@mcp-abap-adt/llm-agent-server-libs 19.2.0 → 20.4.0

package/dist/smart-agent/workers/worker-registry.js

@@ -0,0 +1,203 @@
+/**
+ * WorkerRegistry — owns the per-worker LLM/embedder/MCP cache and the
+ * per-session worker-registry build loop (relocated from SmartServer R6).
+ *
+ * The cache map, the three free functions (`WorkerLlmSet`, `resolveWorkerLlmSet`,
+ * `drainWorkerCache`, `backfillWorkerCacheFromHandle`), and the `buildWorkerRegistry`
+ * loop live here. `buildSubAgent` (200-line, deeply coupled to SmartServer) stays
+ * in SmartServer and is called through the injected `buildSubAgent` callback.
+ */
+import { SmartAgentSubAgent, } from '@mcp-abap-adt/llm-agent-libs';
+// ---------------------------------------------------------------------------
+// Free functions — relocated verbatim from smart-server.ts.
+// ---------------------------------------------------------------------------
+/**
+ * Drain every cached worker's `close` (if any), then clear the cache map.
+ * Used by config-reload (PUT /v1/config + hot-reload — Fix #14/18/21) and by
+ * server `close()` to release per-worker MCP connections that were attached
+ * to the discarded `SmartAgentHandle`s.
+ *
+ * IMPORTANT — in-flight caveat: this aborts any request that is mid-call on
+ * a worker's MCP client. That is acceptable for an admin action (config
+ * reload, server shutdown) where the alternative is leaking connections.
+ * Server `close()` calls this AFTER `lifecycle.disposeAll()` so per-session
+ * graphs that reference worker clients are torn down first.
+ *
+ * Uses `Promise.allSettled` so one failing close cannot block the others.
+ */
+export async function drainWorkerCache(cache) {
+    const closers = [];
+    for (const entry of cache.values()) {
+        if (entry.close) {
+            try {
+                closers.push(entry.close());
+            }
+            catch {
+                // sync throw (defensive — close is async by contract)
+            }
+        }
+    }
+    cache.clear();
+    if (closers.length > 0) {
+        await Promise.allSettled(closers);
+    }
+}
+/**
+ * Build-once-per-worker resolver. The first time a worker name is seen, it
+ * constructs the worker's main/classifier/(optional helper) LLM + embedder and
+ * caches the set; every later call (e.g. each per-session worker re-wire)
+ * returns the SAME set by reference — never reconstructing LLM clients
+ * (locked invariant: LLM/embedder clients are global, built once).
+ *
+ * Accepts optional `makeToolsRag`/`makeHistoryRag`/`makeMcpClients` factories;
+ * when provided, the resolver builds them ONCE on the first miss and caches
+ * them on the returned set. Subsequent calls return the cached resources by
+ * reference — never re-vectorizing or re-connecting MCP.
+ */
+export async function resolveWorkerLlmSet(input) {
+    const hit = input.cache.get(input.name);
+    if (hit)
+        return hit;
+    const mainLlm = await input.makeMain();
+    const classifierLlm = await input.makeClassifier();
+    const helperLlm = input.makeHelper ? await input.makeHelper() : undefined;
+    const embedder = input.makeEmbedder ? await input.makeEmbedder() : undefined;
+    const toolsRag = input.makeToolsRag ? await input.makeToolsRag() : undefined;
+    const historyRag = input.makeHistoryRag
+        ? await input.makeHistoryRag()
+        : undefined;
+    const mcpClients = input.makeMcpClients
+        ? await input.makeMcpClients()
+        : undefined;
+    const set = {
+        mainLlm,
+        classifierLlm,
+        helperLlm,
+        embedder,
+        toolsRag,
+        historyRag,
+        mcpClients,
+    };
+    input.cache.set(input.name, set);
+    return set;
+}
+/**
+ * Backfill the per-worker cache entry from the BUILT handle (review HIGH #7).
+ *
+ * The primary `buildSubAgent` populates `cached.mcpClients`/`toolsRag`/
+ * `historyRag` only when the worker config provided DI factories. Workers
+ * configured with `subCfg.mcp: ...` (regular config that triggers the
+ * builder's own auto-connect) or with `subCfg.rag: ...` whose RAG is owned
+ * by the builder leave those slots empty — so per-session re-wires would
+ * fall back to the PARENT's MCP/RAG, losing the worker's own connection.
+ *
+ * After the builder finishes, this helper captures what the handle actually
+ * holds and stores it BY REFERENCE on the cache entry. Subsequent per-session
+ * re-wires read the same slots and find the worker's own resources.
