Website Console MCP probe counts through --next-actions, --bundle-check, --bundle-compare, and --bundle-handoff JSON, bundled Website Console mcp-probes.json saved probe evidence payload instead of the full site --mcp-check --probes --json response, generated bundle contract verification with per-file diagnostics through bundle-check/compare/handoff metadata, repair guidance through bundle-check and bundle-handoff metadata, repair preview/apply through --bundle-repair with repair report --out file output-file persistence, packed-tarball evidence preservation smoke for non-empty implementationEvidence, linked preview inspection in installed-bin and one-shot npm exec paths, generated contract counts, generated drift diagnostics, repair guidance, and repair preview/apply through verified bundle paths, website-improvement route/skill/command, design-ai site sample workspace generation, prompt template listing, MCP readiness check through --mcp-check, read-only MCP readiness probes through --mcp-check --probes, Markdown MCP action plan export through --mcp-plan and --mcp-plan --probes, prioritized local operator checklists with read-only probe readiness and probe counts through --next-actions, portable workflow graph export through --graph --json, complete handoff bundle export with mcp-probes.json and downstream MCP probe count JSON through --bundle --out, handoff bundle fingerprint verification through --bundle-check --strict --json, handoff bundle comparison through --bundle-compare --strict --json, target-repo handoff prompt generation through --bundle-handoff --strict --json, local handoff bundle repair through --bundle-repair --yes --json, refactor task generation, single prompt template export with task selection, plus JSON validation/report/prompt generation, Site Profile, audit checklist, MCP readiness matrix, refactor plan generator, prompt generator, and handoff report export	Future phases can add real MCP connection probes, approved Playwright/Lighthouse/axe automation, and VS Code Webview reuse
Local release confidence	Complete for v5.0.0	npm run ci:local now passes for v5.0.0, including npm run release:check, Python syntax, the knowledge-size budget, VS Code extension compile/unit tests, MkDocs build, and the 0/0 warning policy. npm run release:check passes after the Website Console bundle mcp-probes.json saved-payload guard phases, after the Product Readiness and release-facing policy docs bundle boundary metadata guards for bundle-check JSON/human and bundle-handoff JSON/prompt boundary metadata plus full release:self-test evidence recording, after the release-facing policy docs guard for Website Console bundle boundary metadata full release:check evidence, and after the release-facing policy docs Product Readiness release policy full gate evidence guard, covering unit tests, strict audits, whitespace checks, package contents, release metadata, release self-tests, the full npm run release:self-test chain across shared smoke assertions, package smoke, registry smoke, release metadata, local CI, and token extractor self-tests, and packed-tarball smoke including workspace strict failure/success readiness checks plus workspace --learning-usage sidecar summaries and workspace --learning-eval checkpoint summaries and design-ai workspace workspace learning restore-backups readiness with restore rollback backup inventory, Website Console export validation through design-ai site --stdin --json, Website Console sample workspace generation through design-ai site --sample, Website Console from-intake workspace JSON stdout, JSON --out, and handoff bundle coverage through design-ai site --from-intake, Website Console project init workspace coverage through design-ai site --init in installed-bin and one-shot paths, Website Console init handoff bundle coverage through design-ai site --init --bundle --out <dir>, Website Console prompt template listing through design-ai site --prompt-list --json, Website Console MCP readiness check through design-ai site --stdin --mcp-check --json, Website Console MCP readiness probe check through design-ai site --stdin --mcp-check --probes --json, Website Console MCP action plan export through design-ai site --stdin --mcp-plan, Website Console workflow graph export through design-ai site --stdin --graph --json, Website Console handoff bundle export through design-ai site --stdin --bundle --out <dir>, Website Console non-empty evidence preservation through design-ai site --stdin --report, --tasks, and --bundle --out <dir>, Website Console handoff bundle check through design-ai site <bundle-dir> --bundle-check --strict --json with SHA-256 checksum verification, bundle digest/fingerprint verification, and generated bundle contract verification plus repair guidance, Website Console handoff bundle compare through design-ai site <bundle-dir> --bundle-compare <other-bundle-dir> --strict --json with bundle digest comparison plus warning-state strict smoke coverage that keeps identical warning bundles at sameBundle: true while exiting non-zero under --strict, Website Console target-repo handoff prompt through design-ai site <bundle-dir> --bundle-handoff --strict --json from a verified bundle digest, Website Console bundle repair preview/apply through design-ai site <bundle-dir> --bundle-repair --yes --json with repair report --out file output-file persistence, Website Console refactor task generation through design-ai site --stdin --tasks, Website Console task-selected single prompt generation through design-ai site --stdin --prompt codex-implementation --task task-homepage-cta, check learning capture, learning feedback, learn feedback --out output-file persistence, backup, redaction, learn JSON --out file writes, learn verify --out output-file persistence, learn diff JSON output, learn restore JSON output plus learn restore --out file-write confirmation, restore rollback backup inventory coverage, restore rollback backup prune preview/apply coverage through design-ai learn --restore-backups --prune --keep N, learn import --out output-file persistence, learn stats --out output-file persistence, human / JSON design-ai learn --usage usage sidecar report plus learn usage --out file-write confirmation, human / JSON design-ai learn --eval-template checkpoint generation plus generated checkpoint strict validation, human / JSON design-ai learn --eval checkpoint report plus learn eval --out file-write confirmation plus learn eval --strict failure gate, learn audit --out output-file