Skip to content
DevAgentRadarDevAgentRadar
Sign in
Guidance/Daily Briefing
digesteveryonedigestevidenceradarmodelcli8 min read · Jul 28, 2026

Radar brief — 2026-07-28

Security patches and agent-control defaults land hard: Goose ships a critical CLI RCE fix, Claude Code splits /fork from /subtask, and Antigravity makes review-before-write the default.

What the evidence shows

Highest-signal versioned releases this window cluster on agent surface control, model catalog churn, and security hardening—not greenfield product launches.

  • Claude Code v2.1.212 (score 251): /fork now copies the conversation into a new background session (own row in claude agents); the old in-session subagent path is /subtask. Also adds claude auto-mode reset (confirm, or --yes), and a session-wide WebSearch cap (default 200, CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION) aimed at runaway loops. Tagged breaking across subagents, MCP, permissions, hooks, sandbox.
  • OpenAI Codex CLI rust-v0.145.0 (score 239): experimental paginated thread history (resume, search, names, sub-agents, memories); /import expands to Cursor + Claude Code settings, MCP servers, plugins, sessions, commands, project memories; experimental Amazon Bedrock login/custom endpoints with GPT-5.6 Sol as default Bedrock model. Hotfix rust-v0.144.6 refreshes bundled GPT-5.6 Sol/Terra/Luna instructions and sets context windows to 272,000 tokens.
  • Google Antigravity 1.1.0 (score 232): public execution mode cycling default → accept-edits → plan; request-review is now default—pauses before file writes with interactive line-level diff preview (f to review/accept/reject). Breaking on settings/commands/subagents.
  • Goose v1.44.0 (score 226): GHSA-r5pp-p5r8-466r — arbitrary command execution via goose review through git core.fsmonitor. Also latest Gemini models, Stop-hook working_dir, declarative Sakana/Fugu provider.
  • Gemini CLI v0.51.0 (score 194): security-focused — case-insensitive sensitive path blocklist + VS Code HITL, defensive at-reference path resolution, symlink directory-escape fix in memory import, ~/.gitconfig read-only in macOS sandbox. Preview twin at v0.51.0-preview.0 carries the same notes; nightlies are thinner.
  • xAI Grok CLI (Grok Build) 8adf901 (score 156): SSRF bypass via HTTP redirect in hook runner fixed; OAuth workspace scopes; sandbox/workspace-server legacy ready-file arm removed; SDK-spawned agents blocked from staging self-updates they cannot adopt.
  • Cline trio: v4.0.7 ClinePass limit UX + free-model tabs, GLM 5.1 references removed; desktop-v0.0.5 major UI perf (60fps bg, composer/keystroke fixes, boot catalog once); sdk/v0.0.65 Claude Code/Codex SDKs become optional peer deps, team runs hide spawn from teammates and report errored teammates as failed.
  • OpenCode v1.18.5: adaptive-thinking / Responses-phase / Mistral cache and MiniMax thinking fixes—stability, not net-new agent surface.
  • Zed AI v1.13.0-pre: prerelease — branch picker UX + GPT-5.6 Luna for ChatGPT subscriptions; breaking/hooks tags present but light on migration detail in the summary.
  • Windsurf v3.5.17: Agent Command Center UI, instant worktree-backed sessions, devin.* settings honored—HTML changelog, lower score (102).

Trusted mix is ~55% GitHub/changelog; 0 sources reporting fetch errors. Themes in evidence: model×13, cli×11, agent×10, security×8, mcp×6, breaking×5, prerelease×6.

For developers

Patch first if you run these CLIs locally

  1. Goose ≤ pre-v1.44.0 — treat GHSA-r5pp-p5r8-466r as drop-everything; upgrade to v1.44.0 before using goose review on untrusted repos.
  2. Gemini CLI → v0.51.0 for symlink-escape, sensitive-path blocklist, and sandbox ~/.gitconfig hardening.
  3. Grok Build — pull through 8adf901 for the hook-runner SSRF redirect fix.