+ *
+ * Pure helper, mutates `entry` in place. No-op when the corresponding slot
+ * is already populated (DI path wins) or when the handle has no resource for
+ * that slot (worker simply didn't declare one).
+ */
+export async function backfillWorkerCacheFromHandle(entry, handle) {
+    if ((!entry.mcpClients || entry.mcpClients.length === 0) &&
+        handle.mcpClients &&
+        handle.mcpClients.length > 0) {
+        entry.mcpClients = handle.mcpClients;
+    }
+    if (!entry.toolsRag) {
+        const t = handle.ragRegistry.get('tools');
+        if (t)
+            entry.toolsRag = t;
+    }
+    if (!entry.historyRag) {
+        const h = handle.ragRegistry.get('history');
+        if (h)
+            entry.historyRag = h;
+    }
+    // Capture the per-worker shutdown function (Fix #21). If the entry already
+    // had a close from a previous build (e.g. the same worker name was rebuilt
+    // WITHOUT going through `drainWorkerCache` first — defence in depth), await
+    // the prior close before overwriting so its MCP connections do not leak.
+    if (handle.close) {
+        if (entry.close) {
+            try {
+                await entry.close();
+            }
+            catch {
+                // Best-effort; never block the new build on a stale close failure.
+            }
+        }
+        entry.close = handle.close;
+    }
+}
+export class WorkerRegistry {
+    deps;
+    cache = new Map();
+    constructor(deps) {
+        this.deps = deps;
+    }
+    async drain() {
+        await drainWorkerCache(this.cache);
+    }
+    /**
+     * Build the FRESH per-session worker (sub-agent) registry from the SAME
+     * `subagents:` configs the primary build() used, injecting globals + this
+     * session's logger + the CACHED per-worker LLM/embedder (this.cache).
+     * NEVER reconstructs LLM clients; NEVER reuses the global registry.
+     */
+    async build(parts) {
+        const registry = new Map();
+        const subAgentConfigs = this.deps.subAgentConfigs;
+        if (!subAgentConfigs || subAgentConfigs.length === 0) {
+            return registry;
+        }
+        const fileLogger = this.deps.getFileLogger();
+        if (!fileLogger) {
+            throw new Error('buildWorkerRegistry invoked before primary build() captured globals');
+        }
+        const embedderFactories = this.deps.getEmbedderFactories();
+        for (const sub of subAgentConfigs) {
+            // Lazy build-on-miss (Fix #18). After PUT /v1/config or hot-reload
+            // clears the cache, the next session build used to throw
+            // "worker LLM set not cached" because the cache was assumed
+            // pre-populated by the primary build(). buildSubAgent itself routes
+            // through `resolveWorkerLlmSet` which is build-on-miss, so calling
+            // it without an `injected` arg rebuilds the cache entry. We then
+            // re-read the entry to honour the per-worker slot priority below.
+            if (!this.cache.has(sub.name)) {
+                await this.deps.buildSubAgent(sub.name, sub.config, fileLogger, embedderFactories);
+            }
+            const cached = this.cache.get(sub.name);
+            if (!cached) {
+                // Defence in depth — should be impossible after the lazy build
+                // above unless buildSubAgent's contract changes.
+                throw new Error(`worker LLM set not cached for '${sub.name}'`);
+            }
+            // Per-worker injected slot priority (review HIGH #7):
+            //   worker-cached (from the primary build, includes backfilled
+            //   subCfg.mcp / subCfg.rag results) → parent's session-scoped
+            //   fallback. Encoded HERE so buildSubAgent does not need to know
+            //   the difference; it just consumes injected.mcpClients/toolsRag.
