Radar brief — 2026-08-24

What the evidence shows

Fifteen curated signals across 23 tracked assistants. About half the mix is GitHub/changelog; the rest is HTML docs that often splice several versions into one page. One source in this window reported errors. Large bodyOmitted tails (Codex +154/+151, Goose +103, Zed +54, Claude Code +42) are not treated as known changes.

GitHub release notes (Codex, Claude Code, Goose, Antigravity, Cline SDK, OpenCode, Zed) are the high-confidence core. Windsurf, Kiro, Factory, and Tabnine HTML changelogs are noisier: they interleave older patches and product lines.

For developers

• Upgrade Codex through rust-v0.149.0 if you live in the TUI. Use /export before you kill a terminal, codex exec fork / resume picker for branch experiments, and codex queue instead of opening a second session. After resume or fork, check the restored permission profile — 0.149.0 stops silently falling back to whatever is in your current config. Point AWS shops at the built-in Bedrock provider rather than a custom wrapper. Run codex doctor before filing “it cannot reach the model” tickets.
• Patch Claude Code to v2.1.234 on Windows first. NT-namespace path rejection is a real credential-leak close, not chrome. If hosts isolate config dirs, set CLAUDE_CODE_PROJECT_DIR_NAME. Turn off “Continue automatically at usage limit” in /config if you do not want a session to wake itself. Bind selection:clear if the agents view leaves stale highlights.
• Treat Cline SDK v0.0.76 as a skills/hooks behavior change. /my-skill no longer pastes SKILL.md into the user message; workflows still expand as text. Re-test PreToolUse contextModification and PostToolUse cancel — post hooks are awaited (120s) instead of fire-and-forget. Scheduled tasks are workspace-scoped; do not assume they follow you to another repo.
• Goose v1.46.0 is the cost-visibility release. Watch per-message cost and cache tokens before you blame the model for a “slow/expensive” session. Disable unused built-in skills. If you compact aggressively, switch to the structured summary path instead of hoping the default rollup is enough.
• Antigravity 1.1.12 is useful in scripts. antigravity -p "/permissions" (and /hooks, /help, /changelog, /config) now answers without starting a turn or burning quota. Mark library skills with disable-slash-command: true so / stays usable while the model can still invoke them.
• Small but sharp breaks: Factory Droid v0.193.0 retires MiniMax M2.5 — pick a replacement before a session fails to start. OpenCode v1.18.22 is worth it if Bedrock or OpenAI-compatible proxies were rejecting textVerbosity. Zed v1.15.0’s git.diff_base changes what “dirty” means in the gutter; set it explicitly if your team reviews against the default-branch merge base.

For teams

• Permission restore is now a policy surface. Codex 0.149.0 reapplies the thread’s permission profile on resume/fork. Claude Code 2.1.234 keeps session-scoped denies on background subagent prompts. Document the intended profile in the repo, or agents will look “randomly locked down” after a weekend resume.
• Windows + MCP are the security tickets this week. Roll Claude Code 2.1.234 for the \\??\ path block. If you run self-hosted GitLab or older Atlassian MCP behind Windsurf/Devin Local, treat v3.7.25 as an auth fix, not a feature upgrade. Codex MCP servers that die after OAuth should recover without a full restart — still verify that in a staging workspace.
• Enterprise model control landed in two places. Factory Droid v0.199.0 lets admins set BYOK custom models as managed defaults for sessions and subagents, and attach model policies to service accounts (open proposal: Set BYOK custom models as managed defaults). Tabnine v6.5.3 finally applies per-team model assignment and lets admins add org-level repositories without switching teams; if you rolled 6.5.0 CLI governance (Run Command / Web Fetch, YOLO gated on PAT scope), re-check those gates after the assignment fix.
• Do not put Cline Cloud handoff in the standard operating procedure yet. desktop-v0.0.15-beta.1 requires a pushed commit and a feature flag; interrupted transfers retry, but it is still a beta path that keeps working after the laptop lid closes. Prefer it for spike jobs, not compliance-bound repos.
• Cost and identity knobs are now visible enough to govern. Codex shows estimated thread credits/cost in /status. Goose tracks cache tokens. Claude Code will auto-resume when a claude.ai limit resets unless you disable it. Kiro Web SSO via Okta/Entra is in the same HTML changelog as v2.19.1 — confirm the product line (Web vs CLI vs IDE 1.0.337) before you change IdP tickets (Kiro v2.19.1 proposal).

