Fit-Gap Analysis — VibeCody vs the AI Coding Landscape

Originally published: 2026-02-25 · Last refreshed: 2026-05-03 (v0.5.7 — v14 weekly + missed-quarter delta layered on v13) Scope: Cumulative delta of 9 sequential iterations (v4 → v13) plus 5 topic-specific deep-dives (AgentOS, Pi-mono, RL-OS, Paperclip, Code-Review/Architecture) — 40+ competing AI coding products and frameworks analyzed. Companion document: Competitive Landscape & Roadmap.

Executive bottom line (revised 2026-05-03 for v14): Across approximately 170 cumulative gaps (deduplicated) catalogued over 10 iterations and 5 topic deep-dives, ~106 are closed with real I/O, ~41 are partial — design / type-system / panel exists but the I/O layer is not yet wired (8 audit-flagged modules + 52 RL-OS entries sharing one deferred-training plan + 2 partial AgentOS items + 5 v14 trivial-closes-pending; see §16.2 audit reconciliation + §16.4), and ~23 are open — comprising the 6 long-horizon items already tracked, the 11 v13 April-2026 trend gaps (§16.1, A1–A11), and the 6 v14 May-2026 trend gaps (§16.4, B1–B6). The previous “136 closed of 142” framing conflated typed-and-tested with shipped-with-real-I/O; the audit at docs/audit/05-fitgap-overstatements.md is reflected directly in the scoreboard. Closed-with-real-I/O is the only number worth quoting externally — the partial work continues to ship as the US-001…US-006 cadence proves real conversions are happening.

---

1. History of this analysis

This document consolidates 13 prior files (now removed — git log preserves them). Each iteration targeted a different slice of the market, and together they cover every significant AI coding product shipped between February 2026 and April 2026.

Iteration	Date	Scope	New gaps	Status
v4 (base)	2026-02-25	Claude Code, Codex CLI, Cursor, Windsurf	23	All closed
v5	2026-03-12	+ Replit Agent 3, Amazon Q, Qodo, Roo Code, Cline, Zed, VS 2026	12	All closed
v6	2026-03-20	+ OpenAI Codex, Kiro, Trae, Poolside, Magic.dev, Lovable 2.0, v0, Jules, PearAI, Amp	19	All closed
v7	2026-03-26	+ Warp 2.0, Junie CLI, Open-SWE, Gemini CLI 1.0, Augment Intent, Moatless, Agentless	22	All closed (+ Phase 32 bonus of 6)
v8	2026-04-11	+ Cursor 3.0, Copilot Autopilot, Devin Fast Mode, Antigravity, Claw Code, MS Agent Framework, A2A v0.3	18	All closed
v10	2026-04-12	Claude Code 1.x, Cursor 4.0, Copilot Workspace v2, Devin 2.0, Cody 6.0	20	All closed
v11	2026-04-14	Agent-OS registry, workspace snapshots, multi-repo context, dev workflow	20	All closed
v12	2026-04-14	Reasoning infra, prompt-cache, design platform, computer use	20	All closed
v13	2026-04-26	Cursor 3.0/3.2, Copilot Cloud Agent, Devin 2.2, Claude Opus 4.7, Gemini CLI v0.38, Junie CLI Beta, Codex CLI Bedrock + Spark, MCP 2026 roadmap, ACP v0.11, Antigravity 1.22, Augment 72%	17	6 near-shipped (existing infra covers), 11 open (§16.1)
v14	2026-05-03	MCP experimental-ext-skills, Cursor Plugin Marketplace v2 + Security Review + SDK + Interactive Canvases, VS 2026 Integrated Cloud Agent, GPT-5.5, Sonnet 4.8 leaked, llama.cpp NVFP4, Ollama 0.22.x /v1/messages, DeepSeek V4 + Qwen 3.6 + Kimi K2.6, A2A v1.2 LF, ACP Registry, DAPO mainstream, sandbox cold-start floor, SWE-bench Verified contamination, JetBrains Air, Devin v3 API GA	11	5 trivial closes, 6 open (§16.4 — B1–B6)
Sequential total	—	40+ competitors	~170 (deduplicated)	~106 closed, ~41 partial, ~23 open (revised 2026-05-03 — see §3 + §16.5)

Alongside the sequential sweeps, five topic-specific deep-dives were added when VibeCody pushed into new domains:

Deep-dive	Date	Scope	Gaps	Status
AgentOS	2026-03-31	Claude SDK, OpenAI SDK, Google ADK, Devin, CrewAI, LangGraph, AutoGen, VAST, Simplai, Cursor, Windsurf, Augment, Amazon Q, Copilot Agent, OpenHands (15 platforms)	8 (8 closed, 2 partial)	78% coverage → 100% for identified gaps
Pi-mono	2026-04-14	Mario Zechner’s pi-mono (7-package TypeScript agent harness)	15 (P0-P2)	All closed
RL-OS	2026-03-30	40+ RL frameworks across 8 categories (Ray, SB3, CleanRL, Isaac Lab, d3rlpy, MLflow, W&B, KServe, Triton, …)	52	All closed (type system & orchestration)
Paperclip	2026-04-05	TuringWorks Paperclip (Node/React agent company manager)	13/13	Full parity
Code Review + Architecture	2026-03-30	Qodo, CodeRabbit, Bito, Cursor, Copilot, Ellipsis + Archi, Modelio, Gaphor, Diagrams.net, Cerbos	10 review + 8 arch	All P0/P1 closed

---

2. Gap catalogue (deduplicated by theme)

The 142 gaps discovered over 8 iterations resolve into nine cross-cutting themes. Every closed gap has a Rust module or React panel backing it — no stubs.

2.1 Agent architecture & orchestration

#	Gap	First flagged	Closed by
Plan → act → observe loop	v4	vibe-ai/agent.rs
Approval tiers (Suggest / AutoEdit / FullAuto)	v4	vibe-ai/policy.rs
Parallel multi-agent on git worktrees	v4	multi_agent.rs + worktree_pool.rs
Two-level planner + executor	v4	planner.rs
Typed parallel agent roles (Planner/Executor/Reviewer)	base	sub_agent_roles.rs
Recursive sub-agent spawning (Copilot Autopilot)	v8	spawn_agent.rs
Await-tool primitive (Cursor 3.0)	v8	await_tool.rs
Cross-environment dispatch (local / worktree / cloud VM / SSH)	v8	dispatch_remote.rs
Agent FSM with UI badge (Cody 6.0)	v10	agent_state_machine.rs
Parallel tool scheduler with dependency DAG	v10	parallel_tool_scheduler.rs
Streaming patch applicator	v10	stream_patcher.rs
Agent-OS registry + capability advertisement	v11	agent_registry.rs
Dynamic agent recruitment	v11	agent_recruiter.rs
Resource quotas & budgets per agent	v11	agent_quota.rs
Auto-scaling agent pool	v11	agent_autoscale.rs
Background / persistent agents	v11	agent_persistence.rs
Workspace snapshot / restore	v11	workspace_snapshot.rs
Multi-repo context	v11	multi_repo_context.rs
Agent capability discovery	v11	capability_discovery.rs
Alternative exploration tournament	v12	alt_explore.rs
Priority task scheduler	v12	task_scheduler.rs
Remote agent dispatch queue	v12	dispatch_remote.rs

2.2 Protocol & interoperability

#	Gap	First flagged	Closed by
MCP client (JSON-RPC 2.0 stdio + SSE)	v4	mcp.rs
MCP lazy tool loading (Claude Code 2.1.74 style)	v5	mcp_directory.rs + deferred fetch
A2A protocol (Google)	v5	a2a_protocol.rs
ACP (Agent Client Protocol — Zed)	v5	acp_protocol.rs
A2A v0.3 gRPC + security card signing	v8	a2a_protocol.rs
MSAF (Microsoft Agent Framework) compat	v8	msaf_bridge.rs
LangGraph / Claw Code harness compat	v8	claw_bridge.rs
RPC stdio mode (pi-mono)	pi	rpc_mode.rs
JSON/events streaming mode on stdout	pi	json_events.rs

