VS Code 1.121 tightens AI context

Microsoft’s Visual Studio Code and GitHub’s Copilot team shipped two related changes this week aimed at making AI coding sessions cheaper and more controlled. Visual Studio Code 1.121, published on May 20, adds new guardrails around what agent workflows can see and handle inside the terminal. Three days earlier, GitHub said Copilot cloud agent added lower-cost model options for simpler jobs. The two updates land as coding assistants move from one-off autocomplete into longer-running agent sessions. In that setup, the practical constraints are not only model quality but also how much context gets pulled in, how secrets are handled, and what each task costs. GitHub’s own documentation says cloud-agent users can now choose models in supported entry points rather than relying only on automatic selection. ### What changed inside VS Code 1.121? Visual Studio Code 1.121 introduced a feature that keeps sensitive terminal prompts in the terminal instead of feeding them back into the chat loop. Microsoft said password, passphrase, PIN and verification-code prompts can create a risk if an agent tries to capture or replay them, so VS Code now intercepts those prompts when a chat terminal command reaches them. (code.visualstudio.com) The May 20 release also continued Microsoft’s work on the Agents window, which runs a long-lived remote “agent host” process built on the Copilot SDK. That matters because persistent agent sessions can accumulate noise and state over time, making terminal hygiene and process cleanup more important than in a single prompt-response exchange. ### Why does terminal prompt interception matter? Password prompts are a concrete failure case for agent tooling because they look like ordinary text input but carry credentials or verification data. (code.visualstudio.com) Microsoft’s release note says VS Code now steps in when those prompts appear, rather than letting the agent try to answer them itself. Chat sessions in VS Code already bundle prompts, responses and relevant code context into a single conversation history, according to Microsoft’s documentation. (code.visualstudio.com) Keeping sensitive terminal interactions out of that flow reduces the chance that secrets become part of a reusable chat context, saved session or follow-on task. That is an inference from how Microsoft describes chat sessions and prompt interception, rather than a separate company statement. ### What did GitHub change in Copilot cloud agent? GitHub said on May 18 that Copilot cloud agent added two “fast, cost-efficient” models for simple tasks: Claude Haiku 4.5 and GPT-5.4-mini. The company listed both at a 0.33x multiplier, compared with larger models used for more complex work. GitHub’s documentation says model selection for Copilot cloud agent is supported when assigning an issue to Copilot on GitHub.com, mentioning @copilot in a pull request comment, or starting a task from the agents tab, agents panel, GitHub Mobile or the Raycast launcher. (code.visualstudio.com) Where no model picker is available, GitHub says Auto is used. ### Is Claude Haiku 4.5 already broader than cloud agent? (github.blog) GitHub had already been rolling out Claude Haiku 4.5 in Copilot before this week’s cloud-agent update. A GitHub changelog entry from October 2025 said the model was available in public preview across Copilot Chat entry points including Visual Studio Code chat, ask, edit and agent modes, with gradual rollout to paid plans. (docs.github.com) GitHub’s supported-models documentation now lists available AI models and says default models in Copilot run prompts and completions through content filters, including harmful-content and public-code-matching checks when enabled. That places the new cloud-agent pricing option inside a broader governance framework GitHub already applies across Copilot models. (github.blog) ### Where does this leave developers using these tools next? GitHub’s Copilot plans page says paid tiers include cloud agent and access to multiple model families from Anthropic, Google, OpenAI and others, with GPT-5 mini listed for unlimited agent mode and chats on some plans. Microsoft’s release archive shows VS Code 1.121 is now the current monthly release line, and GitHub’s cloud-agent model picker documentation points users to change models at supported entry points as those tasks are created. (docs.github.com) The next concrete step for users is operational rather than promotional: choosing when a cheaper model is enough, and when a terminal prompt should stay outside the chat context. (code.visualstudio.com) (github.com)