Radar brief — 2026-09-14
CLI coding agents shipped isolated worktrees, plugin eval, and GPT-6-Astra — while MCP auth, sign-out, and Cascade-era commands are the real upgrade tax.
What the evidence shows
Fifteen signals across 23 tracked assistants, with a 49% GitHub/changelog vs HTML mix. Prefer the versioned GitHub rows. Several releases are tagged breaking with large omitted bodies — treat listed bullets as the only confirmed facts here. One source reported errors; HTML changelogs (Windsurf, Kiro, Tabnine, Factory) are noisier and sometimes bundle older patches.
GitHub / versioned releases
claude plugin eval scores a plugin suite (JSON + HTML). /output-style works in Remote Control, cloud, and headless. Bash file edits can return a diff (bashEditDiffEnabled). Workflow concurrency is tunable (CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS, 1–256). Prompt-cache reuse fixed after token-limit cutoffs and mid-thought interrupts. OTEL can tag vcs.*; gateway model discovery timeout is configurable.on_failure, PreToolUseResult, and a stable tool_call_id.--worktree / /worktree. Inline questions while the agent keeps working. Windows can share a background Codex server. MCP coordinates OAuth refresh and login challenges (no auto-replay of rejected tools). Startup waits for trust before workspace helpers; macOS sandbox blocks terminal input injection. Remote resume/fork keep saved permissions.tui.auto_recap = false. TUI reconnect after app-server drops. Plus/Team get an earlier usage warning. Full Access skips Guardian reviews for confirmation-only actions.excludeDefaultComponents: true (keep post-invocation hooks). Transient 502/503/504/429 and mid-stream cuts retry in-process with backoff. Open MCP/provider object schemas are no longer rejected. --continue no longer starts a fresh conversation from a subdirectory/crash. Remote Control companion UIs keep the signed-in session.none reasoning effort; Bedrock replay is more reliable. apply_patch no longer emits an empty move path in permission metadata. Azure CLI sign-in asks for the resource name directly.gh + GitHub.com origin; no push/submit). Skills can be disabled in-place. Customize tabs: Tools, Plugins, Skills, Rules, MCP, Hooks. New Task → New Session.HTML changelogs (lower confidence)
/sessions (or kiro-cli chat --sessions) dashboard; /config lists agents, MCP, Powers, Steering, Skills, Hooks (local vs cloud). Optional apply of uploaded cloud config to new local CLI sessions without writing ~/.kiro.tools.enableWebFetchTool is on (avoids LLM-mediated fetch hallucinations). Notes also bundle 6.5.3 org-level repos / per-team models and a JetBrains 262 startup crash fix./archive (Ctrl+X). Default model becomes GPT-5.6 Sol when none is selected. CLI starts even if org settings are slow. Out-of-credits suggests a live Droid Core model.For developers
- If you ship Claude Code plugins: run
claude plugin evaland keep the JSON/HTML report. Use/output-stylein headless/Remote Control. Turn onbashEditDiffEnabledwhen you need Bash edit diffs. RaiseCLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTSonly for inference-bound fan-outs (cap 256). - Codex CLI: treat
--worktree//worktreeas experimental isolated checkouts — browse/resume, don’t bet a release process on it yet. Try GPT-6-Astra in the picker/Bedrock catalog. Use inline Q&A instead of aborting a long run. On 0.153, settui.auto_recap = falseif auto recaps are noise; use the plugin CLI for marketplace install/remove. - Goose 1.48.0: if you pinned Z.ai GLM-5.2, you are on GLM-5.3. Wire
on_failureon PreToolUse hooks; rely on stabletool_call_id. Declarative providers (TrustedRouter, OpenCode Zen, Gondola, SayGM, Lynkr, PleumRouter) are how new gateways land — don’t hand-roll a one-off provider unless you must. - Antigravity: set
excludeDefaultComponents: trueon custom-agent Markdown when you want a stripped tool/prompt surface but still need post-invocation hooks. Prefer--continueagain after crash/subdir; that path was wrong before 1.2.1. - OpenCode: Claude 5 + tool/prompt edits should no longer die on stale thinking blocks. On Bedrock GPT-5.6,
nonereasoning effort is valid. Recheckapply_patchpermission metadata if you automate on move paths. - Cline Desktop: after 0.0.27, sign out of Cline and Cline Pass again if the classic extension still re-imports credentials (open proposal, P5). Use 0.0.26’s PR row only with
ghauthenticated to GitHub.com; Cline will not open the PR for you. Skills can now be disabled without disappearing. - Kiro V3:
/sessionsto resume/clean local+cloud;/configbefore you assume which MCP/Skills/Hooks are in play. Enable “Apply your cloud configuration to local sessions” only if you want cloud Steering/agents/Skills/Powers/Hooks without writing them into local.kiro. - Zed
v1.19.0-preis a prerelease. Use it for Agent-session memory/streaming checks, not as a team default. - Tabnine: if
tools.enableWebFetchToolis on, you now get direct fetch — verify network policy. Don’t treat the bundled 6.5.3/6.5.2 notes as today’s sole delta.