+            const injectedMcpClients = cached.mcpClients && cached.mcpClients.length > 0
+                ? cached.mcpClients
+                : parts.mcpClients;
+            const injectedToolsRag = cached.toolsRag ?? parts.toolsRag;
+            const subAgent = await this.deps.buildSubAgent(sub.name, sub.config, fileLogger, embedderFactories, {
+                ragRegistry: parts.ragRegistry,
+                toolsRag: injectedToolsRag,
+                mcpClients: injectedMcpClients,
+                requestLogger: parts.logger,
+                mainLlm: cached.mainLlm,
+                classifierLlm: cached.classifierLlm,
+                helperLlm: cached.helperLlm,
+                embedder: cached.embedder,
+            });
+            registry.set(sub.name, new SmartAgentSubAgent(sub.name, subAgent, {
+                description: sub.description,
+            }));
+        }
+        return registry;
+    }
+}
+//# sourceMappingURL=worker-registry.js.map

package/dist/smart-agent/workers/worker-registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"worker-registry.js","sourceRoot":"","sources":["../../../src/smart-agent/workers/worker-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAaH,OAAO,EAGL,kBAAkB,GACnB,MAAM,8BAA8B,CAAC;AA0CtC,8EAA8E;AAC9E,4DAA4D;AAC5D,8EAA8E;AAE9E;;;;;;;;;;;;;GAaG;AACH,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,KAAgC;IAEhC,MAAM,OAAO,GAAyB,EAAE,CAAC;IACzC,KAAK,MAAM,KAAK,IAAI,KAAK,CAAC,MAAM,EAAE,EAAE,CAAC;QACnC,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC;YAChB,IAAI,CAAC;gBACH,OAAO,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,EAAE,CAAC,CAAC;YAC9B,CAAC;YAAC,MAAM,CAAC;gBACP,sDAAsD;YACxD,CAAC;QACH,CAAC;IACH,CAAC;IACD,KAAK,CAAC,KAAK,EAAE,CAAC;IACd,IAAI,OAAO,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QACvB,MAAM,OAAO,CAAC,UAAU,CAAC,OAAO,CAAC,CAAC;IACpC,CAAC;AACH,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,KAAK,UAAU,mBAAmB,CAAC,KAUzC;IACC,MAAM,GAAG,GAAG,KAAK,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IACxC,IAAI,GAAG;QAAE,OAAO,GAAG,CAAC;IACpB,MAAM,OAAO,GAAG,MAAM,KAAK,CAAC,QAAQ,EAAE,CAAC;IACvC,MAAM,aAAa,GAAG,MAAM,KAAK,CAAC,cAAc,EAAE,CAAC;IACnD,MAAM,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC,UAAU,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IAC1E,MAAM,QAAQ,GAAG,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IAC7E,MAAM,QAAQ,GAAG,KAAK,CAAC,YAAY,CAAC,CAAC,CAAC,MAAM,KAAK,CAAC,YAAY,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;IAC7E,MAAM,UAAU,GAAG,KAAK,CAAC,cAAc;QACrC,CAAC,CAAC,MAAM,KAAK,CAAC,cAAc,EAAE;QAC9B,CAAC,CAAC,SAAS,CAAC;IACd,MAAM,UAAU,GAAG,KAAK,CAAC,cAAc;QACrC,CAAC,CAAC,MAAM,KAAK,CAAC,cAAc,EAAE;QAC9B,CAAC,CAAC,SAAS,CAAC;IACd,MAAM,GAAG,GAAiB;QACxB,OAAO;QACP,aAAa;QACb,SAAS;QACT,QAAQ;QACR,QAAQ;QACR,UAAU;QACV,UAAU;KACX,CAAC;IACF,KAAK,CAAC,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,EAAE,GAAG,CAAC,CAAC;IACjC,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;;;;;;;;;;;;;;;GAiBG;AACH,MAAM,CAAC,KAAK,UAAU,6BAA6B,CACjD,KAAmB,EACnB,MAIC;IAED,IACE,CAAC,CAAC,KAAK,CAAC,UAAU,IAAI,KAAK,CAAC,UAAU,CAAC,MAAM,KAAK,CAAC,CAAC;QACpD,MAAM,CAAC,UAAU;QACjB,MAAM,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC,EAC5B,CAAC;QACD,KAAK,CAAC,UAAU,GAAG,MAAM,CAAC,UAAU,CAAC;IACvC,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,QAAQ,EAAE,CAAC;QACpB,MAAM,CAAC,GAAG,MAAM,CAAC,WAAW,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;QAC1C,IAAI,CAAC;YAAE,KAAK,CAAC,QAAQ,GAAG,CAAC,CAAC;IAC5B,CAAC;IACD,IAAI,CAAC,KAAK,CAAC,UAAU