Muscle-memory / workflow breaks

  • Claude Code: retrain /fork vs /subtask. Background forks show up as separate agent rows; in-session subagents are explicitly /subtask. Set CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION if you hit the new 200 default. Use claude auto-mode reset after experimenting with auto-mode.
  • Antigravity 1.1.0: expect pauses before writes under default request-review; learn the f diff preview and mode cycle (default / accept-edits / plan) before assuming unattended apply.
  • Codex CLI 0.145.0: try experimental thread history + /import if migrating from Cursor or Claude Code (settings, MCP, sessions, memories). On Bedrock, confirm GPT-5.6 Sol default and 272k context metadata from 0.144.6.
  • Cline: v4.0.7 removes GLM 5.1 naming and reshapes ClinePass free vs subscribed model tabs; desktop-v0.0.5 is the perf fix if the app felt sticky; SDK consumers should expect optional peer deps for Claude Code/Codex providers.

Concrete personal actions

  • After Claude Code upgrade: run one deliberate /fork and one /subtask on a throwaway session so the new split is muscle memory.
  • If you script Antigravity or headless applies: pin execution mode explicitly—default is no longer silent write.
  • Codex migrators: run /import once into a clean config, then diff MCP server entries before pointing at prod keys.

For teams

Security / platform

  • Mandate Goose v1.44.0+ on any machine that runs goose review (fsmonitor RCE path). Track advisory GHSA-r5pp-p5r8-466r in your vuln process.
  • Gemini CLI v0.51.0 and Grok hook-runner SSRF fix belong in the same “agent toolchain patch window” as dependency CVEs—especially shared CI runners and sandbox images.
  • Cline SDK v0.0.65 team runs: spawn is no longer exposed to teammate agents; errored teammates report failed not completed—update any orchestration that assumed old status semantics.

Rollout / policy

  • Antigravity request-review default is a governance win for regulated or shared repos: standardize on it for juniors/contractors; allow accept-edits only where audit trail already exists.
  • Claude Code background forks change observability: decide whether background sessions count against seat/agent limits and how they appear in shared claude agents views; document /subtask for in-thread delegation.
  • Codex /import + Bedrock: if you standardize on Codex, treat import as a controlled migration (MCP and memory can carry secrets/paths). Validate Bedrock auth and GPT-5.6 Sol default before org-wide Bedrock cutover.
  • ClinePass / pricing UX (v4.0.7): update internal runbooks for limit-reached → usage-based switch; remove GLM 5.1 from approved-model lists.

Open improvement proposals (all P5) exist for Goose 1.44.0, Codex 0.145.0 & 0.144.6, Cline 4.0.7, Grok 8adf901, Zed pre, and OpenCode PR artifacts—use them as checklist seeds, not as completed reviews.

What to ignore

  • OpenCode pr-38252-videos / pr-37967-screenshots-final — verification media tags, not product releases.
  • Codex rust-v0.146.0-alpha.14 — empty “Release …” stub; no actionable notes.
  • Gemini v0.54.0-nightly.* — two narrow fixes; prefer v0.51.0 stable signal unless you are dogfooding nightlies.
  • Duplicate Gemini preview.0 changelog body (mirrors 0.51.0 notes)—don’t double-count.
  • Windsurf HTML changelog at score 102 vs GitHub-versioned peers—useful UX color, weaker for security/compliance evidence.
  • Zed v1.13.0-pre breaking tag without a crisp migration list in-summary—watch, don’t fleet-upgrade on the pre label alone.
  • Marketing tone around “snappy” desktop perf (Cline)—the concrete fixes (60fps, keystroke path, single catalog fetch) are the signal; the adjective isn’t.

Watch next

  • Whether Claude Code follow-ups clarify permissions/hooks/sandbox for background /fork sessions vs /subtask (heavy surface tags this release).
  • Codex graduation of experimental thread history + Bedrock login out of experimental; any GA notes after 0.146 alphas.
  • Antigravity adoption friction from default request-review (teams may demand policy knobs or non-interactive CI mode).
  • Goose post-advisory residual risk: any related git-hook/fsmonitor guidance beyond v1.44.0.
  • Zed v1.13.0 stable notes for real breaking/hooks/MCP deltas after the pre.
  • Cline desktop + SDK convergence (optional provider peers, team spawn lockdown) into the main extension line.
  • Model ID churn: GPT-5.6 Sol/Terra/Luna metadata and “Luna for ChatGPT subs” in Zed—confirm availability in your billing tier before standardizing prompts.

