Radar brief — 2026-09-12
CLI coding agents shipped plugin evals, isolated worktrees, and GPT-6-Astra — while MCP auth, session resume, and a few hard UI removals matter more than the model names.
What the evidence shows
GitHub release notes carry the load this window (Claude Code, Codex CLI, Goose, Cline, Antigravity, OpenCode). HTML changelogs for Windsurf, Kiro, Tabnine, and Factory Droid are noisier: duplicate Kiro pages, mixed older patches, and one source in the mix reported errors. Several high-score notes omit most of the changelog (bodyOmitted 97–147), so only the listed items below are fair game.
claude plugin eval scores a plugin with JSON + HTML; /output-style works in remote/headless sessions; Bash edit diffs; OTEL vcs.* tags; workflow concurrent-agent cap 1–256; prompt-cache resume fixes.--worktree / /worktree; inline Q&A while the agent keeps going; Windows shared background server; MCP OAuth refresh without replaying rejected tools; no workspace helpers until trust; macOS sandbox blocks terminal input injection.tui.auto_recap = false; TUI reconnect after app-server drops; Plus/Team usage warning; Full Access skips Guardian on confirmation-only actions.on_failure PreToolUse hooks and stable tool_call_id.excludeDefaultComponents: true; retries for 502/503/504/429 and mid-stream cuts; MCP schemas keep open objects; --continue no longer starts a fresh chat from a subdirectory/crash.gh + GitHub.com origin required); Skills can be disabled in-place; sidebar New Task → New Session.none reasoning; apply_patch no longer sends an empty move path in permission metadata./sessions dashboard, /config panel, and optional cloud config applied to new local CLI sessions without writing .kiro.tools.enableWebFetchTool is on./archive; CLI starts even if org settings are slow. Prior note: limited-time 60% off GPT and Grok.For developers
- Score plugins, don’t eyeball them. On Claude Code v2.1.269 run
claude plugin eval --helpand keep the JSON/HTML report next to the plugin. Use/output-stylein remote/cloud sessions so logs match local ones. Turn on Bash edit diffs (bashEditDiffEnabled) when the agent is rewriting files through the shell. - Isolate Codex experiments. rust-v0.154.0: pick GPT-6-Astra only if you want that catalog entry; use
--worktree//worktreefor forked sessions instead of dirtying the main checkout. rust-v0.153.0: install plugins via the marketplace CLI, settui.auto_recap = falseif recaps eat the buffer, and expect Vimu/Ctrl+Rto keep pasted drafts. - Upgrade Cline CLI before MCP. cli-v3.0.61 is a reliability patch: stale Hub prompt (Enter replaces, Esc keeps), 10s MCP connect budget, and restored tools/images for catalogs that advertised empty capabilities. Desktop-v0.0.26 PR/CI in the composer needs
ghsigned in and a GitHub.comorigin; Cline still will not push or open the PR for you. Install that CLI if you want the live PR row. - Pin Goose providers and hooks. v1.48.0 is a routing release: GLM-5.3 replaces 5.2, GPT-5.6 can go through Bedrock, and PreToolUse gains
on_failureplus a stabletool_call_id. Treat new declarative providers as config you review, not defaults you inherit. - Antigravity / OpenCode / Zed. Antigravity 1.2.1:
excludeDefaultComponents: truefor skinny custom agents;--continuenow prefers the last non-empty workspace conversation. OpenCode v1.18.26: Claude 5 thinking-block failures after prompt/tool edits should stop; Bedrock GPT-5.6 can usenoneeffort. Zed v1.19.0-pre is a preview — try OpenRouter effort and long-session memory, don’t standardize on it. - Kiro V3 hygiene.
/sessions(orkiro-cli chat --sessions) to resume/clean local and cloud sessions;/configto see whether agents/MCP/Skills/Hooks are local, cloud, or both. Only enable “Apply your cloud configuration to local sessions” if you want cloud Steering/agents/Skills without writing them into local.kiro. - Windsurf muscle memory. Saved SSH hosts and transcript reconnect are the feature. Drop Cascade-era commands (new-agent option, annotations, conversation mentions, Vibe & Replace, Deploy, workflow commands, automatic Lifeguard). ACP is no longer optional.
For teams
- Treat Codex rust-v0.154.0 as a trust + MCP release, not just a model drop. Startup now waits for trust before workspace-controlled helpers; macOS sandbox blocks terminal input injection; MCP coordinates OAuth refresh and will not auto-replay rejected tools. Remote resume/fork keeps saved permissions; forks follow server model defaults unless overridden. Decide explicitly whether GPT-6-Astra is an org default. rust-v0.153.0 Full Access skipping Guardian on confirmation-only actions is a policy change — don’t assume reviews still fire.
- Instrument Claude Code before raising concurrency.
CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS(1–256) is for inference-bound fan-out. Pair it withOTEL_METRICS_INCLUDE_REPOSITORY/OTEL_LOG_TOOL_DETAILSso spend and tool events carryvcs.*.CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MSonly if/v1/modelsis slow (default 3s). Useclaude plugin evalas a gate for internal plugins. - Goose routing is now a cost and safety surface. Declarative providers plus custom cost fields mean finance can track non-catalog models; GLM-5.3 is a breaking model swap from 5.2. Review PreToolUse
on_failurebefore rolling hooks org-wide. - Cline Hub and Windows. Require cli-v3.0.61 so old Hubs drain in-flight turns instead of serving stale code, and so Windows binaries are Authenticode-signed (blocked launches now print a real error). Desktop PR status is
gh-only and GitHub.com-only. - Windsurf + Kiro config sync are team decisions. Windsurf: MCP auth opens the provider; Docker MCP from the registry should launch; cloud “Allowlist N domains” matches the web app. Archiving a cloud session with open PRs offers to close them and archives children. Kiro: cloud → local config apply avoids writing files locally — useful for managed Steering/Skills, risky if laptops should stay air-gapped from cloud agent defs.