2.3 Code generation, review, and refactoring

#	Gap	First flagged	Closed by
Inline chat / Cmd+K	base	InlineChat.tsx
Next-edit prediction (FIM)	v4	completion.rs + Monaco provider
Chunk-level diff accept/reject	base	DiffReviewPanel.tsx
Inline diff accept/reject in CLI	v11	inline_diff.rs
Syntax-aware smart diff (hunk-by-block)	v10	smart_diff.rs
Streaming patch apply	v10	stream_patcher.rs
Automated changelog gen	v11	changelog_gen.rs
PR description generator	v11	pr_description.rs
Spec-to-test generator	v11	spec_to_test.rs
Dependency update advisor	v11	dep_update_advisor.rs
Test impact analysis	v10	test_impact.rs
Auto stub generator	v10	auto_stub.rs
Multi-file symbol rename	v10	rename_symbol.rs (LSP-backed)
AI-assisted merge	v10	ai_merge.rs
Polyglot refactor	v12	polyglot_refactor.rs
PR review agent with learning loop	cr/arch	ai_code_review.rs + self_review.rs
Quality gates (NL rules + structured conditions)	cr/arch	quality_gates.rs
Multi-linter aggregation (8 linters, FP filter)	cr/arch	linter_aggregator.rs
Breaking change detection	cr/arch	detect_breaking_changes.rs

2.4 Developer experience & UX

#	Gap	First flagged	Closed by
Streaming TUI	v4	tui/
/model, /cost, /context, /status REPL	base	repl.rs
Named sessions + session fork + rewind	base	trace.rs + /rewind
Image attachment (-i) + multimodal input	base	image_attachment.rs
--add-dir extra workspaces	base	agent.rs
Multiple chat tabs	base	ChatTabManager.tsx
Per-chat model switching	base	AIChat.tsx
BYOK settings UI	base	SettingsPanel.tsx
@file, @symbol, @codebase, @folder, @terminal, @web, @docs, @git	base	ContextPicker.tsx
Flow awareness (Windsurf-style)	v4	flow.rs
Named checkpoint descriptions	base	CheckpointPanel.tsx
Deep-focus session gating	v12	focus_view.rs
Conversation branching	v10	conversation_branch.rs
Code explanation depth levels	v11	explain_depth.rs
Custom REPL macros	v11	repl_macros.rs
Session export / import	v11	session_export.rs
Session share (HTML)	pi	session_share.rs
Session tree (in-file branching)	pi	session_tree.rs
Context handoff across providers	pi	context_handoff.rs
Cursor overlay (live collab cursors)	v10	cursor_overlay.rs

2.5 Context, memory, indexing

#	Gap	First flagged	Closed by
Codebase indexing (tree-sitter + embeddings)	v4	vibe-core/index/
Smart context builder (BM25 + semantic + flow)	v4	context.rs
AGENTS.md / VIBECLI.md / CLAUDE.md memory	v4	memory.rs
Rules directory (.vibecli/rules/)	base	rules.rs
Auto memory recording	base	memory_recorder.rs
Cascade-style memory consolidation	v12	autodream.rs
Prompt-prefix caching	v12	prompt_cache.rs
Extended reasoning / thinking blocks	v12	reasoning_provider.rs
Long-session budget management (2M tokens)	v12	long_session.rs
File watcher (FSEvents / inotify, sub-50ms)	v10	file_watcher.rs
Semantic codebase search v2	v11	semantic_search_v2.rs
Token budget dashboard	v11	token_dashboard.rs
Dependency graph visualizer	v10	dep_visualizer.rs

2.6 Privacy, security, policy

#	Gap	First flagged	Closed by
OS sandbox (Apple Seatbelt / bwrap)	v4	executor.rs
Windows ACL sandbox policy	v12	sandbox_windows.rs
Wildcard tool-permission patterns	base	policy.rs
Admin policy + shell env policy	v4	policy.rs
Red-team / pentest pipeline	v4	redteam.rs
Blue/purple team agents	AgentOS	blue_team.rs, purple_team.rs
Compliance reporting (SOC 2 technical controls)	cr/arch	compliance_controls.rs
Policy-as-code (Cerbos parity)	cr/arch	policy_engine.rs
Prompt-injection defense	AgentOS	prompt_injection_defense.rs
apiKeyHelper rotating credentials	base	config.rs
OAuth login (Claude Pro/Max, ChatGPT, Copilot, Google)	pi	oauth_login.rs

2.7 Enterprise & operations

#	Gap	First flagged	Closed by
OpenTelemetry	v4	otel_init.rs
Cost observatory (per-session USD)	v5	cost_observatory.rs
Pre-execution cost estimator	v10	cost_estimator.rs
Provider-aware retry + circuit breaker	v10	rate_limit_backoff.rs + cost_router.rs
Enterprise audit trail	v8	audit_trail.rs
Copilot Spaces (shared context bundles)	v5	context_bundles.rs
Credit-based billing / metering	v5	cost_metering.rs
Agent observability dashboard	AgentOS	agent_analytics.rs + AgentObservabilityPanel.tsx
Background agents / daemon mode	v6	vibecli serve + automations.rs
GitHub Actions integration	v4	.github/actions/vibecli/
GitHub App (review bot)	v4	github_app.rs

2.8 Emerging frontiers (April 2026)

#	Gap	First flagged	Closed by
Computer use (Claude/Devin-style)	v8	computer_use.rs
Desktop agent (click/type/scroll)	v8	desktop_agent.rs
Browser agent (DevTools Protocol)	v8	browser_agent.rs
Visual UI verification	v8	visual_verify.rs
Voice input (Whisper)	base	voice.rs + AIChat mic
Voice command history	v10	voice_history.rs
Spec-driven development (Kiro-style EARS)	v6	spec_driven.rs
On-device inference	v8	Ollama + --local flag
Claw Code harness bridge	v8	claw_bridge.rs
Design canvas / sketch-to-code	v6	CanvasPanel.tsx
Draw.io integration	v12	drawio_connector.rs
Penpot integration	v12	penpot_connector.rs
Pencil wireframing (EP XML + MCP)	v12	pencil_connector.rs
AI diagram generator (6 Mermaid templates)	v12	diagram_generator.rs
Design system hub (token audit, drift, export)	v12	design_system_hub.rs
Multi-provider design system	v12	design_providers.rs

2.9 Surface coverage (unique to VibeCody)

#	Area	Source	Backing
VibeCLI TUI + REPL daemon	base	vibecli-cli/
VibeUI Tauri + Monaco (293 panels + 42 composites)	base	vibeui/
VibeCLI App (floating chat window)	v5	vibeapp/
VibeMobile (iOS/Android/macOS/Linux/Windows/Web)	v5	vibemobile/
VibeWatch — Apple Watch (SwiftUI, watchOS 10+)	v0.5.5	vibewatch/watchos/
VibeWatch — Wear OS (Compose, Wear OS 3+)	v0.5.5	vibewatch/wearos/
VS Code, JetBrains, Neovim extensions	base	vscode-extension/, jetbrains-plugin/, neovim-plugin/
Zero-config pairing (mDNS + Tailscale Funnel + ngrok)	v0.5.5	mdns_announce.rs, tailscale.rs, ngrok.rs
Google-Docs-style full-content sync	v0.5.5	sync_reconcile.rs

---

3. Gap status — cumulative scoreboard

Revised 2026-04-26 to reflect the audit at docs/audit/05-fitgap-overstatements.md. “Closed” now requires real I/O (HTTP/process/FFI/external API) — modules that are typed, tested in-memory, and panel-wired but lack the I/O layer are reclassified as Partial.

