Radar brief — 2026-09-30

What the evidence shows

15 curated signals across 23 tracked assistants. About half the mix is GitHub/changelog-grade; the rest is HTML/docs copy, and 1 source reported errors, so treat non-GitHub rows as noisier. Several GitHub notes omit large tails (bodyOmitted); those tails are not in this brief.

GitHub / versioned releases (prefer these)

HTML / docs changelogs (noisier; some recap older lines)

For developers

• Pin your model. On Cline v4.1.21, 19 providers without a pinned model change defaults; 11 (including GitHub Copilot and Vertex) resolve to Claude Opus 5.5. If the last week felt like a different assistant, check the pin first.
• Upgrade Claude Code before trusting resume. v2.1.282 is the evidence-backed fix for --continue/--resume rewriting history, slash commands (/model, /rename, /artifacts, …) dropping extended thinking, and a reduced --tools list wiping earlier thinking. Also run /status or claude doctor and read which project telemetry vars were ignored.
• Treat Codex 0.156→0.157 as a UX + model jump. Voice is on by default (F8). /tui is optional in 0.156; fullscreen transcripts are default in 0.157. Worktrees default on. New SKUs: GPT-6 Sol and Luna, with Bedrock and migration prompts. Use /usage before assuming spend/token totals. Network policy now applies to redirects and live HTTP/WebSocket traffic.
• Zed v1.22.0-pre only: add BYOK keys if you want Opus 5.5 or GPT-6 Astra/Sol/Luna in-editor, and pass model on spawn_agent instead of inheriting the parent. Do not treat this as a stable IDE train.
• Goose recipes will fail closed at 32 params / 200 select options / 128 KiB. Shrink them before v1.52.0, not after. Custom provider files are now protected — stop editing them as if they were scratch JSON.
• Windows + proxy: Cline SDK v0.0.87 is the loopback/NO_PROXY fix if the hub dies with “No compatible hub runtime.” Antigravity 1.2.13 will stop on long retry delays or daily/billing caps instead of spinning every 5s — surface that in your retry logic.
• Factory Droid: add custom models from the terminal (v0.229.0); stop a runaway subagent from its panel (v0.225.0).
• OpenCode v1.18.33: if MCP browser launch exits immediately you should now get an error; don’t paste debug config dumps assuming secrets are still inlined.
• Tabnine: skip 6.6.0/6.6.1. If you enabled Chat only to unstick Agent-only self-managed models, you can disable Chat again after 6.6.2.

For teams

• Default-model drift is a fleet problem. Cline’s unpinned Copilot/Vertex → Opus 5.5 shift changes cost, latency, and policy without a developer choosing a model. Require pinned model IDs in shared Cline/Zed/Goose configs.
• Managed MCP vs Chrome. Claude Code’s allowClaudeInChromeWithManagedMcp is the explicit gate for claude --chrome beside exclusive managed-mcp.json. Decide the org answer; don’t leave individuals to discover the renamed error.
• Telemetry and secrets hygiene. Claude Code now shows ignored/off telemetry vars at startup. OpenCode redacts debug credentials. Goose protects custom provider files. Cline bumped js-yaml to 4.3.2 because it parses rule/skill frontmatter — patch that parser wherever you vendor it.
• Billing paths are multiplying. Codex /usage (tokens, plugins, skills). Windsurf/Devin Desktop can attach eligible ChatGPT Plus/Pro usage (changelog recommendation: Fusion GPT-6 Astra + SWE-2 — that pairing is vendor copy, not a measured result here). Tabnine 6.6.2 restores org per-user/per-team caps; the install-wide monthly cap still needs an installation admin.
• Quota behavior: Antigravity will halt rather than retry past daily/billing caps or >30s delays. Codex 0.157 cancels in-flight connections when network policy revokes access. Plan for hard stops, not infinite backoff.
• Skip unstable Tabnine 6.6.0/6.6.1 on managed fleets. Kiro enterprise region routing and Crew live-settings/Gateway durability matter if you run Kiro as a service, but those notes are HTML-heavy.
• Open team proposal already on the board: shrink Goose recipes to the new limits.

What to ignore

• Omitted tails. Claude Code (+79), Codex (+128 / +126), Zed (+52), Goose (+25), and others. Do not infer features, breaks, or CVEs from “+N more changes.”

• HTML changelog padding. Kiro v1.1.70 restates 1.0 highlights. Windsurf’s “best value” Fusion pairing is marketing inside a changelog. Factory/Tabnine/Kiro Crew pages are lower-trust than the GitHub tags above.

• Zed v1.22.0-pre and Tabnine 6.6.0/6.6.1 — prerelease / explicitly unstable.

• The P2 OpenCode videos proposal (PR #38252 verification videos) — process noise, not a product signal in this window.

• Grok CLI prompt/MCP cleanup proposals are open tickets, not shipped Grok CLI evidence in this JSON. Don’t upgrade anyone on the back of those links.

• One ingested source errored this window; if a row looks thin, it may be incomplete rather than “no change.”