,EAAE,CAAC;QACtB,MAAM,CAAC,GAAG,MAAM,CAAC,WAAW,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QAC5C,IAAI,CAAC;YAAE,KAAK,CAAC,UAAU,GAAG,CAAC,CAAC;IAC9B,CAAC;IACD,2EAA2E;IAC3E,2EAA2E;IAC3E,4EAA4E;IAC5E,yEAAyE;IACzE,IAAI,MAAM,CAAC,KAAK,EAAE,CAAC;QACjB,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC;YAChB,IAAI,CAAC;gBACH,MAAM,KAAK,CAAC,KAAK,EAAE,CAAC;YACtB,CAAC;YAAC,MAAM,CAAC;gBACP,mEAAmE;YACrE,CAAC;QACH,CAAC;QACD,KAAK,CAAC,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC;IAC7B,CAAC;AACH,CAAC;AAoDD,MAAM,OAAO,cAAc;IAGI;IAFpB,KAAK,GAAG,IAAI,GAAG,EAAwB,CAAC;IAEjD,YAA6B,IAAwB;QAAxB,SAAI,GAAJ,IAAI,CAAoB;IAAG,CAAC;IAEzD,KAAK,CAAC,KAAK;QACT,MAAM,gBAAgB,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC;IACrC,CAAC;IAED;;;;;OAKG;IACH,KAAK,CAAC,KAAK,CAAC,KAAwB;QAClC,MAAM,QAAQ,GAAqB,IAAI,GAAG,EAAE,CAAC;QAC7C,MAAM,eAAe,GAAG,IAAI,CAAC,IAAI,CAAC,eAAe,CAAC;QAClD,IAAI,CAAC,eAAe,IAAI,eAAe,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YACrD,OAAO,QAAQ,CAAC;QAClB,CAAC;QACD,MAAM,UAAU,GAAG,IAAI,CAAC,IAAI,CAAC,aAAa,EAAE,CAAC;QAC7C,IAAI,CAAC,UAAU,EAAE,CAAC;YAChB,MAAM,IAAI,KAAK,CACb,qEAAqE,CACtE,CAAC;QACJ,CAAC;QACD,MAAM,iBAAiB,GAAG,IAAI,CAAC,IAAI,CAAC,oBAAoB,EAAE,CAAC;QAC3D,KAAK,MAAM,GAAG,IAAI,eAAe,EAAE,CAAC;YAClC,mEAAmE;YACnE,yDAAyD;YACzD,4DAA4D;YAC5D,oEAAoE;YACpE,mEAAmE;YACnE,iEAAiE;YACjE,kEAAkE;YAClE,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;gBAC9B,MAAM,IAAI,CAAC,IAAI,CAAC,aAAa,CAC3B,GAAG,CAAC,IAAI,EACR,GAAG,CAAC,MAAM,EACV,UAAU,EACV,iBAAiB,CAIlB,CAAC;YACJ,CAAC;YACD,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;YACxC,IAAI,CAAC,MAAM,EAAE,CAAC;gBACZ,+DAA+D;gBAC/D,iDAAiD;gBACjD,MAAM,IAAI,KAAK,CAAC,kCAAkC,GAAG,CAAC,IAAI,GAAG,CAAC,CAAC;YACjE,CAAC;YACD,sDAAsD;YACtD,+DAA+D;YAC/D,+DAA+D;YAC/D,kEAAkE;YAClE,mEAAmE;YACnE,MAAM,kBAAkB,GACtB,MAAM,CAAC,UAAU,IAAI,MAAM,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC;gBAC/C,CAAC,CAAC,MAAM,CAAC,UAAU;gBACnB,CAAC,CAAC,KAAK,CAAC,UAAU,CAAC;YACvB,MAAM,gBAAgB,GAAG,MAAM,CAAC,QAAQ,IAAI,KAAK,CAAC,QAAQ,CAAC;YAC3D,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,aAAa,CAC5C,GAAG,CAAC,IAAI,EACR,GAAG,CAAC,MAAM,EACV,UAAU,EACV,iBAAiB,EACjB;gBACE,WAAW,EAAE,KAAK,CAAC,WAAW;gBAC9B,QAAQ,EAAE,gBAAgB;gBAC1B,UAAU,EAAE,kBAAkB;gBAC9B,aAAa,EAAE,KAAK,CAAC,MAAM;gBAC3B,OAAO,EAAE,MAAM,CAAC,OAAO;gBACvB,aAAa,EAAE,MAAM,CAAC,aAAa;gBACnC,SAAS,EAAE,MAAM,CAAC,SAAS;gBAC3B,QAAQ,EAAE,MAAM,CAAC,QAAQ;aAC1B,CACF,CAAC;YACF,QAAQ,CAAC,GAAG,CACV,GAAG,CAAC,IAAI,EACR,IAAI,kBAAkB,CAAC,GAAG,CAAC,IAAI,EAAE,QAAQ,EAAE;gBACzC,WAAW,EAAE,GAAG,CAAC,WAAW;aAC7B,CAAC,CACH,CAAC;QACJ,CAAC;QACD,OAAO,QAAQ,CAAC;IAClB,CAAC;CACF"}

package/dist/smart-agent/yaml-loader.d.ts

@@ -0,0 +1,7 @@
+export type YamlConfig = Record<string, unknown>;
+export declare const YAML_TEMPLATE = "port: 4004\nhost: 0.0.0.0\n\n# Request routing mode:\n#   hard        \u2014 Fully managed context. Ignores client history/system prompt. Uses RAG + internal MCP tools only.\n#   pass        \u2014 Transparent proxy. Logs everything but modifies nothing.