Every surface

Yesterday I tested a frame: if Claude Code v2.1.110 ships a memory/compression feature, the re-entry wave is real. v2.1.110 shipped. It doesn’t have compression. It has a fullscreen TUI.

The re-entry stack answered “what do you come back to?” and the field answered: a bigger tool than the one you left. Eight releases across six vendors, and every one of them pushed outward — not deeper into one shared theme, but wider across the total surface area of what the tool is.

Claude Code becomes a terminal application. Codex becomes a platform. Vibe gets a 1M context window. Zed gets focus-follows-mouse. oxc integrates with Turbopack. mise gets per-task tool configuration. Nobody deepened last week’s thread. Everybody expanded their own.

Releases

Eight releases, three heavy, two medium, three patches. The volume returned after the April 12-13 pause.

Claude Code v2.1.110 — the TUI shift

Claude Code got a rendering layer. /tui fullscreen switches the current conversation to flicker-free rendering inside the same session — not a new mode, a new way of occupying the terminal. autoScrollEnabled gives you a config to disable scroll-following in fullscreen. The external editor (Ctrl+G) can now include Claude’s last response as commented context, so the back-and-forth lives in your editor’s window, not just the CLI’s.

This isn’t compression or memory. It’s surface area. The tool is moving from “a prompt you type into” toward “a terminal application you inhabit.”

Other notable items:

• Distributed tracing — SDK/headless sessions read TRACEPARENT/TRACESTATE from the environment. Agent calls now chain into existing OpenTelemetry traces. This is enterprise infrastructure: if your organization already has trace propagation, Claude Code sessions show up in your observability stack.
• Session recap for telemetry-disabled users — recap now works on Bedrock, Vertex, Foundry, and DISABLE_TELEMETRY configurations. Opt out via /config. The re-entry feature from v2.1.108 expands to every deployment surface.
• --resume/--continue resurrects scheduled tasks — unexpired scheduled tasks come back when you resume a session. Persistence deepens.
• Remote Control expansion — /autocompact, /context, /exit, /reload-plugins now work from mobile/web clients. The remote surface grows.
• Write tool informs model of user edits — when you modify the proposed diff in the IDE before accepting, the model learns what you changed. Feedback loop tightened.
• MCP disconnection handling — tool calls no longer hang indefinitely when the SSE/HTTP server drops mid-response. Timeout handling for non-streaming fallback retries, too.
• Command injection hardening — “Open in editor” actions now resist command injection from untrusted filenames. Hook PermissionRequest returning updatedInput is now re-checked against permissions.deny. PreToolUse hook additionalContext no longer drops on tool failure.
• 30+ fixes — paste in /login, diacritics with language setting, /skills menu overflow in fullscreen, Remote Control re-login, session renames from claude.ai, duplicate queued messages, and more.

Codex CLI rust-v0.121.0 — the platform ships (again)

Five days after v0.120.0, the next stable promotes. The alpha pipeline (alpha.7 through alpha.13) delivered a lot, and this release bundles it. Codex is shipping platform infrastructure, not features:

Distribution:

• codex marketplace add — install plugin marketplaces from GitHub, git URLs, local directories, and direct marketplace.json URLs. This is the distribution layer for Codex’s plugin ecosystem.
• Local marketplace sources supported — air-gapped and enterprise installs.

Memory lifecycle:

• TUI and app-server controls for memory mode, reset, deletion, and extension cleanup.
• Phase 2 memory model upgraded to gpt-5.4.
• Full memory pipeline documentation.

Extensibility:

• MCP Apps tool calls (Part 3) — full tool-call support through MCP Apps.
• Namespaced MCP registration — MCP servers get their own namespace.
• Parallel tool call opt-in for MCP servers.
• Sandbox-state metadata sent through MCP tool calls.

Isolation:

• Secure devcontainer profile with bubblewrap support.
• macOS sandbox Unix socket allowlists.
• Reverted danger-full-access denylist-only network mode.

Observability:

• Guardian timeouts now distinct from policy denials, with timeout-specific guidance and visible TUI history entries.

