Skip to content
Guidance/Daily Briefing
digesteveryonedigestevidenceradarmodelcli11 min read · Sep 1, 2026

Radar brief — 2026-09-01

Restricted sandboxes, declarative providers, and agent dashboards land together — MCP and hooks are the real surface area

What the evidence shows

GitHub / changelog tags (preferred):

Goosev1.48.0
Key Signal & Impact
Declarative providers (TrustedRouter, OpenCode Zen, Gondola, SayGM, Lynkr, PleumRouter); GLM-5.2 → GLM-5.3; Bedrock GPT-5.6 routing; on_failure PreToolUse hooks and stable tool_call_id. High blast radius; 98 changes omitted from the tag body.
Claude Codev2.1.248
Key Signal & Impact
--restricted / CLAUDE_CODE_RESTRICTED=1 strips shell/code/WebFetch, refuses bypassPermissions, ignores user/project/local settings. Plus per-agent experimental.cacheTtl, self-hosted runner labels, settings diagnostics, Enterprise /usage-credits. Tagged breaking.
OpenAI Codex CLIrust-v0.149.0
Key Signal & Impact
codex agents dashboard, codex queue, /cd /pwd /cwd, deeper Vim, richer codex doctor. Resumed/forked threads restore the saved permission profile instead of silently using current defaults. 151 changes omitted.
OpenAI Codex CLIrust-v0.152.0
Key Signal & Impact
Package-style MCP names (: @ / .); per-tool output_token_limit; Bedrock/credential refresh in TUI/exec; approval reviews keep instructions across compaction; MCP tools survive cache/plugin refresh. 94 changes omitted.
Google Antigravity1.1.23
Key Signal & Impact
Subagent stream overhead cut; hanging models/agents subcommands fixed; hook panics no longer crash the CLI; Gemini tool-call IDs restored; subagents with enable_mcp_tools=true get the MCP dispatcher again.
Zed AIv1.17.2
Key Signal & Impact
CSV/TSV/PSV/SSV tabular previews; Git blame/stash; lower memory on large files; low reasoning effort for DeepSeek V4 Flash and V4 Pro. 52 changes omitted.
Clinesdk/sdk/v0.0.76
Key Signal & Impact
Image generation in-session; workspace-scoped scheduled tasks + durable todos; skill slash commands load via the skills tool (transcript stores /my-skill, not SKILL.md). Hooks: PreToolUse context reaches the model again; PostToolUse is awaited (120s). Tagged breaking.
Clinedesktop-v0.0.15-beta.1
Key Signal & Impact
Prerelease /handoff to Cline Cloud (needs pushed repo/branch/commit). Cloud sits in the Local/Remote menu behind a flag. Interrupted handoff retries. Not a production control plane.
OpenCodev1.18.22
Key Signal & Impact
Small, concrete: drop stale Go discount copy; device-login URL fixes; stop sending textVerbosity to OpenAI-compatible providers that reject it; Bedrock compatibility.

HTML / docs changelogs (noisier; several bundle older versions):

Kirov2.19.2
Key Signal & Impact
Spec-review mouse support; stream idle watchdog, retries, 60‑minute stream timeout (api.streamIdleSoftTimeout / HardTimeout / api.timeout). Notes also mix 2.19.1 and older IDE permission text.
Kiro2.20.1
Key Signal & Impact
Empty patch page (date only). Do not infer features.
Windsurfv3.8.20
Key Signal & Impact
Agent Command Center follows the selected space; devin.agentWindow.location; faster Agent/Editor switch; untrusted-workspace warning before local agents.
Tabninev6.5.4
Key Signal & Impact
Attribution feedback no longer dropped after diff/external edits; direct web fetch when tools.enableWebFetchTool is on (avoids LLM-mediated hallucination). CLI 0.35.0. Body also recaps 6.5.3/6.5.2.
Factory Droidv0.200.0
Key Signal & Impact
Confirm before mutating the tree after a failed Git pull; MCP stays authenticated across reconnects; background commands report real PIDs; subagents can inherit session reasoning (app).
Factory Droidv0.208.1
Key Signal & Impact
Time-limited 60% off GPT and Grok in the selector — commercial copy, not a capability change.

Themes in this window: model, cli, and agent dominate; security, ide, mcp, and breaking are the ones that change how you run these tools tomorrow.

