# AgentStack Daily EP067 — Codex rust-v0.138, Claude Code .169, MCP Spec RC, Apple WWDC, and Qwen3.7 Flagship Drop

[NOVA]: I'm NOVA.

[ALLOY]: I'm ALLOY, and this is AgentStack Daily...

[NOVA]: Today, two fresh agent-stack CLI releases drop within 24 hours of each other. OpenAI Codex jumps to zero point one thirty eight with a CLI-to-desktop handoff mechanism that eliminates manual context transfer, and Claude Code CLI ticks to point one sixty nine on npm, continuing the rapid iteration cadence that has defined the two point one series since EP064. The Model Context Protocol working group published the July 2026 Release Candidate, pivoting the spec to a stateless architecture with a first-class extensions framework. Apple WWDC 2026 delivered a working Siri AI powered by Google Gemini, a Shortcuts app that assembles automations from natural-language descriptions, and a one-tap compromised-password updater driven by AI browser automation. And Alibaba's Qwen3.7 flagship line breaks cover with Qwen3.7-Max, a 1M-token reasoning model that ran 1,000 tool calls and improved a kernel by 10x entirely on its own over hours, alongside Qwen3.7-Plus, the multimodal agent variant with unified vision and language reasoning. Let's break it all down. [PAUSE]

## [00:00] Opening: the harness cycle, the model lane, and why this week matters

[ALLOY]: The cycle puts the harness layer first because that is where agent reliability gets built or broken. A CLI-to-desktop handoff that removes manual context transfer, a patch-level stability release on npm, and a stateless protocol redesign that removes the sticky-session constraint are the operational changes that make the agent stack production-grade. The model lane is dense: two Qwen3.7 flagship models, a real Siri AI with a Gemini partnership, and a Shortcuts app that non-technical users can now author directly. The project radar closes the loop with three MCP servers that cover enterprise data, protocol governance, and the open-weight local model layer.

[NOVA]: Three layers are moving at once, and the most useful way to keep them separate is to lead with the harness layer because that is where operators feel the work. The agent-harness releases this cycle touched every primary surface that operators actually touch: the CLI session, the desktop client, the protocol transport, and the npm distribution track. The model layer is what determines the quality of what the agent produces, and the model lane is heavy. The tooling layer is the project radar that fills in the gaps between the major releases.

[ALLOY]: The OpenAI Codex release is the headline this cycle, and it lands with three user-facing features that directly change how builders work with the CLI. The CLI-to-desktop handoff is the one that operators have been asking for since the desktop client shipped. The image path exposure fixes a real breakage that has been in the tool since image generation shipped. And the reasoning effort fixes address a keyboard-layout problem that blocked non-US developers from using the feature correctly. The Claude Code CLI release is a hygiene patch, but the gap between the latest track and the stable track is now sixteen releases, which is a meaningful signal for fleet operators. The MCP RC is the architectural story of the cycle, and it affects every operator running MCP servers behind a load balancer. [PAUSE]

## [02:00] OpenAI Codex rust-v0.138.0: CLI-to-desktop handoff, image path exposure, and reasoning effort flexibility

[NOVA]: OpenAI shipped Codex rust-v0.138.0 on June 8, 2026 as a stable release, available from the GitHub releases page under the rust-v0.138.0 tag. This is the first stable release since rust-v0.137.0 appeared in EP065, and it lands with three user-facing features that directly change how builders work with Codex. The headline addition is the slash app command and its companion desktop handoff mechanism. On macOS and native Windows, running slash app in a Codex CLI session transfers the current thread context into Codex Desktop, authenticated by the existing session token with no re-prompt. Windows workspace launches can now open directly into Desktop instead of stopping at a manual prompt, which means a builder can start a task in a terminal, realize they need a GUI browser or multi-window context, and hand off without losing state. The transport uses a named-pipe bus over the local workstation. The desktop client picks up the session token from the CLI's auth store, which means the handoff is authenticated by the existing session rather than requiring a fresh login.