persistence, learn --curate archive-first curation preview/apply plus --report --out Markdown report artifacts and usage-aware curation JSON review, verify, diff, restore, import, query-filtered learn list explanation/export, brief-relevant prompt/pack learning selection, prompt/pack learning usage sidecar recording, and audit cleanup guidance	Public registry smoke after publish now also verifies public registry design-ai site Website Console export validation, sample workspace coverage, prompt template listing, MCP readiness, MCP action plan, handoff bundle, bundle-check/compare/handoff/repair, refactor task generation, task-selected prompt generation, workspace --learning-usage sidecar summaries, workspace --learning-eval checkpoint summaries, public registry design-ai workspace workspace restore-backups readiness with restore rollback backup inventory, public registry design-ai learn --eval-template checkpoint generation plus public registry generated checkpoint strict validation, public registry JSON design-ai learn --restore preview/apply output plus public registry learn restore --out file-write confirmation, public registry learn restore rollback backup verification, public registry learn restore --backup-file path coverage, public registry design-ai learn --restore-backups restore rollback backup inventory coverage, and public registry design-ai learn --restore-backups --prune restore rollback backup pruning coverage
Internal dogfood readiness	Complete for v5.0.0	design-ai workspace reports read-only git cleanliness/sync, separates active status from ignored untracked local portfolio/evidence artifacts in JSON, canonical repository remote/metadata alignment, local learning profile audit state, usage sidecar event counts and stale selected-id readiness, optional or auto-detected sibling learning usage sidecar status, optional or auto-detected sibling learning eval checkpoint status, freshness warnings for checkpoint metadata that predates or mismatches the active learning profile, release-script availability, shell-safe learning profile/usage/checkpoint/report path commands in workspace next actions, usage-aware curation next actions plus companion learn --curate --report --out learning-curation-report.md report artifact commands for learning profile audit warnings and usage sidecar drift, and eval-template bootstrap next-action hints when a clean learning profile has entries but no checkpoint, sibling restore rollback backup inventory readiness, latest backup metadata, and prune next-action hints when older rollback backups exceed the default keep count; --strict exits non-zero on readiness warnings/failures; package smoke verifies strict JSON failure and clean-success behavior plus learning-usage, learning-eval, and workspace restore-backups summaries in installed-bin and one-shot paths, and registry smoke verifies the same learning-eval and workspace restore-backups contracts plus public registry Website Console coverage and public registry eval-template checkpoint generation after publish	Use findings to decide whether the next surface should be CLI, web UI, VS Code, Figma, or SDK
AI chat / conversational design consulting	Complete for v5.0.0	conversational-ui-designer, /conversational, and conversational knowledge cover voice, chatbot, and AI chat UX	Keep Korean platform conventions current
Local AI learning preferences	Complete for v5.0.0	design-ai learn, preview-first learn --init starter profile bootstrap, explicit learn --feedback keep/improve/avoid guidance with JSON --out artifact persistence, explicit check --learn --yes capture for local QA warning/failure results, full learn --backup --json export, redacted learn --redact --json sharing export from local profile / --from-file / --stdin, safe --out file output with --force overwrite control for JSON artifacts and export Markdown, non-mutating learn --verify, read-only learn --diff profile comparison against portable JSON, preview-first learn --restore full-profile replacement from portable backups with automatic rollback backup and optional --backup-file path, read-only learn --restore-backups sibling rollback backup inventory, preview-first learn --restore-backups --prune --keep N cleanup that deletes only older rollback backup files after --yes, confirmed learn --import, query-filtered learn --list --explain / learn --export without recency fallback, brief-relevant filtered prompt --with-learning / pack --with-learning with selection scoring metadata, local learning.usage.json sidecar events that store selected ids and short brief hashes, read-only learn --usage reports for sidecar activity, read-only learn --signals --report --out learning-signals.md Markdown signal handoff artifacts, focused read-only learn --agent-backlog --report --out agent-backlog.md agent backlog artifacts, preview-only learn --propose-skills reports for repeated check-capture skill deltas with --report --out skill-proposals.md Markdown review artifacts, read-only --review-file decision joins, read-only --apply-plan accepted proposal manual apply plans, --review-template --out skill-proposals.review.json JSON decision scaffolds, and --patch --out skill-proposals.patch unified diff handoffs, learn --eval-template runnable checkpoint generation from the active profile, read-only learn --eval checkpoint reports for deterministic learning selection QA with --strict failure gating, confirmed learn --forget/--clear, non-mutating learn --audit cleanup suggestions / learn --stats, safe learn --audit --fix --dry-run previews plus confirmed --fix --yes cleanup, archive-first learn --curate preview/apply with sibling *.archive.json preservation, learn --curate --report --out Markdown review artifacts, workspace report next actions for saving those artifacts before cleanup, and usage-aware profile-mismatch/stale/unused review hints, and learned-context audit summaries provide explicit local preference memory	Keep privacy boundaries clear as learning scope expands
AI model training	Not shipped scope	README states fine-tuning is outside shipped scope	Define a separate product phase if embeddings or fine-tuning becomes a goal
External launch	v5.1.0 distribution complete; adoption evidence pending	v5.1.0 is public across npm and GitHub Release, the Homebrew formula targets the same tag and passed install/test verification, and GitHub Pages plus the VS Code extension remain public	Collect consented adoption evidence before making external outcome claims

