Skip to content
Guidance/Daily Briefing
digesteveryonedigestevidenceradarmodelcli9 min read · Sep 15, 2026

Radar brief — 2026-09-15

GPT-6-Astra lands in Codex and Goose while coding CLIs ship plugin eval, experimental worktrees, and trust gates — treat HTML changelogs as noisier than GitHub tags.

What the evidence shows

Window: 15 curated signals across 23 tracked assistants (49% GitHub/changelog vs HTML). Prefer the versioned GitHub tags below; several HTML notes mix multiple versions and omitted bodies. One source reported errors, so this is not a complete census. Large bodyOmitted counts mean we only use listed bullets.

GitHub-tagged releases (higher confidence)

Claude Codev2.1.269
Key Signal & Impact
claude plugin eval scores a plugin suite (JSON + HTML). /output-style works in Remote Control, cloud, and headless. Bash file-edit diffs when bashEditDiffEnabled. New knobs: OTEL_METRICS_INCLUDE_REPOSITORY, CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS, CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1–256). Cache-reuse fixes after token-limit cutoff and mid-thought interrupt. 97 further changes omitted.
OpenAI Codex CLIrust-v0.154.0
Key Signal & Impact
GPT-6-Astra in the model picker and Amazon Bedrock catalogs. Experimental isolated checkouts via --worktree / /worktree. Inline answers while the agent keeps working. Windows can share a background Codex server. MCP OAuth refresh is coordinated (no auto-replay of rejected tools). Startup skips workspace-controlled helpers until trust; macOS sandbox blocks terminal input injection. Remote resume/fork keep saved permissions. 147 further changes omitted.
OpenAI Codex CLIrust-v0.153.0
Key Signal & Impact
Plugin CLI can list/install/remove from remote marketplaces. Vim undo/redo on full drafts. tui.auto_recap = false. TUI reconnect after app-server drop. Plus/Team get an earlier usage warning. Full Access skips Guardian reviews for confirmation-only actions. 95 further changes omitted.
Goosev1.50.0
Key Signal & Impact
GPT-6 Astra model support. Tool calling for goose-agent. HTTPS required for Snowflake. CLI session naming bound to a trusted path. Subagent platform guards, preserved permission revocations, content-capture opt-out in tracing, prefer latest MCP version. 7 further changes omitted.
Google Antigravity1.2.1
Key Signal & Impact
Custom agents can set excludeDefaultComponents: true in Markdown frontmatter (keep post-invocation hooks). Transient 502/503/504/per-minute 429 and mid-stream interruptions retry in-process. MCP/provider schemas keep open objects. --continue no longer starts a new conversation from a subdirectory/crash. Remote Control companion UI shows the signed-in session.
Clinedesktop-v0.0.27
Key Signal & Impact
Expired sign-in is one actionable error (with Sign in / Open model settings). Sign-out of Cline and Cline Pass now sticks; the runtime had been re-importing classic-extension credentials.
Clinedesktop-v0.0.26
Key Signal & Impact
Composer shows the current branch’s GitHub PR (number, merge, diffs, CI) if gh is signed in to a GitHub.com origin. Skills can be disabled in-place. Tabs: Tools, Plugins, Skills, Rules, MCP, Hooks. Does not push or open the PR for you. 36 further changes omitted.
OpenCodev1.18.26
Key Signal & Impact
Claude 5 sessions tolerate stale thinking blocks after prompt/tool changes. Bedrock GPT-5.6 accepts none reasoning effort. apply_patch no longer emits an empty move path in permission metadata. Azure CLI sign-in asks for the resource name directly.
Zed AIv1.19.0-pre
Key Signal & Impact
Prerelease. Call hierarchy, Git Panel multi-select, untitled-buffer language detection. OpenRouter effort-based reasoning. Lower memory on long Agent sessions; better OpenAI-compatible streaming. 50 further changes omitted.

HTML changelogs (noisier; 49% mix)

Windsurfv3.10.27
Key Signal & Impact
Desktop: drag-to-composer and SSH-after-reload fixes; Devin Fusion can activate without Claude Fable 5.1 as lead. Same page also recaps v3.10.23: SSH remote agent hosts, ACP always on, several legacy Cascade entry points removed.
Kirov2.21.4 (and overlapping 2.21.3)
Key Signal & Impact
V3 /sessions dashboard (local + cloud), /config panel for agents/MCP/Powers/Steering/Skills/Hooks, and optional apply-cloud-config-to-local-CLI without writing .kiro.
Factory Droidv0.217.0 / v0.215.0
Key Signal & Impact
MCP allowlist autocomplete; org managers can remove Web Search and block individual models; terminal streaming of replies/thinking; oversized-request retries.
Tabninev6.5.4
Key Signal & Impact
CLI provenance/attribution no longer dropped after diff/external edits. Direct web fetch is default when tools.enableWebFetchTool is on. Same note also covers 6.5.3 org-level repos and per-team model assignment.