[ALLOY]: The second feature is local image path exposure. When Codex generates an image or attaches a local image to the conversation, the model now receives the saved file path as a first-class reference rather than an ephemeral URL. Previously, follow-up edits required the model to track a URL that could expire or become inaccessible. Now the model holds the exact file reference and can reference it in subsequent tool calls reliably. This fixes the image-follow-up-edit workflow that broke in every release since image generation shipped. The technical mechanism is straightforward: the local file URI is surfaced at the materialize boundary before the model processes the tool result, so Codex receives the path directly rather than holding a separate image URL. The practical impact is that image generation pipelines no longer need URL-tracking logic. The model holds the file path across turns.

[NOVA]: The third feature is reasoning effort flexibility. The TUI now has fallback shortcuts for terminals that miss Alt key bindings, which is a real problem on non-US keyboard layouts where the Alt modifier does not behave as the TUI expects. The model-defined effort levels now flow through in the order the model advertises them rather than being re-ranked by the client, which makes effort-level selection deterministic across Codex deployments. Previously, a developer on a German or French keyboard layout had no reliable way to select reasoning effort because the Alt-based shortcuts did not fire. The fallback shortcuts bypass the Alt binding requirement entirely.

[ALLOY]: Two bug fixes address real operational pain. Clipboard race conditions in high-latency environments were a problem when Codex runs over remote SSH or high-network-jitter connections, where the clipboard write and the model read could happen in the wrong order and corrupt the context. Intermittent authentication failures during long-running sessions were a session token expiry edge case that triggered re-auth prompts mid-task. Neither fix is dramatic on its own, but both were user-reported in the OpenAI developer forum, and both are exactly the class of failure that erodes trust in agent tooling. Operators who run Codex over remote connections or in long-running background sessions will notice the difference immediately.

[NOVA]: The surfaces that changed are the slash app command in the CLI, image path handling in the model interface layer, and reasoning effort fallback shortcuts in the TUI. There are no new config flags or environment variables. Migration risk is low. This is an additive release with backward-compatible defaults. What agents can now do that was previously impossible or brittle: hand off a CLI session to a desktop client mid-task without losing context; reliably reference a generated image file in follow-up edits across turns; select reasoning effort levels on non-US keyboard layouts without workarounds. The named-pipe transport over the local workstation bus is the architecture that makes the handoff reliable. It is not a cloud sync or a file-based state transfer. It is a direct pipe from the CLI process to the desktop client, secured by the existing session token, which means the context transfer is instantaneous and authenticated. [PAUSE]

## [06:30] Claude Code CLI: npm latest tick, stable track gap, and the iteration signal

[NOVA]: Anthropic's Claude Code CLI published version two point one point one sixty nine to npm on June 8, 2026 at 18:11:20 UTC, becoming the latest entry in the latest dist-tag track. The stable dist-tag remains at two point one point one fifty three, leaving a gap of sixteen patch versions between the two tracks. For operators, this means: if you are on the latest tag, you are already on two point one point one sixty nine if you have updated since June 8. If you are on stable, you are on two point one point one fifty three, and the gap is intentional. The stable track is curated for regressions. The latest track is where the rapid iteration lands.

[ALLOY]: The delta from the prior release to this one is a single-patch publish within 24 hours, suggesting a hotfix for a narrow regression. The npm metadata does not advertise new public API surfaces or config flags for this release. The rapid iteration cadence, five releases between EP064 and now, reflects the internal stability-improvement and tool-call reliability focus that has characterized the two point one series. No new CLI flags. No new config files. Migration risk is minimal. This is a patch-level update in a well-established release track.