\n#   smart       \u2014 Hybrid. Preserves client history but enriches it with RAG context and MCP tools based on analysis. (default)\nmode: smart\n\nllm:\n  provider: deepseek                  # deepseek | openai | anthropic | sap-ai-sdk | ollama\n  apiKey: ${DEEPSEEK_API_KEY}        # not required for ollama / sap-ai-sdk\n  model: deepseek-chat\n  temperature: 0.7\n  classifierTemperature: 0.1\n\nrag:\n  type: in-memory                     # in-memory | qdrant | hana-vector | pg-vector\n  embedder: ollama                    # Embedder to use: ollama | openai | sap-ai-core | <custom>\n  url: http://localhost:11434\n  model: bge-m3\n  # resourceGroup: default            # SAP AI Core resource group (sap-ai-core embedder)\n  # scenario: orchestration           # SAP AI Core scenario: orchestration (default) | foundation-models\n  # collectionName: llm-agent         # Collection/table name (qdrant | hana-vector | pg-vector)\n  dedupThreshold: 0.92\n  vectorWeight: 0.7                   # Semantic similarity weight (0..1)\n  keywordWeight: 0.3                  # Lexical matching weight (0..1)\n\nmcp:\n  # type: none | http | stdio\n  # To disable MCP, set type to 'none'\n  type: http\n  url: http://localhost:3001/mcp/stream/http\n\nagent:\n  externalToolsValidationMode: permissive  # permissive | strict\n  maxIterations: 10\n  maxToolCalls: 30\n  toolUnavailableTtlMs: 600000       # Temporary tool blacklist TTL (ms)\n  ragQueryK: 10\n  # contextBudgetTokens: 4000          # Max tokens for RAG context in system prompt (0 = no limit)\n  # semanticHistoryEnabled: false      # Enable semantic history via RAG\n  # historyRecencyWindow: 4            # Last N messages from client history in LLM context\n  # historyTurnSummaryPrompt: \"...\"    # LLM prompt for turn summarization\n  showReasoning: false                # Explain strategy at start of response\n  historyAutoSummarizeLimit: 10       # History length to trigger compression\n  queryExpansionEnabled: false        # Expand RAG queries with LLM-generated synonyms\n  toolResultCacheTtlMs: 300000       # Tool result cache TTL (ms); 0 to disable\n  sessionTokenBudget: 0              # Multi-turn token budget; 0 to disable\n  # ragTranslateEnabled: true        # Translate non-ASCII RAG queries to English (default: true)\n  # classificationEnabled: false     # Enable for custom pipelines with multi-store routing\n  # toolReselectPerIteration: false  # Re-select tools via RAG on each tool-loop iteration\n  # llmCallStrategy: streaming       # streaming | non-streaming | fallback\n  # streamMode: full                 # full | final \u2014 streaming behavior for tool loops\n  # heartbeatIntervalMs: 5000       # SSE heartbeat interval during tool execution (ms)\n  # healthTimeoutMs: 5000            # Health check probe timeout (ms); increase for slow providers (SAP AI Core: 15000)\n  # retry:                           # LLM retry config for 429/5xx errors\n  #   maxAttempts: 3\n  #   backoffMs: 1000\n  #   retryOn: [429, 500, 502, 503]\n  #   retryOnMidStream: ['SSE stream']  # Substrings triggering mid-stream retry\n\n# --- Advanced Multi-Model LLM (optional) ------------------------------------\n# Assign different models for different internal roles via a top-level llm: map.\n# (The legacy pipeline.llm override was removed \u2014 use the top-level llm: map now.)\n# llm:\n#   main:\n#     provider: deepseek              # deepseek | openai | anthropic | sap-ai-sdk\n#     apiKey: ${DEEPSEEK_API_KEY}\n#     model: deepseek-chat\n#     temperature: 0.7\n#   classifier:                       # optional; if absent, main config is reused\n#     provider: deepseek\n#     apiKey: ${DEEPSEEK_API_KEY}\n#     model: deepseek-chat\n#     temperature: 0.1\n#   helper:                           # optional; if absent, main config is reused\n#     provider: deepseek\n#     apiKey: ${DEEPSEEK_API_KEY}\n#     model: deepseek-chat\n#     temperature: 0.1\n\n# prompts:\n#   system: \"You are a helpful assistant specialized in SAP ABAP development.