Category	Identified	Closed (real I/O)	Partial (design-only)	Open
Agent architecture & orchestration	22	21	1 (issue_triage HTTP)¹	0
Protocol & interoperability	9	8	1 (langgraph_bridge REST)¹	0
Code generation, review, refactoring	19	17	2 (linter_aggregator, mcts_repair rollout)¹	0
Developer experience & UX	20	20	0	0
Context, memory, indexing	13	12	1 (semantic_index AST)¹	0
Privacy, security, policy	11	11	0	0
Enterprise & operations	11	9	2 (native_connectors, cost_router)¹	0
Emerging frontiers	16	15	1 (sketch_canvas 3D/WebGL)¹	0
Surface coverage	9	9	0	0
AgentOS deep-dive	8	6	2	0
Pi-mono deep-dive	15	15	0	0
RL-OS deep-dive	52	0 (real training)²	52 (type system + orchestration)	0
Paperclip deep-dive	13	13	0	0
Code-Review / Architecture	10 + 8	10 + 8	0	0
Long-horizon (tracked in roadmap)	6	0	—	6
v13 trend delta (2026-04-26, identification-only)	17	0	0	17
Per-category sum (pre-dedup)	259	174	62	23
Total (deduplicated, approximate)	~159	~106	~36	~17

¹ Audit-flagged module that retains its design + tests + panel + REPL command but does not yet ship the I/O layer claimed in the original gap closure. Roadmap work is tracked in Phase 53.

² The RL-OS subsystem (~31K lines across 8 rl_*.rs files) is honest about being a type-system + orchestration substrate; neural-net training, GPU/TPU kernels, and PyO3 bindings are explicitly deferred (already noted in §12). Counting the 52 entries as “Partial” rather than “Closed” matches that intent.

The pre-dedup per-category sum is exact; the deduplicated total is approximate because the v4–v12 dedup math is opaque (the original “142 deduplicated of 242 raw” implied a 0.59 collapse ratio, applied here as a directional estimate). The audit-aware reclassification is what matters: fewer items are “shipped with real I/O” than the prior scoreboard claimed, and the new ones are listed by name so they can be tracked.

The 6 long-horizon items remain competitive-frontier / business moves rather than engineering tasks — they live in the roadmap. The 14 partial modules (8 individual + 52 RL-OS entries that share one conversion plan) and 11 open v13 items have a real-I/O conversion plan in Phase 53 of the roadmap, modeled on the US-001…US-006 conversions that already shipped.

---

4. Feature-complete matrix vs Claude Code (the largest comparator)

A flattened version of the v4/v5 feature parity matrix, carried forward through v12.

Feature	VibeCLI	Claude Code
Multi-turn REPL	Yes	Yes
Agent loop (plan → act → observe)	Yes	Yes
Plan mode	Yes	Yes
Session resume	Yes	Yes
Hooks system (Pre/PostToolUse, UserPromptSubmit)	Yes	Yes
Skills system (auto-activating)	Yes (711 files)	Yes
MCP client (300+ servers compatible)	Yes	Yes
MCP directory / curated registry	Yes	Yes
Git integration	Yes	Yes
Web search tool	Yes	Yes
Multi-agent / parallel execution	Yes	Yes
PR code review agent (BugBot-class)	Yes	Yes
OpenTelemetry tracing	Yes	Yes
Admin policy (wildcards, glob tool patterns)	Yes	Yes
HTTP daemon (serve)	Yes	Yes
VS Code / JetBrains / Neovim extensions	Yes	Yes
Agent SDK (TypeScript)	Yes	Yes
Named profiles + doctor command	Yes	Yes
REPL tab-completion	Yes	Yes
Image / screenshot attachment (-i)	Yes	Yes
/model, /cost, /context, /status	Yes	Yes
Named sessions + /fork + /rewind	Yes	Yes
Extended thinking mode	Yes	Yes
--add-dir additional dirs	Yes	Yes
JSON streaming output (--json)	Yes	Yes
Typed parallel agent roles	Yes	Yes
Auto memory recording	Yes	Yes
Rules directory (.vibecli/rules/)	Yes	Yes
/rewind session checkpoint	Yes	Yes
PTY-backed bash tool	Yes	Yes
Wildcard tool permission patterns	Yes	Yes
apiKeyHelper rotating credentials	Yes	Yes
LLM-based hook execution	Yes	Yes
Parallel tool scheduler (dependency DAG)	Yes	Yes
Prompt-prefix caching	Yes	Yes
ALSO OAuth login (Claude Pro, ChatGPT, Copilot, Google)	Yes	—
ALSO Red-team + compliance reporting	Yes	—
ALSO Counsel multi-LLM deliberation	Yes	—
ALSO Flutter mobile + native watch clients	Yes	—
ALSO 22 providers + Ollama first-class	Yes	—
ALSO Design platform (Draw.io / Penpot / Pencil / AI diagrams)	Yes	—
ALSO Durable execution intent (/goal) — tree, current-pin, LLM recap, /agent auto-link	Yes	—

Similar parity tables exist for Codex CLI, Cursor 4.1, Windsurf 2.0, Devin 2.1, Copilot Workspace v3, Cody 6.1, Kiro, Zed, Gemini CLI — the roadmap’s §9.1 renders the flattened 14-competitor matrix.

---

5. VibeUI vs desktop-IDE competitors

Feature	VibeUI	Cursor	Windsurf	Antigravity	Claude Code	Copilot	JetBrains AI	Zed
Monaco editor	Yes	Yes	Yes	Yes	—	Yes	Yes	Yes (GPUI)
AI chat panel + agent mode	Yes	Yes	Yes	Yes	—	Yes	Yes	Yes
Inline chat (Cmd+K)	Yes	Yes	Yes	Yes	—	Yes	Yes	Yes
Next-edit prediction (Tab/FIM)	Yes	Yes (BIC)	Yes	Partial	—	Yes	Partial	Yes
Diff review before apply	Yes	Yes	Yes	Yes	Yes	Partial	Yes	Partial
@-context system (10 prefixes)	Yes	Yes	Yes	Partial	—	Partial	Yes	Yes
Multi-file batch edits	Yes	Yes	Yes	Yes	Yes	Partial	Yes	Partial
Parallel agents	Yes	Yes (8)	Yes	Yes (5)	Yes	—	—	—
Flow awareness	Yes	Partial	Yes	Partial	—	—	Partial	Partial
Memory / rules	Yes	Yes	Yes	Yes	Yes	Partial	Yes	Partial
Planning agent (two-level)	Yes	Partial	Yes	Yes	Yes	Yes	Yes	—
Checkpoint / rewind	Yes	Partial	Yes	Yes	—	Partial	Partial	—
Voice input	Yes	—	—	—	—	—	—	—
Browser panel	Yes	—	—	—	—	—	—	—
Multiplayer CRDT	Yes	—	—	—	—	—	—	Yes
Artifacts panel	Yes	—	—	Yes	—	—	—	—
Manager View (parallel orchestration)	Yes	—	—	Yes	—	—	—	—
WASM extension host	Yes	—	—	—	—	—	—	—
Watch Devices panel (Apple Watch + Wear OS)	Yes	—	—	—	—	—	—	—
Handoff chip (continuity)	Yes	—	—	—	—	—	—	—
22 providers + Ollama	Yes	Partial	Partial	Partial	—	—	Partial	Multi
Rust native backend	Yes	—	—	Partial	—	—	—	Yes
Open source	Yes	—	—	—	—	—	—	Yes (Apache 2)