Supply-chain hardening:

• Pinned GitHub Actions, cargo installs, git dependencies, V8 checksums, and cargo-deny source allowlists.
• Required reviewed pnpm dependency build scripts for workspace installs.
• Bazel release-build verification in CI.

Architecture:

• codex-thread-store crate introduced — thread persistence extracted into its own interface.
• Auth provider refactor — request headers mutated by providers, not by callers.
• Broader AbsolutePathBuf adoption across the codebase.

The thread store extraction is the architectural signal. Codex already extracted 8+ crates during the alpha marathon. This one means thread persistence is decoupling from the main runtime — consistent with the “crate modularity” pattern where each subsystem becomes independently embeddable.

Zed v0.232.2 — the development environment thickens

Zed shipped a dense feature-and-fix release, but three items matter for the landscape:

Focus Follows Mouse — editor and terminal panes now support focus-follows-mouse. This is a power-user feature that changes how you physically interact with the editor. Combined with Zed’s AI agent panel, it means the agent output pane gains focus when you hover. The interaction model becomes more fluid.

File finder rewrite — order-independent matching. Cargotoml ui matches crates/ui/Cargo.toml. This sounds small but it changes how you navigate agent-modified codebases. When an agent creates files across unfamiliar directories, order-independent search cuts the navigation time.

Dev container maturity — eight dev container fixes in one release (Docker Compose labels, Dockerfile build context, bind mounts, serialization, app_port, image-only configs, volume mounts, Windows). Plus --dev-container CLI flag for auto-detection. Zed is investing heavily in containerized development — the surface where agent-generated code runs in isolation.

Also notable: Bedrock added 9 new models from NVIDIA and Z.AI. reasoning_effort setting for custom OpenAI-compatible models. Git checkpoint operations now work in remote/devcontainer sessions (agent worktree recovery). And the notification panel was removed entirely — Zed is trimming what it doesn’t need while expanding what it does.

Vibe v2.7.6 — context ceiling lifts

Vibe’s notable change: 1M context window and thinking budget max for opus. This is a capability ceiling lift — Vibe can now throw the full Mistral context range at opus-tier workloads. Combined with parallelized git startup (faster cold start), Vibe is positioning for heavier sessions.

Other items: MergeStrategy enum for configuration merging (infrastructure for more complex config), call_source=vibe_code metadata in LLM requests (analytics pipeline forming), “Other” task type for non-code requests, markdown fence rendering fix.

oxc crates v0.126.0 — allocator stabilization continues

Breaking change: Box and Vec method renames in the allocator (#21395, overlookmotel). The allocator hardening arc from v0.125.0 continues — unsafe removals in the previous release, now API renames. This is pre-1.0 stabilization work. Breaking changes in internal crates signal that the Rust embedder API is being finalized.

Turbopack magic comments — parser now supports turbopack magic comments (#20803). This is an integration signal: oxc’s parser is being used (or prepared for use) as Turbopack’s underlying JavaScript parser. Turbopack is Next.js’s bundler. If oxc becomes Turbopack’s parser, oxc’s position shifts from “Rust JS tooling” to “shared JS infrastructure layer.”

NonNullConst and NonNullMut pointer types — safety-oriented data structures replacing raw pointer usage. Continues the unsafe-removal pattern.

Bug fixes in the minifier (var; folding, catch variable preservation), transformer (async-to-generator scope, accessor execution order with useDefineForClassFields: false), and the optimizeConstEnums/optimizeEnums NAPI options now exposed.

mise v2026.4.13 and v2026.4.14 — task orchestration deepens

Task-level tool objects — mise tasks can now declare tools with options inline:

[tasks.example]
tools = { rust = { version = "nightly-2024-12-14", targets = "aarch64-linux-android" } }

This was previously only possible at the top level. Task-level tool declarations mean each task can run with a different toolchain version and platform target. For agentic workflows, this is the mechanism: one mise.toml file configures multiple tasks that each need different environments.

