Radar brief — 2026-09-03
Lockdown flags, declarative providers, and silent tool-stripping bugs — not model marketing — are the week’s real coding-agent story.
What the evidence shows
GitHub-tagged releases carry this brief. HTML changelogs (Kiro, Tabnine, Factory) are thinner, mix older patches, and should be weighted lower. Several high-score notes omit dozens of changes (bodyOmitted); those gaps are not inferred below. One tracked source reported errors; the payload does not say which.
on_failure on PreToolUse; stable tool_call_id. Tagged breaking; 98 changes omitted.--restricted / CLAUDE_CODE_RESTRICTED=1 strips command/code tools and WebFetch (unless named in --tools), keeps file tools in the cwd, refuses bypassPermissions, ignores user/project/local settings. Per-agent experimental.cacheTtl. Self-hosted runner client labels. Settings-load diagnostics. Cross-session SendMessage/ListAgents. Tagged breaking; 43 changes omitted.GEMINI_API_KEY. Markdown custom agents inherit ambient skills/rules/subagents by default. MCP OAuth fix for auth codes >1024 chars. Duplicate permission-grant accumulation fixed. Remote Control reverse-tunnel hardened.tui.auto_recap = false. Session reconnect after app-server drop. Plus/Team get an earlier half-allowance warning. Full Access skips Guardian on confirmation-only actions. 95 changes omitted.: @ / .. Per-tool output_token_limit. Rate-limit banners with usage/credit actions. Bedrock reauth progress. Compaction keeps approval reviews and authorizations. 94 changes omitted.none reasoning effort. apply_patch no longer emits an empty move path in permission metadata./archive (or Ctrl+X). Unset sessions default to GPT-5.6 Sol. CLI no longer blocks startup on slow org settings.Dominant themes in the evidence: model routing, CLI/agent runtime, MCP auth and naming, permissions/restricted modes.
For developers
- Need a locked-down coding agent? Prefer Claude Code v2.1.248
--restricted(orCLAUDE_CODE_RESTRICTED=1) over hoping settings files stay honest. File tools stay in the working tree; shell/code/WebFetchare gone unless you name them in--tools. Do not expectbypassPermissionsto work. - On Goose v1.48.0, treat provider choice as config, not a fork: register TrustedRouter / OpenCode Zen / OpenAI-compatible declarative providers only if you need them. Budget with custom provider cost fields. If you were on Z.ai GLM-5.2, you are on GLM-5.3. Wire
on_failureon PreToolUse instead of assuming hooks only see success. - Upgrade Cline CLI v3.0.61 and SDK v0.0.82 before debugging “the model won’t use tools.” Empty capability catalogs were stripping every tool (and images) for Dify, SAP AI Core, opencode, and Codex CLI. Also upgrade if remote MCP being offline killed the TUI — connect now has a 10s budget.
- Codex rust-v0.152.0 / v0.153.0: set
tui.auto_recap = falseif recaps are noise; use the plugin CLI against remote marketplaces; putoutput_token_limiton chatty MCP tools; MCP server names can now be package-style (@scope/name). Don’t leave a dropped app-server connection as a lost draft — TUI reconnects and pauses uncertain submissions. - OpenCode v1.18.26: if Claude 5 sessions died after prompt/tool edits, retry — stale thinking blocks are tolerated. On Bedrock GPT-5.6,
nonereasoning effort is valid. Checkapply_patchpermission metadata if you were seeing empty move paths. - Antigravity 1.1.25: Markdown custom agents now inherit ambient skills/rules/subagents like default agents — audit agent files if you relied on isolation. Gemini 3.8 Flash appears only with
GEMINI_API_KEY. - Zed v1.19.0-pre: effort-based reasoning for OpenRouter is in a prerelease. Use it for experiments, not a team pin.
- Factory v0.209.0:
/archive/ Ctrl+X for session hygiene; new chats with no model pick GPT-5.6 Sol — set the model explicitly if that is not what you want.
Open, high-priority developer proposals on the radar: experimental context management in Codex; OpenRouter effort reasoning in Zed Agent; Cline model pins around the Fable-related proposals (verify against your catalog — the release bodies here do not describe that switch).
For teams
- Policy default: Claude Code
--restrictedis the first changelog-backed “no shell, no web, no settings-file escape” switch in this window. Use it for contractors, CI, or untrusted trees. Pair with the new settings-load warnings (/doctor,/status) so Bedrock/Vertex/customANTHROPIC_BASE_URLfailures are visible. - MCP is still an auth and naming surface. Antigravity fixed OAuth codes longer than 1024 characters. Codex now allows
: @ / .in server names and keeps tools across cache/plugin refreshes. Cline no longer dies when an enabled remote MCP is unreachable. Inventory servers after upgrade. - Permission drift: Antigravity was accumulating duplicate grants across reloads and subagents — rotate configs after 1.1.25. Codex “Full Access skips Guardian reviews for confirmation-only actions” is a control-plane change; decide whether that matches your review policy before rolling Full Access.
