Skip to content

services.md

Latest commit

 

History

History
489 lines (338 loc) · 30.7 KB

services.md

File metadata and controls

489 lines (338 loc) · 30.7 KB

Services and Integrations

The services/ directory contains the core subsystems that support the main REPL loop. Each service is designed as a self-contained module with a clear public API, communicating through well-defined boundaries rather than tight coupling. Services range from infrastructure concerns (analytics, cost tracking, rate limits) to feature-level capabilities (MCP, voice, LSP, compaction).

This document covers each service in depth, including key functions, data flow, and integration points.

1. MCP Integration

Directory: services/mcp/

The Model Context Protocol subsystem connects Claude Code to external tool servers, exposing their capabilities as native tools within the conversation loop. It is one of the largest services, spanning client lifecycle management, configuration resolution, authentication, and tool execution.

Client lifecycle

services/mcp/client.ts is the primary module. It implements the full MCP client using @modelcontextprotocol/sdk:

  • Transport support -- Stdio (StdioClientTransport), SSE (SSEClientTransport), Streamable HTTP (StreamableHTTPClientTransport), WebSocket (WebSocketTransport), and an in-process SDK control transport (SdkControlClientTransport).
  • ensureConnectedClient() -- Lazily connects to an MCP server, caching the client instance per server name. Handles reconnection on auth expiry.
  • getMcpToolsCommandsAndResources() -- After connection, fetches all tools, commands, and resources from a server. Tools are wrapped as MCPTool instances, resources as ServerResource objects. Commands from MCP prompts are also extracted.
  • fetchToolsForClient() / fetchResourcesForClient() / fetchCommandsForClient() -- Individual fetch functions for each capability type.
  • callMCPToolWithUrlElicitationRetry() -- Executes an MCP tool call with automatic retry when the server requests OAuth elicitation.
  • processMCPResult() / transformMCPResult() -- Processes raw MCP tool call results into Claude-compatible content blocks, handling binary blobs, image resizing, large output truncation, and Unicode sanitization.
  • reconnectMcpServerImpl() -- Tears down and re-establishes a server connection, returning fresh tools/commands/resources.
  • clearServerCache() -- Removes cached client state for a server.
  • callIdeRpc() -- Sends RPC calls to IDE-connected MCP servers (used for diagnostics, file sync, etc.).

Configuration

services/mcp/config.ts resolves MCP server configuration from multiple scopes:

  • Scopes: local, user, project, dynamic, enterprise, claudeai, managed (defined in ConfigScope type in types.ts).
  • getAllMcpConfigs() -- Merges server configs from all scopes with precedence rules. Enterprise configs come from managed-mcp.json; project configs from .mcp.json; user configs from global settings.
  • getMcpConfigsByScope() -- Returns servers filtered to a single scope.
  • getEnterpriseMcpFilePath() -- Returns path to the enterprise-managed MCP config file.
  • isMcpServerDisabled() / setMcpServerEnabled() -- Per-server enable/disable toggle.
  • writeMcpjsonFile() -- Atomic write to .mcp.json with permission preservation and temp-file rename.
  • Server config types: McpStdioServerConfig, McpSSEServerConfig, McpHTTPServerConfig, McpWebSocketServerConfig (all defined in types.ts with Zod schemas).

Transport types

services/mcp/types.ts defines transport variants: stdio, sse, sse-ide, http, ws, sdk. Each has its own Zod schema for validation. Server configs include optional OAuth configuration (McpOAuthConfigSchema) and Cross-App Access (XAA) flags.

Connection management (React layer)

services/mcp/MCPConnectionManager.tsx provides a React context wrapping useManageMCPConnections(). Exposes:

  • useMcpReconnect() -- Hook for reconnecting a specific server by name.
  • useMcpToggleEnabled() -- Hook for toggling a server on/off.

services/mcp/useManageMCPConnections.ts contains the connection management logic, including:

  • Listening for ToolListChangedNotification, ResourceListChangedNotification, and PromptListChangedNotification from the MCP SDK to auto-refresh capabilities.
  • Deduplication of Claude.ai MCP servers (dedupClaudeAiMcpServers()).
  • Enterprise MCP config detection (doesEnterpriseMcpConfigExist()).
  • Policy-based server filtering (filterMcpServersByPolicy()).