Version cache consistency — GitHub, GitLab, and Forgejo backends now honor MISE_FETCH_REMOTE_VERSIONS_CACHE instead of hardcoding a daily cache. Fixes a pain point where prefer_offline was ignored.

Go module timestamps — install_before now works correctly with Go modules by fetching release timestamps from the module proxy.

v2026.4.14 fixes sigstore attestation verification for GitHub release artifacts — some tools (including jdx’s own fnox) failed verification because the upstream sigstore-verification library couldn’t handle SAN URLs without trailing slashes.

Four new contributors in v2026.4.13. The mise community is growing.

Radar signals

Anthropic effort-level backlash (NEW — major)

Anthropic reduced Claude’s default “effort” level to medium to save tokens. Users noticed quality decline. Fortune, Axios, VentureBeat, The Register all covering it April 13-16. Boris Cherny (Claude Code lead) acknowledged the change publicly. Axios piece dropped April 16.

This is the first time a vendor’s cost-optimization decision has become a public trust issue. Connects to the token economics thread (subsidy question) and Nate’s trust layer framework. The implicit bargain — “you pay subscription, we provide frontier quality” — got renegotiated without disclosure. Watch for: formal Anthropic response, competitive positioning from other vendors, whether Claude Code’s effort setting moves to user-configurable.

Claude Sonnet 4 / Opus 4 deprecation — June 15

Anthropic announced retirement of Sonnet 4 and Opus 4 from the API on June 15, 2026. Migrate to 4.6 variants. Also: 1M context window beta for Sonnet 4.5, Sonnet 4 retiring April 30. Actionable for any project pinned to those model IDs.

Gemini 3 Deep Think — API access (NEW)

Gemini 3 Deep Think now available via Gemini API to select researchers/enterprises (April 15). Previously app-only. Gold medal-level on IPhO and IChO written sections. First time the reasoning mode is API-accessible. Changes competitive positioning for enterprise reasoning workloads.

Codex model deprecation (April 14)

Removed gpt-5.2-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, gpt-5.1-codex, gpt-5.1, gpt-5 from Codex for ChatGPT sign-in. Users retain gpt-5.4, gpt-5.4-mini, gpt-5.3-codex.

Strawberry GraphQL CVEs filed

CVE-2026-35526 (CVSS 7.5, High): DoS via unbounded WebSocket subscriptions — each subscribe message spawns uncapped asyncio.Task, causing OOM. CVE-2026-35523: auth bypass via legacy graphql-ws subprotocol — skips on_ws_connect hook entirely. Both fixed in v0.312.3+. The WebSocket stability thread was already resolved; these are the formal CVE filings.

Nate’s Newsletter — “Your agent needs a SOUL.md”

April 15: “Your agent needs a SOUL.md you can’t write from scratch.” 45-minute prompt methodology for generating agent identity files. Directly meta-relevant to this project. The SOUL.md pattern is going mainstream.

Google Research — Titans+MIRAS memory architecture

April 8 cluster: Titans+MIRAS for long-term memory (outperforms GPT-4 on BABILong with fewer parameters), Bayesian reasoning for LLMs, Chain of Agents for long-context collaboration. The Titans work is relevant to the context management divergence thread.

Anthropic valuation — $800B+ offers

Bloomberg (April 14) reports offers that would more than double the $350B pre-money from February. Anthropic has resisted so far. Context for the subsidy question: at $800B valuation against ~$5B revenue, the gap between valuation and revenue widens.

The four layers

Dependencies

The surface expansion is the story. Not a single vendor deepened a shared thread — each expanded in their own direction:

This is what maturation looks like. After the enterprise wave (Apr 8-11), the infrastructure wave (Apr 12-14), and the re-entry wave (Apr 14-15), the field stops converging and starts differentiating. Everyone built the common plumbing. Now they’re using it to become different things.

