AI 代理／開發工具
Cline SDK 0.0.78 Preserves Agent Work Through Event Replay; Unpinned Models May Switch to New Defaults
Cline has added draining, a persistent event log, and a run queue to the Hub, preventing agent work from being lost during upgrades or reconnections. The release also fixes a bug that silently removed all tools from custom OpenAI-compatible models, while catalog updates have changed the default models for some providers.
Cline released SDK 0.0.78 on August 22, adding the recovery semantics needed for long-running agent executions to the Hub. After an administrator initiates a drain, the Hub rejects new state-changing work so that in-progress tasks can finish first. Durable runs that have not yet started are queued instead of disappearing during a restart or upgrade. Reconnecting clients can catch up on messages missed during the interruption from a persistent event log, while the system deduplicates them by event ID to prevent the same event from being delivered twice through both the replay and live streams.
This change is more significant than ordinary reconnection handling: the Cline SDK Hub is the shared authoritative service for sessions, approvals, scheduling, and multiple frontends. If events exist only in an in-memory stream, a CLI, desktop app, or custom interface that reconnects through a different connection may fail to see tool results or even incorrectly conclude that an agent has stopped. Version 0.0.78 at least establishes a clear recovery path for planned upgrades and brief connection interruptions. However, the release notes do not specify the event-log retention period or the atomicity guarantees if the Hub process crashes during a write, nor do they report multi-node failover testing. It therefore should not be treated as a complete high-availability execution engine.
Another fix concerns custom OpenAI-compatible models. Previous versions treated a capability list inferred from convenience flags such as `supportsReasoning` as authoritative. If that list did not include tool support, every tool in the request was silently stripped. The new version distinguishes inferred capabilities from explicitly declared capability settings; explicit settings can still intentionally disable tools. Teams using private gateways or local models should rerun tool-calling and approval tests after upgrading.
The release also adds Hub and delegated-agent session and client identities to Langfuse traces, making it easier to associate subagent traces with their parent sessions. Engineering teams should also note the synchronized model-catalog update: for providers including DeepSeek, Crof, CrossModel, Eden AI, Kilo, and NanoGPT, the resolved default model may change when no model ID is pinned, potentially affecting costs, output, and regression results.