[NOVA]: The operational signal here is the dist-tag gap. Six episodes in a row with the stable track stuck at two point one point one fifty three while the latest track races through point one sixty nine tells you that the stable track is on a deliberate slow cadence. That is by design. Stable is curated. Latest is where the work lands. For production deployments where predictability matters more than freshness, pin to stable explicitly with npm install -g at-sign anthropic-ai slash claude-code at stable. For active development where you want the latest fixes and tool-call improvements, stay on latest and update regularly. The gap between the two tracks is not a bug. It is the feature. It is the mechanism that lets operators choose their risk posture.

[ALLOY]: What agents can now do: this patch likely addresses a specific tool-call edge case reported since the prior release shipped. Operators on latest should update and monitor for any session-binding anomalies. If none appear, the release is a clean stability patch. The iteration cadence tells you that the Claude Code CLI team is working through a backlog of edge cases in the tool-call interface, which is exactly where background agent work goes wrong silently. Session attachment that fails cleanly. Stream-json event ordering that handles the model output correctly. An interrupt handler that accepts input without hanging. These are the failures that make long-running agents lose work, and the rapid iteration cadence is the signal that the team is closing them systematically. [PAUSE]

## [09:00] MCP July 2026 Release Candidate: stateless protocol, server-initiated requests, and the extensions framework

[NOVA]: The Model Context Protocol working group published the July 28, 2026 Release Candidate on June 8, 2026, marking the most significant architectural change in MCP's history. The core shift is from a stateful session model to a stateless transport with optional stateful applications layered on top. In the previous model, MCP clients maintained a session store on the server side. That was a sticky-session constraint that made horizontal scaling impossible without session affinity at the load balancer. Every request from a client had to route back to the same server that held its session state, which meant you could not put an MCP server fleet behind a standard load balancer and scale out. The RC removes this by moving all session context into the request payload itself. Any MCP client can now route to any server replica with no session store required. That is a fundamental change in how the protocol operates in production, and it is the change that every operator running MCP servers behind a load balancer has been waiting for.

[ALLOY]: Server-to-client requests are restructured. Previously, a server could push data to the client only through the sampling callback, which was a narrow, specialized path designed for one specific use case. The RC adds a general server-request primitive to the JSON-RPC layer, enabling servers to push prompts, resource updates, or tool results without a client poll. This is architecturally similar to how Server-Sent Events work in HTTP. A server-initiated stream with client-side subscription. The server does not wait for the client to ask. It sends the update when it has something to send. For agents that rely on real-time server state, this removes the polling loop that added latency and overhead.

[NOVA]: The Extensions framework is now first-class. New capabilities ship as opt-in extensions with their own versioned namespace, stabilize in extension form, and only move into the core spec if and when they earn broad adoption. This is a governance change as much as a technical change. It means the MCP spec is no longer the only vehicle for new functionality. Extensions can ship independently of the core spec cycle, which gives the ecosystem a faster path to deliver new capabilities. The MCP Apps extension, which is SEP-1865, lets servers ship interactive HTML interfaces rendered in sandboxed iframes with a controlled postMessage bridge. This is a server-rendered UI pattern that could replace custom prompt-injection user interfaces in agent tooling. Servers can author an HTML interface, ship it to the client, and the client renders it in a sandboxed iframe. The agent can interact with that interface without the client needing to understand the interface's logic. The Tasks extension graduates from experimental to a named extension, which means it has a stable interface and a defined behavior contract.

[ALLOY]: Three previously experimental features are deprecated under the new feature lifecycle policy, which is SEP-2577. Full JSON Schema 2020-12 validation for tools is now required rather than recommended, which means MCP tool implementations need to produce schema-compliant output or the client will reject them. The spec is stricter now, which is the right direction for a protocol that is being adopted in production environments where malformed tool responses can break agent sessions.

[NOVA]: The operator surfaces that change are the MCP server implementations. Server implementors need to audit their code for server-side session state and migrate it to the client or an external store. The stateless model means the server no longer holds per-client state between requests. If your server has any in-memory state that represents a client session, that state must move to the client side, encoded in the request payload, or to an external store like Redis. MCP client hosts should update to handle the new server-request primitive, which means handling server-initiated pushes that arrive without a client poll. The Extensions framework means new capabilities will ship as opt-in additions rather than spec changes. Watch extension versions as leading indicators of new MCP functionality. The extension namespace is versioned independently of the core spec, which means you can pin to a specific extension version without upgrading the core protocol.

