Radar brief — 2026-09-04

What the evidence shows

Window: 14 signals across 23 tracked assistants. About half of this mix is versioned GitHub/changelog; the rest is HTML docs that sometimes smear several versions into one page. Eight open improvement proposals are queued. One source in the window is flagged as reporting errors.

GitHub-tagged releases (preferred):

HTML/docs (noisier; do not treat as equivalent to a GitHub tag):

For developers

• Upgrade Claude Code to v2.1.260 before you trust path deny rules. Parentheses in Edit/Write/Read paths were ignored; a single bad [ in a deny rule blocked all edits; zsh could smuggle a substitution through auto-approval. Then turn on fullscreen /diff while Claude edits (open proposal, P5). Use /cost when prompt_cache looks wrong.
• Cline CLI v3.0.61 + SDK v0.0.82 are a paired fix. If you route through Dify, SAP AI Core, opencode, or Codex CLI, tools and images may have been stripped with no error. After upgrade, confirm tool-calling and vision on those providers. Radar also has pin-model proposals for both artifacts (P4) — pin before defaults move under you.
• Goose v1.48.0: treat it as a provider-and-hooks release, not a quiet patch. If you use Z.ai, you are on GLM-5.3. If you bill custom providers, fill in cost fields. If you write PreToolUse hooks, adopt on_failure and the stable tool_call_id. Do not assume the omitted 98 changes are cosmetic (breaking is tagged).
• Codex rust-v0.152.0 then rust-v0.153.0: package-style MCP names, per-tool output_token_limit, reconnect-after-drop, and a plugin marketplace CLI. Set tui.auto_recap = false if automatic recaps get in the way. Vim users get search in 0.152.0 and undo/redo in 0.153.0.
• Antigravity 1.1.25: pull Gemini 3.8 Flash with GEMINI_API_KEY. Re-read Markdown custom agents — they now inherit ambient skills/rules/subagents (behavior change). Retry MCP OAuth if the IdP issues long codes. Windows: re-check /skills grouping.
• OpenCode v1.18.26: if Claude 5 sessions died after a prompt or tool-definition change, this is the stale-thinking-block fix. Bedrock GPT-5.6 can send none reasoning effort.
• Zed v1.19.0-pre only: try OpenRouter effort-based reasoning and watch Agent memory on long sessions. Do not standardize a prerelease (open proposal, P4).
• Tabnine: upgrade CLI to 0.35.0 and enable tools.enableWebFetchTool so fetch is direct, not LLM-invented (open proposal, P3).
• Factory Droid v0.209.0: /archive (or Ctrl+X); new sessions with no model pick GPT-5.6 Sol.
• Two Grok CLI proposals are open (prompt tag rename; drop legacy managed MCP / shadowed platform skills) but no Grok release is in this evidence window — do not invent one.

For teams

• Treat Claude Code v2.1.260 as a sandbox/policy patch. Re-test folder allow/deny lists that use parentheses, character classes, or “read-only” trees. Re-test zsh-heavy developer machines. Roll /reload-plugins and /advisor into Desktop / Remote Control / Agent SDK runbooks.
• Cline Hub replacement can interrupt live sessions — the new prompt tells you how many. Do not force-kill a Hub that will not drain. Windows: Authenticode via Azure Trusted Signing should reduce Smart App Control false blocks; keep the new error text in support macros.
• Codex Full Access skipping Guardian on confirmation-only actions is a control-plane change. Decide whether that matches your review policy before a fleet bump to rust-v0.153.0. Plus/Team usage warnings and rate-limit banners are the cost-control surface.
• Goose cost tracking and Bedrock GPT-5.6 routing will move spend if developers add the new declarative providers. Inventory provider configs; GLM-5.2 is gone in this tag.
• Antigravity: duplicate permission grants no longer accumulate across reloads and subagents — audit existing configs for stacked grants. Custom Markdown agents inheriting ambient skills is a default-behavior change for anyone who used isolation-by-omission. Remote Control reverse-tunnel routing was hardened.
• Factory: org settings can apply after CLI start; do not assume a slow policy fetch blocked launch. v0.208.1’s GPT/Grok 60% discount is time-limited billing, not a model-quality signal. Open proposal (P3, team): archive chats and expect GPT-5.6 Sol as the empty default.
• Tabnine 6.5.3 notes on the same page (org-level repos, per-team model assignment) are admin-relevant but not cleanly isolated from the 6.5.4 tag — verify on your tenant rather than trusting the concatenated HTML.

What to ignore

• Kiro 2.20.2 — empty stub; no changes to act on.
• Kiro v2.19.2 HTML — several versions concatenated; do not attribute 1.0.337 IDE notes to 2.19.2.
• Tabnine docs — and 6.5.1 share one page; only attribution + web-fetch are clearly 6.5.4.
• Omitted bodies (Goose +98, Claude Code +66, Codex +95/+94, Zed +50, Cline +18/+34, and smaller HTML tails). Do not invent those diffs.
• Zed v1.19.0-pre as a production pin.
• Factory 60% off as a lasting pricing baseline.
• Marketing copy, unversioned blogs, and any assistant not in the tables above. One source in this window reported errors — keep HTML changelogs on a short leash.

