V4 Scope

V4 should build on the complete local-first V1-V3 core. It must not include hosted service work, database/storage work, GitHub/PR adapters, CLI UX, or a typed SDK. Those wait until the backend shape is concrete.

V4 stays inside the existing MCP/domain architecture: richer standards, richer file-summary/local-scan review, stronger agent workflow outputs, and more deterministic evals.

V4 Standards Depth

Goal: make architect-mcp better at loading, applying, and reviewing broad engineering quality standards, not only architecture structure.

Source-Backed Stack Packs

• Promote more ingested llms.txt sources into versioned packs only when they produce concrete, reviewable rules.
• Prioritize security, accessibility, performance, testing, dependency hygiene, agent instruction quality, verification honesty, and framework-specific production readiness.
• Require every promoted rule to include source provenance, detector mapping, examples, verification expectations, and conflict review.
• Keep live fetching explicit; default to local ingested snapshots.

Broader Quality Rules

Security:

• Secret handling, auth/session boundaries, authorization checks, webhook verification, SSRF/path traversal risk, MCP config safety, and dependency risk.

Accessibility:

• Semantic UI structure, keyboard flows, form labeling, focus handling, color/contrast policy, and mobile/touch ergonomics for frontend stack packs.

Performance:

• Large render surfaces, unnecessary client/server boundaries, heavyweight dependencies, bundle-risk hints, and avoidable repeated work.

Testing:

• Test proximity, behavior coverage, regression proof, eval fixture coverage, mutation/migration tests, and verification-command alignment.

Dependency Hygiene:

• Upgrade reason, compatibility risk, lockfile impact, rollback plan, audit state, package-surface changes, and dependency-boundary drift.

V4 Standards Deliverables

• Versioned V4 policy/foundation packs for security, accessibility, performance, testing, dependency hygiene, verification honesty, and agent handoff quality.
• More executable detector families where signals are available from file summaries or local scans.
• Review findings with stable codes, confidence, and remediation text.
• Fixture coverage for clean and failing examples per policy area.
• Stack-pack quality matrix that distinguishes source depth, executable detector coverage, docs, and tests.

V4 Agent Workflow

Goal: improve how clients run coding agents through a whole work session, from vague intent to final proof, without slowing low-risk work.

Harness Loops

• Strengthen review_agent_session into the primary end-of-work gate for clients.
• Add richer session summaries that capture changed scope, assumptions, triggered standards, skipped checks, unresolved risks, and next-agent handoff.
• Detect intent debt when too many assumptions or yellow decisions accumulate across a session.
• Make confidence-cliff downgrades explicit after repo inspection changes the original risk picture.

Change Review

• Define a change-review input shape using changed file summaries, intended contract, verification output, and optional baseline.
• Compare changes against pre-edit contracts, stack packs, policy packs, and generated agent instructions.
• Flag unrelated files, scope upgrades, missing tests, skipped checks, dependency churn, and public API drift.

Memory Proposals

• Keep memory stateless in core V4.
• Improve proposal quality: scope, sensitivity, expiry/staleness, invalidation condition, source summary, and token-cost estimate.
• Add memory-review recipes that stay storage-neutral and return structured proposals only.

Agent Handoff

• Generate concise handoff blocks after intent confirmation, after implementation review, and after final-response review.
• Include current contract, loaded standards, assumptions, verification status, remaining risks, and next recommended tool call.

YOLO Budgets

• Formalize strict, guided-yolo, and full-yolo as client-facing policy profiles.
• Add budget counters for assumptions, skipped checks, widened files, risky terms, and unverified claims.
• Let low-risk reversible changes proceed quickly while medium/high-risk work asks one focused question.

Policy Packs

• Treat policy packs as composable standards applied before stack packs.
• Keep policy packs local-first and versioned like foundation/stack packs.
• Candidate V4 policy packs: security, accessibility, performance, testing, dependency hygiene, verification honesty, and agent handoff quality.

V4 Workflow Deliverables

• Stronger review_agent_session coverage and examples.
• Change-review recipe and fixtures.
• Expanded memory proposal/relevance tests.
• YOLO budget outputs and tests.
• Handoff summary output contract.
• V4 eval suites for standards depth and agent workflow.

Non-Goals

• Hosted service.
• Database/storage work, including schema, persistence, migrations, review history, or memory storage.
• GitHub or PR adapter implementation.
• Hosted API, hosted policy, hosted auth, hosted billing, or hosted dashboard work.
• CLI UX.
• Typed client SDK.

Those stay deferred until a later backend/productization phase after the local MCP architecture is concrete.