- Goose provider sprawl is a team decision, not a personal one. v1.48.0 adds multiple declarative gateways plus cost fields. Register TrustedRouter / OpenCode Zen centrally (there is an open team proposal) rather than letting every developer add a gateway.
- Runtime integrity: Cline now prompts before replacing an older Hub and drains in-flight turns; Windows binaries are Authenticode-signed. Require that CLI version if you run long-lived hubs.
- Spend: Codex Plus/Team warn before the five-hour window is half gone. Claude Code adds
/usage-creditsfor some Enterprise billing paths. Factory’s 60% GPT/Grok cut is time-limited and selector-visible — treat as a campaign, not a forecast. Goose cost fields only help if you fill them in. - Factory v0.209.0 applies org settings asynchronously so CLI startup is not blocked — good for large orgs, means early commands may race settings. Default model is now GPT-5.6 Sol; pin if you standardize elsewhere.
What to ignore
- Kiro 2.20.2 — title-only HTML stub. No changes to act on.
- Kiro v2.19.2 and Tabnine v6.5.4 — HTML pages that splice older versions (2.19.1 / 1.0.337; 6.5.3–6.5.1). Do not treat the bundle as a single upgrade.
- Factory v0.208.1 “60% off” — pricing promo, not an agent-quality signal.
- Any
+N more changes/ omitted bodies — Goose 98, Claude 43, Zed 50, Codex 95+94, Cline 18+34. Do not guess breaking MCP/hook/sandbox detail from tags alone. - Zed
v1.19.0-preas production IDE policy. Editor features (call hierarchy, Git Panel) are real; Agent/OpenRouter bits are still pre. - Proposal titles as changelog facts. Cline “Claude Fable 5.1” default-switch claims live in open proposals, not in the release bodies quoted here.
Watch next
- Goose follow-up that publishes the omitted 98 items — tags include
breaking,surface:mcp!, permissions, hooks, skills. - Claude Code restricted-mode fallout: which
--toolsallowlists teams actually ship, and whether ignored settings files surprise existing projects. - Codex plugin marketplace + Guardian skip on Full Access — supply chain and review coverage.
- Whether Cline’s shared capability translator changes default models in the wild; the pin-Fable proposals (P4, everyone) need catalog confirmation.
- Zed 1.19 stable: keep or drop OpenRouter effort-based reasoning.
- Antigravity custom-agent inheritance and reverse-tunnel hardening — confirm no extra skill/subagent blast radius.
- Remaining open proposals: Goose TrustedRouter/Zen registration (team, P5); Codex experimental context management (developer, P5); Grok CLI prompt/MCP cleanup (developer, P4); Factory
/archive+ GPT-5.6 Sol default (team, P3).
How to use DevAgentRadar
DevAgentRadar is a public, evidence-first feed of AI coding assistants. Read the GitHub/versioned row before the HTML changelog. Act on named flags, CVEs-in-all-but-name (silent tool stripping, OAuth length, restricted mode), and pinned versions. Ignore omitted tails, promo pricing, and empty stubs. Developers take the CLI flags and bugfix upgrades; teams take permission, MCP, and default-model decisions. Re-check the eight open proposals before you copy anyone else’s pin list.
Evidence appendix
- Evidence window: top curated signals (versioned releases preferred)
- Assistants tracked: 23
- Signals in this brief: 14
- Trusted source mix (GitHub/changelog vs HTML): 49%
- Open improvement proposals: 8
- 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
v2.1.248: Added--restricted(orCLAUDE_CODE_RESTRICTED=1): removes the built-in tools that run commands or code andWebFetch(unless named in--tools), keeps fil — 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
1.1.25: Added an opt-in workspace-grouped view to the/resumeconversation picker, allowing users to toggle between a flat list and conversations grouped by directory — 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— 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 -
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 -
Kiro
2.20.2— Aug 31, 2026: 2.20.2 Aug 31, 2026 — source
Open proposals
- Enable experimental context management in Codex CLI (audience: developer, P5)
- Register TrustedRouter and OpenCode Zen as Goose declarative providers (audience: team, P5)
- Enable effort-based reasoning for OpenRouter models in Zed Agent (audience: developer, P4)
- Pin Cline models after the v3.0.61 switch to Claude Fable 5.1 (audience: everyone, P4)
- Pin Cline provider models before SDK v0.0.82 switches 57 defaults 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)
- Archive Factory Droid chats with /archive and default new sessions to GPT-5.6 Sol (audience: team, P3)
Anonymous · one vote per visitor · re-click to clear. Helps improve digests—not a product ranking.