diff --git a/extensions/copilot/package.json b/extensions/copilot/package.json index 329d6124a89..a12de99ce35 100644 --- a/extensions/copilot/package.json +++ b/extensions/copilot/package.json @@ -4113,6 +4113,16 @@ ], "description": "%github.copilot.config.anthropic.promptCaching.extendedTtl%" }, + "github.copilot.chat.freezeCustomizationsIndex": { + "type": "boolean", + "default": false, + "tags": [ + "advanced", + "experimental", + "onExp" + ], + "description": "%github.copilot.config.freezeCustomizationsIndex%" + }, "github.copilot.chat.installExtensionSkill.enabled": { "type": "boolean", "default": false, diff --git a/extensions/copilot/package.nls.json b/extensions/copilot/package.nls.json index c1dd11a1807..bb207a64817 100644 --- a/extensions/copilot/package.nls.json +++ b/extensions/copilot/package.nls.json @@ -340,6 +340,7 @@ "github.copilot.config.useMessagesApi": "Use the Messages API instead of the Chat Completions API when supported.", "github.copilot.config.anthropic.contextEditing.mode": "Select the context editing mode for Anthropic models. Automatically manages conversation context as it grows, helping optimize costs and stay within context window limits.\n\n- `off`: Context editing is disabled.\n- `clear-thinking`: Clears thinking blocks while preserving tool uses.\n- `clear-tooluse`: Clears tool uses while preserving thinking blocks.\n- `clear-both`: Clears both thinking blocks and tool uses.\n\n**Note**: This is an experimental feature. Context editing may cause additional cache rewrites. Enable with caution.", "github.copilot.config.anthropic.promptCaching.extendedTtl": "Use the extended (1 hour) prompt cache TTL on tools and system blocks for the Anthropic Messages API. Only applied to 1M context Claude variants; other models keep the default 5 minute TTL even when this setting is enabled.\n\n**Note**: This is an experimental feature. Only the main agent conversation is eligible — inline chat, terminal chat, notebook chat, and subagent requests are excluded.", + "github.copilot.config.freezeCustomizationsIndex": "Freeze the bundled ``, ``, and `` listing in the system prompt at the first turn of a conversation and reuse it on every subsequent turn. Prevents per-turn churn (e.g. the active mode swapping which subagent entry is listed, or async experimentation flipping a skill in or out) from invalidating the prompt cache. When the listing changes mid-conversation, the updated set is appended to the latest user message so the model still sees instructions, skills, or agents that became available or were removed.\n\n**Note**: This is an experimental feature.", "github.copilot.config.useResponsesApi": "Use the Responses API instead of the Chat Completions API when supported. Enables reasoning and reasoning summaries.\n\n**Note**: This is an experimental feature that is not yet activated for all users.\n\n**Important**: URL API path resolution for custom OpenAI-compatible and Azure models is independent of this setting and fully determined by `url` property of `#github.copilot.chat.customOAIModels#` or `#github.copilot.chat.azureModels#` respectively.", "github.copilot.config.responsesApiReasoningSummary": "Sets the reasoning summary style used for the Responses API. Requires `#github.copilot.chat.useResponsesApi#`.", "github.copilot.config.responsesApiContextManagement.enabled": "Enables context management for the Responses API. Requires `#github.copilot.chat.useResponsesApi#`.", diff --git a/extensions/copilot/src/extension/prompt/common/conversation.ts b/extensions/copilot/src/extension/prompt/common/conversation.ts index 2bb640ef003..b4ec9eea178 100644 --- a/extensions/copilot/src/extension/prompt/common/conversation.ts +++ b/extensions/copilot/src/extension/prompt/common/conversation.ts @@ -4,7 +4,7 @@ *--------------------------------------------------------------------------------------------*/ import { PromptReference, Raw } from '@vscode/prompt-tsx'; -import type { ChatRequest, ChatRequestEditedFileEvent, ChatResponseStream, ChatResult, LanguageModelToolResult } from 'vscode'; +import type { ChatLanguageModelToolReference, ChatRequest, ChatRequestEditedFileEvent, ChatResponseStream, ChatResult, LanguageModelToolResult } from 'vscode'; import { FilterReason } from '../../../platform/networking/common/openai'; import { IWorkspaceService } from '../../../platform/workspace/common/workspaceService'; import { isLocation, toLocation } from '../../../util/common/types'; @@ -445,6 +445,27 @@ export class GlobalContextMessageMetadata { ) { } } +/** + * Captures the