โ† All digests
AI Developer Digest

Tue, Nov 10, 2026

6 signals that cleared the gate18 min read
The Signal โ€” start here
Tuesday, November 10: a light day after yesterday's major Fable 5.2 release. The main shipping event is Claude Code v2.1.322 extending fast mode to Vertex AI (global, us, eu regions). More importantly, a close read of the full Fable 5.2 "What's New" docs surfaced two additional breaking changes that weren't captured in yesterday's digest: forced tool_choice (any/tool) now returns 400 on Fable 5.2, and the thinking: {"type": "enabled"} parameter with budget_tokens also returns 400 โ€” teams using either pattern need to update before switching to Fable 5.2. The other story: DIYgod/pinky, an MCP-native phone-automation agent that reached GitHub Trending today, is the first cross-platform tool to bring human-in-the-loop phone control to Claude Code, Codex, and Cursor via a gateway-enforced approval model.
Must-reads today
1
Fable 5.2 forced tool_choice + thinking type [BREAKING]: two more Fable 5.2 400-error conditions missed yesterday โ€” check before migrating
2
Claude Code v2.1.322 [NOTABLE]: fast mode lands on Vertex AI (global/us/eu) with standard-speed fallback

๐Ÿ† Repo of the Day

1

DIYgod/pinky โ€” MCP-native phone control for AI agents, with gateway-enforced human-in-the-loop

macOS via Homebrew

brew install --cask diygod/tap/pinky

or download the .dmg from github.com/DIYgod/pinky/releases

Connect your phone

Android: enable USB debugging, plug in

iPhone: enable Developer Mode, plug in, let Pinky install WebDriverAgent

Add Pinky MCP to Claude Code

In ~/.claude/settings.json, add under "mcpServers":

"pinky": { "command": "pinkyctl", "args": ["mcp"] }

Try a fake phone session first (no hardware needed)

pinkyctl demo --fake


Breaking Changes

1
โ—Breaking

Fable 5.2: forced tool_choice and explicit thinking type both return 400 errors

What changed
Two additional Fable 5.2 breaking changes not surfaced in yesterday's digest (from the same Nov 9 release):

  1. Forced tool use blocked: tool_choice: {"type": "any"} or tool_choice: {"type": "tool", "name": "..."} returns a 400 invalid_request_error on Fable 5.2. Only tool_choice: {"type": "auto"} is accepted.

  1. Thinking type parameter invalid: Sending thinking: {"type": "enabled"} with budget_tokens, or thinking: {"type": "disabled"}, returns a 400 error. Adaptive thinking is always on โ€” omit thinking entirely or send {"type": "adaptive"} and use the effort parameter to control thinking depth.

TL;DR
Two more Fable 5.2 400-error conditions: forced any/tool tool_choice is gone, and the explicit thinking enable/disable toggle is gone โ€” both are hard errors, not graceful fallbacks.
Developer signal
Before switching model to claude-fable-5-2, grep your codebase for tool_choice with type: "any" or type: "tool" and migrate to {"type": "auto"} with strict: true (strict tool use) for schema-valid output, or use JSON outputs mode. Also grep for thinking: {type: "enabled"} or thinking: {type: "disabled"} and remove those fields โ€” on Fable 5.2, thinking depth is controlled via effort: "low" | "medium" | "high" | "xhigh" | "max" (API default: high). Teams migrating from Claude Fable 5 or Claude Opus models are most likely to hit these, since those models required explicit thinking toggles. Teams already on Fable 5.1 with neither pattern are unaffected. Note: CMEK (customer-managed encryption key) organizations cannot use structured outputs or strict tool use on Fable 5.2 โ€” validate tool input manually in that case.


Affects you ifYou set tool_choice to any or a specific named tool on any Fable model; you pass thinking: {type: "enabled", budget_tokens: N} or thinking: {type: "disabled"} in your request (common for Fable 5 or Claude Opus codepaths); you are on CMEK and use strict tool use or JSON outputs.EffortQuick (grep-and-replace on tool_choice and thinking params) / Moderate (CMEK orgs need a workaround)

API & SDK Changes

1
Medium

Fable 5.2 extended: server-side fallback for refusals (beta), account-bound thinking, fast mode on three more platforms

What changed
Additional Fable 5.2 API details not covered in yesterday's digest (all from the same Nov 9 release):

  1. Server-side fallback for refusals (beta): Pass fallbacks: "default" to have a declined request automatically retried on Anthropic's recommended fallback model for that refusal category. For Fable 5.2, permitted fallback targets are Claude Opus 4.8 and Claude Opus 5. For categories with no recommended fallback (e.g., reasoning_extraction), the refusal stands.

  1. Thinking blocks are account-bound: Thinking blocks produced by Fable 5.2/Mythos 5.2 work only in the account that produced them (or a linked account). Sending them from another account results in a silent drop, not a 400 error. With thinking-binding-controls-2026-08-01 beta header, the response's input_transformations array reports the drop with reason: "organization_binding_mismatch".

  1. Fast mode extended to three more platforms: Fast mode for Fable 5.2 is now also available on Google Cloud (Anthropic managed), Google Cloud (Google managed), and Microsoft Foundry (as separate model claude-fable-5-2-fast). Previously only on the Claude API.

  1. New beta headers: compact-2026-09-04 (compaction on demand โ€” control when older turns are summarized), inline-tools-2026-09-15 (mid-conversation tool changes via role: "system" messages instead of editing tools), thinking-binding-controls-2026-08-01 (get input_transformations on thinking block drops/mismatches).