Auth approvals

services/mcpServerApproval.tsx renders Ink-based approval dialogs when project-scoped MCP servers need user consent:

  • handleMcpjsonServerApprovals() -- Checks for pending project servers and renders either MCPServerApprovalDialog (single server) or MCPServerMultiselectDialog (multiple servers).

services/mcp/auth.ts implements ClaudeAuthProvider for OAuth-based MCP server authentication, including step-up detection and token refresh.

Other MCP modules

File Purpose
channelPermissions.ts Permission checks for MCP channels
channelAllowlist.ts Allowlist management for MCP server channels
channelNotification.ts Notification routing for channel events
elicitationHandler.ts Handles MCP elicitation requests (OAuth prompts to users)
envExpansion.ts Environment variable expansion in server configs
headersHelper.ts Custom header resolution for HTTP-based MCP servers
mcpStringUtils.ts buildMcpToolName() and name formatting utilities
normalization.ts normalizeNameForMCP() -- sanitizes server/tool names
officialRegistry.ts Integration with the official MCP server registry
oauthPort.ts OAuth callback port management
InProcessTransport.ts Transport for in-process MCP servers
SdkControlTransport.ts Transport for SDK-controlled MCP connections
vscodeSdkMcp.ts VS Code SDK MCP integration
xaa.ts / xaaIdpLogin.ts Cross-App Access (XAA) identity provider login
claudeai.ts Claude.ai-specific MCP config fetching
utils.ts getProjectMcpServerStatus(), getLoggingSafeMcpBaseUrl()

2. Analytics and Diagnostics

Directory: services/analytics/

The analytics subsystem provides sanitized, privacy-aware event logging with zero-dependency design to avoid import cycles.

Event logging API

services/analytics/index.ts is the public entry point. It exposes:

  • logEvent(eventName, metadata) -- Synchronous event logging. Events are queued in memory until a sink is attached.
  • logEventAsync(eventName, metadata) -- Async variant for events that need delivery confirmation.
  • attachAnalyticsSink(sink) -- Called during app startup to attach the routing backend. Idempotent. Queued events are drained via queueMicrotask to avoid blocking startup.
  • stripProtoFields(metadata) -- Removes _PROTO_* keys from metadata before routing to general-access storage. Proto-prefixed keys are only sent to the first-party exporter where they map to PII-tagged proto columns.

Safety types:

  • AnalyticsMetadata_I_VERIFIED_THIS_IS_NOT_CODE_OR_FILEPATHS -- A never-typed brand that forces developers to explicitly cast string values, confirming they do not contain code or file paths.
  • AnalyticsMetadata_I_VERIFIED_THIS_IS_PII_TAGGED -- Brand for values destined for PII-tagged proto columns with restricted access.

Sink implementation

services/analytics/sink.ts routes events to two backends:

  • Datadog (datadog.ts) -- Gated by a feature flag (tengu_log_datadog_events). Uses trackDatadogEvent().
  • First-party event logger (firstPartyEventLogger.ts) -- Logs to Anthropic's internal event pipeline. Applies sampling via shouldSampleEvent().
  • Kill switch (sinkKillswitch.ts) -- Can disable individual sinks at runtime.

initializeAnalyticsSink() wires up the sink during app startup.

Supporting modules

File Purpose
config.ts Analytics configuration
growthbook.ts Feature flag integration via GrowthBook. getFeatureValue_CACHED_MAY_BE_STALE() is used throughout the codebase for gating features
metadata.ts Event metadata enrichment

Diagnostic tracking

services/diagnosticTracking.ts provides DiagnosticTrackingService, a singleton that tracks IDE diagnostics (errors/warnings) across files:

  • initialize(mcpClient) -- Connects to an IDE MCP client to fetch diagnostics.
  • shutdown() / reset() -- Lifecycle management for diagnostic state.
  • Maintains a baseline of diagnostics and tracks changes introduced by Claude's edits.
  • Normalizes file URIs and caps diagnostic summaries at 4000 characters.

3. Cost Tracking

Files: cost-tracker.ts, costHook.ts

Cost tracking aggregates token usage and USD costs across the entire session, persisting state between sessions and displaying summaries on exit.

