Radar brief — 2026-10-01

What the evidence shows

Window: 15 signals across 23 tracked assistants. Versioned GitHub/changelogs are 49% of the mix, so treat HTML vendor pages as weaker. Eight open improvement proposals sit on top of this window. One source in the mix reported errors.

Preferred GitHub-style releases:

HTML/docs changelogs (noisier; prefer the GitHub rows above):

For developers

• Cline CLI v3.0.67 / Desktop v0.0.40: Retry tasks that used Add Provider or providers.json (they were picker-only before). If you pinned Vultr, re-select the new ids. When an MCP tool overflows context, follow the preview link / read_files page instead of assuming the tail is gone. Open proposal: re-select Vultr + page MCP overflow.
• Claude Code v2.1.282: Upgrade before relying on --continue / --resume, /model, or a reduced --tools list — this drop is the evidence that earlier reasoning/thinking was being stripped. Set maxProseWidth on ultrawide terminals. Run /status or claude doctor if project settings files silently ignore or disable telemetry.
• Codex CLI rust-v0.156.0 then rust-v0.157.0: Expect voice on (F8 to toggle) and, after 0.157, fullscreen transcripts on by default. Use /usage for token and skill spend. Worktrees are default. If you still need the old model, watch the GPT-6 Sol/Luna migration prompts (Bedrock included). /tui in 0.156 becomes the default transcript mode in 0.157.
• Zed AI v1.23.1-pre: Prerelease only. If you already have an OpenAI key, BYOK GPT-6.1 Sol is in this tag; cap idle agent threads with agent.max_idle_retained_threads. Open proposal: enable BYOK + idle-thread cap.
• Goose v1.52.0: Count recipe params/options/size before the new caps reject them. New model ids in this tag: Opus 5.5, GPT-6-sol, GPT-6-luna.
• Factory Droid v0.229.0: Add custom models from the terminal; pause/resume/stop scheduled loops from the chat pill. Open proposal: custom models + loop pill. v0.225.0 already lets you stop a running subagent from the transcript.
• OpenCode v1.18.33: If an MCP browser never appears, this version reports immediate launcher exit. Do not paste debug config dumps from older builds — this tag redacts credentials.
• Google Antigravity 1.2.13: Rate-limit loops now wait the server’s delay and will not retry a daily/billing cap or a delay longer than 30 seconds — stop hammering when it exits.
• Windsurf v3.10.48 (HTML): ChatGPT plan sign-in for Devin Desktop GPT usage is documented here, not on GitHub. Open proposal: ChatGPT sign-in + Fusion GPT-6 Astra / SWE-2. Verify on your account before treating it as a billing source of truth.

Two open proposals target Grok CLI prompt/MCP cleanup. They are not backed by a versioned signal in this evidence JSON — do not act as if those releases shipped here.

For teams

• Cline Desktop v0.0.40: Re-pin renamed Vultr models org-wide and re-test Add Provider paths that looked configured but failed at run time. Unify MCP file location: with CLINE_MCP_SETTINGS_PATH, CLINE_DATA_DIR, or CLINE_DIR, reads, OAuth, and Open MCP settings now share one file. Defaults moved for 302.AI (Claude Sonnet 5.5), NanoGPT (Ling 3.1 Flash), Vivgrid (GPT-6.1 Sol), and Vultr (MiMo V2.6 Flash RL). Open proposal: re-pin Vultr + retry Add Provider.
• Claude Code v2.1.282: If you ship exclusive managed-mcp.json, decide explicitly on allowClaudeInChromeWithManagedMcp — Chrome is blocked unless that managed setting is on, and the error now names it. Treat project settings files as telemetry policy: ignored or disabling vars are now visible to every developer at startup.
• Goose v1.52.0: Recipe limits are enforced, not documented-only (32 params, 200 select options, 128 KiB). Oversized shared recipes will fail. Open proposal: shrink Goose recipes.
• Codex CLI 0.156.0 / 0.157.0: Voice-default and fullscreen-default change the workstation baseline. 0.157 also enforces network restrictions across redirects and live HTTP/WebSocket traffic, including cancel-on-policy-revoke. Budget via /usage rather than vendor dashboards.
• Tabnine v6.6.2: Skip 6.6.0 and 6.6.1 (vendor-marked unstable). 6.6.2 restores org-admin per-user and per-team caps; the org-wide monthly total cap still needs an installation administrator. If you enabled Chat as a workaround for the 6.6.0 “Unsupported model” Agent/CLI bug, you can disable Chat again after 6.6.2.
• Kiro v1.1.70 (HTML): Enterprise region routing and clearer MCP failures are the team-relevant bits; do not roll 1.0 recap items as if they were new in 1.1.70.
• Antigravity 1.2.13: Daily/billing caps no longer get a retry loop — incident runbooks should treat an immediate stop as quota, not as a transient 5s backoff.