[ALLOY]: What agents can now do that was previously impossible: scale MCP server infrastructure horizontally without sticky sessions, which means a standard load-balanced deployment with multiple server replicas is now a supported topology rather than a workaround. Receive server-initiated pushes without polling, which means agents can respond to server state changes in real time rather than waiting for the next polling interval. Use server-rendered UI surfaces as interactive components within agent sessions, which opens a new class of human-in-the-loop workflows where the server provides a structured interface that the agent navigates. [PAUSE]

## [13:00] Apple WWDC 2026: Siri AI, Gemini partnership, and natural-language Shortcuts

[NOVA]: Apple's WWDC 2026 keynote delivered the most substantive AI update in the company's history. The centerpiece is Siri AI, a full natural-language agent deeply integrated into iOS 27, iPadOS 27, and macOS 27, powered by a Google Gemini partnership for cloud inference with on-device Apple Intelligence running on the A19 Pro and M4 chip families. The partnership is architecturally a two-tier inference stack. Lightweight, privacy-sensitive tasks run on-device on the Apple Silicon Neural Engine. Complex reasoning requests route to Google's Gemini API. The routing logic is built into the Apple Intelligence stack, which means developers and users do not need to decide which model handles which request. The system makes that determination based on task complexity, privacy sensitivity, and available compute.

[ALLOY]: The new Shortcuts app accepts natural-language workflow descriptions. A user types notify my partner when I leave work and give them my ETA, and the system decomposes this into a sequence of system and app actions, pulling a stored address from contacts, calculating travel time via Apple Maps, sending a message via Messages, then assembling and saving the resulting automation graph. This is the same underlying primitive set that powers the existing Shortcuts automation layer, now accessible via natural language rather than a drag-and-drop UI. The planner takes a high-level intent and produces a structured automation that the user can review, modify, and save. Non-technical users who could never assemble a Shortcuts workflow by hand can now author complex automations by describing what they want. This is the shift that matters most for the agent tooling ecosystem. It is the first time Apple has exposed its automation primitives through a natural-language interface, and it is a preview of how the broader consumer market will interact with agentic workflows.

[NOVA]: Safari gains AI tab management and a page-change monitor that alerts when a tracked page updates. This is useful for monitoring prices, news stories, or any time-sensitive web content where manual refresh is the current workaround. The one-tap compromised password updater uses AI to access the login flow through Safari automation, handling multi-step credential change processes without manual user input. When Apple detects that a stored password appears in a breach database, it can navigate the change-password flow on the user's behalf, enter the current password, navigate to the change-password field, generate a new secure password, submit the form, and update the keychain. Messages surfaces photos by text description, and Calendar accepts natural-language event creation with people and time context pulled from email during active calls. Cross-app context awareness lets Siri surface relevant information like flight details from email during an active phone call.

[ALLOY]: The operator surfaces that change are the Shortcuts natural-language planner, the Gemini partnership architecture, and the one-tap password updater. The Shortcuts natural-language planner is a new surface for workflow automation that non-technical users can now author directly. If you build tools that target the Apple ecosystem, this is a new authoring surface to consider. The Gemini partnership means iOS now has a two-tier inference stack, architecturally similar to how Claude Code routes between local and remote models. The one-tap password updater is likely to be among the first third-party-capable AI actions as Apple opens the Shortcuts API, which means the password-updater pattern may become a standard AI action that third-party developers can invoke.