Watch next

• Full Goose v1.48.0 and Claude Code v2.1.260 changelogs when the omitted slices land — both are tagged breaking with MCP/permissions/hooks surface flags.
• Whether Cline’s shared capability translator holds for every gateway producer, and whether the Hub drain path leaves wedged processes behind.
• Codex plugin marketplaces + reconnect behavior under real app-server drops; Guardian skip scope in Full Access.
• Zed 1.19.0 stable vs this -pre.
• Antigravity custom-agent inheritance and MCP OAuth in non-happy-path IdPs.
• Disposition of the eight open proposals, especially Claude /diff, Cline model pins, Grok CLI MCP cleanup, Tabnine web fetch, and Factory archive/default-model.

How to use DevAgentRadar

DevAgentRadar is a public, evidence-first feed for AI coding assistants: versioned GitHub releases outrank HTML, and omitted changelog tails are not filled in. Use For developers as individual upgrade and pin actions; use For teams for permission, spend, and fleet policy. Follow linked open proposals when you want a concrete next step; ignore What to ignore. Re-read on the next window — 23 assistants are tracked, and this brief is only the 14 signals that cleared the current cut.

Evidence appendix

• Evidence window: top curated signals (versioned releases preferred)
• Assistants tracked: 23
• Signals in this brief: 14
• Trusted source mix (GitHub/changelog vs HTML): 49%
• Open improvement proposals: 8
• Sources reporting errors: 1

High-signal releases used

• Goose v1.48.0: TrustedRouter declarative provider #11422 · OpenCode Zen gateway as a declarative provider [#10939](https://gi — source

• Claude Code v2.1.260: Added a diff panel that opens beside the conversation in fullscreen mode and shows your uncommitted changes as Claude edits; toggle it with /diff · Added a li — source

• Zed AI v1.19.0-pre: This week's release includes call hierarchy support, multi-select in the Git Panel, automatic language detection for untitled buffers, and project search on typ — source

• Google Antigravity 1.1.25: Added an opt-in workspace-grouped view to the /resume conversation picker, allowing users to toggle between a flat list and conversations grouped by directory — source

• OpenAI Codex CLI rust-v0.153.0: Vim mode now supports undo with u and redo with Ctrl+R, preserving complete drafts including pasted content and attachments. ([#41941](https://github.com/op — source

• Cline cli-v3.0.61: Cline now handles a running Hub that is older than your CLI. · Instead of quietly talking to a hub executing stale code, you get a prompt showing how many activ — source

• Cline sdk/sdk/v0.0.82 — SDK v0.0.82: Fixed tool calling being silently disabled for gateway models whose catalog entry declares no capabilities. · Three separate producers built gateway model defin — source

• OpenAI Codex CLI rust-v0.152.0: Vim mode supports / and ? searches within drafts, highlighted matches, and repeat navigation with n and N. (#41586) · Rate-limit banners offer actions f — source

• OpenCode v1.18.26: Claude 5 sessions now tolerate stale thinking blocks instead of failing after prompt or tool changes. · Bedrock GPT-5.6 models now accept none reasoning effor — source

• Kiro v2.19.2: 2.19.2 Aug 25, 2026 · 2.19.1 Aug 21, 2026 Spec Review Mouse Support and Automatic Stream Recovery This release brings mouse support to the spec review screen. · — source

• Tabnine v6.5.4: v6.5.4 August 30, 2026 Bug Fixes Tabnine CLI: Provenance & Attribution : Fixed attribution feedback being dropped after tool inputs were edited via VS Code diff — source

• Factory Droid v0.209.0 — · Desktop v0.166.0 (September 1, 2026): Session archiving from chat, a new default model, faster startup, and steadier app views · Archive sessions from chat - Archive the current session with /ar — source

• Factory Droid v0.208.1 — · Desktop v0.165.1 (August 29, 2026): 60% off GPT and Grok models · GPT and Grok discount - GPT and Grok models now bill at 60% off for a limited time, with the discount shown in the model selec — source

• Kiro 2.20.2 — Aug 31, 2026: 2.20.2 Aug 31, 2026 — source

Open proposals

• Toggle /diff in fullscreen to watch uncommitted changes as Claude edits (audience: developer, P5)
• Enable effort-based reasoning for OpenRouter models in Zed Agent (audience: developer, P4)
• Pin Cline models after the v3.0.61 switch to Claude Fable 5.1 (audience: everyone, P4)
• Pin Cline provider models before SDK v0.0.82 switches 57 defaults to Claude Fable 5.1 (audience: everyone, 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)
• Archive Factory Droid chats with /archive and default new sessions to GPT-5.6 Sol (audience: team, P3)
• Enable tools.enableWebFetchTool and upgrade Tabnine CLI to 0.35.0 (audience: developer, P3)