$ cat ai-agents/week/2026-10-19.log
the week in review · Oct 19 – Oct 25, 2026
Goose 1.56 ships subagent concurrency caps and resume
By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology
Gemini CLI 0.65.0 enforces read-only workspace settings in untrusted folders, and Cline stopped sending native Google Search on the Gemini API.
v1.56.0 aaif-goose/goose
The headline release: automatic thinking level for model reasoning, resumable delegates, and a view-only subagent session view in the UI. Operators running multi-agent work get caps and recovery in the same shipped feature set.
subagent: Limit concurrent subagents with GOOSE_MAX_CONCURRENT_SUBAGENTS aaif-goose/goose
A new GOOSE_MAX_CONCURRENT_SUBAGENTS setting defaults to five per parent session; extra tasks queue until a running subagent finishes, and stopping the parent cancels running and waiting tasks. The goose-agent crate keeps its existing unlimited default, so check which surface your workload runs on.
fix(desktop): bound long conversation rendering aaif-goose/goose
Long conversations now render a bounded transcript window: the first 20 and latest 200 messages immediately, with the hidden middle expandable on request and conversation search expanding it automatically. Tool-call context stays full, so browsing a long session no longer stalls the desktop app.
fix(llms): stop sending native Google Search on the Gemini API (#15194) cline/cline
Web search had become default-on, so every Gemini API request carried the googleSearch grounding tool, and free-tier keys without grounding quota on Gemini 3+ models failed each turn with a 429. If your Cline sessions error out on Gemini, this is the fix.
Release v0.65.0 google-gemini/gemini-cli
Gemini CLI now enforces read-only workspace settings in untrusted folders and avoids duplicate tool response turns when resuming sessions. It also adds an oauth.allowPrivateNetwork option for MCP servers and OAuth 2.0 endpoints on RFC 1918, RFC 6598 and RFC 4193 ranges, plus .internal and .local, for teams keeping MCP behind an internal network.
Add observation-only model request lifecycle hooks openai/codex
PreModelRequest and PostModelRequest hooks fire around model request attempts, retries included, carrying model, session, turn and optional subagent context plus token usage. Anyone instrumenting Codex usage or auditing requests can now observe every attempt without proxying traffic.
feat(core): edit saved permissions anomalyco/opencode
Saved "Always allow" permissions can finally be edited rather than deleted and recreated. Narrowing a broad approval such as a saved shell: * no longer requires dropping it entirely, which unblocks a UI for reviewing saved approvals.
v2.1.304 anthropics/claude-code
Transcripts now name the event and the plugin when a plugin that runs first skips one of your mod's hooks, and a mod's tool.check answer can carry decider: "person" so only a permission dialog or host callback approves the ask. Both make hook-driven permission setups auditable.
$ ls ai-agents/week/ # the briefings behind this review
Keep up with Agentic Coding in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.
fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.
Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.