What to ignore

• Any claim drawn from omitted bodies (+79, +128, +126, +46, +25, and the HTML +N tails). Those changes are not in evidence.
• Zed v1.23.1-pre as a stable rollout.
• Tabnine v6.6.0 and v6.6.1 (explicitly unstable).
• Kiro Crew 0.7.2: empty tags, mostly 0.7.0 copy, large omission, HTML.
• Windsurf “best value” Fusion pairing and other billing advice on a non-GitHub changelog.
• Factory Droid and Tabnine docs pages as equivalent to GitHub tags — same window, weaker provenance.
• Breaking/MCP tags on Claude Code and Codex where the breaking detail lives in omitted text. Visible, citable breaks in this brief are Cline’s Vultr id rename, Goose recipe caps, Codex voice/fullscreen defaults, and Claude’s managed Chrome+MCP gate.
• Grok CLI proposals in this window — no matching assistant/version row in the evidence JSON.

Watch next

• Whether Codex keeps fullscreen transcripts and voice as defaults after rust-v0.157.0, and how GPT-6 Sol/Luna migration prompts behave on Bedrock.
• Whether GPT-6.1 Sol / GPT-6 Sol / Luna stay catalog-only (Cline, OpenCode docs, Goose) or become defaults beyond Vivgrid.
• MCP overflow paging in Cline spreading to other agents; Claude managed-MCP + Chrome policy in real enterprise managed-mcp.json fleets.
• Goose recipe-limit fallout after v1.52.0 enforcement.
• Zed 1.23.1 stable vs this -pre.
• Tabnine CLI (OpenCode-powered) after the 6.6.2 auth and cost-control fixes.
• The eight open proposals, especially the two Cline Vultr/MCP items (P5) versus the two Grok CLI items that still lack a versioned signal here.

How to use DevAgentRadar

DevAgentRadar is a free public radar: judgment from versioned evidence, not launch posts. Read What the evidence shows first — GitHub tags outrank HTML changelogs. Use For developers and For teams as checklists tied to exact versions. If a row says bodyOmitted, do not infer the missing notes. Cross-check the open proposals before rolling a default, pin, or managed setting. This brief is 15 curated signals from 23 tracked assistants; it is not a complete catalog of every vendor blog.

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: 8
• 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

• Cline desktop-v0.0.40: Custom providers added with Add Provider now work when you run a task. · They showed up in the provider and model pickers but failed with `Unknown or disabl — source

• Zed AI v1.23.1-pre: This week's release includes BYOK support for GPT-6.1 Sol and the ability to preview JSONL and NDJSON files as tables. · Added BYOK support for GPT-6.1 Sol with — source

• Cline cli-v3.0.67: When an MCP tool returns more output than fits in context, the agent can now read the rest. · It gets a preview plus a link it can page through with `read_files — 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

• 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

• Re-pin renamed Vultr models and retry Add Provider tasks in Cline Desktop v0.0.40 (audience: team, P5)
• Re-select pinned Vultr models and page MCP overflow with read_files (audience: developer, P5)
• Add custom models from the Droid terminal and pause scheduled loops from the chat pill (audience: developer, P5)
• Enable GPT-6.1 Sol BYOK and set agent.max_idle_retained_threads in Zed (audience: developer, P4)
• 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)