For teams
- Trust and MCP first. Codex 0.154 delays workspace-controlled helpers until trust, blocks macOS terminal input injection, and surfaces MCP OAuth refresh/login failure instead of silently replaying rejected tools. Remote resume/fork keep saved permissions; forks follow server model defaults unless you override. Windows shared background server + managed updates need an owner.
- Windsurf/Devin runbooks break on purpose. ACP is mandatory. Cascade-era commands and UI entry points are gone. Add saved SSH agent hosts, and train on activation failures (last stderr + output channel). Cloud archive may close PRs — decide whether that is allowed. There is a team P5 proposal to add those SSH hosts and drop Cascade commands.
- Model defaults will move under people. Factory Droid sessions with no model now land on GPT-5.6 Sol. Goose routes GPT-5.6 variants on Bedrock and replaced GLM-5.2. Codex exposes GPT-6-Astra. Pin models in team config if “whatever the picker says” is not an approved spend/quality bar. Factory’s 60% GPT/Grok cut is explicitly time-limited.
- Cost and observability. Goose custom provider cost fields drive tracking — fill them in. Claude Code can tag OTEL with
vcs.*viaOTEL_METRICS_INCLUDE_REPOSITORY. Codex Plus/Team warn earlier in a ~5-hour usage window (0.153). - Identity hygiene. Cline 0.0.27: rejected refresh = signed out; sign-out must not bounce via classic-extension secrets. Antigravity Remote Control no longer shows a fake signed-out companion. Tabnine 6.5.3 notes (bundled) claim org-level repo add and per-team model assignment — verify on your tenant; this is not a GitHub release body.
- Do not execute proposals as if they shipped. Open items still include Goose v1.50.0 / GPT-6 Astra (not in this evidence), Cline CLI v3.0.61 pinning, and Grok CLI prompt/MCP cleanups. Those versions are not in the 15 signals.
What to ignore
- Omitted bodies (
+97,+98,+147,+95, …). Do not invent break lists frombreakingtags alone. Claude Code, Goose, and both Codex releases hide most of the diff. - Kiro 2.21.3 as a separate story — it repeats the 2.21.4 V3 dashboard/cloud-sync copy at lower score.
- Zed v1.19.0-pre as a stable IDE drop.
- Factory 60% off as a lasting price. Tabnine’s multi-version docs blob (6.5.4 down to 6.5.1) as a single coherent ship.
- Marketing pages and HTML changelogs when they conflict with GitHub tags. Mix is only 49% trusted changelog/GitHub; one source in this window reported errors.
- Open proposals that name unreleased evidence (Goose v1.50.0, Cline CLI v3.0.61, Grok CLI prompt/MCP). They are backlog, not radar facts.
Watch next
- Whether Codex worktrees leave experimental (
--worktree//worktree) and whether GPT-6-Astra stays picker-only or shows up in other CLIs (a Goose upgrade proposal already wants it — not evidenced). - Adoption of
claude plugin evaland whether plugin authors publish the HTML scores. - Goose declarative provider sprawl vs lock-in, plus GLM-5.3 behavior vs 5.2.
- Windsurf after Cascade removal and always-on ACP — support load, MCP Docker registry, SSH host UX.
- Cline Desktop sign-out vs classic extension re-import (developer P5 proposal).
- Kiro cloud config applied to local V3 sessions without writing
.kiro(developer P4). - Antigravity custom agents that strip default tools (
excludeDefaultComponents) and still keep hooks. - The other open P4s: Grok CLI
<response_guidelines>rename and dropping legacy managed MCP / shadowed platform skills — only if those releases enter evidence.
How to use DevAgentRadar
This brief is an evidence cut, not a vendor recap. Use GitHub-tagged versions as the source of truth; treat HTML docs as leads. Developers should copy the commands and flags above into a personal checklist (plugin eval, --worktree, /sessions, sign-out). Teams should pin models, decide MCP/OAuth and SSH policy, and update runbooks where Cascade/ACP/sign-in behavior changed. Ignore omitted +N more changes. Open proposals are optional follow-ups — they do not mean a version shipped. Re-read the next brief from scratch; older marketing claims never outrank a new release 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: 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 -
Goose
v1.48.0: TrustedRouter declarative provider #11422 · OpenCode Zen gateway as a declarative provider [#10939](https://gi — 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 -
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-v0.0.27: An expired Cline sign-in now produces one actionable error instead of two dead ends. · A turn that fails before the runtime takes the prompt was reported twice — 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 -
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
v2.21.4: 2.21.4 Sep 11, 2026 · 2.21.3 Sep 10, 2026 · 2.21.2 Sep 8, 2026 +1 more Session Dashboard, Configuration Panel, and Cloud Sync Browse and manage local and cloud — 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 -
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 -
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
- Re-sign out of Cline and Cline Pass so classic extension credentials stop re-importing (audience: developer, P5)
- Add saved SSH agent hosts in Windsurf and drop Cascade commands (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 sessions in Kiro V3 (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)
- Drop legacy managed MCP configs and shadowed platform skills in Grok CLI (audience: developer, P4)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.