Watch next

• Whether GPT-6 Sol/Luna Bedrock + migration prompts in Codex 0.157 show up as defaults or remain opt-in in the next GitHub tag.
• Whether Cline’s 19-provider default retarget sticks or gets a pin-on-upgrade safeguard.
• Codex voice-default and worktree-default backlash (both are behavior changes, not just features).
• Zed v1.22.0 stable: BYOK + per-spawn model only count once they leave -pre.
• Claude Code managed Chrome+MCP and telemetry notices — likely the enterprise-policy surface to watch, not maxProseWidth.
• Open proposals still unmerged: Factory custom models / loop pill (P5), Grok CLI response_guidelines (P4), Grok CLI drop legacy managed MCP (P4), Windsurf ChatGPT sign-in + Fusion (P3), Zed explicit spawn_agent model + BYOK (P3).

How to use DevAgentRadar

DevAgentRadar is a free, evidence-first radar for AI coding assistants. Read Headline for the window, What the evidence shows for versioned facts only, then take For developers or For teams as the action list. Use What to ignore to drop marketing, omitted bodies, and unstable trains. Watch next is the change-detection queue, including the 7 open improvement proposals — none of those are releases until they ship in evidence. If a vendor blog and a GitHub tag disagree, keep the tag.

Evidence appendix

• Evidence window: top curated signals (versioned releases preferred)
• Assistants tracked: 23
• Signals in this brief: 15
• Trusted source mix (GitHub/changelog vs HTML): 49%
• Open improvement proposals: 7
• Sources reporting errors: 1

High-signal releases used

• Claude Code v2.1.282: Added a maxProseWidth setting that caps the width of Claude's prose in wide terminals while tables and code blocks keep the full width · Added a startup notic — source

• OpenAI Codex CLI rust-v0.156.0: Choose an optional fullscreen UI with /tui for your next launch, including transcript search, mouse selection, and right-click copying. (#46732, #46734, #4688 — source

• OpenAI Codex CLI rust-v0.157.0: Added GPT-6 Sol and Luna, including Amazon Bedrock support and migration prompts for older models. (#47332, #47347) · Enabled fullscreen transcripts by default — source

• Google Antigravity 1.2.13: Improved rate-limit handling when the model API returns a retry delay: the CLI now waits the delay the server asks for instead of a fixed 5 seconds, and stops r — source

• Cline sdk/sdk/v0.0.87 — The hub now starts reliably on Windows machines with a system proxy. · Bun's fetch sends loopback requests through HTTP(S)_PROXY with no localhost bypass, so — source

• Zed AI v1.22.0-pre: This week's release includes BYOK support for Claude Opus 5.5 (Anthropic) and GPT-6 Astra, Sol, and Luna (OpenAI), on-the-fly model selection for new subagents, — source

• Cline v4.1.21: New provider: ai&, an OpenAI-compatible endpoint serving open-weight models from Japan. · Refreshed the model catalog to 6,386 models across 209 providers. · Th — source

• OpenCode v1.18.33: Cloudflare AI Gateway models now honor provider response and stream timeouts. (@danlapid) · MCP browser launch failures are now reported when the launcher exits — source

• Goose v1.52.0: Live voice conversations in the desktop app #12093 · Decisions provider crate with OpenRouter and Jev implemen — source

• Kiro v1.1.70: 1.1.70 Sep 24, 2026 Agent Artifacts and Native ARM64 Builds IDE 1.1 lets agents produce durable artifacts you can review and return to, adds native ARM64 builds — source

• Windsurf v3.10.48: v3.10.48 September 29, 2026 Sign in with ChatGPT You can now sign in to Devin with ChatGPT and use your eligible ChatGPT Plus or Pro plan for GPT model usage in — source

• Factory Droid v0.229.0 — · Desktop v0.186.0 (September 29, 2026): Custom model management, PDF viewer, quoted replies, voice dictation key, plus terminal and app fixes · Custom models in terminal - You can now add, edit, a — source

• Factory Droid v0.225.0 — · Desktop v0.182.0 (September 22, 2026): Default app file opening, subagent controls, Droid Shield accuracy, and desktop sign-in and menu fixes · Open files in default app - You can now open a sess — source

• Tabnine v6.6.2: v6.6.2 September 22, 2026 Bug Fixes Accounts: Fixed requests made with an expired or invalid sign-in token returning a server error instead of prompting you to — source

Open proposals

• Add custom models from the Droid terminal and pause scheduled loops from the chat pill (audience: developer, P5)
• Replace <output_efficiency> with <response_guidelines> in Grok CLI prompts (audience: developer, P4)
• Drop legacy managed MCP configs and shadowed platform skills in Grok CLI (audience: developer, P4)
• Sign in to Devin Desktop with ChatGPT and set Fusion to GPT-6 Astra + SWE-2 (audience: developer, P3)
• Shrink Goose recipes to 32 params, 200 select options, and 128 KiB (audience: team, P3)
• Pass an explicit model to every spawn_agent call and add BYOK keys for Opus 5.5 and GPT-6 (audience: developer, P3)
• Improve with OpenCode pr-38252-videos: pr-38252-videos PR #38252 verification videos (audience: everyone, P2)