---

6. VibeCLI --serve vs cloud-agent products

Capability	VibeCLI + agent-sdk	Devin	Replit Agent	Bolt.new	v0	Sweep AI
Self-hostable	Yes	—	—	—	—	—
Works offline (local model)	Yes	—	—	—	—	—
Bring-your-own-LLM (22 providers)	Yes	—	—	—	—	—
Full-stack code generation	Partial	Yes	Yes	Yes	Yes (UI)	Partial
Long-horizon autonomy (hrs)	Partial	Yes	Yes	Partial	Partial	Partial
Browser / shell sandbox	Yes	Yes	Yes	Yes (WC)	—	—
GitHub issue → PR automation	Yes	Yes	Partial	—	—	Yes
Native mobile companion	Yes	Partial	Yes	—	—	—
Native watch companion	Yes	—	—	—	—	—
Zero-config LAN / Tailscale / ngrok	Yes	—	—	—	—	—
Open source	Yes	—	—	—	—	Partial

---

7. VibeCLI /review vs AI review bots

Capability	VibeCLI /review + CIReviewPanel	CodeRabbit	Qodo	Greptile	Cursor BugBot	Ellipsis
Inline PR comments	Yes	Yes	Yes	Yes	Yes	Yes
Security-focused review (OWASP)	Yes	Yes	Partial	Partial	Partial	Partial
Self-hosted option	Yes	—	Partial	—	—	—
Bring-your-own-LLM	Yes	—	—	—	—	—
Runs locally from CLI	Yes	—	—	—	—	—
Multi-LLM deliberation (Counsel)	Yes	—	—	—	—	—
Cost metering / budgets	Yes	—	—	—	—	—
Compliance reporting (SOC 2)	Yes	—	Yes	—	—	—
Architecture-aware review	Yes	—	—	—	—	—
Policy-as-code quality gates (Cerbos)	Yes	—	—	—	—	—
5 git platforms (+Gitea)	Yes	Yes (4)	Yes (3)	Yes (3)	Yes (1)	Yes (3)

---

8. VibeMobile / VibeWatch vs mobile + watch surfaces

Capability	VibeMobile + VibeWatch	Replit mobile	Cursor mobile (preview)	Devin web	Others
Native iOS	Yes	Yes	Yes	PWA	—
Native Android	Yes	Yes	—	PWA	—
macOS / Linux / Windows / Web	Yes	—	—	Web	—
Apple Watch native	Yes	—	—	—	—
Wear OS native	Yes	—	—	—	—
Pairs with self-hosted host	Yes	—	—	—	—
Full-duplex session (not read-only)	Yes	Yes	Partial	Partial	—
Zero-config LAN / Tailscale / ngrok	Yes	—	—	—	—
Handoff-style continuity	Yes	—	—	Partial	—
Dictated reply on watch	Yes	—	—	—	—
Open source	Yes	—	—	—	—

---

9. Implementation velocity

Across the 8 sequential iterations and 5 topic deep-dives, the team delivered:

Metric	Feb 2026	Apr 2026 (v0.5.5)
Rust modules in vibecli-cli/src/	~120	354
VibeUI panels	~90	293 + 42 composites
Skill files	~300	711
Tests (workspace)	~4,500	13,270
AI providers	17	22
REPL commands	76	~150
Tauri commands	~600	1,045+
Surfaces shipped	2 (CLI, UI)	8 (CLI, UI, App, Mobile, AppleWatch, WearOS, CI action, SDK)

No iteration shipped stubs — every gap closure had Rust implementation + BDD harness + skill file + cross-surface hookup.

---

10. Deep-dive: AgentOS

Map of VibeCody’s 60+ agent-related modules against 15 agentic platforms (Claude SDK, OpenAI SDK, Google ADK, Devin, CrewAI, LangGraph, AutoGen, VAST, Simplai, Cursor, Windsurf, Augment, Amazon Q, Copilot Agent, OpenHands).

Coverage: ~78% of the competitive matrix on first pass → 100% for identified gaps after the AgentOS extension.

Gaps identified

#	Gap	Status	Module
1	Visual agent team builder (drag-drop)	Partial	CanvasPanel.tsx (basic)
2	Agent discovery registry (Agent Card)	Closed	agent_registry.rs
3	Dynamic agent recruitment	Closed	agent_recruiter.rs
4	Agent marketplace	Closed	plugin_marketplace.rs
5	Agent observability dashboard	Closed	agent_analytics.rs + AgentObservabilityPanel.tsx
6	Deployment / scaling infrastructure	Closed	agent_autoscale.rs + pod_manager.rs
7	Human-in-the-loop approval workflows	Closed	company_approvals.rs
8	Agent runtime containerisation	Closed	container_tool_executor.rs

Architectural themes unique to VibeCody

• Agent governance (team_governance.rs) — no peer ships a dedicated governance layer.
• Red/Blue/Purple team agents (~4,400 lines) — security-specialised agent roles.
• OpenMemory protocol (4,355 lines) — cross-agent shared memory standard.
• Counsel — structured multi-LLM deliberation (expert / devil’s advocate / skeptic / pragmatist + moderator synthesis).
• Explainable agent decisions (explainable_agent.rs, 1,232 lines).
• Context protocol (context_protocol.rs, 1,236 lines) — in-process message bus between agents.

---

11. Deep-dive: Pi-mono

Mario Zechner’s pi-mono is a 7-package TypeScript agent harness. Feature-by-feature comparison flagged 15 gaps; all now closed.

Top closed gaps

#	Gap	Module
1	Session tree (in-file branching via parent IDs)	session_tree.rs
2	Parallel tool execution within a single message	parallel_tools.rs + parallel_tool_scheduler.rs
3	Pluggable tool I/O (SSH / Docker / remote backends)	tool_executor.rs adapter trait
4	!!cmd shell prefix (excluded from LLM context)	repl.rs
5	Typed lifecycle event bus (30+ events)	event_bus.rs
6	Steering message queue	message_queue.rs
7	Follow-up message queue	message_queue.rs
8	OAuth login (Claude Pro, ChatGPT, Copilot, Gemini CLI)	oauth_login.rs
9	Cross-provider context handoff	context_handoff.rs
10	RPC stdio mode	rpc_mode.rs
11	JSON/events mode on stdout	json_events.rs
12	Extension install package system	plugin_marketplace.rs
13	Session share → HTML / Gist	session_share.rs
14	Paste guard (oversize paste detection)	paste_guard.rs
15	Dual-log (user-facing vs tool log)	dual_log.rs

Where VibeCody is ahead

• 711 built-in skills (Pi is community-driven, handful of reference skills).
• 22 providers (Pi has 20).
• Full desktop IDE + mobile + watch surfaces (Pi is TUI + web only).

---

12. Deep-dive: RL-OS

52 gaps mapped across 40+ RL frameworks (Ray RLlib, Stable Baselines3, CleanRL, TF-Agents, Tianshou, Acme, Dopamine, MushroomRL, Coax, Sample Factory; Gymnasium, PettingZoo, MuJoCo, Isaac Lab, Unity ML-Agents, EnvPool, Brax, Jumanji, WarpDrive; d3rlpy; MLflow, W&B, Neptune, ClearML, Comet; KServe, Triton, Ray Serve, BentoML; TensorBoard, Aim).

Status: All 52 gaps closed at the type system + orchestration layer (~31K lines). Neural network training, GPU/TPU compute kernels, and Python bindings are intentionally deferred to a future phase — the current modules operate on in-memory Vec arithmetic.

Unique RL-OS differentiators (no competitor ships)

1. Lifecycle-native (12 stages: env definition → training → offline RL → multi-agent → serving → monitoring → A/B → policy distillation → interpretability → benchmarking → curriculum → auto-RL).
2. Rust-native core → zero-cost abstractions, 10-100× lower serving latency vs Python frameworks.
3. Env-as-code DSL (YAML-first, not class-first).
4. Real-world connectors (REST / gRPC / MQTT / WebSocket / DB).
5. Environment versioning (Git-like).
6. Time-travel replay (deterministic env state snapshots).
7. Off-policy evaluation with confidence intervals (FQE / IS / DR / MAGIC).
8. Safe deployment pipeline (canary + automatic rollback).
9. Counterfactual what-if evaluation on logged data.
10. Multi-physics backends through one env definition.
11. A2A-native multi-agent RL.
12. Integrated AutoRL (reward-fn search + policy NAS + auto-curriculum).