TL;DR
Fable 5.2 adds a fallbacks: "default" parameter for automatic refusal fallback to Opus 4.8/Opus 5, account-bound thinking that drops silently cross-account, fast mode on three additional platforms, and three new beta headers for context management.
Developer signal
Add fallbacks: "default" to production Fable 5.2 requests if you want graceful handling of safety refusals โ€” the API retries automatically on Opus 4.8 or Opus 5. If you're building multi-tenant systems that share thinking blocks across accounts, this will silently fail: thinking blocks must stay in the originating account's API calls. Enable thinking-binding-controls-2026-08-01 during development to surface drops explicitly. If you use compaction (long agentic sessions that need to trim context without breaking the prefix check), compact-2026-09-04 lets you trigger compaction at a point you choose rather than having the model decide. If you're on Google Cloud or Microsoft Foundry, fast mode is now available โ€” use speed: "fast" or the claude-fable-5-2-fast model ID on Foundry.


Affects you ifYou want auto-retry on refusals; you share thinking blocks across API accounts or orgs; you deploy on Google Cloud or Microsoft Foundry; you manage context windows in long agentic sessions.EffortQuick (add fallbacks: "default" parameter) / Moderate (refactor multi-tenant thinking-block sharing)

Research

arxiv.org DNS-blocked (ninth consecutive day). huggingface.co/papers DNS-blocked (tenth consecutive day). No research items passed the quality gate from primary sources.


Tooling

1
Notable

Claude Code v2.1.322 โ€” fast mode now available on Vertex AI (global, us, eu regions)

What changed
v2.1.321 (Nov 9) โ†’ v2.1.322 (Nov 10, 02:29 UTC): one change โ€” fast mode added for Vertex AI in the global, us, and eu regions, with automatic fallback to standard speed if Vertex declines fast mode.
TL;DR
Claude Code v2.1.322 makes fast mode (up to 2.5x output tokens/second) available to Vertex AI users in three regions, falling back gracefully if the region is overloaded.
Developer signal
Run claude update to pick up v2.1.322. If you're using Claude Code via Vertex AI (via claude remote-control or VS Code extension with Vertex credentials), fast mode is now available. Set speed: "fast" in your configuration or pass it in API calls โ€” Claude Code will fall back to standard if Vertex refuses it rather than erroring. This is particularly relevant for teams running Claude Code in GCP-native infrastructure via Vertex rather than the direct Anthropic API.


Affects you ifYou run Claude Code via Vertex AI in the global, us, or eu regions; you want faster output throughput on GCP infrastructure.EffortQuick (claude update)

Rising Dev Tools

1
Notable

DIYgod/pinky โ€” MCP-native phone control for AI agents with gateway-enforced human approval

Developer signal
brew install --cask diygod/tap/pinky on macOS; deb/tar.gz for Linux; installer for Windows. After installing, run Settings โ†’ Agents โ†’ Connect for Claude Code, or add pinkyctl mcp as an MCP server in ~/.claude/settings.json. Use --fake mode for demos without hardware. iOS setup requires Xcode + WebDriverAgent (~15min one-time setup); Android is USB-debug + plug-in.


Affects you ifYou build agents that need to interact with mobile apps (testing, automation, QA, demos); you want phone-control in Claude Code sessions; you're building multi-step workflows that cross desktop and mobile.
GitHub Trending (all languages, daily, Nov 10) | Stars: ~1.7k | Link: github.com/DIYgod/pinky

Benchmarks & Leaderboards

lmarena.ai DNS-blocked (ninth consecutive day). swebench.com DNS-blocked. No benchmark data confirmed from primary sources this period.


Technical Discussions

simonwillison.net DNS-blocked. hnrss.org inaccessible. Nothing cleared the quality bar this period.


Quick Hits

  • Claude Desktop policy change โ€” November 10 (today): Warnings now active for Cowork policies using coworkTabEnabled: false or chatSessionRetentionDays without a desktopHome field. Source: https://github.com/anthropics/claude-code/releases/tag/v2.1.319 (noted in that release; behavior effective today).
  • Fable 5.2 effort default is high, not max: per the migration docs, start evals at high โ€” most capability gain is already present there. Reserve xhigh for hardest coding/agentic work where evals show a measurable gain. Reserve max for ceiling measurement only.
  • Fable 5.2 fallback targets confirmed: Claude Opus 4.8 and Claude Opus 5 (server-side fallback, beta) โ€” not Fable 5.1. If your code retries on refusal and falls back to Fable 5.1, migrate that path.

Worth Watching (Announced, Not Yet Shipped)

  • Gemini API preview model deprecations โ€” November 17 (7 days): Gemini 3.1 Flash TTS preview, Gemini Live API preview models, Gemini 2.5 TTS previews retire. Replacement: gemini-3.8-live and gemini-3.8-flash-tts. Source: ai.google.dev/gemini-api/docs/deprecations (DNS-blocked โ€” verify at official URL).
  • Fast mode removal for Claude Opus 4.8 โ€” November 30 (20 days): speed: "fast" with claude-opus-4-8 will error. Migrate to Claude Opus 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview
  • Claude Sonnet 4.5 retirement โ€” November 30 (20 days): claude-sonnet-4-5-20250929 retires. Migrate to Claude Sonnet 5.5. Source: https://platform.claude.com/docs/en/release-notes/overview


Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff โ€” only what changes what you build.