\"\n#   classifier: |\n#     You are an intent classifier... (see source for full default prompt)\n#   reasoning: |\n#     IMPORTANT: Always start your response with a brief <reasoning> block...\n#   ragTranslate: |\n#     Translate the user request to English for search purposes...\n#   historySummary: |\n#     Summarize the conversation so far...\n\nlog: smart-server.log                 # path to log file; omit for stdout\n# logDir: sessions                    # Directory for detailed session debug logs\n# pluginDir: ./my-plugins             # Additional plugin directory (loaded after defaults)\n# plugins:                            # Explicit plugin module specifiers (npm packages or paths)\n#   - \"@scope/my-pipeline-plugin\"     #   bare specifier resolved from cwd\n#   - ./local-plugin.mjs              #   relative path resolved against cwd\n\n# subagents:                          # Optional: nested agents callable from pipeline\n#   - name: code-reviewer             # Used as stage config: { agent: code-reviewer }\n#     description: |                  # Optional. Shown to the Coordinator planner LLM\n#       Reviews code and returns      # so it can pick this agent for the right step.\n#       structured JSON.\n#     config: ./agents/code-reviewer.yaml\n\n# pipeline:                           # Optional: select the request pipeline.\n#   name: flat                        # flat (default) | linear | dag | stepper | controller | controller-weak | <plugin>\n#   config:                           # Opaque per-pipeline dialect (validated by the plugin)\n#     mode: planned-react             # e.g. stepper: cyclic-react | planned-react | deep-stepper\n#     knowledgeSeed: []               # stepper: deployment-supplied tool guidance\n#\n# NOTE: the legacy 'coordinator:' block and the old 'pipeline:' shape\n# (mcp/rag/stages/llm overrides) were REMOVED in this major. A config that\n# still uses them fails loud at startup. Top-level 'llm:', 'mcp:', 'rag:' now\n# own those concerns; pipeline behavior moves under 'pipeline.config'.\n";
+export declare const get: (obj: unknown, ...keys: string[]) => unknown;
+export declare function resolveEnvVars(value: unknown, env?: NodeJS.ProcessEnv): unknown;
+export declare function loadYamlConfig(filePath: string, env?: NodeJS.ProcessEnv): YamlConfig;
+export declare function generateConfigTemplate(outputPath: string): void;
+//# sourceMappingURL=yaml-loader.d.ts.map

package/dist/smart-agent/yaml-loader.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"yaml-loader.d.ts","sourceRoot":"","sources":["../../src/smart-agent/yaml-loader.ts"],"names":[],"mappings":"AAGA,MAAM,MAAM,UAAU,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AAEjD,eAAO,MAAM,aAAa,8tMAqHzB,CAAC;AAEF,eAAO,MAAM,GAAG,GAAI,KAAK,OAAO,EAAE,GAAG,MAAM,MAAM,EAAE,KAAG,OAM7C,CAAC;AAEV,wBAAgB,cAAc,CAC5B,KAAK,EAAE,OAAO,EACd,GAAG,GAAE,MAAM,CAAC,UAAwB,GACnC,OAAO,CAeT;AAED,wBAAgB,cAAc,CAC5B,QAAQ,EAAE,MAAM,EAChB,GAAG,GAAE,MAAM,CAAC,UAAwB,GACnC,UAAU,CAGZ;AAED,wBAAgB,sBAAsB,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAE/D"}

package/dist/smart-agent/yaml-loader.js

@@ -0,0 +1,146 @@
+import fs from 'node:fs';
+import { parse as parseYaml } from 'yaml';
+export const YAML_TEMPLATE = `port: 4004
+host: 0.0.0.0
+
+# Request routing mode:
+#   hard        — Fully managed context. Ignores client history/system prompt. Uses RAG + internal MCP tools only.
+#   pass        — Transparent proxy. Logs everything but modifies nothing.
