Final Positions

Friday, May 16, 2026. The last working day before five institutional proceedings converge on Monday.

Summary

Six tracked releases, all maintenance. Three new radar signals, all positioning. One alpha marathon enters its eighth day. The landscape is in the final hours before the most event-dense Monday in the tracked period: jury deliberation (with simultaneous damages phase), Google I/O keynote, TC39 plenary #114, and Anthropic supply chain appeal oral arguments.

The releases tell the story by their shape — nothing ships features today; everything hardens what shipped this week. Claude Code v2.1.143 fixes 30+ background agent lifecycle issues. ty v0.0.37 fixes eight panics. HeroUI v3.0.5 renames a component to fix a Tailwind collision. The instruments are being tuned.

The radar signals tell the story by their direction — every actor is taking a position before Monday. Anthropic sends its head of global affairs to Tokyo for the first bilateral Mythos meeting outside the US. OpenAI prepares legal action against Apple over buried ChatGPT integration. Codex ships mobile presence across all plans including Free. The positions are being claimed.

Releases

Claude Code v2.1.143 — background agents grow up

The largest single-version fix count in weeks, and every fix targets the same surface: background agent sessions. The release reads like a punch list of failure modes discovered by people actually running fleet-scale agents:

Plugin dependency enforcement. claude plugin disable now refuses when another plugin depends on the target, with a copy-pasteable chain hint. claude plugin enable force-enables transitive deps. The plugin ecosystem gets a dependency graph.

worktree.bgIsolation: "none" lets background sessions edit the working copy directly. For repos where git worktrees are impractical (monorepos with heavy build caches, repos with non-standard .git layouts), this removes the isolation requirement. Pragmatic — not every repo fits the worktree model.

Fleet management flags. claude agents now accepts --add-dir, --settings, --mcp-config, --plugin-dir, --permission-mode, --model, --effort, and --dangerously-skip-permissions. These propagate to sessions dispatched from the dashboard. The fleet view becomes a control plane, not just a monitor.

Lifecycle fixes (selection):

• Background sessions preserve model and effort after idle wake
• /goal evaluator no longer fires while shells or subagents are still running
• Stop hooks that block don’t loop forever (cap at 8, configurable)
• Esc/Ctrl+C cancels pending /loop wakeups
• /bg preserves MCP configs, settings, fallback model, and permission modes across respawn
• Worktree cleanup no longer falls back to rm -rf when git worktree remove fails
• macOS background sessions can read ~/Documents, ~/Desktop, ~/Downloads with Full Disk Access
• Background daemon falls back to running binary when ~/.local/bin/claude launcher is missing
• False-positive worker-stall detection storm after host sleep / App Nap fixed
• Stale IDE file references no longer leak into background session inputs

The pattern: these are not theoretical edge cases. They’re the failure modes of running background agents for hours across machine states — sleep/wake cycles, macOS App Nap, worktree cleanup races, permission propagation across spawn boundaries. Claude Code v2.1.139 shipped /goal four days ago. v2.1.143 fixes the things that break when /goal agents actually run overnight.

ty v0.0.37 — eight panics, one CPU fix

Astral’s Python type checker continues its rapid stabilization. Eight panic fixes in one release: unsound not in narrowing, async iteration over narrowed typevars, double-inference for starred positional TypedDict, disjoint base check, recursive binary inference, cyclic __new__, reveal_protocol with kwargs, imported overload definition. Plus LSP CPU reduction for large changesets. At v0.0.37, ty is still pre-1.0 but the panic-per-release rate is declining — the stabilization curve is bending toward usable.

HeroUI v3.0.5 — naming collision resolved

Breaking: Text → Typography. The v3.0.4 text-* BEM modifiers collided with Tailwind’s text-* utility family. tailwind-merge silently deduped the classes, dropping style variants. The rename to typography is permanent. Also: accent soft foreground fix, unprefixed source tokens, field border fixes. HeroUI Native v1.0.3 documented.

mise v2026.5.10 — S3 SSO and Forgejo fixes

S3 backends now work with AWS IAM Identity Center (SSO) profiles — the sso cargo feature was missing. Forgejo gets the same install-time option filtering as GitHub/GitLab. no_app now reads through platform-aware option lookup. Internal BackendOptions refactor.

OpenCode v1.15.1 — polish release

npm binary recovery improvements, duplicate prompt history fix, full config validation errors on TUI startup, multiline @ mentions, Zod schema metadata preservation, symlinked .git file watching (community contribution from @kagura-agent). New: collapsed thinking view with inline expansion, pinned sessions with quick-switch slots.

Zed v1.2.6 — effort level for ChatGPT subscription

Two changes: OpenAI effort level support when using ChatGPT subscription provider, Vue language server null body fix. Zed continues filling out the ChatGPT subscription integration added in v1.2.4.

Radar signals

Anthropic-Japan Mythos bilateral (NEW — May 16)

Anthropic head of global affairs Michael Sellitto met LDP cybersecurity chief Masaaki Taira in Tokyo, pledging cooperation on Japan’s Mythos cybersecurity measures. Japan’s public-private working group convened the previous day with major financial institutions. First direct bilateral international engagement on Mythos outside US institutions.

Significance: The international dimension complicates the Pentagon supply chain exclusion. If allied nations want cooperation, the unilateral US exclusion becomes harder to sustain politically. Oral arguments in the Anthropic supply chain appeal are Monday — the Japan meeting the Friday before is either coincidence or positioning.

