$ the-wire · showcase
Compaction triggers on real token counts, LangChain 1.4.1 ships
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
8 people shipped this
Bee
@abeatrix
3 cited
github-actions[bot]
@github-actions[bot]
1 cited
langchain-oss-model-profiles
@langchain-oss-model-profiles
1 cited
Sydney Runkle
@sydney-runkle
1 cited
alex-lum
@alex-lum
1 cited
John Choi
@johnwschoi
1 cited
langchain-oss-model-profiles[bot]
@langchain-oss-model-profiles[bot]
aivong-openhands
@aivong-openhands
Cline re-anchored context compaction to the provider's actual input-token count, and it is the only change today that alters behavior users will feel.
core: trigger compaction on the provider's actual input-token count (#14195) cline/cline
Compaction used to fire off a character-based estimate (roughly 3 chars per token) checked against the model's input budget, so dense transcripts like disassembly, image dumps, or minified sources could hit the true context ceiling while the estimate stayed under threshold and compaction never ran. It now reads the provider-reported input-token count, so long-running agent sessions stop silentl...
release(langchain): 1.4.1 (#40498) langchain-ai/langchain
Sydney Runkle cut 1.4.1 to publish the fixes merged since 1.4.0, notably support for open MCP object arguments; the bump covers the package version, the importable __version__, and the lockfile entry, with unrelated workspace dependency churn deliberately excluded.
langchain==1.4.1 langchain-ai/langchain
The release notes for langchain==1.4.1 list the open MCP object argument fix alongside a correction to the InterruptOnConfig documentation. If you pass MCP tool arguments as open objects through LangChain, pull the patch.
feat(telemetry): enable release traces and preserve Langfuse user/session tracking (#13982) cline/cline
Cline baked OTEL_TRACES_EXPORTER=otlp into stable and nightly extension and CLI publish builds, turning on provider AI SDK trace emission at 100% task sampling with metadata only and no prompt or completion content. Traces can be disabled without a release via the remote config key openTelemetryTracesExporter on the extension.
feat(desktop): add cloud session lifecycle (#13856) cline/cline
The desktop app gained cloud session lifecycle support, built on authenticated remote Hub connections, pending session approval recovery scoped to attached clients, and a boolean cloud rollout flag; the sidecar suite now runs in pull requests.
fix(desktop): keep streamed tool output and stop the spinner when a tool ends without a final output cline/cline
A follow-up fix keeps streamed tool output in place when a tool finishes without emitting a final output field: the desktop chat hook had rebuilt the tool message payload with result: null when the runtime's chat_tool_call_end event carried no output, and buildToolPresentation() reads a null result as a spinning, outputless tool. Tool icons no longer revert to spinning after a completed tool call.
fix(desktop): steer queued prompts with Enter CLINE-3285 cline/cline
Pressing Enter with an empty composer now steers the first queued prompt into the active turn in both the CLI and desktop app, with the runtime interrupting only the in-flight model request while tools and the overall run continue. Enter without any queued prompt still does nothing.
chore(model-profiles): refresh model profile data langchain-ai/langchain
Quietly: an automated refresh added 11 model profiles and changed 3 across two providers, including fireworks-ai entries such as accounts/fireworks/models/minimax-m2p7 at 196,608 context with reasoning and tool support, and the OpenHands desk landed five test-only PRs covering conversation and workspace mocks, MCP OAuth polling and timeout paths, the telemetry consent banner, conversation name ...