+#   smart       — Hybrid. Preserves client history but enriches it with RAG context and MCP tools based on analysis. (default)
+mode: smart
+
+llm:
+  provider: deepseek                  # deepseek | openai | anthropic | sap-ai-sdk | ollama
+  apiKey: \${DEEPSEEK_API_KEY}        # not required for ollama / sap-ai-sdk
+  model: deepseek-chat
+  temperature: 0.7
+  classifierTemperature: 0.1
+
+rag:
+  type: in-memory                     # in-memory | qdrant | hana-vector | pg-vector
+  embedder: ollama                    # Embedder to use: ollama | openai | sap-ai-core | <custom>
+  url: http://localhost:11434
+  model: bge-m3
+  # resourceGroup: default            # SAP AI Core resource group (sap-ai-core embedder)
+  # scenario: orchestration           # SAP AI Core scenario: orchestration (default) | foundation-models
+  # collectionName: llm-agent         # Collection/table name (qdrant | hana-vector | pg-vector)
+  dedupThreshold: 0.92
+  vectorWeight: 0.7                   # Semantic similarity weight (0..1)
+  keywordWeight: 0.3                  # Lexical matching weight (0..1)
+
+mcp:
+  # type: none | http | stdio
+  # To disable MCP, set type to 'none'
+  type: http
+  url: http://localhost:3001/mcp/stream/http
+
+agent:
+  externalToolsValidationMode: permissive  # permissive | strict
+  maxIterations: 10
+  maxToolCalls: 30
+  toolUnavailableTtlMs: 600000       # Temporary tool blacklist TTL (ms)
+  ragQueryK: 10
+  # contextBudgetTokens: 4000          # Max tokens for RAG context in system prompt (0 = no limit)
+  # semanticHistoryEnabled: false      # Enable semantic history via RAG
+  # historyRecencyWindow: 4            # Last N messages from client history in LLM context
+  # historyTurnSummaryPrompt: "..."    # LLM prompt for turn summarization
+  showReasoning: false                # Explain strategy at start of response
+  historyAutoSummarizeLimit: 10       # History length to trigger compression
+  queryExpansionEnabled: false        # Expand RAG queries with LLM-generated synonyms
+  toolResultCacheTtlMs: 300000       # Tool result cache TTL (ms); 0 to disable
+  sessionTokenBudget: 0              # Multi-turn token budget; 0 to disable
+  # ragTranslateEnabled: true        # Translate non-ASCII RAG queries to English (default: true)
+  # classificationEnabled: false     # Enable for custom pipelines with multi-store routing
+  # toolReselectPerIteration: false  # Re-select tools via RAG on each tool-loop iteration
+  # llmCallStrategy: streaming       # streaming | non-streaming | fallback
+  # streamMode: full                 # full | final — streaming behavior for tool loops
+  # heartbeatIntervalMs: 5000       # SSE heartbeat interval during tool execution (ms)
+  # healthTimeoutMs: 5000            # Health check probe timeout (ms); increase for slow providers (SAP AI Core: 15000)
+  # retry:                           # LLM retry config for 429/5xx errors
+  #   maxAttempts: 3
+  #   backoffMs: 1000
+  #   retryOn: [429, 500, 502, 503]
+  #   retryOnMidStream: ['SSE stream']  # Substrings triggering mid-stream retry
+
+# --- Advanced Multi-Model LLM (optional) ------------------------------------
+# Assign different models for different internal roles via a top-level llm: map.
+# (The legacy pipeline.llm override was removed — use the top-level llm: map now.)
+# llm:
+#   main:
+#     provider: deepseek              # deepseek | openai | anthropic | sap-ai-sdk
+#     apiKey: \${DEEPSEEK_API_KEY}
+#     model: deepseek-chat
+#     temperature: 0.7
+#   classifier:                       # optional; if absent, main config is reused
+#     provider: deepseek
+#     apiKey: \${DEEPSEEK_API_KEY}
+#     model: deepseek-chat
+#     temperature: 0.1
+#   helper:                           # optional; if absent, main config is reused
+#     provider: deepseek
+#     apiKey: \${DEEPSEEK_API_KEY}
+#     model: deepseek-chat
+#     temperature: 0.1
+
+# prompts:
+#   system: "You are a helpful assistant specialized in SAP ABAP development."
+#   classifier: |
+#     You are an intent classifier... (see source for full default prompt)
+#   reasoning: |
+#     IMPORTANT: Always start your response with a brief <reasoning> block...
+#   ragTranslate: |
+#     Translate the user request to English for search purposes...
+#   historySummary: |
+#     Summarize the conversation so far...
