Radar brief — 2026-08-25
CLI coding agents are turning into session operators: fork, queue, export, and handoff land in versioned releases, while Claude Code hardens Windows path access and several tools finally restore the permission profile you thought you already set.
What the evidence shows
15 signals across 23 tracked assistants. GitHub/changelog-style sources are only 49% of the mix, one feed reported errors, and several high-score notes omit large tails (bodyOmitted up to 154). Treat the table as the shipped surface, not a complete changelog.
Higher-confidence GitHub releases
CLAUDE_CODE_PROJECT_DIR_NAME; GitLab MR badge via authenticated glab; sessions can auto-continue when a claude.ai usage limit resets; Windows NT-namespace (\??) paths are rejected on remaining pre-approval file access to close an NTLM credential-leak path; permission answers no longer drop on background subagent prompts. 42 later changes are omitted./export dumps a TUI thread to Markdown; codex exec fork plus archive/restore; draft-while-startup; estimated thread credits/cost in /status; Amazon Bedrock Runtime is a built-in provider with AWS profile/region and GPT-5.6 routing; hooks can run async and call MCP tools; resume restores cwd and approval policy; MCP recovers after OAuth reauth. 154 later changes are omitted.codex agents dashboard; /cd /pwd /cwd; codex queue to local/remote sessions; richer Vim motions; codex doctor covers endpoint protection, proxy, desktop app, and updates; SDK can pass exact CLI overrides and max/ultra reasoning; resume/fork restore the active permission profile instead of silently using current defaults. 151 later changes are omitted.git.diff_base, Project Panel drag-out, JSX/TSX linked editing, Emmet in return/arrow bodies. Agent-specific evidence is thin: first line of terminal output no longer hidden in the Agent Panel. 54 later changes are omitted.working_dir on Stop hooks; structured compaction summaries. 103 later changes are omitted./my-skill ..., not the whole SKILL.md); provider-executed tools now emit observational events; PreToolUse/PostToolUse hook context is delivered again (Post awaited, 120s bound). 11 later changes are omitted./handoff moves a local session, images, and an optional follow-up to Cline Cloud after a push preflight; Cloud sits in the Local/Remote menu behind a flag; interrupted transfers retry and restore drafts. Bundles 0.0.14 plus Plugins hub, recommended/free model tiers, scheduled tasks, and the “Cline” rename.t); print-mode answers for read-only /permissions, /hooks, /help, /changelog, /config with no agent turn or quota spend; models/agents gain --output-format json/stream-json; disable-slash-command: true hides a skill from / while leaving it model-invocable. 21 later changes are omitted.textVerbosity to OpenAI-compatible providers that reject it.Lower-confidence HTML/docs changelogs
m toggles); stream idle watchdog, retries with backoff, and a 60-minute streaming timeout (api.streamIdleSoftTimeout, api.streamIdleHardTimeout, api.timeout). Same page also mentions earlier IDE permission/Windows-shell and Okta/Entra Web access notes. 7 later changes are omitted.https_proxy; MiniMax M2.5 is no longer selectable.For developers
- Upgrade Claude Code to v2.1.234 before the next long Windows or subagent session. The NT-namespace reject is a real security fix, and the background-permission plus post-compaction sandbox-network fixes stop “I already answered that” loops. If a host isolates each session’s config dir, set
CLAUDE_CODE_PROJECT_DIR_NAMEso transcripts stay short and findable. If you do not want a paused thread to resume itself after a claude.ai limit reset, turn Continue automatically at usage limit off in/configthe same day you upgrade. - If you are on Codex, skip 0.148 and land on rust-v0.149.0. Use
/exportfor reviewable Markdown,codex exec fork/ the resume picker for branchy work, andcodex queuepluscodex agentsinstead of opening another terminal. After resume/fork, verify the restored permission profile — 0.149 says it no longer silently falls back to whatever your current defaults are. Script health withcodex doctorbefore blaming the model. Bedrock users can point at the built-in provider rather than a sidecar. - Goose v1.46.0 is the cost-and-loop release. Watch the per-message cost/TTFT panel, disable built-in skills you do not want injected, and confirm Stop hooks now receive
working_dir. Do not invent behavior from the 103 omitted items. - Cline SDK users should re-test skills and hooks on sdk/v0.0.76.
/my-skillis no longer pasted as a fullSKILL.mdblob;PreToolUsecontext reaches the model again;PostToolUseis awaited (120s) and can cancel. If you wrap Claude Code as a provider, tool activity that used to vanish should now show as observational events — update any log parsers that assumed silence meant “no tools.” - Antigravity 1.1.12 is the scripting release. Prefer
-p "/changelog"/--output-format jsonover driving an agent turn for inventory. Mark helper skills withdisable-slash-command: trueso the/menu stays usable. - Do not treat Zed v1.15.0 as an agent-platform bump. The agent evidence is a terminal-output clip fix; the rest is editor/git/JSX.
- OpenCode v1.18.22 is a compatibility patch, useful if Bedrock or OpenAI-compatible providers were erroring on
textVerbosity. It is not a workflow release. - Ignore the second Kiro 2.19.1 stub and the omitted tails. If a note is not in the visible body, it is not in this brief.