customizations-index variable value (the bundled + * ``/``/`` text) as it appeared on the first + * turn of the conversation. Reused on subsequent turns so per-turn churn in + * any of those listings (e.g. the active mode swapping which subagent entry + * appears in ``) does not invalidate the system prompt cache. The + * cacheKey invalidates the snapshot when something genuinely changed (e.g. + * the user opened a different workspace mid-conversation). + * + * The {@link toolReferences} carry byte offsets into {@link value} and must + * be captured together — current-turn references built against current-turn + * text would mis-slice the frozen value. + */ +export class CustomizationsIndexMetadata { + constructor( + readonly value: string, + readonly toolReferences: readonly ChatLanguageModelToolReference[] | undefined, + readonly cacheKey: string + ) { } +} + /** * Captures `prompt_tokens` and `completion_tokens` from the most recent * successful fetch on a turn. All providers return these values in their diff --git a/extensions/copilot/src/extension/prompts/node/agent/agentPrompt.tsx b/extensions/copilot/src/extension/prompts/node/agent/agentPrompt.tsx index 6a1223abf34..00019fc3a2e 100644 --- a/extensions/copilot/src/extension/prompts/node/agent/agentPrompt.tsx +++ b/extensions/copilot/src/extension/prompts/node/agent/agentPrompt.tsx @@ -4,7 +4,7 @@ *--------------------------------------------------------------------------------------------*/ import { BasePromptElementProps, Chunk, Document, PromptElement, PromptPiece, PromptPieceChild, PromptSizing, Raw, SystemMessage, TokenLimit, UserMessage } from '@vscode/prompt-tsx'; -import type { ChatRequestEditedFileEvent, LanguageModelToolInformation, NotebookEditor, TaskDefinition, TextEditor } from 'vscode'; +import type { ChatLanguageModelToolReference, ChatRequestEditedFileEvent, LanguageModelToolInformation, NotebookEditor, TaskDefinition, TextEditor } from 'vscode'; import { sessionResourceToId } from '../../../../platform/chat/common/chatDebugFileLoggerService'; import { ChatLocation } from '../../../../platform/chat/common/commonTypes'; import { ConfigKey, IConfigurationService } from '../../../../platform/configuration/common/configurationService'; @@ -27,7 +27,7 @@ import { IInstantiationService } from '../../../../util/vs/platform/instantiatio import { ChatRequestEditedFileEventKind, Position, Range } from '../../../../vscodeTypes'; import { GenericBasePromptElementProps } from '../../../context/node/resolvers/genericPanelIntentInvocation'; import { ChatVariablesCollection, extractDebugTargetSessionIds, isCustomizationsIndex } from '../../../prompt/common/chatVariablesCollection'; -import { getGlobalContextCacheKey, GlobalContextMessageMetadata, RenderedUserMessageMetadata, Turn } from '../../../prompt/common/conversation'; +import { CustomizationsIndexMetadata, getGlobalContextCacheKey, GlobalContextMessageMetadata, RenderedUserMessageMetadata, Turn } from '../../../prompt/common/conversation'; import { InternalToolReference } from '../../../prompt/common/intents'; import { IPromptVariablesService } from '../../../prompt/node/promptVariablesService'; import { ToolName } from '../../../tools/common/toolNames'; @@ -131,9 +131,10 @@ export class AgentPrompt extends PromptElement { const sessionId = sessionResource ? sessionResourceToId(sessionResource) : undefined; const debugTargetSessionIds = extractDebugTargetSessionIds([...this.props.promptContext.chatVariables].map(v => v.reference)); const templateVariablesContext = this.promptVariablesService.buildTemplateVariablesContext(sessionId, debugTargetSessionIds); + const customizationsSnapshot = this.getOrFreezeCustomizationsIndex(); const baseInstructions = <> {!omitBaseAgentInstructions && baseAgentInstructions} - {await this.getAgentCustomInstructions()} + {await this.getAgentCustomInstructions(customizationsSnapshot?.frozen)} {isAutopilot && When you have fully completed the task, call the task_complete tool to signal that you are done.