What is complete

The shipped product is a model-agnostic design intelligence layer for AI coding agents. It is complete when an agent can:

• Route a design request to the right workflow.
• Load the relevant design knowledge and examples.
• Produce design-system, component, UX, motion, illustration, print, video, game UI, conversational, spatial, document, and slide-deck artifacts.
• Review web/app interface craft across purpose and frequency, response, spatial continuity, interruptibility, motion cohesion, performance, accessibility, and responsive resilience with code/runtime evidence.
• Prepare one canonical HTML review and pass its exact evidence to another agent without claiming transport, acceptance, target-repository inspection, or implementation.
• Build the same read-only implementation-plan, critique-loop, or agent-readable DESIGN.md contract through CLI, SDK, MCP, or Website Console, with source files, explicit approval gates, and verification evidence preserved.
• Prepare website improvement and homepage implementation/refactor work through a local Site Profile, audit checklist, MCP readiness matrix, refactor plan, prompt generator, browser-local handoff evidence tracker, CLI/bundle evidence export, verified bundle evidence metadata, handoff readiness guidance in bundle README and summary.json.handoff, MCP probe count telemetry, generated bundle contract verification with per-file diagnostics and repair preview/apply, and a read-only MCP handoff that remains pending until explicit human approval without mutating the target website repo from design-ai.
• Check those artifacts for grounding, accessibility, responsive coverage, route-specific quality, and unresolved markers, with optional check --learn preview/application for non-pass QA feedback.
• Inspect the current local dogfood workspace with a read-only design-ai workspace snapshot, optionally include --learning-usage path and --learning-eval path, rely on automatic sibling learning.usage.json sidecar and learning-eval.json checkpoint detection when available, treat stale or source-mismatched usage sidecars and checkpoint freshness as a readiness warning, follow usage-aware learn --curate --usage-file next actions when profile audit or usage drift needs review, save companion learn --curate --report --out learning-curation-report.md curation report artifacts before applying archive cleanup, use the eval-template bootstrap next action when profile entries exist without a checkpoint, rely on shell-safe path quoting in learning usage/eval/report next-action commands, or use design-ai workspace --strict when warnings should block handoff, before deciding whether to continue development, run verification, capture learning feedback, commit, or push.
• Store, initialize starter dogfood preferences through preview-first learn --init, record explicit keep/improve/avoid feedback, capture local check warning/failure feedback after --yes, query-filter with optional list explanations, audit with cleanup suggestions, preview/apply safe audit cleanup, archive duplicate/sensitive entries through learn --curate --yes without losing audit history, save learn --curate --report --out Markdown curation reports, review usage sidecar profile mismatch, stale selected ids, and unused active entries through advisory usage-aware curation, summarize profile and usage activity, generate runnable eval checkpoint templates, evaluate deterministic learning-selection checkpoints, export, write JSON/export artifacts through safe --out file output, create redacted shareable backups from local profiles or portable JSON sources, verify/diff/restore with rollback backup, list restore rollback backups, preview/apply older rollback backup pruning without mutating the active profile, and import portable profiles, forget, and clear explicit local learning preferences, and inject brief-relevant scoped category/limit subsets into prompts/packs only when requested with selection scoring, audit-summary metadata, and privacy-preserving usage sidecar events attached.
• Install, update, status-check, and uninstall through the packaged CLI.
• Pass the release gate without relying on manual inspection.

