Radar brief — 2026-08-16
Agent CLIs tighten plan-mode guardrails and spend attribution while plugin catalogs and enterprise sign-in land in the same window—security and MCP reliability outrank flashy model drops.
What the evidence shows
Versioned GitHub releases dominate this window (Codex, Claude Code, Cline, Goose, Gemini CLI). Themes skew to model / MCP / agent / CLI, with clear security and breaking edges on approvals, plan mode, and path controls.
--worktree / agents view; opt-in forward_user_identity on Anthropic upstreams for per-user spend headers; opt-in Linux Bash memory cgroups (CLAUDE_CODE_TOOL_MEMORY_LIMIT) so runaway builds don’t stall sessions--approve-for-me auto-reviewed approvals; import/sync Cursor & Claude skills without duplicate threadsrun.failedgit.diff_base (HEAD vs default-branch merge base) for gutters/status/git: diff; Project Panel drag-out; JSX/TSX linked editing + Emmet in expressionsLower-confidence / noisy in this set: Gemini CLI v0.56.0-preview.1 and nightly v0.56.0-nightly.20260815… (prerelease/caretaker churn), Codex rust-v0.148.0-alpha.20 (tag-only alpha), and Grok’s commit-synced monorepo blob. Useful as trajectory, weak as “ship it Friday” pins.
For developers
- On Codex
rust-v0.147.0: try Agent Plugins from local/personal/workspace catalogs; use conversation sections for long threads; only enable--approve-for-mewhen you accept auto-reviewed approvals—pair with therust-v0.146.1cyber-model review defaults, don’t run blind full-auto on untrusted trees. - On Cline
v4.1.4: plan mode is now human-gated and shell file mutation is hard-blocked—use it for read-only investigation; switch to act yourself when you want edits. Skills appear beside workflows; disambiguate same-name commands instead of assuming shadow order. - On Cline Desktop
v0.0.8: treat message edit as a real fork+checkpoint rewind; rely on transactional restore rather than manual half-reverts after a bad turn. - On Claude Code
v2.1.233: pass GitLab MR URLs into--worktree; on Linux, setCLAUDE_CODE_TOOL_MEMORY_LIMITfor heavy Bash/build agents; leaveforward_user_identityoff unless your proxy needs per-user spend headers. - On Goose
v1.46.0: use streaming shell + per-message cost/TTFT to catch runaway loops early; cache-safe turn assembly matters if you depend on prompt caching. - On Gemini CLI
v0.55.1: pull for the sensitive-path blocklist and path-resolution fixes before broad MCP/tool-registry experiments; skip nightlies/previews unless you’re reproducing caretaker work. - On OpenCode
v1.18.11: upgrade if MCP SSE servers were sticky-reconnecting after errors. - On Zed
v1.15.0: setgit.diff_baseexplicitly so gutter/branch diff matches your review habit (HEAD vs merge-base).
For teams
- Spend & identity: Claude Code’s opt-in
forward_user_identityis the concrete lever for proxy-side per-user attribution on Anthropic upstreams—pilot with gateway owners before org-wide on. - Enterprise auth: Antigravity
1.1.10Business sign-in (Gemini Enterprise seats, region, admin controls) is the team path; validate license assignment and regional inference with your Cloud admin before migrating personal logins. - Approval policy: Codex’s safer cyber-model auto-review defaults (
rust-v0.146.1) plus--approve-for-me(rust-v0.147.0) need a written policy—who may auto-approve, on which repos, under which sandbox. - Plan/act control: Cline’s removal of model-initiated plan→act and hard blocks on mutating shell in plan mode reduce “agent surprised us into writes”—update runbooks and training so humans own mode switches.
- Session durability: Cline Desktop checkpoints + SDK queue/abort semantics (
sdk/v0.0.72) improve recoverability for long agent sessions; standardize “restore checkpoint” over ad-hoc git cleanups. - MCP reliability: OpenCode SSE fix and Windsurf MCP auth fixes (GitLab/Atlassian) are ops-relevant where self-hosted MCP is common; verify against your server matrix (Windsurf signal is HTML changelog—confirm in your install).
- Cost observability: Goose’s per-message usage/cost UI supports chargeback conversations better than session-end anecdotes—decide whether to require it in shared agent hosts.
What to ignore
- Gemini CLI nightlies/previews and Codex
rust-v0.148.0-alpha.20as production pins—tag noise, minimal actionable notes. - Grok Build
be71313“Synced from monorepo” as a stability milestone—broad commit bundle without a clean version narrative. - Open proposals still aimed at older pins (Codex 0.144/0.145, Cline v4.0.7, Goose v1.44.0, OpenCode v1.18.5, Zed v1.13.0-pre): superseded by newer evidence in this brief; don’t prioritize them over current releases.
- Marketing tone around “+N more changes” without the security/MCP/breaking lines above.
Watch next
- Whether Codex plugin catalogs +
--approve-for-mepick up follow-on sandbox/permission defaults after the cyber-model review hardening. - Cline ecosystem alignment: Desktop fork/rewind vs core plan-mode hard blocks vs SDK abort/queue durability in one team workflow.