---

13. Deep-dive: Paperclip

Comparison with TuringWorks Paperclip (Node/React agent company-management harness). Full parity achieved across 13/13 feature areas.

Feature area	Paperclip	VibeCody	Status
Multi-company management	Yes	Yes (company_store)	Parity
Org chart (reports_to tree)	Yes	Yes (company_store)	Parity
Hierarchical goal alignment	Yes	Yes (company_goals)	Parity
Full task lifecycle (Kanban)	Yes	Yes (company_tasks)	Parity
Approval workflows	Yes	Yes (company_approvals)	Parity
Per-agent monthly budgets	Yes	Yes (company_budget)	Parity
Agent heartbeat system	Yes	Yes (company_heartbeat)	Parity
Encrypted secrets vault	Yes	Yes (company_secrets)	Parity
Company portability (export/import)	Yes	Yes (company_portability)	Parity
Recurring routines	Yes	Yes (company_routines)	Parity
BYOA adapter registry	Yes	Yes (adapter_registry)	Parity
Documents with revision history	Yes	Yes (company_documents)	Parity
Real-time dashboard	Yes	Yes (company_orchestrator)	Parity

VibeCody advantages: Rust memory safety, single binary, Tauri2 desktop app with 12 dedicated panels, SQLite session tree with replay, branch-per-task git worktree integration, /company REPL command suite (18 subcommand groups, 60+ leaves), local-first privacy, 22 AI providers.

Residual BYOA work: Claude/Codex/Cursor adapters currently covered via generic HttpAdapter; dedicated adapters will come when upstream tools publish stable agent APIs.

---

14. Deep-dive: Code Review & Architecture

Feature-by-feature match against Qodo Merge, CodeRabbit, Bito, Cursor, Copilot, Ellipsis (review) and Archi, Modelio, Gaphor, Diagrams.net, Cerbos (architecture).

Review matrix — VibeCody vs top AI review bots

Feature	Qodo Merge	CodeRabbit	Bito	Cursor	Copilot	VibeCody
Automated PR review bot	15+ workflows	Full	Yes	Basic	PR summaries	Full
Line-by-line findings	F1 64.3%	Yes + 1-click fix	Yes	Inline	Inline	Yes (severity + category + confidence)
OWASP Top 10 scanning	Yes	40+ linters	Partial	No	Basic	Yes (6 detectors)
Complexity analysis	Partial	Via linters	No	No	No	Yes (cyclomatic, deep nesting, long fns)
Duplication detection	No	Via linters	No	No	No	Yes (cross-file)
Test gap analysis	Coverage delta	Test gen	No	No	No	Yes (suggest_tests)
Breaking change detection	Multi-repo (10+)	Partial	Yes	No	No	Yes
PR summary + risk score	Auto-describe	Walkthroughs	No	No	Yes	Yes
Architectural diagrams from diff	No	Mermaid	No	No	No	Yes
NL quality gates	Live rules	YAML + NL	No	No	No	Yes
Learning loop (precision/recall)	Yes	Yes	Graph	No	No	Yes (ReviewLearning)
Multi-linter aggregation (FP filter)	OWASP-only	40+	No	No	No	Yes (8 linters)
Git platforms supported	3	4	3	1	2	5 (+Gitea)
On-prem / air-gapped	Enterprise	No	No	No	Enterprise	Free (Docker+Ollama)
SOC 2 compliance	Enterprise	Type II	No	No	Enterprise	Yes (controls)

Architecture matrix — VibeCody vs spec tools

Feature	Archi	Modelio	Gaphor	Diagrams.net	Cerbos	VibeCody
TOGAF ADM phases	ArchiMate only	Full	No	Templates	No	Full (9 phases)
Zachman Framework (6×6)	No	Partial	No	Templates	No	Full
C4 Model	No	No	Yes	Templates	No	Full (4 levels + Mermaid/PlantUML)
ADRs (Decision Records)	No	No	No	No	No	Full (CRUD + markdown export)
Governance engine	Basic	Scripts	No	No	Auth	Full (rule-based, violations, recs)
Policy-as-code (RBAC/ABAC)	No	No	No	No	Full	Full (Cerbos parity)
Text-based (code-first)	GUI	GUI	GUI	GUI	YAML	CLI + GUI
Export formats	ArchiMate XML	XMI, HTML	PNG/SVG	Multiple	JSON	JSON, MD, Mermaid, PlantUML
Air-gapped	Yes	Yes	Yes	Yes	Yes	Yes (+free)

Policy engine vs Cerbos / OPA / Cedar / Casbin

VibeCody matches all RBAC/ABAC/derived-roles/conditions/audit-trail features and uniquely adds:

• Conflict detection (overlapping rules with different effects).
• Coverage analysis (which resources/actions are covered).
• Unused-rule detection (via audit log replay).
• Starter-policy templates for any resource.

Enterprise readiness

Capability	Required for	VibeCody status
TOGAF ADM compliance	Enterprise IT, gov, banking	Full
Zachman Framework	Defense, healthcare	Full
C4 Model	Modern software architecture	Full
ADRs	All teams	Full
Policy-as-code	Finance, healthcare, gov	Full (Cerbos parity)
SOC 2 controls	SaaS enterprise	Full (compliance_controls.rs)
Air-gapped deployment	Gov, defense, banking	Full (Docker + Ollama)
Multi-provider AI	No vendor lock-in	Full (22 providers)

---

15. Remaining parity gaps (honest list)

Six items remain open from the v4–v12 cycles; none are engineering-blocked, and each is a conscious trade-off tracked in the roadmap. Eleven additional items from the v13 April-2026 trend survey are listed in §16.1.

1. Cursor’s proprietary Tab model — next-edit prediction quality. We ship FIM completions via Ollama + cloud models; Cursor trains their own Behavior-Informed Completion model. (Note 2026-04-26: Cursor 3.0 added an Agents Window and Cursor 3.2 added async subagents — the Tab model itself is unchanged but the surrounding multi-agent UX has widened the gap; tracked separately as v13 items A8/A9.)
2. Devin-level hours-long autonomy — Devin 2.2 (Apr 2026) added computer-use self-verification + Linux-desktop access; our agent loop still tops out at ~50 steps before compaction / re-plan, and we don’t run UI tests against our own output. Cognition’s 2025 acquisition of Windsurf consolidated three previously-tracked competitors (Devin + Windsurf + Cascade) into one entity — competitive positioning §1 should treat them as a single “Cognition family” going forward.
3. Claude Code’s MCP catalog — our MCP client is spec-compliant and we ship a directory, but the Anthropic-curated server catalog continues to grow and now includes the MCP Apps extension (interactive UI in conversations) + MCPB bundle format from the 2026 MCP roadmap. Tracked as v13 items A1–A4.
4. SWE-bench leaderboard position — Augment Code leads open agent systems at 72.0% pass@1 SWE-bench Verified (April 2026), Claude Opus 4.7 hits 87.6% Verified / 64.3% Pro, Claude Mythos Preview tops the provisional board at 93.9% Verified, GPT-5.3-Codex sits at 85.0%. Caveat (2026-05-03): OpenAI stopped reporting Verified scores after a contamination audit found 59.4% of hard tasks have flawed tests and all frontier models test as contaminated; SWE-bench Pro, SWE-rebench, and SWE-bench-Live are now the primary references. We track all four boards in the benchmark panel but haven’t entered the leaderboards ourselves.
5. Enterprise SSO / audit packaging — Cody Enterprise and Copilot for Business are further along on SOC 2 Type II, SAML SSO, central policy distribution. MCP enterprise readiness is now a first-class roadmap workstream for the protocol (audit/SSO/gateway as extensions, not core), opening a path for VibeCody to lead on open-source MCP enterprise tooling.
6. Polished BYOA adapters for Claude/Codex/Cursor — covered today by the generic HttpAdapter; dedicated adapters arrive when upstream APIs stabilise. JetBrains Junie CLI’s “1-click migration from Claude Code, Codex” (Mar 2026) is the bar for what users now expect; tracked as v13 item A17.