What to ignore

• The second Kiro 2.19.1 signal is a date stamp, not a second release.
• OpenCode’s removed first-month discount copy is marketing cleanup, not a price-card change you can plan around.
• Zed v1.15.0 is mostly editor/git UX; do not read the pricing/mcp/breaking tags as a Zed agent platform shift — those tags outrun the listed notes.
• Windsurf, Kiro, Tabnine, and Factory HTML pages mix older versions in one body. Do not attribute 3.7.16, IDE 1.0.337, or Tabnine 6.5.0 features to the headline version without checking the date line.
• Anything in a +N more changes tail. Codex and Goose omitted 100+ items; inventing them is how changelogs turn into folklore.
• Cline Cloud, “Coming soon” environment rows, and recommended/free model-tier copy until they ship outside the beta tag.

Watch next

• Codex is shipping on a very short rust-v0.148.0 → rust-v0.149.0 cadence. Next notes should show whether codex agents + codex queue become the default multi-session workflow or stay power-user commands.

• Cline: whether /handoff leaves beta, and whether skill-via-tool (not paste) stays compatible with existing SKILL.md libraries.

• Factory: what replaces MiniMax M2.5, and whether BYOK managed defaults actually stick on subagents and service accounts.

• Open proposals still open: Grok CLI <response_guidelines> (developer), drop legacy managed MCP configs / shadowed platform skills (everyone), Kiro v2.19.1 (team), Factory BYOK defaults (team), OpenCode PR #38252 verification videos (everyone).

• The one error-reporting source in this window: treat any undated HTML changelog line as unverified until a GitHub tag repeats it.

How to use DevAgentRadar

DevAgentRadar is a public evidence radar, not a vendor scoreboard. Read the GitHub-tagged rows first, treat HTML changelogs as leads, and ignore omitted tails. Developers should map each row to one local action (upgrade, flag, or config). Teams should map the same row to a policy (permissions, MCP, model defaults, Windows path rules). If a claim is not in the table, it is not in today’s evidence.

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: 5
• Sources reporting errors: 1

High-signal releases used

• OpenAI Codex CLI rust-v0.148.0: Export complete TUI conversations to Markdown with /export, either to the clipboard or a new file. (#37358) · Fork sessions with codex exec fork, and archiv — source

• Claude Code v2.1.234: Added the optional CLAUDE_CODE_PROJECT_DIR_NAME environment variable: hosts that give each session its own config directory can choose a short name for the pe — 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

• Zed AI v1.15.0: This week's release includes a new git.diff_base setting for choosing whether editor gutters, file-status colors, and git: diff show uncommitted changes aga — source

• Goose v1.46.0: Unrolled agent loop #9574 · Cache-safe request assembly with append-only turn context and declared cache semant — source

• Google Antigravity 1.1.12: Added a heading outline to the artifact viewer, opened with t, so a long markdown document's structure is visible at a glance and you can jump straight to a s — source

• Cline sdk/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

• Windsurf v3.7.25: v3.7.25 August 13, 2026 Devin Desktop Faster sidebar for users with thousands of cached sessions: filtering, space grouping, and sorting now only process the se — source

• Kiro v2.19.1: 2.19.1 Aug 21, 2026 Spec Review Mouse Support and Automatic Stream Recovery This release brings mouse support to the spec review screen. · Under the hood, a str — source

• Factory Droid v0.193.0 — (August 11, 2026): Faster session search, fixes for repeated tool calls, lowercase proxy variables, the credits usage chart, and diff loading, plus the MiniMax M2.5 retirement · * — source

• Tabnine v6.5.3: v6.5.3 August 17, 2026 Improvements Administrators can now add repositories at the organization level regardless of which team is selected Bug Fixes · Fixed tea — source

• Factory Droid v0.199.0 — (August 18, 2026): The explorer droid can use every read-only tool, agent questions are easier to answer from the keyboard, and diagrams and long sessions render better in the app — source

Open proposals

• Replace <output_efficiency> with <response_guidelines> in Grok CLI prompts (audience: developer, P5)
• Drop legacy managed MCP configs and shadowed platform skills in Grok CLI (audience: everyone, P5)
• Kiro v2.19.1 (audience: team, P4)
• Set BYOK custom models as managed defaults for sessions and subagents (audience: team, P3)
• Improve with OpenCode pr-38252-videos: pr-38252-videos PR #38252 verification videos (audience: everyone, P2)