Who contributed to Agentic Coding on September 11, 2026?

9 developers shipped this update, including github-actions[bot], wanhakim, langchain-oss-model-profiles[bot], all-hands-bot, enyst, juanmichelini, hieptl, and Bee, and 1 more.

What were the notable Agentic Coding updates?

langchain-anthropic==1.7.2, feat(huggingface): use torch.accelerator for device-agnostic device count detection, and fix(huggingface): remove retired IPEX backend.

Sep 11, 2026

$ the-wire · showcase

Anthropic 1.7.2 preserves invalid tool blocks, Cline stops resurrecting deleted prompts

By RepoJournal · Filed 07:07 UTC on September 11, 2026 · About Agentic Coding · Composed from the cited sources · methodology

9 people shipped this

wanhakim @wanhakim 2 cited

hieptl @hieptl 2 cited

Bee @abeatrix 2 cited

Saoud Rizwan @saoudrizwan 2 cited

github-actions[bot] @github-actions[bot] 1 cited

langchain-oss-model-profiles[bot] @langchain-oss-model-profiles[bot] 1 cited

all-hands-bot @all-hands-bot 1 cited

juanmichelini @juanmichelini 1 cited

enyst @enyst

langchain-anthropic 1.7.2 ships a fix that keeps malformed tool-use blocks intact instead of dropping them, and Cline closed a desktop bug where deleting a queued prompt made it reappear in the transcript as a message that never ran.

langchain-anthropic 1.7.2 is out with one behavioral fix: invalid tool use blocks are now preserved rather than discarded, so callers that inspect or repair malformed tool calls downstream keep the original payload instead of getting nothing back [1]. The release is versioned 1.7.2 with no other entries.

In the Hugging Face integration, device detection moved off `torch.cuda.device_count()` and onto the device-agnostic `torch.accelerator.device_count()` API added in PyTorch 2.6, with a fallback to the CUDA call on older PyTorch. Previously, users on non-CUDA accelerators such as Intel XPU had their hardware go undetected and could not select it via `device=`. The same contributor removed the retired IPEX backend, since Intel stopped official releases after IPEX 2.8 and archived the project on March 30, 2026, while Hugging Face removed its IPEX integration in Optimum Intel v2.0.0 [2][3]. Model profile data also refreshed, adding `deepseek/deepseek-v4.1-flash` on openrouter at a 1,048,576-token context window [4].

Cline fixed a queue-rendering bug: the sidecar inferred that a queued prompt had started running whenever a pending snapshot arrived with fewer items and a different head, and emitted `chat_queued_prompt_start` for the old head. Deleting the first queued prompt produced the same snapshot shape, so the removed prompt surfaced in the transcript as a user message that never executed [5]. Separately, Codex credentials were being reinstated after ChatGPT sign-out because every ProviderSettingsManager construction re-ran the legacy globalState.json/secrets.json import and re-added any provider missing from providers.json; each legacy provider is now imported only once [6]. Cline Pass and free models now report zero cost, since the API reports upstream cost on those responses, and live model catalogs refresh for all shared providers with private catalogs excluded [7][8].

OpenHands made its Canvas settings read free and default flags from the database rather than static configuration, with model search returning `free` and `default` and the logical `Default` profile resolving to the concrete DB default model [9]. Two delivery fixes landed alongside: Cloud settings links open in the same tab when the canvas is locked to a Cloud host and in a new tab otherwise [10], and automations now display which identity they run as [11]. The PR artifacts workflow was also brought to parity with the software-agent-sdk implementation, inspecting fork content without checking out untrusted code under write credentials [12].

Action items

→ Upgrade langchain-anthropic to 1.7.2 if you parse or repair tool use blocks downstream langchain-ai/langchain [plan]
→ Upgrade Cline desktop past the Codex sign-out fix so ChatGPT credentials are not re-imported on every sidecar command cline/cline [plan]
→ Drop IPEX configuration and verify device selection still resolves on non-CUDA accelerators after the huggingface changes langchain-ai/langchain [monitor]

References

[1] langchain-anthropic==1.7.2 ↗ langchain-ai/langchain
[2] feat(huggingface): use torch.accelerator for device-agnostic device count detection ↗ langchain-ai/langchain
[3] fix(huggingface): remove retired IPEX backend ↗ langchain-ai/langchain
[4] chore(model-profiles): refresh model profile data (#40358) ↗ langchain-ai/langchain
[5] fix(desktop): stop rendering a deleted queued prompt in the chat (#14053) ↗ cline/cline
[6] fix(desktop): clear legacy Codex credentials on ChatGPT sign-out (#14040) ↗ cline/cline
[7] fix(llms): report zero cost for Cline Pass and free models (#14012) ↗ cline/cline
[8] fix(desktop): refresh live model catalogs for all shared providers CLINE-3239 (#14003) ↗ cline/cline
[9] feat: use DB-driven free/default model flags in Canvas ↗ OpenHands/OpenHands
[10] fix: open Cloud settings in the same tab when locked to a Cloud host ↗ OpenHands/OpenHands
[11] feat: show which identity an automation runs as ↗ OpenHands/OpenHands
[12] ci: add .pr/ cleanup-after-merge and fork-aware notice ↗ OpenHands/OpenHands