Core functions in cost-tracker.ts

  • getStoredSessionCosts(sessionId) -- Reads persisted cost state from project config. Only returns data if the session ID matches the last saved session.
  • restoreCostStateForSession(sessionId) -- Restores accumulated cost state when resuming a session. Returns true if restoration succeeded.
  • saveCurrentSessionCosts(fpsMetrics?) -- Persists current cost counters to project config. Saves: total cost, API duration, tool duration, lines added/removed, per-model token breakdowns, web search request counts, and FPS metrics.
  • formatTotalCost() -- Renders a chalk-formatted summary including cost, API/wall duration, code change stats, and per-model usage breakdown.
  • formatCost(cost, maxDecimalPlaces) -- Formats a USD value, switching between 2 and 4 decimal places based on magnitude.

State management

Cost state lives in bootstrap/state.ts with accessor functions re-exported from cost-tracker.ts:

Function Description
getTotalCostUSD() Accumulated USD cost for the session
getTotalInputTokens() / getTotalOutputTokens() Total token counts
getTotalCacheReadInputTokens() / getTotalCacheCreationInputTokens() Prompt cache statistics
getTotalAPIDuration() / getTotalAPIDurationWithoutRetries() API call timing
getTotalLinesAdded() / getTotalLinesRemoved() Code change tracking
getModelUsage() / getUsageForModel(model) Per-model ModelUsage records (input/output/cache tokens, cost, web searches)
hasUnknownModelCost() Whether any model lacked a pricing entry

Actual USD calculation is performed by calculateUSDCost() in utils/modelCost.ts, which maps model names to per-token prices for input, output, cache read, and cache creation.

StoredCostState

The StoredCostState type persisted to project config includes: totalCostUSD, totalAPIDuration, totalAPIDurationWithoutRetries, totalToolDuration, totalLinesAdded, totalLinesRemoved, lastDuration, and modelUsage (keyed by model name).

The cost hook

costHook.ts contains useCostSummary(), a React hook that:

  1. Registers a process.on('exit') handler.
  2. On exit, prints the formatted cost summary (only if the user has console billing access via hasConsoleBillingAccess()).
  3. Calls saveCurrentSessionCosts() to persist state for session resumption.

4. Voice I/O

Files: services/voice.ts, services/voiceStreamSTT.ts, services/voiceKeyterms.ts

Voice input provides push-to-talk speech-to-text for the CLI, using native audio capture with fallback to external tools.

Audio recording (voice.ts)

Recording uses a layered capture strategy:

  1. Native audio capture (audio-capture-napi) -- Preferred on macOS, Linux, and Windows. Links against CoreAudio/AudioUnit frameworks. Lazy-loaded on first voice keypress via loadAudioNapi() to avoid blocking startup (dlopen takes 1-8s).
  2. SoX rec -- Fallback on Linux. Detected via hasCommand('rec').
  3. ALSA arecord -- Second fallback on Linux. Validated with probeArecord() which spawns arecord for 150ms to test device availability (handles WSL/headless scenarios).

Constants: 16kHz sample rate, mono channel, 2.0s silence detection duration, 3% silence threshold.

checkRecordingAvailability() probes all capture backends in priority order and returns whether recording is possible on the current platform.

Speech-to-text streaming (voiceStreamSTT.ts)

Connects to Anthropic's voice_stream WebSocket endpoint at /api/ws/speech_to_text/voice_stream:

  • Wire protocol: JSON control messages (KeepAlive, CloseStream) and binary audio frames. Server responds with TranscriptText (partial/final text) and TranscriptEndpoint (end-of-utterance) messages.
  • VoiceStreamConnection -- The connection interface with send(audioChunk), finalize(), close(), and isConnected().
  • VoiceStreamCallbacks -- Callback interface: onTranscript(text, isFinal), onError(error), onClose(), onReady(connection).
  • isVoiceStreamAvailable() -- Checks if voice streaming is available (requires Anthropic OAuth authentication).
  • FinalizeSource -- Indicates how transcription ended: post_closestream_endpoint, no_data_timeout (1500ms, silent-drop detection), safety_timeout (5000ms), ws_close, or ws_already_closed.
  • Keepalive: 8-second interval to maintain the WebSocket connection.