[NOVA]: What Apple shipped at WWDC 2026 is a working Siri AI, not a conceptual demo or a research preview. The Gemini backing is live. The natural-language Shortcuts planner assembles real automations. The one-tap password updater navigates real login flows. For the agent tooling ecosystem, the WWDC story is a signal that the consumer AI landscape is catching up to what developers have been building with agent frameworks. The natural-language workflow authoring paradigm that Apple is now shipping to a billion devices is the same paradigm that agent tooling builders have been working toward. The difference is that Apple can ship it at a scale that changes user expectations overnight. [PAUSE]

## [17:00] Qwen3.7-Max: 1,000-step autonomous demo, 1M-token context, and the Anthropic-compatible API

[NOVA]: Alibaba released Qwen3.7-Max on May 21, 2026 as the proprietary text-only reasoning flagship of the Qwen 3.7 generation, available API-only via Alibaba Cloud Model Studio. The model was designed specifically for long-running agent workloads with upgraded reasoning and coding capabilities. In an internal test, Qwen3.7-Max autonomously performed more than 1,000 tool calls and iterative code modifications to optimize a SGLang Extend Attention kernel on a T-Head ZW-M890 PPU, which is a hardware architecture not in the model's training data. The optimization process improved inference speed by approximately 10x over the baseline across a multi-hour run. The model wrote code, executed it, measured the result, revised the implementation, and repeated that cycle over more than 1,000 steps. That is not a single-shot benchmark. That is a sustained multi-hour agentic task, and the fact that the model maintained coherent reasoning across 1,000 steps on hardware it had never seen during training is a meaningful data point for evaluating long-horizon agent capability.

[ALLOY]: Qwen3.7-Max features a 1M-token context window and exposes an Anthropic-compatible API endpoint, meaning it works with the same tool-call interface that Claude Code uses. Integration requires a three-line config change pointing to Alibaba Cloud's Anthropic-compatible API endpoint, setting the model name to qwen3.7-max, and updating the base URL. No proprietary SDK required. The model is not open-sourced. The Max tier has never been open-sourced across any Qwen generation. This is an API-only proprietary release, which means operators who want to evaluate it need an Alibaba Cloud account and the Model Studio credentials.

[NOVA]: The Anthropic-compatible API is the integration angle that matters for the agent tooling ecosystem. It means Qwen3.7-Max slots into Claude Code, OpenClaw, and any MCP-compatible agent without a migration. You point the base URL to Alibaba Cloud, you set the model name, and the agent tooling talks to Qwen3.7-Max the same way it talks to Claude Opus 4.7 or 4.8. For operators running long-horizon agent workloads where the 1M-token context window and the sustained reasoning capability are the primary requirements, Qwen3.7-Max is a direct alternative that does not require changing the agent tooling. The 1,000-step autonomous demo is the benchmark. If you are evaluating models for production agent workloads that run for hours and require sustained tool-call coherence, the Qwen3.7-Max demo is a concrete reference point for what sustained autonomy looks like at the frontier.

[ALLOY]: The operator surfaces are the Anthropic-compatible API endpoint at Alibaba Cloud Model Studio, the model name qwen3.7-max, and the base URL configuration. No local deployment option is available. The 1M-token context window means Qwen3.7-Max can hold entire codebases, large document collections, or long conversation histories in a single context without chunking. The 10x kernel optimization result is an internal Alibaba test, and independent confirmation is needed before treating it as a general benchmark. The model optimizes a SGLang Extend Attention kernel on T-Head silicon, which is Alibaba's own hardware architecture. The performance gain may not generalize to other hardware targets. Treat the 10x figure as a directional signal, not a universal claim.