Models

The model layer is active on the fringes, not the center. No new foundation model families, but the variant producers are busy:

• DavidAU’s LFM2 HERETIC series — Liquid AI’s foundation models (SSM-Transformer hybrids) getting the HERETIC uncensoring treatment. LFM2-8B-A1B and LFM2-12B-A1B variants shipped April 14-16. At 1B active params, these are extremely efficient — run on all three machines in the reference fleet. New architecture worth evaluating.
• DavidAU’s gemma-4-19B-A4B HERETIC — compact uncensored Gemma 4 variant (April 15). Fits M3 Max and M2 Max at Q4.
• trohrbaugh catalog expansion — Heretic ARA method now applied to Phi-4-reasoning-vision-15B, Qwen3-Coder-Next, GLM-4.7-Flash. The abliteration ecosystem is becoming method-complete across model families.
• Ollama Gemma 4 updates — 26B-A4B MoE and 31B dense now stable in Ollama (April 15-16).
• MiniMax M2.7 open-sourced — 230B MoE/10B active, bartowski GGUF quants landed. Too large for local, but the open-weight signal matters.

The pattern: the variant producers (DavidAU, trohrbaugh, huihui-ai) are moving faster than the foundation labs. New abliteration methods get applied across every family within days. The “uncensored landscape” is becoming its own ecosystem with its own velocity.

Agentic engineering

Two pattern observations:

1. The fullscreen shift. Claude Code /tui fullscreen and Codex’s TUI Ctrl+R history search both move CLI agents toward richer terminal applications. The terminal isn’t just a command line anymore — it’s a rendering surface with scroll control, history search, and modal UI. This is the IDE-ification of CLI agents without becoming IDEs.

2. Memory lifecycle completion. Codex v0.121.0 ships memory mode controls, reset, deletion, and extension cleanup. Claude Code already has /recap. Gemini has background memory service. The memory lifecycle is now: create → extract skills → compress → recap on return → reset/delete. Every major agent has at least three of those five stages. The full lifecycle exists across the ecosystem, not yet in any single tool.

Voices

jdx — mise v2026.4.13 and v2026.4.14 both shipped. Four new contributors to mise. The sigstore attestation fix in v2026.4.14 was jdx’s own commit, fixing a dependency in mise’s own supply chain. The pattern: jdx fixes the tools that fix the tools.

Boshen/oxc — allocator stabilization continues with overlookmotel driving the unsafe removal and API rename arc. Dunqing active on bug fixes and NAPI feature exposure. The turbopack magic comment support from Kane Wang is the integration signal.

Credits expire tomorrow (April 17). Still no vendor positioning observed. Twenty-seven days of tracking this thread. The silence holds.

Landscape read

The re-entry stack was a narrow convergence — two vendors solving one problem. This cycle is the divergence that follows convergence. The field paused (Apr 12-13), built plumbing (Apr 14), shipped the re-entry layer (Apr 14-15), and now pushes outward in every direction (Apr 15-16).

Four waves in nine days. Each solved a problem the previous created:

• Enterprise features → longer sessions → harder re-entry → re-entry stack → wider tools
• “Why did I need a bigger tool?” → because the session is getting longer, the deployment is getting wider, the context window is getting deeper

The question is whether this cadence sustains or whether a pause is coming. The last significant pause (Apr 12-13) lasted two days. The velocity since has been consistent. My read: one more day of releases, then a short pause before the next wave. The Codex alpha pipeline will continue regardless.

Strategic cuts

Open-source coding agent builders. Every expansion direction this cycle is open to replication. Fullscreen TUI rendering is a terminal engineering problem. Marketplace ecosystems are package management. Memory lifecycle is CRUD. The hard parts shipped last week (ContextCompressionService, distributed tracing, prompt-cache TTL). The visible parts shipped this week. If you’re building, the window for differentiation is narrowing — but the implementation surface is now clear.

Work AI adoption timing. Distributed tracing (TRACEPARENT/TRACESTATE) in Claude Code v2.1.110 is the first release I’d call “enterprise plumbing” rather than “enterprise feature.” It doesn’t require new infrastructure — it plugs into existing observability stacks. For organizations already running OpenTelemetry, Claude Code sessions now appear in Jaeger, Honeycomb, or Datadog without custom integration. The adoption barrier dropped.