Cline SDK v0.0.67 Normalizes Reasoning Controls Across Providers, Fixes Auto-Compaction Bug

Cline released SDK version 0.0.67, an incremental update to the coding agent toolkit that overhauls how reasoning parameters are matched to model providers and fixes a bug that silently doubled summarizer calls during long sessions.

Reasoning controls now driven by models.dev

The SDK now sources reasoning controls — effort level, token budget, and on/off state — from the models.dev catalog and normalizes them once before encoding requests for each provider. According to Cline, this ensures requests match what each provider actually advertises rather than relying on hardcoded assumptions. Anthropic's mandatory and impossible thinking modes are now handled explicitly, and out-of-range budgets are clamped instead of causing request failures.

OpenRouter now defaults to anthropic/claude-sonnet-5.

MCP timeout configuration

The per-server timeout field in cline_mcp_settings.json is now honored by the SDK's MCP clients for initialize, tools/list, and tools/call calls. Previously these operations used hardcoded limits of 1.5 and 5 seconds regardless of configuration. The new default is 60 seconds, clamped to a range of 1 to 3,600 seconds.

Bug fixes

The release fixes several provider-specific and session-management issues:

• China and international endpoint toggles were being ignored for Qwen, Moonshot, and Z AI — now fixed.
• Legacy API keys are now migrated for every secret-backed provider, not just a subset.
• Legacy OpenAI Compatible model-info overrides now carry into the seeded models.json instead of being dropped.
• The "Enable R1 messages format" option has been removed from the OpenAI Compatible provider.
• Checkpoint restores across session resumes are fixed.
• Auto-compaction state was being rejected as stale on every save, forcing an unnecessary full re-compaction (an extra summarizer call) on every turn past the trigger point, and could leave a dead sidecar permanently blocking replacements after a resume. Both issues are fixed.

New capabilities

SDK v0.0.67 adds session forking and user-run message APIs, letting a host application edit an earlier prompt by forking the session before a selected user run, trimming checkpoint history, and restoring prior messages. A new ClineCore.readLiveMessages method reads a resident session's in-memory transcript, so rebuilding plan/act state during an in-flight turn no longer starts from empty history.

Additional changes include: insert_line and read_files line bounds now accept numbers emitted as JSON strings instead of failing the tool call; plugins can emit telemetry through ctx.telemetry from both subprocess sandbox and in-process execution; a legacy single-file .clinerules no longer aborts the config scan; telemetry events now carry a device_id; and a malformed OTEL header entry no longer discards valid ones.

What this means

This release is largely infrastructure hardening rather than a feature showcase. The shift to models.dev-driven reasoning controls reduces the maintenance burden of tracking each provider's reasoning API quirks manually, which matters as the number of supported providers grows. The auto-compaction fix is the most consequential bug fix here — silently doubling summarizer API calls on long sessions would have inflated costs and latency for any user running extended agent sessions without them noticing. Session forking opens the door for IDE integrations to let users branch conversations from an earlier point, a workflow pattern increasingly common in coding assistants.