---

16. v13 — April 2026 trend delta + audit reconciliation

This iteration splits into two independent passes against the same fitgap. §16.1 is the external delta — a survey of what shipped in the AI-coding ecosystem between the v0.5.5 refresh (2026-04-17) and today (2026-04-26). §16.2 is the internal delta — a reclassification of previously-claimed gap closures against the audit at docs/audit/05-fitgap-overstatements.md. Both feed into Phase 53 of the roadmap.

16.1 External delta — what the industry shipped

Sources surveyed (web, 2026-04-26): Cursor changelog, Anthropic Claude Code changelog, GitHub Copilot blog, Cognition Devin blog, OpenAI Codex changelog, Google Antigravity changelog, Gemini CLI release notes, JetBrains Junie blog, MCP 2026 roadmap, ACP repo, A2A specification, SWE-bench leaderboards, sandbox provider coverage (E2B / Northflank / Cloudflare / Modal / Vercel / Docker).

Headline shifts in the eight days since v0.5.5:

• Cursor 3.0 (Apr 2) + 3.2 (Apr 24) — Agents Window, Design Mode (UI-element annotation in browser), Agent Tabs (side-by-side / grid view), async subagents, multi-root workspaces, multi-repo agent context.
• GitHub Copilot Cloud Agent (Apr 1) — formerly “Copilot coding agent” — no longer PR-only; can branch-only work; CLI sessions remote-controllable from GitHub.com or GitHub Mobile. Inline Agent Mode public preview in JetBrains IDEs (Apr 24). Claude Opus 4.7 GA on Copilot for Pro+/Business/Enterprise.
• Devin 2.2 — agent now has full Linux desktop, tests its work via computer use, self-verifies, auto-fixes. Cognition raising at $25B valuation; Windsurf folded into Cognition family.
• OpenAI Codex CLI (Apr 2026) — GPT-5.3-Codex-Spark lightweight model at 1000+ TPS, hooks GA, plugin marketplace browsing, multi-environment app-server sessions, Amazon Bedrock auth + AWS SigV4 signing as a first-class provider.
• Claude Code — /agents tabbed UI (Running / Library tabs with Run + View instance from Library), parallel MCP server reconnect, plugin-skill hot-reload, isolated-worktree subagent permission fix, YAML-list globs in skill paths, real-time skill progress display.
• Gemini CLI v0.38 — Subagents (delegating orchestrator pattern), Chapters (intent-grouped interactions), Context Compression Service, generalist agent task delegation.
• JetBrains Junie CLI (Beta, Mar 2026) — LLM-agnostic, runs in IDE / terminal / CI/CD / GitHub / GitLab; connects to running JetBrains IDE for full code intelligence; one-click migration from Claude Code + Codex configs.
• Antigravity 1.20.3 → 1.22.2 — AGENTS.md fallback (in addition to GEMINI.md), Linux sandboxing, MCP authentication improvements, conversation load-time improvements, Auto-continue default-on (deprecated as setting).
• MCP 2026 roadmap — stateless transport for horizontal scale, .well-known capability metadata, MCP Apps (interactive UI components in conversations), MCPB bundle distribution format, enterprise SSO/audit/gateway as extensions.
• ACP v0.11.0 (Mar 4) — Zed + JetBrains official partnership Oct 2025; Anthropic, OpenAI, GitHub, Google all ship implementations; Gemini CLI is the reference implementation. JSON-RPC 2.0 over stdio; >40% lower prompt response latency vs. ad-hoc bridges per OpenClaw measurements.
• Augment Code SWE-bench Verified 72.0% pass@1 — highest open-system score, no best-of-N tricks.
• Sandbox infrastructure mainstreamed — Cloudflare Sandboxes GA, Vercel/Ramp/Modal/Docker/E2B/Northflank/Together all shipped microVM AI-execution platforms in 2026; isolation tier (microVM > gVisor > containers) is now table-stakes for cloud agents.

Eleven new gaps surfaced by this delta (none yet implemented in VibeCody):

#	Gap	Surfaced by	Notes
A1	MCP Apps extension — interactive UI components in conversations	MCP 2026 roadmap	New extension; would render dashboards/forms/multi-step workflows directly inside the chat panel.
A2	MCPB bundle distribution format	MCP 2026 roadmap	Local-server packaging; analogous to VS Code .vsix for MCP.
A3	MCP .well-known capability discovery + stateless transport	MCP 2026 roadmap	Lets vibecli serve announce MCP endpoints without a live connection; required for horizontal scale.
A4	ACP server mode (Zed + JetBrains + Neovim editor protocol)	ACP v0.11	VibeCLI/VibeUI as ACP servers callable from Zed/JetBrains/Neovim — different from being an ACP client.
A5	Async subagents (long-running, check-back-later)	Cursor 3.2	Distinct from our current parallel-agent worktree pool, which assumes synchronous oversight.
A6	Multi-root workspace agent — agent that targets several working dirs per turn	Cursor 3.2, Codex CLI	Our --add-dir is read-only; this is write across roots in one agent invocation.
A7	Browser-native UI-element annotation Design Mode	Cursor 3.0	Existing design_mode.rs annotates static screenshots; this is live DOM annotation in a controlled browser.
A8	Self-verifying agent loop (UI/desktop tests against own output, auto-fix)	Devin 2.2	Closes the verification loop our visual_verify.rs opened — currently we screenshot-diff but don’t feed failures back into the agent.
A9	Cloud-agent remote-control protocol (start local, resume from web/mobile)	Copilot Cloud Agent	VibeMobile pairs with a host but doesn’t resume an in-flight CLI session the way Copilot’s new flow does.
A10	Skills hot-reload + real-time progress display	Claude Code	Our skill loader requires restart for new skills; no streaming progress UI.
A11	One-click migration from Claude Code / Codex configs	Junie CLI	Read existing CLAUDE.md, codex.toml, MCP server lists → emit VIBECLI.md + ~/.vibecli/config.toml. Lowers switching cost.

Six items from the same survey are already in flight or partially shipped and don’t count as new gaps:

• Bedrock auth for our Claude provider — provider.rs already accepts SigV4-signed bearer; needs explicit doc + vibecli config provider claude --aws UX.
• GPT-5.3-Codex-Spark-class fast inference — covered by our existing routing layer; needs the model added to useModelRegistry.ts once OpenAI exposes it via API.
• Generalist routing layer (Gemini CLI Chapters / generalist agent) — partially covered by our cost_router.rs (data-structure-only — see §16.2) and next_task.rs.
• AGENTS.md ↔ GEMINI.md fallback parser — our memory.rs already reads AGENTS.md / VIBECLI.md / CLAUDE.md; adding GEMINI.md is one-line.
• Plugin marketplace listing (Codex CLI) — our existing plugin_marketplace.rs covers this; needs remote browsing UX in VibeUI.
• Manager/Agents Window UI consolidation — our ManagerView.tsx covers parallel agents; we should explicitly stay distant from Cursor 3’s “Agents Window” and Antigravity’s “Manager Surface” layout choices on patent grounds (ties to the patent-distance posture in notes/PATENT_AUDIT_INLINE.md).