How to use DevAgentRadar

  • Developers: patch security-tagged CLIs the same day; rehearse one breaking command path (/fork//subtask, Antigravity mode cycle, Codex /import) in a scratch repo before weekday work.
  • Teams: fold Goose/Gemini/Grok fixes into the vuln cadence; set explicit defaults for review-before-write and multi-agent spawn; treat /import and model-default flips as config migrations with secret review.
  • Evidence discipline: prefer the GitHub release URLs above; discount prerelease stubs, screenshot tags, and lower-score HTML changelogs when prioritizing.
  • Proposals: eight open improvement proposals are queued (mostly team/P5)—promote only those tied to Goose 1.44.0, Codex 0.145.0, and the security commits if you need a short executive list.
  • This brief is judgment from 19 curated signals / 22 tracked assistants in-window; it does not invent releases beyond the JSON evidence.

Evidence appendix

  • Evidence window: top curated signals (versioned releases preferred)
  • Assistants tracked: 22
  • Signals in this brief: 19
  • Trusted source mix (GitHub/changelog vs HTML): 55%
  • Open improvement proposals: 8
  • Sources reporting errors: 0

High-signal releases used

  • Claude Code v2.1.212 — /fork now copies your conversation into a new background session (its own row in claude agents) while you keep working; the in-session subagent it used to l — source

  • OpenAI Codex CLI rust-v0.145.0 — Added experimental paginated thread history with efficient resume, search, persisted names, sub-agent support, and memories. (#33364, #33907, #34085, #34229, #3 — source

  • Google Antigravity 1.1.0 — Agent execution mode cycling is now publicly available: default -> accept-edits -> plan) · Added request-review (default) mode as the default execution — source

  • Goose v1.44.0 — Arbitrary command execution in goose CLI via goose review (via git core.fsmonitor config) - GHSA-r5pp-p5r8-466r

  • Zed AI v1.13.0-pre — This week's release includes improved branch picker filtering and grouping, run-status indicators for runnable gutter controls, editable image-viewer zoom level — source

  • Cline v4.0.7 — Add a ClinePass limit-reached error with a one-click option to switch to Cline usage-based billing. · Allow selecting Cline free models on the ClinePass provide — source

  • Cline desktop-v0.0.5 — Major performance overhaul: the app now feels snappy end-to-end. · The animated background renders at a locked 60fps instead of ~10fps, typing in the composer n — source

  • Gemini CLI v0.51.0 — Changelog for v0.50.0-preview.1 · Fix no_proxy test · chore(release): bump version to 0.51.0-nightly.20260625.g3fbf93e26 · Vertex base url update · fix(security — source

  • OpenCode v1.18.5 — Improve Claude adaptive thinking handling across more response shapes. · Avoid OpenAI Responses phase handling that could break some conversations. · Preserve g — source

  • Gemini CLI v0.51.0-preview.0 — Changelog for v0.50.0-preview.1 · Fix no_proxy test · chore(release): bump version to 0.51.0-nightly.20260625.g3fbf93e26 · Vertex base url update · fix(security — source

  • Cline sdk/sdk/v0.0.65 — Claude Code and Codex provider SDKs are now optional peer dependencies loaded on demand, dramatically cutting install size · Added Kimi K3 to the bundled ClineP — source

  • Gemini CLI v0.54.0-nightly.20260728.gbef611950 — fix(a2a-server): normalize CRLF line endings to LF in getProposedContent · fix(core): enforce explicit tag length and validation in file keychain — source

  • OpenAI Codex CLI rust-v0.144.6 — Refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna, and corrected their context windows to 272,000 tokens. (#33972, #34009) · #33972 Backport refre — source

  • xAI Grok CLI (Grok Build) 8adf901 — Synced from monorepo: Synced from monorepo · Changes: · grok-shell: request workspaces:read/write OAuth2 scopes · security: fix SSRF bypass via HTTP redirect in hook runner · fix(gro — source

Open proposals

Was this guidance useful?

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