Voice keyterms (voiceKeyterms.ts)

Improves STT accuracy by providing domain-specific vocabulary hints (Deepgram keywords):

  • getVoiceKeyterms(recentFiles?) -- Builds a list of up to MAX_KEYTERMS (50) terms from:
    • Global keyterms: Hardcoded coding terms that Deepgram consistently mangles (MCP, symlink, grep, regex, localhost, TypeScript, JSON, OAuth, webhook, gRPC, dotfiles, subagent, worktree).
    • Project root basename -- e.g., "claude-code" as a phrase.
    • Git branch words -- splitIdentifier() breaks feat/voice-keyterms into individual words.
    • Recent file names -- Extracts words from file stems.
  • splitIdentifier(name) -- Splits camelCase, PascalCase, kebab-case, snake_case, and path segments into words. Filters to 3-20 character length.

5. Rate Limiting and Policy

Files: services/rateLimitMessages.ts, services/rateLimitMocking.ts, services/mockRateLimits.ts, services/claudeAiLimits.ts

Rate limit messages

services/rateLimitMessages.ts is the single source of truth for rate limit user messages:

  • getRateLimitMessage(limits, model) -- Returns the appropriate warning/error message based on ClaudeAILimits state (overage status, weekly limits, session limits).
  • isRateLimitErrorMessage(text) -- Detects rate limit error messages by prefix matching against RATE_LIMIT_ERROR_PREFIXES.
  • Handles overage scenarios, fallback model availability, and different subscription tiers.

Mock rate limits (internal testing)

services/mockRateLimits.ts provides test scenarios for rate limit UI without hitting real API limits:

  • MockScenario type covers scenarios: normal, session-limit-reached, approaching-weekly-limit, weekly-limit-reached, overage-active, overage-warning, overage-exhausted, out-of-credits, opus-limit, sonnet-limit, fast-mode-limit, and more.
  • Generates mock anthropic-ratelimit-unified-* headers for each scenario.

services/rateLimitMocking.ts is the facade layer:

  • processRateLimitHeaders(headers) -- Applies mock headers if the /mock-limits command is active, otherwise passes through real headers.
  • shouldProcessRateLimits(isSubscriber) -- Returns true for subscribers or when mock limits are active.
  • checkMockRateLimitError(currentModel, isFastModeActive) -- Generates synthetic 429 errors for mock scenarios.

Policy limits

Directory: services/policyLimits/

Fetches organization-level policy restrictions from the API:

  • initializePolicyLimitsLoadingPromise() -- Creates a loading promise early in startup for other systems to await. Includes a 30-second timeout to prevent deadlocks.
  • loadPolicyLimits() -- Fetches restrictions from the API with ETag caching. Eligible users: all console (API key) users; Team and Enterprise OAuth users.
  • Polling: Background polling every hour (POLLING_INTERVAL_MS = 3,600,000ms), with retry logic up to 5 attempts.
  • Fail-open: If the API call fails, the CLI continues without restrictions.
  • Cache: Session-level in-memory cache plus disk cache at policy-limits.json in the config home directory.

services/policyLimits/types.ts defines the response schema:

type PolicyLimitsResponse = {
  restrictions: Record<string, { allowed: boolean }>
}

Only blocked policies are included in the response. Absent keys are treated as allowed.

6. Compaction Service

Directory: services/compact/

Compaction manages conversation context length by summarizing older messages when the token count approaches the model's context window limit.

Auto-compaction trigger (autoCompact.ts)

  • getEffectiveContextWindowSize(model) -- Returns context window minus reserved tokens for summary output (capped at 20,000 tokens based on p99.99 data).
  • getAutoCompactThreshold(model) -- Effective context window minus AUTOCOMPACT_BUFFER_TOKENS (13,000). Overridable via CLAUDE_AUTOCOMPACT_PCT_OVERRIDE env var.
  • calculateTokenWarningState(tokenUsage, model) -- Returns percentLeft, isAboveWarningThreshold, isAboveErrorThreshold, and isAboveAutoCompactThreshold.
  • Circuit breaker: MAX_CONSECUTIVE_AUTOCOMPACT_FAILURES = 3 -- stops retrying after 3 consecutive failures to prevent wasting API calls.
  • Buffer constants: WARNING_THRESHOLD_BUFFER_TOKENS = 20,000, ERROR_THRESHOLD_BUFFER_TOKENS = 20,000, MANUAL_COMPACT_BUFFER_TOKENS = 3,000.

