Radar brief — 2026-09-09
CLI coding agents spent this window on permission correctness, MCP/provider plumbing, and headless reliability — while Windsurf removed Cascade from Devin Desktop and GPT-5.6 routing showed up in multiple GitHub changelogs.
What the evidence shows
15 curated signals across 23 tracked assistants. GitHub/changelog share is only 49%, so HTML docs pages are called out separately. One source in the window reported errors. Large bodyOmitted counts mean the tables below cover only listed changes, not the hidden rest.
GitHub-backed releases (prefer these)
on_failure, PreToolUseResult, and a stable tool_call_id. 98 further changes omitted; tags include breaking./diff panel for uncommitted edits. /cost and the status line now explain prompt-cache misses. Headless/Desktop/SDK get /reload-plugins and text /advisor. Permission/sandbox fixes: parentheses in paths, uncompilable deny regexes, and zsh assignments that hid command substitution. 66 changes omitted; permissions! / sandbox! / hooks!.-p: wait for background work (bounded by --print-timeout), keep daemons, print fatal errors on stderr with a stable error: marker, drop ~200 ms idle latency per turn. 11 omitted.tui.auto_recap = false. TUI reconnects after app-server drops. Plus/Team get an earlier warning when < half of a ~5-hour allowance remains. Full Access skips Guardian for confirmation-only actions. 95 omitted./ ? search in drafts. Rate-limit banners with usage/credit/plan actions. MCP names may include : @ / .. Per-tool output_token_limit. App-server thread/shellCommand timeouts can exceed one hour. Auto-approval reviews keep more transcript across compaction. MCP tools survive cache/plugin refreshes. 94 omitted.none reasoning effort; Bedrock replay is more reliable. apply_patch no longer emits an empty move path in permission metadata. Azure CLI sign-in asks for the resource name instead of hitting management APIs.HTML changelogs (noisier; mixed older patches)
api.streamIdleSoftTimeout / HardTimeout / api.timeout). Page also mixes 2.19.1 and older IDE notes. 8 omitted.tools.enableWebFetchTool is on (avoids LLM-mediated hallucination). Page bundles 6.5.3–6.5.1 plugin matrices. 7 omitted./archive (or Ctrl+X) from chat. Unset sessions default to GPT-5.6 Sol. CLI starts even if org settings are slow. Credit exhaustion suggests a live Droid Core model, not a retired one. Desktop noted as v0.166.0.For developers
- Upgrade Cline CLI
cli-v3.0.61and SDKv0.0.82before debugging “the model won’t use tools.” Dify, SAP AI Core, OpenCode, and Codex CLI routes were shipping with every tool (and images) stripped. If those providers still look dumb after upgrade, then debug prompts — not before. - Re-test Claude Code
v2.1.260permission rules if paths contain(), deny globs with broken character classes, or you rely on zshREPORTTIME/REPORTMEMORY/DIRSTACKSIZEassignments. Previously, “read-only” folders could stay writable and some substitutions were auto-approved. - Goose
v1.48.0: expect GLM-5.3 instead of 5.2; wire cost fields on custom providers; tryon_failureon PreToolUse hooks. Do not assume the other 98 omitted commits are cosmetic — the release is taggedbreaking. - Codex
rust-v0.152.0+v0.153.0: settui.auto_recap = falseif recaps are noise; use plugin marketplace commands; putoutput_token_limiton chatty MCP tools; MCP server names can now be package-style (@scope/name). - OpenCode
v1.18.26: Claude 5 sessions should survive prompt/tool edits that used to fail on stale thinking; Bedrock GPT-5.6 can sendnoneeffort. Recheckapply_patchpermission prompts (empty move path is gone). - Antigravity
1.1.28headless: parse stderr forerror:; set--print-timeout; do not expect the CLI to kill dev-server daemons when-pfinishes. - Factory
v0.209.0: new sessions with no model pick GPT-5.6 Sol. Use/archiveinstead of deleting threads. Do not treat the 60% GPT/Grok cut inv0.208.1as permanent. - Windsurf
v3.9.19: Cascade threads on Devin Desktop will not continue as Cascade. Use Continue in Devin Local before you need the history. - Zed
v1.19.0-preis a prerelease. Long-session memory and OpenRouter effort are worth a trial profile, not a fleet default. - Tabnine: if
tools.enableWebFetchToolis on, fetches are now direct — verify allowlists/proxies, because the model is no longer in the middle inventing page text.
For teams
- Treat Claude Code
v2.1.260as a security patch, not a UI release. Parentheses paths, bad regex deny rules, and zsh-assignment smuggling changed what the sandbox actually blocks. Re-run policy tests on “read-only” trees. - Review Codex Full Access.
rust-v0.153.0skips Guardian reviews for confirmation-only actions. If Guardian was your audit trail, that hole is now product behavior. Also budget-alert Plus/Team users on the ~5-hour window. - Cline Hub upgrades are now interactive. Rolling the CLI will ask before replacing a Hub and will drain live turns. Windows application-control policies get a real Authenticode signature — update allowlists so launches stop failing closed.
- Pin models where defaults moved. Open proposal: pin Cline models after the v3.0.61 switch to Claude Fable 5.1. Factory now defaults unset sessions to GPT-5.6 Sol — archive + default-model proposal. Goose replaced GLM-5.2 with 5.3.
- Windsurf desktop agent inventory changed. Cascade is gone; Devin Local is the only desktop agent. Track Continue Cascade conversations in Devin Local as a migration, not an optional theme.