16.2 Internal delta — audit reconciliation

The audit at docs/audit/05-fitgap-overstatements.md catalogued modules previously claimed as “closed” that ship data structures + in-memory tests + a panel + a REPL command but lack the I/O layer the gap closure implied. Six of those have already been converted to real I/O (US-001 web grounding, US-002 A2A, US-003 worktree, US-004 MCP streamable, US-005 voice/whisper, US-006 proactive scanner). The remaining 8 modules + the RL-OS subsystem are reclassified as Partial in §3 and queued in roadmap Phase 53 for the same conversion treatment:

Module	Original gap	What’s missing	Conversion approach
issue_triage.rs	v7 Gap 10 — autonomous issue classification with GitHub/Linear integration	No HTTP calls to GitHub/Linear	octocrab + Linear SDK; gate behind VIBECLI_GITHUB_TOKEN / VIBECLI_LINEAR_TOKEN; mock-server BDD harness like US-001.
native_connectors.rs	v7 Gap 14 — connector trait + 20 service implementations + OAuth	Endpoint URL strings only; no reqwest, no async, no OAuth	Phase the 20 connectors; ship 4–5 first (Stripe, Slack, Linear, Notion, GitHub) with real OAuth + oauth2 crate; defer the remaining 15 to a later slice.
langgraph_bridge.rs	v7 Gap 19 — LangGraph-compatible REST API + checkpoint format interop	No HTTP/REST implementation	axum server exposing LangGraph’s documented routes; checkpoint JSON schema validation; LangGraph Python SDK conformance test.
mcts_repair.rs	v7 Gap 8 — MCTS with UCB1 + rollout via test execution	Has select/expand/backpropagate; rollout never runs actual tests	Wire rollout to cargo test / pytest / npm test per language; cap with per-rollout time budget; record outcome as the reward signal.
sketch_canvas.rs	v7 Gap 20 — wireframe → React/HTML/SwiftUI; 3D scene export	Basic shape data; no WebGL, no three.js, no 3D	Defer 3D entirely; ship the 2D wireframe → React JSX path against tldraw or an existing OSS recognizer; mark 3D as out of scope.
cost_router.rs	v7 — intelligent cost-aware request routing	Data structures only	Wire to provider.rs retry + circuit breaker; track per-(provider, model) latency/cost in agent_analytics.rs; routing decision becomes a real function of observed data.
semantic_index.rs	v7 Gap 5 — AST-level codebase understanding + call graph + type hierarchy	Line-by-line regex (trimmed.starts_with("pub fn")); no tree-sitter; no call graph	Replace regex with tree-sitter + per-language grammars (Rust, TS, Python, Go); reuse the index from vibe-core/src/index/symbol.rs which already uses tree-sitter.
linter_aggregator (in ai_code_review.rs)	cr/arch — 8 linters: clippy, eslint, pylint, …	simulate_linter() returns canned “Linter check passed” for every file	Spawn each linter as a subprocess; parse stdout; map findings to the existing Finding schema; the FP-filter LLM pass already exists.
rl_*.rs (8 files, ~31K lines)	RL-OS deep-dive — 30+ algorithms, JIT GPU/TPU kernels, Python bindings	No tch/candle/onnxruntime; “gradient sync” is Vec<f64> averaging; no GPU compute; no PyO3	Ship one algorithm end-to-end first (PPO with candle on CPU), then expose via PyO3; the 52 type-system entries become real once one training loop is real.

This is the same playbook that produced the US-001…US-006 conversions — design exists, tests exist, panel exists, REPL command exists; the conversion is purely “wire up the I/O layer + add a mock-server BDD harness”. Phase 53 in the roadmap groups these as US-007…US-015 for tracking parity with the prior conversions.

16.3 Updated remaining-parity-gaps list

Combining §15 (six long-horizon items, refreshed for v13) with §16.1 (eleven new external gaps), the honest open-gaps total is 17, not 6. The 14 partial items are tracked separately because they have shipped UX surface area — the work to complete them is well-scoped, not open-ended.

16.4 v14 — May 2026 weekly delta + missed-quarter items (added 2026-05-03)

This is a one-week refresh on top of v13, plus a small set of Q1-Q2 2026 items v13 missed. Sources surveyed (web, 2026-04-26 → 2026-05-03): cursor.com/changelog, GitHub Copilot blog, Anthropic Claude Code releases, OpenAI Codex / ChatGPT release notes, Cognition Devin docs, blog.modelcontextprotocol.io, a2a-protocol.org, Linux Foundation press, JetBrains Junie + Air blogs, Ollama releases, ggml-org/llama.cpp, vLLM releases, SWE-bench leaderboards, sandbox provider coverage (E2B / Daytona / Modal / Blaxel / SmolVM / Hyperlight), and OSS coding-agent repos (Cline / OpenHands / Aider / Continue).

Headline shifts in the seven days since v13 (most are also surfaced in Roadmap §1ter):

• MCP experimental-ext-skills (May 4) — skills discovery + distribution as MCP primitives. The single highest-leverage signal of the week for VibeCody — our 711 skill files could become MCP-discoverable across every host that speaks MCP, without per-host plugin work.
• Cursor Plugin Marketplace v2 (May 1) — plugins now bundle MCP servers + skills + subagents + rules + hooks; admin install policy (Default Off / On / Required); Team Marketplace decoupled from any specific repo.
• Cursor Security Review (Apr 30, beta) — always-on Security Reviewer + Vulnerability Scanner agents on Teams / Enterprise plans.
• VS 2026 + VS Code Integrated Cloud Agent (Apr 29) — “assign a task, close the IDE, get a PR” — Copilot Cloud Agent now controllable from inside the editor.
• OpenAI GPT-5.5 GA (Apr 23) — recommended Codex default (replaces 5.4); GPT-5 latency at higher intelligence; fewer tokens per Codex task; computer-use focus.
• Cursor SDK / @cursor/sdk (TypeScript) — same agent runtime / harness / models as desktop, CLI, and web exposed as a TS SDK; direct competitor to packages/agent-sdk/.
• llama.cpp NVFP4 (PR #22196 reposted Apr 21) — Blackwell-native FP4 path merged; MXFP4 progressing in ik_llama.cpp; b8196+ runs MXFP4 MoE on Blackwell tensor cores.
• Ollama 0.22.x (Apr–May) — /v1/messages (Anthropic Messages API compat — Claude Code can drive Ollama-hosted open models); ollama launch registers Claude Desktop / Cowork / Code; Gemma 4 thinking + tool calls; MLX runner gains logprobs + fused top-P/K + repeat-penalty-in-sampler.
• Chinese frontier wave (Apr) — DeepSeek V4-Flash $0.14 / $0.28 per 1M (~7.7× cheaper than Qwen 3.6-Plus on chatbot loads); Qwen 3.6-Plus + Qwen 3.6-35B-A3B (Apache 2.0); Kimi K2.6 long-horizon agentic; MiniMax M2.7; GLM-5.1.
• A2A v1.2 (Linux Foundation Agentic AI Foundation, Q1) — 150+ orgs in production; signed agent cards (cryptographic signatures for domain verification); GA across Google / Microsoft / AWS.
• ACP Registry live (Q1) — built into Zed + JetBrains; lists Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Gemini CLI. VibeCLI is not yet registered.
• DAPO mainstreamed (Q1) — OpenRLHF, verl, NeMo-RL all ship DAPO as default reasoning RL alongside PPO / GRPO; ByteDance paper open-sourced (50% fewer training steps for AIME-class tasks vs DeepSeek-R1-Zero-Qwen-32B).
• Sandbox cold-start floor (Q1) — Blaxel 25 ms; Daytona 27–90 ms (Docker); E2B Firecracker microVMs ~150 ms; Modal gVisor; SmolVM debuted 2026-04-17; Hyperlight Wasm 1–2 ms (still experimental, CNCF Sandbox).
• SWE-bench Verified contamination (Q1) — OpenAI stopped reporting Verified after audit found 59.4% of hard tasks have flawed tests; all frontier models contaminated. Reflected in §15.4 above; SWE-bench Pro / SWE-rebench / SWE-bench-Live are the new primary references.
• Google I/O 2026 (May 19, planned) — Gemini 4 + Android 17 + Agentic Coding Developer Preview; Gemini 3.1 Pro Preview already shipping ahead.