- Claude Code
forward_user_identityand memory cgroups adoption in shared gateway/Linux CI agents. - Gemini CLI post-
v0.55.1stabilization (tool registry + path blocklist) once previews stop dominating the feed. - Goose cache-semantics + unrolled loop under long MCP tool chains (cost and compaction behavior).
- Windsurf MCP auth fixes validated against self-hosted GitLab/Atlassian (need stronger than changelog HTML if you gate on it).
How to use DevAgentRadar
- Read What the evidence shows as the only source of truth for versions—this brief does not invent releases.
- Developers: apply the concrete flag/setting/upgrade actions; prefer GitHub-tagged builds over nightlies and HTML-only notes when pinning.
- Teams: turn approval, identity, plan-mode, and MCP auth items into policy and gateway config—not slideware.
- Ignore superseded open proposals and prerelease noise until they show up as stable, versioned signals.
- Re-check the radar when your assistants cross the versions named above; drop older improve/review proposals that lag the current tags.
Evidence appendix
- Evidence window: top curated signals (versioned releases preferred)
- Assistants tracked: 22
- Signals in this brief: 16
- Trusted source mix (GitHub/changelog vs HTML): 55%
- Open improvement proposals: 8
- Sources reporting errors: 0
High-signal releases used
-
Claude Code
v2.1.233: Added GitLab merge request URL support to the--worktreeflag and theclaude agentsview (where MRs display as!N) · Added an opt-in `forward_user_identit — source -
OpenAI Codex CLI
rust-v0.147.0: Install portable Agent Plugins and search across local, personal, workspace, and remote plugin catalogs. (#36544, #36409, #36919, #36796) · Organize conversatio — source -
Zed AI
v1.15.0: This week's release includes a newgit.diff_basesetting for choosing whether editor gutters, file-status colors, andgit: diffshow uncommitted changes aga — source -
Google Antigravity
1.1.10: Added Business sign-in for Gemini Enterprise accounts, so you can authenticate with a Google Cloud project under Google Cloud terms, use a license seat allocate — source -
Goose
v1.46.0: Unrolled agent loop #9574 · Cache-safe request assembly with append-only turn context and declared cache semant — source -
Gemini CLI
v0.55.1: fix/verify release npm ci ignore scripts · fix(ci): prevent workspace binary shadowing in release verification · Feat/tool registry discovery · fix(ci): prevent — source -
Cline
desktop-v0.0.8: Edit any earlier message in a conversation — the app forks the session at that point, rewinds the workspace to that run's checkpoint, and re-runs from your edit — source -
Cline
v4.1.4: Recognize Chutes as a provider. · Show skills alongside workflows in the slash command menu, and disambiguate commands that share a name instead of letting one — source -
Cline
sdk/sdk/v0.0.72— SDK v0.0.72: Prompts queued during a turn now survive being interrupted: they are preserved across user-initiated aborts, drained after a turn aborts itself, and edits made — source -
Gemini CLI
v0.56.0-preview.1: Changelog for v0.55.0-preview.1 · chore(release): bump version to 0.56.0-nightly.20260806.g761f604c1 · Changelog for v0.54.0 · Reclassifying Capacity Exhaustion — source -
OpenAI Codex CLI
rust-v0.146.1: Apply safer automatic-review defaults for cyber-capable models and explain permission changes in the terminal interface. (#37057) · #37057 [0.146] Backport safe — source -
Gemini CLI
v0.56.0-nightly.20260815.g2a87e7be1: [SSR Agent] Issue Fix (19826): Migrate process.env to vi.stubEnv in a2a-server tests — source -
xAI Grok CLI (Grok Build)
be71313— Synced from monorepo: Synced from monorepo · Presence protocol end-to-end: live presence updates through the gateway to client presence tiers · Worktree: narrow standalone origin fet — source -
OpenCode
v1.18.11: Stopped MCP SSE connections from getting stuck in reconnect loops after server error responses. · Fixed provider model configs that use interleaved reasoning fi — source
Open proposals
-
Improve with OpenAI Codex CLI rust-v0.145.0: rust-v0.145.0 0.145.0 (audience: team, P5)
-
Improve with Cline v4.0.7: v4.0.7 (audience: team, P5)
-
Improve with xAI Grok CLI (Grok Build) 8adf901: Synced from monorepo (audience: team, P5)
-
Improve with OpenCode pr-38252-videos: pr-38252-videos PR #38252 verification videos (audience: everyone, P5)
-
Improve with OpenAI Codex CLI rust-v0.144.6: rust-v0.144.6 0.144.6 (audience: developer, P5)
-
Review breaking changes in Zed AI v1.13.0-pre (audience: team, P5)
-
Improve with OpenCode v1.18.5: v1.18.5 (audience: team, P5)
-
Improve with Goose v1.44.0: v1.44.0 (audience: team, P5)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.