IMPORTANT: Before calling task_complete, you MUST provide a brief text summary of what was accomplished in your message. The task is not complete until both the summary and the task_complete call are present. @@ -166,13 +167,14 @@ export class AgentPrompt extends PromptElement { userQueryTagName={userQueryTagName} ReminderInstructionsClass={ReminderInstructionsClass} ToolReferencesHintClass={ToolReferencesHintClass} + customizationsIndexUpdate={customizationsSnapshot?.drift} /> ; } else { return <> {baseInstructions} - + ; } @@ -197,13 +199,15 @@ export class AgentPrompt extends PromptElement { />; } - private async getAgentCustomInstructions() { + private async getAgentCustomInstructions(frozenCustomizationsIndex?: { value: string; toolReferences: readonly ChatLanguageModelToolReference[] | undefined }) { const putCustomInstructionsInSystemMessage = this.configurationService.getConfig(ConfigKey.CustomInstructionsInSystemMessage); const customInstructionsBodyParts: PromptPiece[] = []; customInstructionsBodyParts.push( @@ -225,6 +229,67 @@ export class AgentPrompt extends PromptElement { {customInstructionsBodyParts}; } + /** + * When the experimental `FreezeCustomizationsIndex` setting is enabled, + * snapshot the customizations-index variable on the first turn of the + * conversation and reuse it for every subsequent turn. Stops per-turn + * churn in the bundled ``/``/`` text (e.g. + * the active mode swapping which subagent entry is listed in ``) + * from invalidating the system prompt cache. + * + * Returns: + * - `frozen`: the value (and matching tool-reference offsets) to substitute + * in the system prompt. Always present when the setting is enabled and a + * variable is available. + * - `drift`: the live current-turn value (and offsets) when it differs from + * `frozen`. Rendered in the latest user message so the model sees the + * up-to-date listing without busting the system prompt cache. Also + * emitted as an empty value when the live variable disappears, so the + * model gets a signal that previously listed entries are no longer + * available. + * + * Returns `undefined` overall if no override should apply (setting off, + * no first turn available, or no snapshot yet and the variable is absent + * on this turn). + */ + private getOrFreezeCustomizationsIndex(): { + frozen: { value: string; toolReferences: readonly ChatLanguageModelToolReference[] | undefined }; + drift?: { value: string; toolReferences: readonly ChatLanguageModelToolReference[] | undefined }; + } | undefined { + const enabled = this.configurationService.getExperimentBasedConfig(ConfigKey.Advanced.FreezeCustomizationsIndex, this.experimentationService); + if (!enabled) { + return undefined; + } + const firstTurn = this.props.promptContext.conversation?.turns.at(0); + if (!firstTurn) { + return undefined; + } + const variable = this.props.promptContext.chatVariables.find(isCustomizationsIndex); + const currentValue = variable && typeof variable.value === 'string' ? variable.value : undefined; + const currentToolReferences = variable?.reference.toolReferences; + + const currentCacheKey = this.instantiationService.invokeFunction(getGlobalContextCacheKey); + const existing = firstTurn.getMetadata(CustomizationsIndexMetadata); + if (existing && existing.cacheKey === currentCacheKey) { + const frozen = { value: existing.value, toolReferences: existing.toolReferences }; + // Surface drift in either direction: a different live value, or the + // live variable disappearing entirely (treated as an empty listing). + // Without the second case the model is left looking at the stale + // frozen ``/``/`` block with no signal + // that entries have been removed. + const effectiveCurrent = currentValue ?? ''; + if (effectiveCurrent !== existing.value) { + return { frozen, drift: { value: effectiveCurrent, toolReferences: currentToolReferences } }; + } + return { frozen }; + } + if (currentValue === undefined) { + return undefined; + } + firstTurn.setMetadata(new CustomizationsIndexMetadata(currentValue, currentToolReferences, currentCacheKey)); + return { frozen: { value: currentValue, toolReferences: currentToolReferences } }; + } + private async getOrCreateGlobalAgentContext(endpoint: IChatEndpoint): Promise { const globalContext = await this.getOrCreateGlobalAgentContextContent(endpoint); const isNewChat = this.props.promptContext.history?.length === 0; @@ -325,6 +390,25 @@ export interface AgentUserMessageProps extends BasePromptElementProps, AgentUser readonly additionalHookContext?: string; /** When true, this request was system-initiated (e.g. terminal completion notification) and should skip context/wrapping. */ readonly isSystemInitiated?: boolean; + /** + * Live customizations-index text rendered into the latest user message + * when it has drifted from the frozen snapshot in the system prompt. + * Lets the model see mid-conversation updates (new skill, mode swap, etc.) + * without invalidating the system prompt cache. + * + * Rendered inside this message's `` (rather than as a + * sibling `UserMessage`) so it becomes part of the captured + * `RenderedUserMessageMetadata`. Emitting