Core compaction (compact.ts)

  • compactConversation() -- The main compaction function. Uses a forked subagent (runForkedAgent()) to summarize conversation history. Produces a CompactionResult containing the summary and boundary markers.
  • partialCompactConversation() -- Partial compaction for incremental summarization (keeps recent messages intact).
  • buildPostCompactMessages(result) -- Constructs the post-compaction message array with boundary markers.
  • stripImagesFromMessages() -- Removes image content before compaction to save tokens.
  • truncateHeadForPTLRetry() -- Truncates early messages when hitting prompt-too-long errors.
  • createCompactCanUseTool() -- Returns a restricted CanUseToolFn for the compaction agent.

Message grouping (grouping.ts)

  • groupMessagesByApiRound(messages) -- Groups messages at API-round boundaries (one group per round-trip). Uses assistant message IDs as boundary markers. This enables reactive compaction on single-prompt agentic sessions where the entire workload is one human turn.

Micro-compaction (microCompact.ts)

Time-based micro-compaction that clears old tool results without a full summarization pass:

  • COMPACTABLE_TOOLS -- Only compacts results from: FileRead, Shell tools, Grep, Glob, WebSearch, WebFetch, FileEdit, FileWrite.
  • Time-based MC clears tool result content with [Old tool result content cleared] markers.
  • Cached MC (feature-gated) -- Uses a separate cachedMicrocompact.js module for cache-aware clearing.

Supporting modules

File Purpose
prompt.ts Compaction prompt templates. Includes NO_TOOLS_PREAMBLE to prevent the summarizer from calling tools, plus DETAILED_ANALYSIS_INSTRUCTION_BASE / DETAILED_ANALYSIS_INSTRUCTION_PARTIAL for thorough summary generation
postCompactCleanup.ts runPostCompactCleanup() -- cleanup after compaction
sessionMemoryCompact.ts trySessionMemoryCompaction() -- compacts session memory alongside conversation
compactWarningState.ts Manages warning suppression state (pure, React-free)
compactWarningHook.ts useCompactWarningSuppression() -- React hook for warning state
timeBasedMCConfig.ts getTimeBasedMCConfig() -- configuration for time-based micro-compaction
apiMicrocompact.ts API-level micro-compaction implementation

7. Magic Docs

Directory: services/MagicDocs/

Magic Docs automatically maintains markdown documentation files that are marked with a special header. When Claude reads a file containing # MAGIC DOC: [title], the system registers it and periodically updates it in the background using a forked subagent.

How it works

  1. Detection -- detectMagicDocHeader(content) scans for the pattern # MAGIC DOC: [title] at the start of a file. Optionally detects italicized instructions on the next line.
  2. Registration -- registerMagicDoc(filePath) adds the file to the trackedMagicDocs map (idempotent per path).
  3. Listening -- A registerFileReadListener() hook detects Magic Doc headers when files are read during normal tool use.
  4. Update cycle -- A registerPostSamplingHook() fires after each assistant turn. For each tracked doc:
    • Re-reads the file (with a cloned FileStateCache to bypass dedup).
    • Verifies the header is still present.
    • Runs a forked magic-docs agent (Sonnet model) with buildMagicDocsUpdatePrompt().
    • The agent has access only to the Edit tool to modify the doc.
  5. Cleanup -- clearTrackedMagicDocs() resets tracking state.

8. Prompt Suggestions

Directory: services/PromptSuggestion/

Generates follow-up prompt suggestions after Claude completes a response, helping users continue their workflow.

Suggestion generation (promptSuggestion.ts)

  • shouldEnablePromptSuggestion() -- Gating logic that checks (in order): env var override, GrowthBook feature flag (tengu_chomp_inflection), non-interactive mode (disabled), swarm teammate status (disabled), user settings, and rate limits.
  • getPromptVariant() -- Returns the prompt variant ('user_intent').
  • generateSuggestion() -- Runs a forked agent to generate a suggestion based on conversation context.
  • Uses a dedicated abort controller (currentAbortController) so suggestions can be cancelled when the user types.

Speculation (speculation.ts)

Speculative execution that pre-runs the suggested prompt in an overlay filesystem:

  • isSpeculationEnabled() / startSpeculation() -- Controls speculative pre-execution.
  • Limits: MAX_SPECULATION_TURNS = 20, MAX_SPECULATION_MESSAGES = 100.
  • Safety: Categorizes tools into WRITE_TOOLS (Edit, Write, NotebookEdit) and SAFE_READ_ONLY_TOOLS (Read, Glob, Grep, ToolSearch, LSP, TaskGet, TaskList). Write operations go to an overlay directory.
  • safeRemoveOverlay(overlayPath) -- Cleans up overlay directories after speculation completes or is cancelled.
  • Speculation results are stored as SpeculationState / SpeculationResult in the app state.