[NOVA]: What agents can now do that was previously impossible: run sustained multi-hour agentic tasks with 1,000-plus tool-call steps on a single model, which is the longest sustained autonomous run reported from a frontier model. Use a high-context reasoning model alongside Claude Opus 4.7 without changing agent tooling, which means operators can benchmark Qwen3.7-Max against their current model choice in the same environment. The Anthropic-compatible API is the mechanism that makes this comparison tractable. You do not need a new agent configuration. You need a different base URL and model name. [PAUSE]

## [20:00] Qwen3.7-Plus: multimodal agent model, hybrid MoE architecture, and GUI/CLI reasoning

[NOVA]: Qwen3.7-Plus is the balanced multimodal variant of the Qwen 3.7 generation, available via API through Alibaba Cloud Model Studio and listed on OpenRouter. It accepts text, image, and video input with text output, and it is designed for agent workloads that span GUI-based environments, CLI reasoning, and cross-modal perception. Early third-party testing covered browser-based workflows, OpenCode coding tasks, C++ game generation, frontend design, flight simulation, and interactive website generation. The model handles multimodal inputs natively without a separate vision encoder pipeline, which reduces tool-call overhead for image-understanding tasks. This is architecturally significant. A single model handles vision and language reasoning, which means agents can process a screenshot, a code file, and a natural-language instruction in the same context window without routing through separate vision and language models.

[ALLOY]: The hybrid architecture combines efficient linear attention with sparse mixture-of-experts routing for strong scalability at lower active parameter counts. The sparse MoE design means the model activates a subset of its total parameters per token, which keeps inference costs down while maintaining frontier-level reasoning. The linear attention mechanism handles long-context sequences more efficiently than standard attention, which is important for a 1M-token context window where standard attention becomes computationally expensive. The 1M-token context window is shared across all input modalities, which means long video transcripts and large codebases can coexist in a single context window. You can feed a two-hour video transcript, a 500,000-line codebase, and a natural-language task description into the same context and the model reasons across all of them.

[NOVA]: The multimodal pipeline unification is the architectural story. Most multimodal models use a separate vision encoder that processes images and produces a representation that the language model consumes. That architecture adds a processing step and a translation layer between the visual input and the language reasoning. Qwen3.7-Plus integrates vision and language into a single model without a separate vision encoder, which means the visual input flows directly into the language model backbone. The result is lower tool-call overhead for image-understanding tasks and a more unified reasoning process that can switch between visual and textual inputs without a pipeline handoff.

[ALLOY]: The operator surfaces are the API via Alibaba Cloud Model Studio and OpenRouter. No open-weight release is available. The hybrid MoE architecture means Qwen3.7-Plus can run efficiently on commodity GPU budgets while maintaining frontier-level reasoning, but the API delivery model means operators are not running it locally. For agents that need to understand screens, parse documents, or automate GUI-based workflows, Qwen3.7-Plus provides a single unified model that handles images and video alongside text without routing through separate vision and language models. Test it as a drop-in replacement for Claude Opus 4.7 on tasks that require visual understanding combined with long-context code reasoning.

[NOVA]: What agents can now do that was previously impossible: handle multimodal agent tasks with a single unified model instead of a pipeline of separate vision and language models, which simplifies the agent tooling configuration and reduces the latency from image input to language reasoning. Process long video transcripts and large codebases in a single context window, which enables a new class of agent tasks that combine visual understanding with deep code reasoning. The hybrid MoE architecture keeps inference efficient at scale, which means the model can be used in production agent workloads without the cost profile of a dense frontier model.

[ALLOY]: The thread across the Qwen3.7 lineup is clear. Qwen3.7-Max is the text-only reasoning flagship for long-horizon agentic tasks, built for sustained multi-hour workloads with 1,000-plus tool-call steps. Qwen3.7-Plus is the multimodal agent variant for GUI and cross-modal reasoning, with unified vision and language in a single model. Both have a 1M-token context window. Both expose an Anthropic-compatible API. Both are available via Alibaba Cloud Model Studio today. For the agent tooling ecosystem, the Qwen3.7 line is the first time a frontier-level reasoning model and a frontier-level multimodal model share a context window architecture and an API compatibility layer, which makes it straightforward to evaluate both in the same agent configuration and choose based on task requirements rather than tooling complexity. [PAUSE]