it as a separate UserMessage + * would cause `result.messages.at(-1)` in `agentIntent.runOne` to point + * at the drift block instead of the user query — the metadata would then + * store just the drift block and historical replays on later turns would + * lose the actual user query, busting cross-turn cache continuity. + * + * Only set when the experimental `FreezeCustomizationsIndex` setting is + * enabled and the current value differs from the snapshot captured on + * the first turn. + */ + readonly customizationsIndexUpdate?: { value: string; toolReferences: readonly ChatLanguageModelToolReference[] | undefined }; } export function getUserMessagePropsFromTurn(turn: Turn, endpoint: IChatEndpoint, customizations?: AgentUserMessageCustomizations): AgentUserMessageProps { @@ -441,6 +525,7 @@ export class AgentUserMessage extends PromptElement { {hasTerminalTool && } {hasTodoTool && } {this.props.additionalHookContext && } + {this.props.customizationsIndexUpdate && } @@ -536,6 +621,47 @@ class AdditionalHookContextPrompt extends PromptElement` alongside `EditedFileEvents`, + * `TerminalStatePromptElement`, etc., so it becomes part of the same + * `UserMessage` and is captured into `RenderedUserMessageMetadata` together + * with the user query. Replays verbatim on subsequent turns as the + * historical user message, preserving cross-turn cache continuity. Emitting + * it as a sibling `UserMessage` would make `result.messages.at(-1)` point at + * the drift block in `agentIntent.runOne`, so the metadata would store just + * the drift block and historical replays on later turns would lose the + * actual user query. + * + * Used only when `FreezeCustomizationsIndex` is on and the live index + * differs from the snapshot captured on the first turn. + */ +class CustomizationsIndexUpdate extends PromptElement { + constructor( + props: CustomizationsIndexUpdateProps, + @IPromptVariablesService private readonly promptVariablesService: IPromptVariablesService, + ) { + super(props); + } + + async render() { + let value = this.props.update.value; + const toolReferences = this.props.update.toolReferences; + if (toolReferences?.length) { + value = await this.promptVariablesService.resolveToolReferencesInPrompt(value, toolReferences); + } + return + The available instructions, skills, and agents have changed since this conversation started. The listings below supersede the ones in the system prompt.
+ {value} + ; + } +} + interface SkillAdherenceReminderProps extends BasePromptElementProps { readonly chatVariables: ChatVariablesCollection; } diff --git a/extensions/copilot/src/extension/prompts/node/agent/summarizedConversationHistory.tsx b/extensions/copilot/src/extension/prompts/node/agent/summarizedConversationHistory.tsx index aaabcb6a965..a368814c9d6 100644 --- a/extensions/copilot/src/extension/prompts/node/agent/summarizedConversationHistory.tsx +++ b/extensions/copilot/src/extension/prompts/node/agent/summarizedConversationHistory.tsx @@ -7,7 +7,7 @@ import * as l10n from '@vscode/l10n'; import { BasePromptElementProps, PrioritizedList, PromptElement, PromptMetadata, PromptSizing, Raw, SystemMessage, UserMessage } from '@vscode/prompt-tsx'; import { BudgetExceededError } from '@vscode/prompt-tsx/dist/base/materialized'; import { ChatMessage } from '@vscode/prompt-tsx/dist/base/output/rawTypes'; -import type { ChatResponsePart, ChatResultPromptTokenDetail, LanguageModelToolInformation, NotebookDocument, Progress } from 'vscode'; +import type { ChatLanguageModelToolReference, ChatResponsePart, ChatResultPromptTokenDetail, LanguageModelToolInformation, NotebookDocument, Progress } from 'vscode'; import { IChatHookService, PreCompactHookInput } from '../../../../platform/chat/common/chatHookService'; import { ChatFetchResponseType, ChatLocation, ChatResponse, FetchSuccess } from '../../../../platform/chat/common/commonTypes'; import { getTextPart } from '../../../../platform/chat/common/globalStringUtils'; @@ -263,7 +263,7 @@ class ConversationHistory extends PromptElement { userQueryTagName: this.props.userQueryTagName, ReminderInstructionsClass: this.props.ReminderInstructionsClass, ToolReferencesHintClass: this.props.ToolReferencesHintClass, - })} />); + })} customizationsIndexUpdate={this.props.customizationsIndexUpdate} />); } // We may have a summary from earlier in the conversation, but skip history if we have a new summary @@ -408,6 +408,11 @@ export interface SummarizedAgentHistoryProps extends BasePromptElementProps, Age readonly summarizationInstructions?: string; /** Skip Full