9. LSP Service

Directory: services/lsp/

Integrates Language Server Protocol servers to provide diagnostics, definitions, and other IDE-like features.

Architecture

The LSP subsystem follows a manager-instance pattern:

  • createLSPServerManager() (LSPServerManager.ts) -- Factory that creates a manager routing requests based on file extensions. Maintains:

    • servers map (server name to instance).
    • extensionMap (file extension to server names).
    • openedFiles map (URI to owning server name).
    • Methods: initialize(), shutdown(), getServerForFile(), ensureServerStarted(), sendRequest(), openFile(), changeFile(), saveFile(), closeFile(), isFileOpen().

  • createLSPServerInstance() (LSPServerInstance.ts) -- Manages a single LSP server lifecycle:

    • State tracking: LspServerState (starting, running, stopped, crashed).
    • Health monitoring with crash callback (onCrash).
    • Transient error retry: MAX_RETRIES_FOR_TRANSIENT_ERRORS = 3 with 500ms base exponential backoff for content modified errors (LSP_ERROR_CONTENT_MODIFIED = -32801).
    • Methods: start(), stop(), restart(), isHealthy(), sendRequest(), sendNotification(), onNotification().

  • createLSPClient() (LSPClient.ts) -- Low-level wrapper around vscode-jsonrpc MessageConnection:

    • Spawns the LSP server process via stdio.
    • Manages connection lifecycle (start, initialize, stop).
    • Supports lazy handler registration (queued before connection is ready).
    • Methods: start(), initialize(), sendRequest(), sendNotification(), onNotification(), onRequest(), stop().

Configuration

services/lsp/config.ts -- LSP servers are only loaded from plugins (not user/project settings):

  • getAllLspServers() -- Loads all enabled plugins and extracts LSP server configurations via getPluginLspServers(). Plugin loading errors are tolerated (fail-open).

Diagnostics

  • LSPDiagnosticRegistry.ts -- Stores asynchronous diagnostics from LSP servers:

    • registerPendingLSPDiagnostic() -- Stores diagnostics received via textDocument/publishDiagnostics.
    • Cross-turn deduplication via LRU cache (MAX_DELIVERED_FILES = 500).
    • Volume limits: MAX_DIAGNOSTICS_PER_FILE = 10, MAX_TOTAL_DIAGNOSTICS = 30.
    • Diagnostics are delivered as attachments in the next query.

  • passiveFeedback.ts -- Converts LSP diagnostics to Claude's DiagnosticFile[] format:

    • formatDiagnosticsForAttachment() -- Maps LSP severity (1-4) to Claude severity strings (Error/Warning/Info/Hint).
    • mapLSPSeverity() -- The severity mapping function.

10. OAuth Service

Directory: services/oauth/

For complete authentication details, see auth.md.

Implements OAuth 2.0 Authorization Code flow with PKCE:

  • OAuthService (index.ts) -- Main service class. startOAuthFlow() supports automatic (browser redirect to localhost) and manual (copy-paste code) flows. Uses AuthCodeListener for the localhost callback server.
  • client.ts -- buildAuthUrl() constructs authorization URLs for both Console and Claude.ai. shouldUseClaudeAIAuth() checks for the CLAUDE_AI_INFERENCE_SCOPE. Token exchange, refresh, and profile fetching.
  • crypto.ts -- PKCE utilities: generateCodeVerifier(), generateCodeChallenge(), generateState().
  • auth-code-listener.ts -- Local HTTP server for receiving OAuth callbacks.
  • getOauthProfile.ts -- Fetches user profile from OAuth tokens.

11. Plugins Service

Directory: services/plugins/

For complete plugin architecture details, see plugins.md.

Core plugin operations for install, uninstall, enable, disable, and update:

  • pluginOperations.ts -- Pure library functions (no process.exit(), no console output). Used by both CLI commands and interactive UI. Handles marketplace integration, dependency resolution (findReverseDependents()), and plugin data directory management.
  • pluginCliCommands.ts -- CLI command implementations for plugin management.
  • PluginInstallationManager.ts -- Manages plugin installation lifecycle.

