feat: add native Anthropic Messages API proxy endpoint

CLAUDE.md

@@ -246,6 +246,7 @@ Domains chain together: competitive research feeds marketing and business strate
 
 ## Recent Changes
 
+- anthropic-proxy-endpoint: Native Anthropic Messages API proxy — `POST /ai/anthropic/v1/messages` pass-through to Cloudflare AI Gateway (`apps/api/src/routes/ai-proxy-anthropic.ts`); receives native Anthropic format, forwards unchanged to AI Gateway's `/anthropic/v1/messages` path (no format translation); auth via `x-api-key` header (workspace callback token — matches Claude Code's auth format); forwards `anthropic-version` and `anthropic-beta` headers; SSE streaming pass-through; model validation (claude-* only); `POST /ai/anthropic/v1/messages/count_tokens` token counting endpoint; shared helpers extracted to `apps/api/src/services/ai-proxy-shared.ts` (`extractCallbackToken`, `verifyAIProxyAuth`, `buildAIGatewayMetadata`, `buildAnthropicGatewayUrl`, `resolveAnthropicApiKey`, `AIProxyAuthError`); existing `ai-proxy.ts` refactored to use shared helpers; per-user RPM rate limiting and daily token budget shared with OpenAI proxy; Anthropic-format error responses (`{ type: "error", error: { type, message } }`); `cf-aig-metadata` header injected for cost attribution; configurable via AI_PROXY_ENABLED (kill switch), AI_PROXY_RATE_LIMIT_RPM (default: 30), AI_PROXY_RATE_LIMIT_WINDOW_SECONDS (default: 60), AI_PROXY_DAILY_INPUT_TOKEN_LIMIT (default: 500000), AI_PROXY_DAILY_OUTPUT_TOKEN_LIMIT (default: 200000), AI_GATEWAY_ID
 - cost-monitoring-dashboard: Admin cost monitoring dashboard — `GET /api/admin/costs` route (`apps/api/src/routes/admin-costs.ts`) aggregates LLM costs from Cloudflare AI Gateway logs API (paginated, per-model/per-day/per-user breakdown, trial cost tracking, cached/error request counts) and compute costs from `getAllUsersNodeUsageSummary` node usage service into a unified `CostSummaryResponse`; monthly projection via daily average extrapolation; `AdminCosts.tsx` page with KPI cards (LLM cost, monthly projection, compute estimate, combined), daily cost trend AreaChart, cost by model horizontal BarChart + table, cost by user table; period selector (current-month, 30d, 90d); admin tab at `/admin/costs`; configurable via COST_MONITORING_ENABLED (default: true, set to 'false' to disable), COMPUTE_VCPU_HOUR_COST_USD (default: 0.003), AI_USAGE_PAGE_SIZE (default: 50), AI_USAGE_MAX_PAGES (default: 20, hard cap: 20)
 - sam-observability-context-tools: SAM observability and codebase context tools — 5 new tools in `apps/api/src/durable-objects/sam-session/tools/` enable SAM to search task messages and browse project codebases: `list_sessions` (browse project chat sessions with status/taskId filters via `projectDataService.listSessions`), `get_session_messages` (retrieve grouped messages from a session via `projectDataService.getMessages` + `groupTokensIntoMessages`), `search_task_messages` (full-text search across task messages via `projectDataService.searchMessages` with FTS5/LIKE fallback, taskId→sessionId resolution), `search_code` (GitHub Code Search API with `repo:owner/name` qualifier, path/extension filters, text_matches snippets), `get_file_content` (GitHub Contents API for file content or directory listing with base64 decode); shared `helpers.ts` extracts `resolveProjectWithOwnership()`, `parseRepository()`, `getUserGitHubToken()` from `get-ci-status.ts`; SAM_SYSTEM_PROMPT updated with "Task Message Search (Observability)" and "Codebase Context" tool sections; configurable via SAM_SESSION_MESSAGES_LIMIT (default: 50), SAM_SESSION_MESSAGES_MAX_LIMIT (default: 200), SAM_SESSION_LIST_LIMIT (default: 20), SAM_SESSION_LIST_MAX_LIMIT (default: 100), SAM_TASK_MESSAGE_SEARCH_LIMIT (default: 10), SAM_TASK_MESSAGE_SEARCH_MAX_LIMIT (default: 50), SAM_CODE_SEARCH_LIMIT (default: 10), SAM_CODE_SEARCH_MAX_LIMIT (default: 30), SAM_FILE_CONTENT_MAX_BYTES (default: 1048576)
 - sam-agent-phase-a-tools: SAM Phase A orchestration tools — 4 new tools in `apps/api/src/durable-objects/sam-session/tools/` transform SAM from read-only to functional orchestrator: `dispatch_task` (provisions workspace, runs agent, resolves config via explicit→profile→project→platform chain, reuses `startTaskRunnerDO`, `generateBranchName`, `generateTaskTitle`, `resolveAgentProfile`, `resolveCredentialSource`, `projectDataService.createSession/persistMessage`), `get_task_details` (full task details with output/PR/error via D1 tasks+projects join), `create_mission` (D1 insert + ProjectOrchestrator DO registration, per-project limit enforcement), `get_mission` (mission status with task summary counts and individual task list via projects join for ownership); all tools registered in `tools/index.ts` SAM_TOOLS array and toolHandlers map; SAM_SYSTEM_PROMPT updated with Observation and Action tool categories; `dispatch_task` accepts optional `missionId` for mission-task association; configurable via SAM_DISPATCH_MAX_DESCRIPTION_LENGTH (default: 32000)

apps/api/src/index.ts

@@ -40,6 +40,7 @@ import { agentProfileRoutes } from './routes/agent-profiles';
 import { agentSettingsRoutes } from './routes/agent-settings';
 import { agentsCatalogRoutes } from './routes/agents-catalog';
 import { aiProxyRoutes } from './routes/ai-proxy';
+import { aiProxyAnthropicRoutes } from './routes/ai-proxy-anthropic';
 import { analyticsIngestRoutes } from './routes/analytics-ingest';
 import { authRoutes } from './routes/auth';
 import { bootstrapRoutes } from './routes/bootstrap';
@@ -320,7 +321,7 @@ app.use('*', cors({
     return null;
   },
   credentials: true,
-  allowHeaders: ['Content-Type', 'Authorization'],
+  allowHeaders: ['Content-Type', 'Authorization', 'x-api-key', 'anthropic-version', 'anthropic-beta'],
   allowMethods: ['GET', 'POST', 'PUT', 'PATCH', 'DELETE', 'OPTIONS'],
 }));
 
@@ -429,6 +430,7 @@ app.route('/api', trialRoutes);
 app.route('/api/trial', trialOnboardingRoutes);
 app.route('/api/gcp', gcpRoutes);
 app.route('/ai/v1', aiProxyRoutes);
+app.route('/ai/anthropic/v1', aiProxyAnthropicRoutes);
 app.route('/auth/google', googleAuthRoutes);
 // MCP endpoint CORS override — MCP uses Bearer token auth (not cookies/sessions),
 // so it needs credentials: false + origin: '*' to allow VM agent requests from any origin.