Bhavya U

This commit **enhances the responsiveness of core language tools** within the **Copilot extension** by making `vscode_renameSymbol` and `vscode_listCodeUsages` non-deferred. Previously, these essential features required a `tool_search` discovery process, introducing potential delays. By explicitly marking them as immediately available in `toolDeferralService.ts`, this **improves the user experience** for symbol renaming and code usage lookup. This **performance enhancement** ensures that fundamental IDE capabilities are always ready, aligning them with other core tools like `run_in_terminal` and `get_errors`.

Remove redundant quick pick titles from memory commands (#4877)

This commit introduces **new internal configuration settings** within the **Copilot extension** to precisely control reasoning and thinking effort during evaluations. Two hidden settings, `chat.advanced.responsesApiReasoningEffort` for the **Responses API** and `chat.advanced.anthropicThinkingEffort` for **Anthropic thinking**, are added to `configurationService.ts`. These settings enable `evals` to configure model behavior, taking priority over per-request model picker values, without affecting end-user experience. The `createMessagesRequestBody` and `createResponsesRequestBody` functions are updated to incorporate these new controls, ensuring their application during internal testing.

This commit provides a crucial **bug fix** for the **Copilot agent intents** system, resolving an issue where `toolTokens` were significantly over-counted when **Anthropic Tool Search** was active. Previously, the `agentIntent` logic incorrectly included deferred tools in its token budget, leading to an over-estimation of approximately 25-30K tokens and causing premature context window compaction. This **fix** now accurately calculates tool tokens by filtering for only non-deferred tools using `IToolDeferralService` in `agentIntent.ts`. Related **refactoring** propagates this new dependency to `askAgentIntent.ts`, `editCodeIntent2.ts`, and `notebookEditorIntent.ts`. This ensures a more precise token budget, preventing unnecessary prompt truncation and improving the overall efficiency of agent interactions.

This commit delivers a **bug fix** for the **Copilot extension's Anthropic API integration**, resolving an issue where `rawContentToAnthropicContent` generated invalid whitespace-only text blocks. Previously, an orphaned `CacheBreakpoint` (often after `prompt-tsx` pruning) would lead to the creation of `{ type: 'text', text: ' ' }` blocks, causing the Anthropic API to reject the request. The fix modifies the `rawContentToAnthropicContent` function in `extensions/copilot/src/platform/endpoint/node/messagesApi.ts` to **defer the `cache_control` from such orphaned breakpoints to the next valid content block**, or silently drop it if no subsequent block exists. This change enhances the **robustness of content processing** and prevents API errors in scenarios involving long conversations and token budget management.

This commit **fixes a bug** in the **Copilot extension's prompt management** where **background compaction** was not consistently applied between tool calls, leading to an outdated prompt context. To address this, the `Turn` class in `conversation.ts` now includes `_pendingSummaries` to temporarily store summaries during mid-tool-call-loops, and the `normalizeSummariesOnRounds` function is updated to utilize these pending summaries when `resultMetadata` is not yet available. This ensures that the **agent's prompt context** remains accurate and consistent throughout multi-turn interactions involving tool invocations. A new test case in `conversation.spec.ts` validates this improved summary restoration logic.

This commit **fixes premature context compaction** within the **Copilot agent's intent handling** by refining its **token budgeting logic**. It **decouples compaction ratios from tool tokens** by computing them against `baseBudget` instead of `budgetThreshold`, preventing early context reduction for users with many tools. The changes also **raise the background compaction kick-off threshold** from 75% to 80% and adjust the safety multiplier, allowing more context to be retained. This **bug fix** and **optimization** in `extensions/copilot/src/extension/intents/node/agentIntent.ts` improves the overall efficiency and user experience of the Copilot extension by ensuring context is compacted only when truly necessary.

This commit introduces a **new capability** to track and expose the **line count of session transcripts**, enhancing the existing summarization process. The `ISessionTranscriptService` now provides a `getLineCount()` method, populated by counting lines during active sessions and on resume from disk in `sessionTranscriptService.ts`. This line count is then incorporated into the **compaction summary** displayed by the `summarizedConversationHistory.tsx` component, providing the AI model with critical information about the original transcript's length. Ultimately, this **feature enhancement** helps the model more effectively **target relevant context for recovery** from compacted transcripts, improving the accuracy of AI responses.

This commit **fixes** an issue where **Anthropic API tool input schemas** were incorrectly stripped, losing critical JSON Schema keywords like `$defs` and `additionalProperties`. It ensures that the full schema, excluding `$schema`, is preserved, thereby enabling proper functionality for tools utilizing `$ref` and complex definitions. To achieve this, a new shared helper, `buildToolInputSchema()`, was extracted and integrated into both the **CAPI** (`messagesApi.ts`) and **BYOK** (`anthropicProvider.ts`) paths, improving consistency in schema generation. This **bug fix** significantly enhances the **robustness** and **compatibility** of **Anthropic tool definitions** within the system, particularly for complex tool specifications.

This commit **introduces telemetry logging** for user interactions within the **chat todo list widget**, a component of the **Chat feature**. It specifically tracks `clear`, `expand`, and `collapse` actions by injecting the telemetry service into `chatTodoListWidget.ts` and defining new event types. This **new capability** provides crucial data for understanding user engagement and informing future enhancements to the chat experience.