12. Key Source Files Reference

File Service Key exports
services/mcp/client.ts MCP ensureConnectedClient(), getMcpToolsCommandsAndResources(), callMCPToolWithUrlElicitationRetry(), processMCPResult(), reconnectMcpServerImpl(), callIdeRpc()
services/mcp/config.ts MCP getAllMcpConfigs(), getMcpConfigsByScope(), getEnterpriseMcpFilePath(), isMcpServerDisabled()
services/mcp/types.ts MCP ConfigScope, McpStdioServerConfig, McpSSEServerConfig, McpHTTPServerConfig, ScopedMcpServerConfig, MCPServerConnection
services/mcp/MCPConnectionManager.tsx MCP MCPConnectionManager, useMcpReconnect(), useMcpToggleEnabled()
services/mcp/useManageMCPConnections.ts MCP Connection lifecycle, tool/resource/command refresh
services/mcpServerApproval.tsx MCP handleMcpjsonServerApprovals()
services/mcp/auth.ts MCP ClaudeAuthProvider, wrapFetchWithStepUpDetection()
services/analytics/index.ts Analytics logEvent(), logEventAsync(), attachAnalyticsSink(), stripProtoFields()
services/analytics/sink.ts Analytics initializeAnalyticsSink()
services/analytics/growthbook.ts Analytics getFeatureValue_CACHED_MAY_BE_STALE()
services/diagnosticTracking.ts Diagnostics DiagnosticTrackingService
cost-tracker.ts Cost getStoredSessionCosts(), restoreCostStateForSession(), saveCurrentSessionCosts(), formatTotalCost()
costHook.ts Cost useCostSummary()
services/voice.ts Voice loadAudioNapi(), checkRecordingAvailability(), probeArecord()
services/voiceStreamSTT.ts Voice VoiceStreamConnection, VoiceStreamCallbacks, isVoiceStreamAvailable()
services/voiceKeyterms.ts Voice getVoiceKeyterms(), splitIdentifier()
services/rateLimitMessages.ts Rate limits getRateLimitMessage(), isRateLimitErrorMessage()
services/rateLimitMocking.ts Rate limits processRateLimitHeaders(), checkMockRateLimitError()
services/mockRateLimits.ts Rate limits MockScenario, mock header generation
services/claudeAiLimits.ts Rate limits ClaudeAILimits state management
services/policyLimits/index.ts Policy initializePolicyLimitsLoadingPromise(), loadPolicyLimits()
services/policyLimits/types.ts Policy PolicyLimitsResponse, PolicyLimitsFetchResult
services/compact/compact.ts Compaction compactConversation(), partialCompactConversation(), buildPostCompactMessages()
services/compact/autoCompact.ts Compaction getAutoCompactThreshold(), calculateTokenWarningState(), getEffectiveContextWindowSize()
services/compact/microCompact.ts Compaction Time-based/cached micro-compaction, COMPACTABLE_TOOLS
services/compact/grouping.ts Compaction groupMessagesByApiRound()
services/compact/prompt.ts Compaction Compaction prompt templates
services/MagicDocs/magicDocs.ts Magic Docs detectMagicDocHeader(), registerMagicDoc(), clearTrackedMagicDocs()
services/PromptSuggestion/promptSuggestion.ts Suggestions shouldEnablePromptSuggestion(), generateSuggestion(), getPromptVariant()
services/PromptSuggestion/speculation.ts Suggestions isSpeculationEnabled(), startSpeculation()
services/lsp/LSPServerManager.ts LSP createLSPServerManager()
services/lsp/LSPServerInstance.ts LSP createLSPServerInstance()
services/lsp/LSPClient.ts LSP createLSPClient()
services/lsp/config.ts LSP getAllLspServers()
services/lsp/LSPDiagnosticRegistry.ts LSP registerPendingLSPDiagnostic()
services/lsp/passiveFeedback.ts LSP formatDiagnosticsForAttachment()
services/oauth/index.ts OAuth OAuthService
services/oauth/client.ts OAuth buildAuthUrl(), shouldUseClaudeAIAuth()
services/plugins/pluginOperations.ts Plugins Plugin install/uninstall/enable/disable operations
services/tokenEstimation.ts Tokens roughTokenCountEstimation()