For developers

  • Untrusted or CI-like sessions: upgrade Claude Code to v2.1.248 and actually run --restricted (or CLAUDE_CODE_RESTRICTED=1). File tools stay inside the working directory; shell, code execution, and WebFetch are gone unless you name them in --tools. Do not expect user/project/local settings or bypassPermissions to apply.
  • Codex as a task board, not a single chat: on rust-v0.149.0+, use codex agents to start/stop/rename, codex queue to poke existing local/remote sessions, and /cd /pwd /cwd instead of hoping the TUI cwd matches the repo. On rust-v0.152.0, cap noisy MCP tools with output_token_limit and allow : @ / . in MCP server names. After resume/fork, check the restored permission profile — it is no longer silently today’s default.
  • Goose providers: v1.48.0 is a provider pack. If you already terminate on TrustedRouter, OpenCode Zen, Gondola, SayGM, Lynkr, or PleumRouter, register them as declarative providers rather than one-off OpenAI-compatible hacks. Budget for GLM-5.3 replacing GLM-5.2 and for custom provider cost fields feeding cost tracking. Hook authors: on_failure on PreToolUse and stable tool_call_id across the tool lifecycle.
  • Antigravity 1.1.23 is a fix release you want if you use subagents + MCP. enable_mcp_tools=true subagents were failing with unknown tools; Gemini history dropped tool-call IDs; models/agents could hang on an inherited stdin pipe; prompt-hook panics crashed the CLI. Tab-complete /model names while you are there.
  • Cline SDK v0.0.76 changes skill and hook contracts. Slash skills go through the skills tool (transcript keeps what you typed). PostToolUse is no longer fire-and-forget — it waits up to 120s and honors contextModification / cancel. Provider-executed tools now show up as observational events instead of silent workspace edits. Desktop /handoff is beta: only after the repo is pushed, and only if you accept a cloud workspace that keeps running after the app closes.
  • Zed v1.17.2: turn on low reasoning effort for DeepSeek V4 Flash / V4 Pro when you want cheaper/faster turns; use the new tabular previews instead of opening large CSV/TSV in a buffer.
  • Tabnine: set tools.enableWebFetchTool and upgrade Tabnine CLI to 0.35.0 so fetch is direct, not an LLM paraphrase. If you rely on provenance, this is the attribution-feedback fix after VS Code diff or external-editor acceptance.
  • Windsurf v3.8.20: set devin.agentWindow.location if the Agent Command Center should be a separate window; trust the workspace when the sidebar warns, or local agents stay off.
  • Kiro: if streams die mid-turn, tune api.streamIdleSoftTimeout, api.streamIdleHardTimeout, and api.timeout on 2.19.x. Do not plan around 2.20.1 until the changelog has a body.

For teams

  • Policy default, not a blog toggle: Claude Code --restricted is the first clearly versioned “no shell, no code exec, no WebFetch, no settings files, no bypass” switch in this set. Use it for contractors, ephemeral runners, and anything that should not inherit a developer’s ~/. config. Pair with the new startup /doctor /status lines when server-managed settings fail to load (Bedrock/Vertex/third-party/ANTHROPIC_BASE_URL).
  • Goose v1.48.0 is a procurement and routing change. A team proposal is already open to register TrustedRouter and OpenCode Zen as declarative providers. Do that in shared config, not in each engineer’s local provider YAML. GLM-5.2 → 5.3 is a breaking model swap; pin before rolling the tag org-wide. Custom cost fields only help if billing actually reads them.
  • Cap MCP output on Codex rust-v0.152.0. Per-tool output_token_limit plus consistent truncation on resume is the control you want before an MCP server floods context. The matching team proposal is to enable update_plan and set that cap. Also: resumed threads restoring permission profiles means a “safe” resume can re-arm a looser profile you thought you left behind — audit profiles in the golden image, not at the keyboard.
  • Do not standardize on Cline Cloud handoff yet. desktop-v0.0.15-beta.1 is a feature-flagged beta; preflight requires a pushed commit. SDK hook/skill behavior in v0.0.76 is worth a team notes pass because it changes transcripts, observability, and hook timing.
  • Factory Droid v0.200.0: failed Git pull now asks before rewriting the working tree. That is the right default for shared repos; keep it on. MCP reauth-across-reconnect reduces prompt fatigue but also means a stale grant can persist — rotate MCP credentials in one place.
  • Enterprise billing UX: Claude Code /usage-credits for AWS Marketplace / self-serve Enterprise / trials; Windsurf adds an ACU-limit request link. These are process features, not capacity.
  • Tabnine 6.5.3 notes in the 6.5.4 doc: org-level repository add, and a fix so per-team model assignment actually applies. If you run multi-team Tabnine, verify team model policy after upgrade — the previous build could ignore it.
  • Windsurf untrusted-workspace warnings are the desktop equivalent of --restricted: train people not to click Trust on random folders just to silence the agent sidebar.