For developers

  • Claude Code v2.1.269: run claude plugin eval --help on any plugin you ship; use /output-style in headless/Remote Control; turn on bashEditDiffEnabled if you need Bash edit diffs; raise CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS only for inference-bound fan-out (cap 256); extend CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS if /v1/models discovery flakes (default 3s).
  • Codex rust-v0.154.0: select GPT-6-Astra in the picker or Bedrock catalog; treat --worktree / /worktree as experimental isolated checkouts; use inline Q&A so the main draft is not lost; on Windows, learn the shared background-server lifecycle. From 0.153.0, set tui.auto_recap = false if auto recaps are noise; use the plugin CLI against remote marketplaces.
  • Goose v1.50.0: point at GPT-6 Astra; expect Snowflake to require HTTPS; do not assume arbitrary CLI session names still work (bound to a trusted path).
  • Antigravity 1.2.1: add excludeDefaultComponents: true only when you want a custom agent without default prompt sections/built-in tools; prefer --continue after crashes/subdirectories — it should resume the latest non-empty workspace conversation.
  • Cline desktop-v0.0.27 / v0.0.26: if sign-out bounced you back in, sign out again on 0.0.27 (classic-extension re-import was the bug). PR row needs gh + GitHub.com origin; it will not push or file the PR. Re-enable previously “missing” skills — they can now be disabled in place instead of hidden.
  • OpenCode v1.18.26: Claude 5 sessions should survive stale thinking blocks; Bedrock GPT-5.6 can use none reasoning effort; Azure login now wants the resource name, not a management-API probe.
  • Zed v1.19.0-pre: useful for long agent sessions and OpenRouter reasoning effort, but it is a prerelease — do not standardize on it yet.

For teams

  • Trust before workspace code: Codex 0.154.0 refuses workspace-controlled helpers until trust and blocks macOS terminal input injection. Goose binds session naming to a trusted path, keeps permission revocations across managers, and enforces subagent platform guards. Review those defaults before rolling out.
  • MCP and plugins are becoming operational surfaces: Codex coordinates MCP OAuth refresh without replaying rejected tools, and 0.153.0 can install plugins from remote marketplaces. Goose prefers the latest MCP version. Antigravity keeps open object schemas ({"type":"object"} / additionalProperties: true). Factory 0.217.0 autocompletes MCP allowlist entries — treat that as an access-control change, not just UX.
  • Org policy knobs in the HTML feeds (verify in-product): Factory 0.215.0 lets managers remove Web Search and block individual models while “Allow All” still admits new ones. Tabnine 6.5.3 (bundled in the 6.5.4 note) adds org-level repositories and per-team model assignment; 6.5.4 makes web fetch direct when enabled. Kiro V3 can load cloud Steering/agents/Skills/Powers/Hooks into local CLI sessions without writing .kiro — decide whether that is a feature or a shadow-config risk.
  • Auth and telemetry: Cline 0.0.27 is the sign-out/credential-hygiene fix (see open proposal). Claude Code can tag OTEL with vcs.* when OTEL_METRICS_INCLUDE_REPOSITORY is set. Goose honors tracing content-capture opt-out.
  • Breaking / policy watch: Windsurf’s HTML recap says ACP is always enabled and some legacy Cascade entry points are gone. Codex Full Access skipping Guardian on confirmation-only actions is a control-plane change — confirm that matches your review policy. Plus/Team Codex users get earlier usage warnings in ~five-hour windows (0.153.0).

What to ignore

  • Omitted bodies. Do not infer the +97 / +147 / +95 / +50 / +36 unnamed changes.

  • Zed v1.19.0-pre. Prerelease; editor features here are not a model-quality signal.

  • Stacked HTML pages. Windsurf v3.10.27 also dumps v3.10.23. Kiro 2.21.4 and 2.21.3 repeat the same V3 dashboard story. Tabnine 6.5.4 embeds 6.5.3–6.5.1. Factory notes mix Desktop app and CLI.

  • Marketing model names without a picker/changelog line. GPT-6-Astra is evidenced only for Codex CLI and Goose. “Claude Fable 5.1” appears only as a Windsurf Fusion availability caveat. OpenCode’s Claude 5 / Bedrock GPT-5.6 lines are session-compat fixes, not new model launches.

  • Low-weight proposals. OpenCode PR #38252 verification videos (P2) and Grok CLI prompt/MCP proposals have no matching release in this evidence JSON — do not treat them as shipped.

  • One source reported errors. Do not assume missing assistants were quiet; coverage is incomplete.

Watch next

  • Whether Codex worktree stays experimental and whether GPT-6-Astra quality matches picker availability (proposal).
  • Whether claude plugin eval becomes a plugin quality gate (JSON/HTML scores).
  • Cline sign-out vs classic-extension credential re-import (proposal).
  • Windsurf SSH remote-agent hosts from the location selector (proposal).
  • Kiro “apply cloud configuration to local sessions” leaking org Steering/Skills/Hooks onto laptops (proposal).
  • MCP OAuth refresh, open-object schema validation, and permission-metadata correctness (apply_patch empty move path).
  • Seven open improvement proposals remain on the board; Grok CLI items still lack a versioned GitHub release in this brief.

How to use DevAgentRadar

DevAgentRadar is a public, evidence-first feed for AI coding assistants: versioned GitHub releases over blogs, listed facts only, no invented ships. Use For developers as local CLI/IDE actions (upgrade, flags, eval commands). Use For teams as trust, MCP, identity, and org-policy checks. Discount HTML changelogs, prereleases, and omitted bodies. Track the linked proposals when you want a concrete follow-up rather than another marketing recap.


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 v2.1.269: Added claude 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 v1.50.0: Support GPT-6 Astra models #11869 · Tool calling for goose-agent [#11627](https://github.com/aaif-goose/goose/ — 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 for excludeDefaultComponents: true in 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 with u and redo with Ctrl+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 accept none reasoning effor — source

  • Windsurf v3.10.27: v3.10.27 September 15, 2026 Devin Desktop Dragging files from the Explorer onto the agent panel, or sending an editor selection to chat, adds them to the compos — 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

  • Factory Droid v0.217.0 — · Desktop v0.174.0 (September 11, 2026): Browser shortcuts, smoother command animations, cleaner retries, and clearer organization controls · Smoother command animations - Long-running commands now — source

  • Factory Droid v0.215.0 — · Desktop v0.172.0 (September 8, 2026): Streaming terminal responses, organization Web Search control, unified connector settings, and steadier sessions · Stream responses as they arrive - A new t — 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

Open proposals

Was this guidance useful?

Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.