Six new gaps surfaced by this delta (B1–B6, all open; A1–A11 from v13 remain unchanged):

#	Gap	Surfaced by	Notes
B1	Skills as MCP primitives — discoverable & distributable across MCP hosts	MCP experimental-ext-skills (May 4)	Re-shape vibecli/vibecli-cli/skills/ to expose each skill via MCP list_skills / get_skill resources; one MCP server, every host benefits. Largest single-leverage item this cycle.
B2	Plugin bundle format with admin install policies	Cursor Plugin Marketplace v2 (May 1)	Define a VibeCody plugin manifest that bundles MCP servers + skills + subagents + rules + hooks; expose Default-Off / Default-On / Required tiers via WorkspaceStore policy + governance panel.
B3	Always-on agent classes (security review, vuln scan) running on every change	Cursor Security Review (Apr 30)	Convert /review from on-demand to a daemon-resident agent class triggered by file-watcher / pre-commit / CI; route findings to existing Finding schema.
B4	Cursor SDK parity audit	Cursor SDK (Apr)	Compare packages/agent-sdk/ to @cursor/sdk along: subagents, hooks, plugins, skills, sandbox tiers, recap/resume, multi-client (mobile/watch). Items where Cursor’s surface is wider become roadmap entries.
B5	NVFP4 (Blackwell native) as a TurboQuant target	llama.cpp PR #22196 (Apr 21)	Add NVFP4 Metal+CUDA kernels alongside MXFP4 + AWQ-Marlin; CubeCL/Burn ban scope unchanged.
B6	A2A signed agent-card façade	A2A v1.2 + LF (Q1)	Serve /.well-known/agent.json with P-256 ECDSA signature reusing watch-pairing’s key infrastructure (Secure Enclave-aligned); register as A2A server, not just client.

Five items already covered or trivially closeable (not new gaps):

• Ollama /v1/messages route — one route handler in vibecli/vibecli-cli/src/serve.rs; the existing Anthropic provider format already matches.
• GPT-5.5 / GPT-5.4 model entries — append to useModelRegistry.ts STATIC_MODELS.openai.
• Sonnet 4.8 entry (when Anthropic exposes it) — same one-file change in useModelRegistry.ts.
• Qwen 3.6 / DeepSeek V4 / Kimi K2.6 entries — append to the Ollama section of useModelRegistry.ts once GGUF / vLLM weights land.
• GEMINI.md fallback in memory.rs — already noted in v13 as one-line; remains pending.

Three positioning signals (informational, no roadmap action):

• Copilot training-default opt-in (Apr) — community backlash drove migration to Cline (58k stars), OpenHands (72k), Aider (27k). VibeCody’s “no training on user code” stance becomes a measurable sales axis; surface in marketing, not engineering.
• Doe v. GitHub Copilot (ongoing) — DMCA dismissed; license / contract claims still proceeding. Reinforces the privacy-first positioning above; informs /review’s open-source-license-scan UX (already shipped).
• JetBrains Air (Mar) — agentic IDE rebuilt on Fleet remnants; supports OpenAI Codex, Anthropic Claude Agent, Gemini CLI, Junie as native agents. Watch item for §1.2; not a direct VibeUI competitor today.

16.5 Updated remaining-parity-gaps list (v14)

Combining §15 (six long-horizon items), §16.1 (eleven v13 external gaps A1–A11), and §16.4 (six v14 external gaps B1–B6), the honest open-gaps total is 23, up from 17 in v13. The 14 partial items continue to be tracked separately. Phase 54 (queued in Roadmap §1ter) targets B1–B6 plus the trivially-closeable items above.

---

17. Headline positioning

VibeCody closes the 136 gaps that matter across 40+ competing AI coding tools — and is the only project that ships competitive entries in every category (terminal, IDE, cloud daemon, review bot, completions, mobile, watch) from a shared Rust + TypeScript monorepo.

See Competitive Landscape & Roadmap for the forward plan, surface-by-surface feature inventory, and differentiators.

18. Patent-distance posture for unshipped UX surfaces (B2 / B3 / A7) — 2026-05-09

Three v14 gaps — B2 (Plugin Marketplace v2), B3 (Always-on Security Review), A7 (Design Mode) — are queued behind a patent-distance audit before any UX work begins. The competitor surfaces (Cursor) are recent enough that filed-or-likely-to-file claims are a real risk; shipping in the same shape would re-introduce the exposure that motivated removing Paths A and C of the inline-completion stack on 2026-04-26.

The audit itself is internal triage, not legal advice, and lives in notes/PATENT_AUDIT_INLINE.md (gitignored). What’s public-facing is the posture — the principles each eventual implementation must honor:

1. No telemetry-driven personalization. Plugin / skill / agent discovery surfaces are query + category, never “for-you” lists fed by usage analytics. (Applies to B2.)
2. Policy enforcement is client-side and admin-authored in WorkspaceStore. No server can flip a workspace plugin or agent class from Off to Required. (B2, B3.)
3. Bundle format is open MCPB. The artifact format is the open MCP-spec bundle already shipped in PR #18 / A2 (mcpb_bundle.rs); lineage to VS Code .vsix and MetaPK keeps prior-art clear. No proprietary container. (B2.)
4. Trust roots are per-publisher P-256 ECDSA keys. Same key infrastructure as the A2A signed agent cards (PR #14 / B6) and watch pairing. The user explicitly trusts a publisher key; no opaque chain. (B2.)
5. Always-on is opt-in per workspace, default off. The trigger is a user-configured file-watcher rule; the daemon ships no system-imposed always-on default. (B3.)
6. Findings flow through the existing generic Finding schema alongside clippy / eslint / semgrep. The LLM is one finding source among many; the UI does not single it out as a privileged “security agent” class with an interactive canvas or one-click apply-fix gesture. To act on a finding the user invokes diffcomplete (⌘.) explicitly — same chord, same modal as Path D. (B3.)
7. No agent-controlled browser, no live DOM mutation by the agent. The user attaches their existing browser via WebDriver / CDP they have authorized; the agent emits a CSS / HTML unified diff into a DiffReviewPanel and the user applies via the existing diffcomplete mechanism. (A7.)
8. No closed-loop hidden iteration. Each refinement is an explicit user chord (the diffcomplete regenerate-with-refinement pattern from Phase 7), not an automatic screenshot-diff retry loop. (A7.)
9. No agent-side hidden state from the source tree beyond what the user explicitly selects or attaches. No automatic embedding RAG, no cross-file taint inference, no scratchpad in the prompt. (B3, A7.)

These principles are gating, not polish. B2 / B3 / A7 stay deferred until a design that satisfies all applicable principles ships as a proposal on this branch and clears the corresponding slice audit in notes/PATENT_AUDIT_INLINE.md.

The shape that does clear the audit, sketched (full breakdown in the gitignored doc):

• B2 — federated plugin index served over HTTPS / Git, MCPB bundles, per-publisher signature keys, query+category UI, policy stored in WorkspaceStore. plugin_marketplace.rs already targets this topology.
• B3 — opt-in workspace flag → file-watcher rule → LLM call → Finding records → existing ReviewPanel list → user invokes diffcomplete (⌘.) to act. Every component already exists; the work is wiring, not new claim-evocative surfaces.
• A7 — diffcomplete extended to the rendered DOM. User clicks an element in their own browser, types an instruction, presses ⌘.; agent emits a CSS / HTML unified diff in DiffReviewPanel; user reviews and applies. Same Path D claim-element posture, same prior-art lineage.

Cross-cutting takeaway: the patent-distance discipline isn’t a tax on shipping these gaps — it’s the only way they ship without re-creating the Path A / Path C exposure we deliberately deleted in April. Items that look like fast follows in competitor announcements often need a different shape on the VibeCody side; the audit is what produces that shape.