What to ignore

  • Factory Droid v0.208.1 “60% off GPT and Grok” and OpenCode’s removal of outdated Go first-month discount copy. Pricing banners are not capability evidence.
  • Kiro 2.20.1 — date stamp, no body. Also distrust the 2.19.2 HTML page where 2.19.1 and even older “1.0.337 IDE permission” notes are concatenated.
  • Anything in bodyOmitted. Goose 98, Codex 151 + 94, Claude Code 43, Zed 52, Cline SDK 11, Antigravity 3, plus smaller HTML omissions. Do not reverse-engineer those as roadmap.
  • Cline desktop-v0.0.15-beta.1 as a platform decision. Prerelease, flag-gated, cloud side-effects after the laptop lid closes.
  • HTML changelogs as complete. Trusted GitHub/changelog mix is only 49% this brief; Kiro, Windsurf, Tabnine, and Factory are bundled, dated August 2026 pages that mix patches. One source reported errors — do not debug from those pages.
  • Grok CLI work is only in open proposals, not in this evidence set. No Grok CLI version is cited here.

Watch next

  • 8 open improvement proposals, including: register TrustedRouter + OpenCode Zen on Goose v1.48.0 (team, P5); Codex update_plan + MCP output_token_limit (team, P4); Kiro /spec run in full-screen task view (developer, P4 — evidence for 2.20.1 is empty, so wait); Zed low-reasoning DeepSeek V4 (developer, P4); Tabnine enableWebFetchTool + CLI 0.35.0 (developer, P3); Windsurf devin.agentWindow.location (developer, P3). Two Grok CLI prompt/MCP cleanup proposals have no matching release in this JSON.
  • Whether Claude --restricted becomes the CI default or stays a flag people forget.
  • Whether Goose’s provider wave is durable routing or a pile of one-off declarative wrappers; GLM-5.3 and GPT-5.6-on-Bedrock need production traffic, not tag notes.
  • Codex 0.149 → 0.152 omitted-change volume (245 combined unpublished bullets). Permission-profile restore and MCP truncation are the parts that will bite in incident review.
  • Cline Cloud leaving the feature flag, and whether SDK hook awaiting (120s) shows up as latency in real PostToolUse packs.
  • Antigravity MCP dispatcher staying available to subagents after the 1.1.23 fix — that class of bug tends to regress.
  • Fill-in of Kiro 2.20.1 and a cleaner split of Kiro HTML notes so 2.19.x is not mixed with unrelated IDE patches.

How to use DevAgentRadar

DevAgentRadar ranks versioned evidence, not launch posts. Prefer the GitHub tags in this brief over HTML changelogs; if a tag sets bodyOmitted, stop at the listed bullets. Track 23 assistants, but act only on the 15 signals here.

Developers: take the concrete flags and commands above (--restricted, codex agents / queue, MCP output_token_limit, Goose declarative providers, Tabnine tools.enableWebFetchTool, Windsurf devin.agentWindow.location, Zed DeepSeek low reasoning). Teams: use the open proposals as the change-control queue (P5 Goose providers and P4 Codex MCP caps first), and treat breaking/prerelease tags as rollout gates. Ignore discounts, empty patches, and omitted tails. Re-read the next brief when those omitted GitHub bodies land as their own versioned notes — not before.


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

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

  • Claude Code v2.1.248: Added --restricted (or CLAUDE_CODE_RESTRICTED=1): removes the built-in tools that run commands or code and WebFetch (unless named in --tools), keeps fil — source

  • OpenAI Codex CLI rust-v0.149.0: Added an interactive codex agents dashboard for searching, starting, opening, renaming, and stopping tasks, with configurable shortcuts. (#39094, #39112, #391 — source

  • Google Antigravity 1.1.23: Improved /model <name> autocompletion to accept the proposed model name ghost text with Tab. · Reduced subagent streaming overhead by sending subagent traje — 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

  • Zed AI v1.17.2: This week's release includes tabular data previews for CSV, TSV, PSV, and SSV files with sortable and resizable columns, value-based row filtering, and right-cl — source

  • Cline sdk/sdk/v0.0.76 — SDK v0.0.76: Added model-driven image generation. · Models that support it can generate images during a turn, and generated images are persisted in session history and expor — source

  • Cline desktop-v0.0.15-beta.1 — beta.1: Beta: hand off a local session to Cline Cloud with /handoff — the conversation, attached images, and an optional follow-up command move to a cloud workspace t — source

  • OpenCode v1.18.22: Removed outdated OpenCode Go first-month discount messaging and pricing. · Fixed OpenCode device login links when servers return relative verification URLs or u — 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

  • Windsurf v3.8.20: v3.8.20 August 21, 2026 Devin Desktop The Agent Command Center now follows whichever space is selected instead of being tied to one folder, converts an editor w — 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 v0.200.0 — · Desktop v0.157.0 (August 20, 2026): Background processes report real process IDs, MCP connections stay authenticated across reconnects, and the model list is searchable in the app · **Failed pull — 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

Open proposals

Was this guidance useful?

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