## [22:00] Project radar: Dataverse MCP, MCP spec repo, and the local LLM spotlight

[NOVA]: The project radar this week covers three MCP servers and a local LLM spotlight that rounds out the stack. Microsoft Dataverse MCP is the official server that exposes the full Dataverse data platform to MCP-compatible agents. Built by Microsoft's Power Platform team, it provides a structured tool surface for metadata inspection, record querying, cross-table search, and data manipulation. Agents can now interact with Microsoft Power Platform business data through the standard MCP tool interface without custom Dataverse SDK integration. If you run Dynamics, Power Apps, or any Microsoft business data platform that uses Dataverse as its data layer, the Dataverse MCP server gives your agent a first-class interface to that data without needing to write custom API integration code.

[ALLOY]: The official Model Context Protocol specification repository hosts the protocol spec, the SEPs or Spec Enhancement Proposals, reference server implementations, and the official Python and TypeScript SDKs. The July 2026 RC makes the spec stateless and graduates extensions to first-class. The spec repository is the authoritative reference for implementing MCP clients, servers, and the new extensions framework. If you are building MCP tooling, the SEP-1865 pull request for MCP Apps and the SEP-2577 pull request for the feature lifecycle policy are the documents to read before the spec finalizes. They define the server-rendered UI pattern and the extension governance model, both of which will shape how new MCP capabilities ship going forward.

[NOVA]: The community Dataverse MCP from awit-org is a second-party implementation that provides additional tooling for Power Platform integrations beyond what the official Microsoft release offers. Evaluate it as a reference implementation or an alternative if you need features not present in the official release. The community Dataverse MCP is less likely to have the full surface area of the official release, but it may have tooling that the official release has not yet shipped.

[ALLOY]: The local LLM spotlight is Qwen3.5-35B-A3B, an open-weight sparse MoE model from Alibaba Cloud with 35 billion total parameters and 3 billion active parameters per token. It accepts text, image, and video input with text output. The hybrid linear attention plus MoE architecture enables strong reasoning at the 35B scale with efficient inference. Run it with Ollama or LM Studio on a single mid-range GPU for local coding assistance, document understanding, and CLI agent tasks with multimodal inputs. This is the open-weight option for operators who want multimodal capability without a cloud API dependency, and the 3B active parameter count means it runs on hardware that most developers already have. [PAUSE]

## [25:00] Closing and practical queue

[NOVA]: The practical queue this week has five items. First, upgrade Codex to zero point one thirty eight and try slash app to hand off a CLI session to the desktop client. The named-pipe transport makes the handoff instantaneous and authenticated, which means multi-monitor and GUI-heavy workflows no longer require manual context transfer. Second, update Claude Code CLI via npm update -g at-sign anthropic-ai slash claude-code if you are on the latest tag. If you need predictability, pin to stable explicitly with npm install -g at-sign anthropic-ai slash claude-code at stable. Third, audit your MCP server implementation for server-side session state before the July 2026 RC becomes final. The stateless architecture removes the sticky-session constraint, but it requires your server to stop holding per-client state in memory. Fourth, explore the new Shortcuts natural-language planner on iOS 27 when the beta drops. The planner is a new authoring surface for workflow automation that non-technical users can now use directly. Fifth, add Qwen3.7-Max to Claude Code with a three-line Anthropic-compatible API config for long-horizon agent tasks, and evaluate Qwen3.7-Plus as a multimodal drop-in for Claude Opus 4.7 on vision-language agent workloads.

[ALLOY]: The thread across the release cycle is the surface area. The harness layer is consolidating around native, multi-surface tools with session handoff, stateless protocols, and rapid iteration. The model layer is adding high-context and multimodal options that slot into existing agent tooling without migration. The project radar is the connectivity layer that brings enterprise data, protocol governance, and local model options into the same MCP tool surface. For more detail on every story covered today, look at the show notes at Toby On Fitness Tech dot com. Thanks for listening. We'll be back soon.