- Spend and MCP hygiene: Goose custom providers can finally drive cost tracking. Codex MCP tools truncate on
output_token_limitacross resumes. Cline no longer dies if one remote MCP is offline. Tabnine org-level repo add and per-team model assignment appear on the same docs page as 6.5.3 — confirm on your tenant before rolling. - Queue the higher-severity open proposals: move Cline agent plugins into
~/.agents/pluginsand Hub-enable them (team, P5); Grok CLI prompt/MCP cleanup is developer-scoped (P4) and has no matching Grok release in this evidence.
What to ignore
-
Kiro
2.20.2: empty patch heading. Not a signal. -
+N more changes/bodyOmitted: Goose 98, Claude Code 66, Codex ~95+94, Zed 50. Do not invent those diffs; open the GitHub tag if you need them. -
Factory
v0.208.160% off GPT/Grok: promotional, time-limited, HTML changelog. -
Zed
v1.19.0-preas production guidance. -
Stacked HTML pages (Kiro 2.19.x mixing 1.0.337; Tabnine 6.5.4 bundling 6.5.3–6.5.1). Easy to mis-attribute an older fix to the latest tag.
-
OpenCode PR #38252 verification videos (P2): process nit, not a user-facing agent change.
-
Marketing tone on Windsurf/Kiro HTML vs versioned GitHub releases. Trusted mix this window is 49%.
Watch next
- Full Goose
v1.48.0and Claude Codev2.1.260omitted bodies — both taggedbreakingwithpermissions/sandbox/mcpsurfaces. - Whether GPT-5.6 (Goose Bedrock sol/terra/luna, OpenCode Bedrock, Factory Sol default) becomes the quiet default across CLIs.
- Cline Fable 5.1 pin vs Hub-upgrade prompts in mixed-version fleets.
- Fallout from Cascade → Devin Local (lost threads, worktree/cloud picker confusion, proxy MCP).
- Codex Guardian skip + plugin marketplaces: new supply-chain surface for teams that enabled Full Access.
- Grok CLI proposals (prompt tag rename; drop legacy managed MCP / shadowed skills) — watch for an actual versioned release; none is in this brief.
- The 1 source reporting errors this window: do not promote that channel until a clean changelog appears.
How to use DevAgentRadar
DevAgentRadar is a public, evidence-first radar for AI coding assistants — not a vendor roundup. This brief is 15 top curated signals over 23 tracked assistants. Prefer versioned GitHub tags over HTML docs; this window’s trusted mix is only 49%, which is why HTML rows are isolated. If bodyOmitted is set, stop at the listed bullets. Use For developers as local upgrade/regression checks and For teams as policy, default-model, and migration work. The 7 open proposals are suggested actions (priority + audience labeled), not shipped facts — click through before you change defaults. Re-read the source URL on anything tagged breaking, permissions!, or sandbox! before you roll a fleet.
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: 7
- 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://gi — source -
Claude Code
v2.1.260: Added a diff panel that opens beside the conversation in fullscreen mode and shows your uncommitted changes as Claude edits; toggle it with/diff· Added a li — source -
Zed AI
v1.19.0-pre: This week's release includes call hierarchy support, multi-select in the Git Panel, automatic language detection for untitled buffers, and project search on typ — source -
Google Antigravity
1.1.28: Improved resilience to transient model API errors: the agent now retries errors such as503 Unavailablefor much longer with exponential backoff, so a brief s — source -
OpenAI Codex CLI
rust-v0.153.0: Vim mode now supports undo withuand redo withCtrl+R, preserving complete drafts including pasted content and attachments. ([#41941](https://github.com/op — source -
OpenAI Codex CLI
rust-v0.152.0: Vim mode supports/and?searches within drafts, highlighted matches, and repeat navigation withnandN. (#41586) · Rate-limit banners offer actions f — source -
Cline
cli-v3.0.61: Cline now handles a running Hub that is older than your CLI. · Instead of quietly talking to a hub executing stale code, you get a prompt showing how many activ — source -
Cline
sdk/sdk/v0.0.82— Fixed tool calling being silently disabled for gateway models whose catalog entry declares no capabilities. · Three separate producers built gateway model defin — source -
OpenCode
v1.18.26: Claude 5 sessions now tolerate stale thinking blocks instead of failing after prompt or tool changes. · Bedrock GPT-5.6 models now acceptnonereasoning effor — 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
v3.9.19: v3.9.19 September 8, 2026 Devin Desktop Cascade has been removed. · Devin Local is now the only agent available in Devin Desktop. · Use the Continue in Devin Lo — 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
v0.209.0— · Desktop v0.166.0 (September 1, 2026): Session archiving from chat, a new default model, faster startup, and steadier app views · Archive sessions from chat - Archive the current session with /ar — 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
- Move Agent Plugins into ~/.agents/plugins and Hub-enable them in Settings → Customize (audience: team, P5)
- Pin Cline models after the v3.0.61 switch to Claude Fable 5.1 (audience: everyone, P4)
- Replace <output_efficiency> with <response_guidelines> in Grok CLI prompts (audience: developer, P4)
- Drop legacy managed MCP configs and shadowed platform skills in Grok CLI (audience: developer, P4)
- Continue Cascade conversations in Devin Local on Desktop 3.9.19 (audience: everyone, P3)
- Archive Factory Droid chats with /archive and default new sessions to GPT-5.6 Sol (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.