$ the-wire · showcase
Codex stops sleeping through your input; gemini-cli 0.65.0 lands
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
Codex now lets queued input break an agent's wait, gemini-cli ships v0.65.0 with read-only settings in untrusted folders, opencode exposes a configurable location idle timeout, and goose lets you resume a dead subagent instead of re-delegating it.
Release v0.65.0 google-gemini/gemini-cli
The v0.65.0 release bundles three fixes an operator will actually feel: read-only workspace settings are now enforced in untrusted folders, resumed sessions no longer duplicate tool response turns, and a quick exit no longer deletes resumed session history.
Make pending injections interrupt agent waits and sleep openai/codex
Input injected before wait_agent or clock.sleep begins now interrupts the wait and reaches the next model request; the fix tracks notifying injections as pending steering activity and resets that state once pending input is consumed, so later passive context does not break waits.
feat(core): add experimental.location_idle_timeout anomalyco/opencode
A location's services were unloaded after 60 minutes without session activity with no way to change it; the new experimental.location_idle_timeout setting makes that window configurable. Session-keeping while locations are busy already landed separately, so this PR only adds the knob.
feat(subagent): resume failed or cancelled delegates aaif-goose/goose
You can now call delegate with resume: "<subagent id>" and the subagent continues from its saved conversation instead of starting over; completed subagents and subagents belonging to other sessions are refused.
fix(core): support opt-in private network endpoints for MCP OAuth flows google-gemini/gemini-cli
The new oauth.allowPrivateNetwork option in settings.json lets MCP servers and OAuth 2.0 endpoints on private ranges (RFC 1918, RFC 6598, RFC 4193, .internal, .local) connect when you explicitly enable it. Also landing in gemini-cli: isAuthenticationError stopped matching a bare '401' substring, so ports like http://localhost:4012 no longer mask real connection errors.
fix(llms): send max_completion_tokens to SAP AI Core cline/cline
Elsewhere, cline now sends max_completion_tokens to SAP AI Core after 51 users hit the max_tokens rejection in 30 days, mostly on gpt-5 models, and the desktop Schedule page stops showing Pending after a failed run and stops truncating the error.
Action items
- → Upgrade gemini-cli to v0.65.0 to pick up the untrusted-folder settings enforcement and session resume fixes google-gemini/gemini-cli [plan]