$ the-wire · showcase
Codex reworks exec-server auth, gemini-cli fixes sandbox and model selection
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
19 people shipped this
diegogodinezr
@diegogodinezr
1 cited
jesussamuel-byte
@jesussamuel-byte
1 cited
mkondratek
@mkondratek
1 cited
maxpaulus43
@maxpaulus43
1 cited
lifeizhou-ap
@lifeizhou-ap
1 cited
GitHub Actions
@actions-user
copyberry
@copyberry
iceweasel-oai
@iceweasel-oai
usrnk1
@usrnk1
rekram1-node
@rekram1-node
tadipavan
@tadipavan
duvanka-prasad
@duvanka-prasad
OpenHands Bot
@all-hands-bot
Alona
@ak684
abeatrix
@abeatrix
johnwschoi
@johnwschoi
tmellor-block
@tmellor-block
github-actions
@github-actions
daveh-beep
@daveh-beep
The day's operator-relevant work is mostly in gemini-cli, where explicit model requests stop being rewritten, and in cline and goose, where long-standing failures in resumed and reasoning history get closed.
fix(core): preserve explicit gemini-2.5-flash model in dynamic model resolution google-gemini/gemini-cli
With dynamic model configuration on and useLatestFlash active, an explicit --model gemini-2.5-flash or /model selection was rewritten to LATEST_GEMINI_FLASH_MODEL; it now stays on gemini-2.5-flash. If you pin that model on Vertex AI or an API-key setup, your requests finally resolve to what you asked for.
fix(sandbox): supply --userns=keep-id for rootless podman google-gemini/gemini-cli
In container sandbox mode under rootless Podman on Linux, files created in the bind-mounted workspace inherited subordinate UID ownership and threw EACCES when the host user touched them. The sandbox now detects rootless Podman and passes --userns=keep-id when running as the host UID.
feat(subagent): resume completed delegates with a follow-up aaif-goose/goose
Completed delegates now accept a follow-up and keep their earlier context; previously goose rejected the request with "already completed". The follow-up must include instructions, and the earlier answer is ignored while a new one is produced.
fix(llms): send reasoning placeholder for models that need interleaved reasoning cline/cline
Reflection Beam returned 400 for requests where any earlier assistant message lacked reasoning_content, which happened after context compaction or a model switch, and the PR cites roughly 21k failed runs from 8.2k users in 14 days with about 79% of those chats never recovering. Cline now sends a reasoning placeholder so later turns keep working.
fix(core): stop re-cutting read_files results to the 8k tool-result cap cline/cline
read_files bounds its own output at 2000 lines or about 48k characters and tells the model so, but results were then re-cut to an 8k tool-result cap it had never matched. File reads now reach the model at the size the tool advertises. Also shipped: Gemini extension downloads now handle stream errors and clean up partial files, and gemini-cli no longer waits on unawaited history replay before an...