In published v5.1.0, the manifest, commands, skills, agents, examples, and release smoke suite cover the workflows summarized above. Detailed release evidence remains in the release and smoke artifacts.

The release gate retains warning-state Website Console bundle-compare strict smoke coverage. See RELEASE-GATES.md for the full command and evidence matrix.

Local learning preferences are documented in AI-LEARNING.md.

What is not complete

The product now includes local learning preferences, but it should not be described as having completed AI model learning unless that scope is explicitly added later.

Not shipped:

• Fine-tuning a model.
• Training a private model on user artifacts.
• Embedding index generation for semantic retrieval.
• Background feedback loops that learn from accepted/rejected design recommendations without an explicit CLI command.

These are valid future product ideas, but they are different from the current architecture. The current architecture is deterministic corpus routing, prompt packing, quality checking, scoped local preference injection, and release-smoked CLI distribution.

Current release blockers

Distribution blockers are closed. Only adoption and future-scope decisions remain:

• v5.1.0 distribution: complete — npm latest, GitHub Release, and the Homebrew formula identify v5.1.0; the package has SLSA provenance and public registry smoke passed.
• Real-CI verification: complete — PR #48, main Audit run 29485559867, Docs run 29485559848, npm Trusted Publishing run 29485715200, and GitHub Release run 29485715193 passed on 2026-07-16.
• External launch: package distribution is complete across npm, GitHub Release, Homebrew, VS Code Marketplace sungjin.design-ai-vscode@0.4.1, and GitHub Pages; the public announcement itself remains a maintainer decision.
• Reference-link policy: decided — refs/ source links moved behind generated reference pages. tools/extractors/reference_pages.py emits docs/reference/{ant-design,mui,shadcn-ui,awesome-design-md}.md, and corpus pages (knowledge/components/INDEX.md, knowledge/patterns/brand-references.md, examples/*.md) link to those pages instead of the gitignored refs/ mirror. Plain-text refs/ provenance mentions (frontmatter source:/sources:, link text) stay as-is, and the MkDocs refs-only warning baseline in tools/audit/local-ci.py is now 0.

Recommended next decision

The next product decision is evidence-led adoption, not another contract or surface:

1. Privately invite three to five trusted project owners with the P16 Korean-first message, while recording only aggregate counts in this repository.
2. Accept candidate intake only through the public P15 Issue Form, assign exactly one P16 lifecycle label, and do not treat an application or label as consent.
3. Activate the public marketing-page slot only after a real project owner records authority, data boundaries, target scope, screenshot policy, and consent.
4. Select one product capability only after the same fixed-key problem appears in completed, source-backed pilot records from at least two distinct owners.