This commit **refactors** the mechanism for determining whether a **Copilot tool** should be deferred or non-deferred, moving from a hardcoded list in the **Anthropic integration** to an **opt-in static property** (`nonDeferred`) on individual `ICopilotToolCtor` classes. A new **`IToolDeferralService`** is introduced and registered via dependency injection, providing a centralized and extensible way to query this status, which impacts the **Anthropic language model provider** and **prompt generation**. This **feature enhancement** formalizes tool behavior, marks 18 frequently-used tools as non-deferred, and **fixes a latent bug** where the `vscode_askQuestions` tool was previously misidentified.

This commit **refactors** the **Anthropic thinking budget configuration** within the **Copilot extension**, graduating it from an experiment-based setting to a standard, simple configuration. It updates the retrieval mechanism for the thinking budget and simplifies the logic for applying reasoning effort within the `createMessagesRequestBody` function in `messagesApi.ts`. This **maintenance** work removes specific `thinking_budget` customization from `customizeCapiBody` in `chatEndpoint.ts`, streamlining how extended thinking parameters are handled for Anthropic models. The change standardizes the thinking budget feature, making it a core capability rather than an experimental one, and prepares for or directly implements support for an explicit effort parameter.

This commit **reverts** the previously added support for an **extended prompt cache TTL for Anthropic models**, effectively disabling this feature. It involves a **refactoring** of the **Messages API layer** and **Anthropic networking utilities** to remove all associated plumbing. Specifically, the `chat.anthropic.promptCaching.extendedTtl` setting is removed, along with the `isExtendedCacheTtlEnabled` function and the `ttl` property from the `AnthropicMessagesTool` interface. This change also cleans up conditional logic and parameter passing related to extended TTL in functions like `rawMessagesToMessagesAPI`, `rawContentToAnthropicContent`, and `addToolsAndSystemCacheControl`.

This commit performs **maintenance** by **removing the unused `agentHistorySummarizationForceGpt41` experiment setting** and its associated logic. Specifically, it eliminates the configuration option from `package.nls.json` and removes the conditional code in `summarizedConversationHistory.tsx` that previously forced the use of GPT-4.1 for **agent history summarization** within the **Copilot extension**. This cleanup simplifies the codebase by removing a deprecated feature and its related dependencies, ensuring that the system no longer attempts to override the summarization model based on this specific flag.

This commit **fixes critical bugs** in the **Copilot memory tool's invocation messages**, preventing the **leakage of session IDs** and ensuring **proper display of file widgets**. It **refactors** the invocation preparation logic by consolidating all scope handling (user, session, repo) through `_prepareLocalInvocation`, which now consistently processes directory paths and correctly renders file widgets. This change improves **data privacy** and the **user experience** by providing accurate and secure invocation messages, removing the redundant `_prepareRepoInvocation` method in the process. A corresponding test case was updated to verify the correct rendering of file widgets for repository paths.

This commit introduces a **bug fix** within the **Copilot extension's tool calling mechanism**, specifically addressing how large tool results are handled. It **exempts the memory tool** from the existing **large tool result disk caching** by adding it to an exclusion list within the caching condition in `toolCalling.tsx`. This change prevents potentially large or temporary data from the memory tool from being unnecessarily persisted to disk. Consequently, this helps to improve performance and reduce disk I/O for temporary tool results.

This commit introduces a **new capability** for the **Anthropic agent** within the **Copilot extension** by enhancing its tool search prompt. It adds a `<dynamicToolDiscovery>` section, instructing the model to actively search for newly available tools after executing one that might enable others on an **MCP server**. Concurrently, the 'do not retry' instruction is softened, preventing the model from prematurely terminating when tools are dynamically added mid-turn. This **prompt adjustment** ensures the agent can effectively discover and utilize tools that become available during a multi-turn interaction, significantly improving its adaptability and overall performance.

This commit **improves the user experience** within the **Copilot extension's conversation feature** by **skipping the reasoning effort picker** when the **`Auto` model endpoint** is selected. Previously, this picker was displayed even though the `Auto` model delegates to various backends, making a fixed effort level semantically incorrect for the user. The change introduces a condition within the `buildConfigurationSchema` function in `extensions/copilot/src/extension/conversation/vscode-node/languageModelAccess.ts` to hide this option specifically for `AutoChatEndpoint` instances. This **refinement** ensures that users are presented with only relevant configuration options, leading to a **more intuitive and accurate language model setup** for the **Copilot** feature.

This commit introduces a significant **refactoring** to the control of AI model **thinking** and **reasoning effort**, making it a **per-request opt-in** mechanism. It adds `enableThinking` and `reasoningEffort` parameters to `IMakeChatRequestOptions` and updates **`IChatEndpoint`** to expose `supportsReasoningEffort`, allowing callers to explicitly enable thinking and specify an effort level. This change impacts various **AI model providers** (Anthropic, Claude, OpenAI), **agent loops**, and **proxy endpoints** (`messagesApi`, `responsesApi`), which are now updated to respect these new parameters and validate the provided effort. Consequently, thinking is now **off by default** for most requests, requiring explicit opt-in, and the **model picker UI** now dynamically offers an effort dropdown for supported models like Claude and GPT, providing **fine-grained control** over AI model behavior.

This commit introduces **enhanced cache control for Anthropic API requests**, specifically targeting 1M context models. It adds a new configuration setting, `AnthropicExtendedCacheTtl`, allowing users to enable an extended Time-To-Live for prompt caching. The **Anthropic message processing functions**, notably `rawMessagesToMessagesAPI` in `messagesApi.ts`, are updated to incorporate and propagate an optional `cacheTtl` parameter, providing more granular control over caching behavior. This **new feature** improves performance and resource utilization by optimizing how responses from Anthropic models are cached, with new tests ensuring correct `cacheTtl` propagation.