For teams
- Decide the usage-limit policy on Claude Code. Auto-continue after a claude.ai reset is convenient for overnight agents and expensive if nobody is watching. Pair that with the new email-use instruction: Claude is told to use account email only to identify the user, not to ship it to unrelated services unless asked — still not a data-governance control.
- Codex is now a multi-session control plane. Budget for Bedrock routing and in-TUI credit estimates; require rust-v0.149.0 if you rely on resumed approval policies. Teach the team
codex queuerather than parallel unmanaged TUIs. - Keep Cline Cloud handoff in a beta lane.
/handoffneeds repo/branch/commit already pushed and can keep working after the laptop closes. That is a productivity win and a secret-sprawl risk. Do not make it default until the feature flag, receipt link, and retry path are in your runbook. - Factory v0.199.0 is the enterprise-controls release to review this week. Set BYOK custom models as managed defaults for sessions and subagents, then add service-account model policies. The matching open proposal is Set BYOK custom models as managed defaults for sessions and subagents (team, P3). Explorer droids can now run read-only commands, web search, and MCP — file edits stay blocked, but “read-only” plus shell is still a data-exfil surface. Also plan for MiniMax M2.5 disappearing in v0.193.0.
- Tabnine v6.5.3 is worth the admin pass even though the source is a rolling docs page: org-level repositories, corrected per-team model assignment, and (from the same page’s 6.5.0 notes) CLI governance that gates YOLO on PAT scope.
- Windsurf MCP auth fixes for self-hosted GitLab / older Atlassian are HTML-only. Validate in a canary org before declaring MCP “fixed.”
- Kiro stream-recovery defaults (60s warn, 300s cancel, 60-minute hard stream timeout) need an explicit team setting review if you run long spec sessions. There is a team proposal: Kiro v2.19.1 (P4). Okta/Entra for Kiro Web is on the same noisy page — confirm against your IdP admin, not this brief alone.
- Open proposals still open: Grok CLI prompt rename (developer, P5), drop legacy managed MCP configs and shadowed platform skills in Grok CLI (everyone, P5), and OpenCode PR #38252 verification videos (everyone, P2). None of those are shipped releases in this evidence pack.
What to ignore
- Marketing tone, “+N more changes,” and any behavior not in the visible release body.
- The duplicate empty Kiro 2.19.1 signal (score 62, no tags). Same version is already covered by the richer changelog entry.
- OpenCode’s removed first-month discount copy — pricing-page cleanup, not a capability change.
- Zed v1.15.0 as an MCP/agent platform story. Tags say agent/mcp; the written evidence does not.
- Windsurf and Tabnine pages that collapse several dates/versions into one blob. Only v3.7.25 / v6.5.3 are the headline versions; older bullets on those pages are mixed in.
- Cline desktop-v0.0.15-beta.1 as production policy until it leaves beta.
- Invented model launches. Codex’s Bedrock note says GPT-5.6 routing support, not that this brief independently verified a new public model drop.
- The one error-reporting source in this window. If a feed failed, do not backfill it from memory or vendor blogs.
Watch next
- Whether Codex
agents+queuebecomes the default multi-agent HUD, and whether permission-profile restore holds up on remote sessions. - Whether Cline
/handoffgraduates from feature-flag beta with a non-push failure mode that is still safe. - Copycat NT-namespace / NTLM hardening in other Windows-capable CLIs after Claude Code v2.1.234.
- Goose cache-token accounting vs Codex thread-credit estimates — first credible way to compare agent $ / turn in the TUI itself.
- Factory BYOK defaults and the Grok CLI legacy MCP / shadowed skills proposals actually merging.
- Cleanup of the Kiro duplicate feed and the unidentified erroring source. Until both are quiet, HTML Kiro/Windsurf/Tabnine dates stay provisional.
How to use DevAgentRadar
DevAgentRadar is a free, evidence-first radar for AI coding assistants. Read the versioned GitHub rows first, then the HTML/docs rows as unconfirmed. If a change is not in the JSON evidence, it did not happen in this brief — including the omitted tails. Use For developers as personal upgrade/test actions and For teams as policy, identity, MCP, and spend actions. Check the five open proposals before filing the same request upstream. Re-run the radar on the next window rather than stacking rumor on top of a 49% trusted mix.
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
-
Claude Code
v2.1.234: Added the optionalCLAUDE_CODE_PROJECT_DIR_NAMEenvironment variable: hosts that give each session its own config directory can choose a short name for the pe — source -
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 withcodex exec fork, and archiv — source -
OpenAI Codex CLI
rust-v0.149.0: Added an interactivecodex agentsdashboard 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 newgit.diff_basesetting for choosing whether editor gutters, file-status colors, andgit: diffshow 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 -
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 -
Google Antigravity
1.1.12: Added a heading outline to the artifact viewer, opened witht, so a long markdown document's structure is visible at a glance and you can jump straight to a s — 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
v0.199.0— (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 -
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
v0.193.0— (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
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)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.