- Factory default model moved. v0.209.0 sessions with no model selected now land on GPT-5.6 Sol; CLI no longer blocks startup on slow org settings (settings apply when they arrive). The v0.208.1 60% GPT/Grok discount is a billing promo, not a capability change. Credit exhaustion now suggests a Droid Core model instead of a retired one.
- Antigravity MCP schemas. Open object schemas (
{"type":"object"}/additionalProperties: true) are no longer rejected for extra args — tighten tool definitions if you relied on that accidental strictness.
What to ignore
- Omitted changelog bulk. Claude (+97), Codex 0.154.0 (+147) / 0.153.0 (+95), Goose (+98), Zed (+50), Cline desktop (+36) / CLI (+18) are not extra features you can plan around.
- Duplicate / HTML noise. Two Kiro 2.21.3 pages tell the same V3 story; Tabnine’s 6.5.4 note also dumps 6.5.3 and 6.5.2; Factory’s discount note is marketing. Trusted GitHub/changelog mix is only 49%, and one source reported errors — don’t treat Windsurf/Kiro/Tabnine/Factory HTML as equivalent to the GitHub tags above.
- Zed v1.19.0-pre as a team baseline. Editor + agent polish, still a prerelease.
- Open proposals that name versions not in this evidence (Goose v1.50.0, a Cline “Claude Fable 5.1” pin, Grok CLI prompt-tag rename). This brief will not invent those releases.
Watch next
- Whether Codex
--worktreestays experimental and whether GPT-6-Astra becomes the Bedrock default rather than a picker entry. - Whether
claude plugin evalreports become an actual plugin CI gate, and whether the omitted Claude/Codex “breaking” tags show up as follow-on notes. - Goose provider sprawl vs. cost-field tracking; GLM-5.3 regressions vs. 5.2.
- Windsurf fallout from deleted Cascade entry points and always-on ACP.
- Cline Hub drain behavior on wedged Hubs (left alone, not killed) and MCP 10s timeouts in large remote-MCP fleets.
- Eight open improvement proposals, including: run
claude plugin eval; Codex GPT-6-Astra + worktrees; Clineghfor PR/CI; Windsurf saved SSH hosts; Kiro cloud-config-on-local.
How to use DevAgentRadar
DevAgentRadar ranks versioned evidence (prefer GitHub tags and changelogs) across coding assistants and writes a public brief from what those notes actually say — not launch posts. Use For developers as local CLI/IDE actions, For teams as defaults, trust, MCP, and billing policy. Ignore omitted changelog tails, duplicate HTML pages, and any proposal that cites a version this brief does not. Re-check the radar when a tagged breaking, surface:mcp, or security release lands; 23 assistants are tracked, this window used 15 signals.
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
v2.1.269: Addedclaude plugin eval: run a plugin's eval suite against Claude Code and get scored, reproducible results (JSON + HTML report); see `claude plugin eval --h — source -
OpenAI Codex CLI
rust-v0.154.0: GPT-6-Astra is now available in the model picker and Amazon Bedrock catalogs. (#42879, #42619) · Experimental worktree support lets you create isolated checkout — source -
Goose
v1.48.0: TrustedRouter declarative provider #11422 · OpenCode Zen gateway as a declarative provider [#10939](https://gi — 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 -
Cline
desktop-v0.0.26: The composer now shows the current branch's GitHub pull request — PR number, merge status, changed-line totals, and CI checks. · Click through to open it in you — source -
Google Antigravity
1.2.1: Added support forexcludeDefaultComponents: truein custom agent Markdown frontmatter, allowing custom agents to opt out of default prompt sections and built- — source -
OpenAI Codex CLI
rust-v0.153.0: Vim mode now supports undo withuand redo withCtrl+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 -
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 acceptnonereasoning effor — source -
Windsurf
v3.10.23: v3.10.23 September 10, 2026 Devin Desktop Remote agent hosts: connect to agents over SSH from the location selector, with saved hosts, status dots, and a server — source -
Kiro
v2.21.3: 2.21.3 Sep 10, 2026 · 2.21.2 Sep 8, 2026 · 2.21.1 Sep 3, 2026 Session Dashboard, Configuration Panel, and Cloud Sync Browse and manage local and cloud sessions, — source -
Kiro
2.21.3— Sep 10, 2026: September 1, 2026 Session Dashboard, Configuration Panel, and Cloud Sync · 2.21.0 · Browse and manage local and cloud sessions, inspect configured resources, an — 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
Open proposals
- Install gh CLI so Cline composer shows live PR and CI status (audience: developer, P5)
- Add saved SSH agent hosts in Windsurf and drop Cascade commands (audience: team, P5)
- Run
claude plugin evalto score plugins with JSON and HTML reports (audience: team, P5) - Upgrade Goose to v1.50.0 and enable GPT-6 Astra with goose-agent tool calling (audience: team, P5)
- Switch Codex CLI to GPT-6-Astra and isolated
--worktreecheckouts (audience: everyone, P5) - Enable Apply your cloud configuration to local Kiro CLI V3 sessions (audience: developer, P4)
- Pin Cline models after the v3.0.61 switch to Claude Fable 5.1 (audience: everyone, P4)
- Replace <output_efficiency> with <response_guidelines> in Grok CLI prompts (audience: developer, P4)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.