mode and go straight to Simple mode for foreground budget-exceeded recovery. */ readonly forceSimpleSummary?: boolean; + /** + * Forwarded to the latest user message when the customizations-index has + * drifted from its frozen snapshot. See {@link AgentUserMessageProps.customizationsIndexUpdate}. + */ + readonly customizationsIndexUpdate?: { value: string; toolReferences: readonly ChatLanguageModelToolReference[] | undefined }; } /** diff --git a/extensions/copilot/src/extension/prompts/node/panel/customInstructions.tsx b/extensions/copilot/src/extension/prompts/node/panel/customInstructions.tsx index dc45b214c33..8d7179db621 100644 --- a/extensions/copilot/src/extension/prompts/node/panel/customInstructions.tsx +++ b/extensions/copilot/src/extension/prompts/node/panel/customInstructions.tsx @@ -20,6 +20,21 @@ import { Tag } from '../base/tag'; export interface CustomInstructionsProps extends BasePromptElementProps { readonly chatVariables: ChatVariablesCollection | undefined; + /** + * When provided, this value is used in place of the live customizations-index + * variable's value (the bundled ``/``/`` text). + * The agent prompt sets this to freeze the index at the first turn of a + * conversation so per-turn churn doesn't invalidate the system prompt + * cache. + * + * Must be paired with {@link customizationsIndexToolReferencesOverride} + * because tool references carry byte offsets into this text — using the + * current turn's offsets against the frozen text would mis-slice. + */ + readonly customizationsIndexOverride?: string; + /** Tool references paired with {@link customizationsIndexOverride}. */ + readonly customizationsIndexToolReferencesOverride?: readonly ChatLanguageModelToolReference[]; + readonly languageId: string | undefined; /** * @default true @@ -71,14 +86,27 @@ export class CustomInstructions extends PromptElement { if (includeCodeGenerationInstructions !== false) { const hasSeen = new ResourceSet(); const hasSeenContent = new Set(); + + const liveIndexVariable = this.props.chatVariables?.find(isCustomizationsIndex); + const overrideActive = this.props.customizationsIndexOverride !== undefined; + const indexValue = overrideActive + ? this.props.customizationsIndexOverride! + : (liveIndexVariable && typeof liveIndexVariable.value === 'string' ? liveIndexVariable.value : undefined); + const indexToolReferences = overrideActive + ? this.props.customizationsIndexToolReferencesOverride + : liveIndexVariable?.reference.toolReferences; + if (indexValue !== undefined) { + let value = indexValue; + if (indexToolReferences?.length) { + value = await this.promptVariablesService.resolveToolReferencesInPrompt(value, indexToolReferences); + } + chunks.push({value}); + } + if (this.props.chatVariables) { for (const variable of this.props.chatVariables) { if (isCustomizationsIndex(variable)) { - let value = variable.value; - if (variable.reference.toolReferences?.length) { - value = await this.promptVariablesService.resolveToolReferencesInPrompt(value, variable.reference.toolReferences); - } - chunks.push({value}); + continue; } else if (isInstructionFile(variable)) { const value = variable.value; if (!hasSeen.has(value)) { diff --git a/extensions/copilot/src/platform/configuration/common/configurationService.ts b/extensions/copilot/src/platform/configuration/common/configurationService.ts index 319fbae28d3..57ea114f940 100644 --- a/extensions/copilot/src/platform/configuration/common/configurationService.ts +++ b/extensions/copilot/src/platform/configuration/common/configurationService.ts @@ -730,6 +730,9 @@ export namespace ConfigKey { /** Enable extended (1 hour) prompt cache TTL on tools and system blocks for the Anthropic Messages API. Only applied to 1M context Claude variants. */ export const AnthropicExtendedCacheTtl = defineSetting('chat.anthropic.promptCaching.extendedTtl', ConfigType.ExperimentBased, false); + /** Freeze the customizations-index variable (the ``/``/`` block) at the first turn of a conversation and reuse it on subsequent turns. Prevents the system prompt cache from being invalidated by per-turn churn — e.g. the active mode swapping which subagent entry appears in ``, or async experimentation flipping a `when`-gated skill. */ + export const FreezeCustomizationsIndex = defineSetting('chat.freezeCustomizationsIndex', ConfigType.ExperimentBased, false); + export const InlineEditsXtabProviderModelConfiguration = (() => { const oldKey = 'chat.advanced.inlineEdits.xtabProvider.modelConfiguration'; const newKey = 'chat.inlineEdits.xtabProvider.modelConfiguration';