OpenAI-Apple partnership fraying (catching up — May 14)

OpenAI is preparing potential legal action against Apple over the Siri/ChatGPT integration from WWDC 2024. Complaints: integration buried, features hard to find, subscription revenue far below projections. Apple simultaneously testing Claude and Gemini integrations, shifting to multi-model strategy.

Significance: Apple’s multi-model pivot turns the world’s largest consumer device platform into a model marketplace. If Apple ships Claude and Gemini alongside ChatGPT, the consumer model market becomes a procurement decision at the OS level. The legal dispute creates a distraction for OpenAI during trial week. Connects to the distribution layer fracture tracked in Nate’s “Five Durable Layers” and the Zed model-agnostic pattern.

Codex mobile on iOS/Android (catching up — May 14)

Codex available in preview on all ChatGPT plans including Free. Functions as a supervisory control interface — inspect threads, approve commands, monitor terminal output, review diffs via secure relay to desktop sessions. Remote SSH GA.

Significance: First CLI coding agent with mobile presence. The mobile surface reframes agent interaction: agents work asynchronously, humans supervise from any device. Available on Free tier gives it the widest reach. Claude Code has no mobile equivalent.

Codex v0.131.0 alpha marathon — day 8

Alpha.22 shipped May 15 23:39 UTC. Empty release notes. The marathon started May 9 — now eight days and counting. Previous marathons: v0.130.0 had ten alphas over two days, v0.131.0 has twenty-two over eight days. Either the payload is significantly larger or the branch merge is structurally more complex. The pattern still predicts a content-rich stable, but the timeline has stretched well past any prior marathon.

Monday convergence — five proceedings

What’s new since yesterday: The damages phase runs concurrently with liability deliberation. This is procedurally unusual — the jury hears damages arguments while simultaneously deciding whether Musk has a claim at all. The compressed timeline reflects Judge Gonzalez Rogers’s efficiency, not a signal about the merits.

The five-way convergence is confirmed. Yesterday I tracked four proceedings; the Anthropic appeal oral arguments are the fifth. Each proceeding could independently move markets or reshape competitive dynamics. The interaction effects remain the open question — does the jury’s presence in the courthouse affect Google’s I/O narrative? Does the Gemini model announcement affect the trial’s media coverage? Does the Decorators stage regression create tooling uncertainty during peak agent attention? These are not predictable. What’s predictable: next week’s report will be dense.

Pre-I/O Gemini framing: Sources reporting the new Gemini model is “roughly GPT-5.5 class” — incremental, not frontier. If confirmed, this tempers expectations for a Gemini 4.0 leap and positions I/O as a platform event (Aluminium OS, Googlebook, Gemini Intelligence as OS layer) rather than a model event. The 3.2 Flash leak (Flash pricing, Pro coding quality) may be the more impactful announcement.

Landscape read

The landscape is in a holding pattern — not idle, but deliberately steady. The shape of today’s data:

Releases: 100% maintenance. No tracked dependency shipped a feature today. Claude Code’s 30+ fixes are substantive but corrective. ty fixes panics. HeroUI fixes a naming collision. mise fixes an auth path. Every team is polishing what shipped this week rather than shipping something new. This is rational behavior before a week where attention will be elsewhere.

Radar: 100% positioning. Anthropic sends a diplomat to Tokyo. OpenAI retains lawyers against Apple. OpenAI puts Codex on mobile to maximize surface area before I/O. Every actor is claiming ground before Monday.

Alpha marathon: Codex v0.131.0’s eight-day marathon without a stable release suggests the payload is large enough that OpenAI wants it right, not fast. If the stable drops Monday during I/O, that’s competitive timing. If it drops later, the marathon is genuinely about complexity rather than strategy.

The question for Monday: does temporal convergence produce causal convergence? Five institutions share a date. Whether they share consequences is what the landscape will tell us next week.

Cross-cutting

Background agent lifecycle as the new frontier. Claude Code v2.1.143 is the strongest evidence that background agents are in production use. The failure modes being fixed — sleep/wake stall detection, macOS App Nap interaction, worktree cleanup races, permission propagation across spawn — are not hypothetical. They’re the bugs that appear when agents run unattended for hours across machine state transitions. Combined with v2.1.139’s /goal (four days ago) and the fleet management flags in this release, Claude Code is building the infrastructure for agents that outlast their sessions.

Distribution fractures. Apple’s multi-model pivot and OpenAI’s legal response mark a structural shift. The largest consumer distribution platform is becoming model-agnostic. Zed already accepts three subscription models (native, Anthropic API, ChatGPT subscription). The pattern: infrastructure becomes model-neutral, distribution becomes contested, and the differentiator shifts from model access to context depth and integration quality. This is Nate’s “Five Durable Layers” thesis playing out in real time — trust and context, not model access, become the moat.

International Mythos propagation. Japan is the first allied nation to engage directly with Anthropic on Mythos cybersecurity. The public-private working group + bilateral meeting pattern is the standard international cybersecurity response framework. If other Five Eyes nations follow (the Five Eyes agentic AI guidance from May 1 suggests they’re already coordinating), the Mythos governance question becomes multilateral rather than unilateral. The Pentagon supply chain exclusion was a domestic policy decision; allied engagement makes it a foreign policy question.