cline sdk/sdk/v0.2.1 — SDK v0.2.1
- Source
- github.com
- Author
- github-actions[bot]Top Viber
- Date

Context
Cline SDK v0.2.1 mostly fixes cases where the agent runtime gave a model less than it should have, or sent a request the model's endpoint refused. The SDK is the runtime behind Cline, the open-source coding agent, and the release notes for Cline's published alongside it list the same fixes. The most direct one concerns file reads. When the agent read a file with the read_files tool, the result was cut back to the 8k cap for tool results on its way to the provider, even when the tool had reported the read as complete. File reads now use the 50k file-content cap, so a read reported as complete reaches the model complete, according to the v0.2.1 release notes. Output from commands, searches, servers and custom tools keeps the 8k cap.
Several other fixes involve reasoning settings, the parameter that tells a model how much to think before answering. Some models expect a reasoning field on every earlier assistant message. After a compaction (the step where the agent summarises older conversation to fit the model's input limit) or a model switch, Reflection Beam failed with a 400 error about a missing reasoning_content parameter. The SDK now adds a placeholder field to fix that. The compaction summarizer also stops sending reasoning_effort "none" to models the catalog does not describe. Some endpoints rejected that value, which forced compaction back to a basic summary. Custom models with saved overrides now pass an explicit reasoning-off setting through to the provider. Before this fix, gpt-5.6 and gpt-6 chat-completions endpoints rejected every tool call. The native OpenAI adapter also stops sending max and Extra High levels to GPT-5.x and o-series models that reject them.
The release changes which model a user gets without choosing one. The cline provider's bundled default is now the first bundled free model. The free list dropped Reflection Beam, Laguna S 2.1 and Solar Mini 4, so that default is Nemotron 3.5 Super VL. Underneath, the code first tries the live recommended feed, then the bundled list, and finally a fixed fallback, which is now Anthropic's Claude Sonnet 5.5. The grows from 6,457 to 7,621 entries, more than reversing the drop from 7,041 in v0.2.0, and default models change for 67 other providers.
For developers building on the SDK, the release adds pieces for checking an agent's work. When tools and subagent spawning are enabled, the core registers a ux_verifier tool, which hands an ordered checklist to a subagent. A separate opt-in tool files structured reports with pass, fail or blocked results and screenshot references. Another optional tool publishes files as artifacts to cloud sessions. The release also removes the CLINE_CAPTURE_* request capture, which wrote full request bodies to disk.
- Only read_files results now skip the 8k tool-result cap and use the 50k file-content cap. Command, search, MCP and custom tool results are still cut at 8k, and an explicit tool-result cap still applies.
- With tools and spawning enabled, core registers a ux_verifier tool that hands an ordered checklist to a subagent through the host's spawn_agent executor. Disabling spawn_agent disables it too.
- The release removes the CLINE_CAPTURE_* provider request capture, which wrote full request bodies to disk. Setting those variables no longer writes request bodies to local files.
- The model catalog grows from 6,457 to 7,621 models, and defaults change for 67 other providers, mostly to Kimi K3.1 Pro, Gemini 4 Neon or DeepSeek V4.1 Flash 1017. Hosts relying on provider defaults rather than pinned IDs get a different model.
- session.resume now starts a runtime from a saved session ID and reads history on the hub instead of requiring initialMessages. The hub can also return history in bounded chunks, including archived cloud history.
- CLI tool — A command-line capability an agent can discover and compose through shell commands.
- Model Context Protocol (MCP) — A published agreement for an AI application to discover what a separate program offers and ask it to run one of those offerings.
- cloud model catalog — A selection of first-party and partner models offered inside a larger cloud platform's security and billing boundary.
- token budget — A cap on how many tokens a task, session, or agent run may consume — the practical control on both cost and how long an agent will grind.
Checking sign-in…
Loading comments…