+
+log: smart-server.log                 # path to log file; omit for stdout
+# logDir: sessions                    # Directory for detailed session debug logs
+# pluginDir: ./my-plugins             # Additional plugin directory (loaded after defaults)
+# plugins:                            # Explicit plugin module specifiers (npm packages or paths)
+#   - "@scope/my-pipeline-plugin"     #   bare specifier resolved from cwd
+#   - ./local-plugin.mjs              #   relative path resolved against cwd
+
+# subagents:                          # Optional: nested agents callable from pipeline
+#   - name: code-reviewer             # Used as stage config: { agent: code-reviewer }
+#     description: |                  # Optional. Shown to the Coordinator planner LLM
+#       Reviews code and returns      # so it can pick this agent for the right step.
+#       structured JSON.
+#     config: ./agents/code-reviewer.yaml
+
+# pipeline:                           # Optional: select the request pipeline.
+#   name: flat                        # flat (default) | linear | dag | stepper | controller | controller-weak | <plugin>
+#   config:                           # Opaque per-pipeline dialect (validated by the plugin)
+#     mode: planned-react             # e.g. stepper: cyclic-react | planned-react | deep-stepper
+#     knowledgeSeed: []               # stepper: deployment-supplied tool guidance
+#
+# NOTE: the legacy 'coordinator:' block and the old 'pipeline:' shape
+# (mcp/rag/stages/llm overrides) were REMOVED in this major. A config that
+# still uses them fails loud at startup. Top-level 'llm:', 'mcp:', 'rag:' now
+# own those concerns; pipeline behavior moves under 'pipeline.config'.
+`;
+export const get = (obj, ...keys) => keys.reduce((o, k) => {
+    if (o !== null && typeof o === 'object' && k in o) {
+        return o[k];
+    }
+    return undefined;
+}, obj);
+export function resolveEnvVars(value, env = process.env) {
+    if (typeof value === 'string')
+        return value.replace(/\$\{([^}:]+)(?::-(.*?))?\}/g, (_, name, fallback) => env[name] || fallback || '');
+    if (Array.isArray(value))
+        return value.map((v) => resolveEnvVars(v, env));
+    if (value !== null && typeof value === 'object')
+        return Object.fromEntries(Object.entries(value).map(([k, v]) => [
+            k,
+            resolveEnvVars(v, env),
+        ]));
+    return value;
+}
+export function loadYamlConfig(filePath, env = process.env) {
+    const raw = fs.readFileSync(filePath, 'utf8');
+    return resolveEnvVars(parseYaml(raw), env);
+}
+export function generateConfigTemplate(outputPath) {
+    fs.writeFileSync(outputPath, YAML_TEMPLATE, 'utf8');
+}
+//# sourceMappingURL=yaml-loader.js.map

package/dist/smart-agent/yaml-loader.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"yaml-loader.js","sourceRoot":"","sources":["../../src/smart-agent/yaml-loader.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,SAAS,CAAC;AACzB,OAAO,EAAE,KAAK,IAAI,SAAS,EAAE,MAAM,MAAM,CAAC;AAI1C,MAAM,CAAC,MAAM,aAAa,GAAG;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CAqH5B,CAAC;AAEF,MAAM,CAAC,MAAM,GAAG,GAAG,CAAC,GAAY,EAAE,GAAG,IAAc,EAAW,EAAE,CAC9D,IAAI,CAAC,MAAM,CAAU,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE;IAC5B,IAAI,CAAC,KAAK,IAAI,IAAI,OAAO,CAAC,KAAK,QAAQ,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC;QAClD,OAAQ,CAA6B,CAAC,CAAC,CAAC,CAAC;IAC3C,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC,EAAE,GAAG,CAAC,CAAC;AAEV,MAAM,UAAU,cAAc,CAC5B,KAAc,EACd,MAAyB,OAAO,CAAC,GAAG;IAEpC,IAAI,OAAO,KAAK,KAAK,QAAQ;QAC3B,OAAO,KAAK,CAAC,OAAO,CAClB,6BAA6B,EAC7B,CAAC,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,EAAE,CAAC,GAAG,CAAC,IAAI,CAAC,IAAI,QAAQ,IAAI,EAAE,CACnD,CAAC;IACJ,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC;QAAE,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,cAAc,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC,CAAC;IAC1E,IAAI,KAAK,KAAK,IAAI,IAAI,OAAO,KAAK,KAAK,QAAQ;QAC7C,OAAO,MAAM,CAAC,WAAW,CACvB,MAAM,CAAC,OAAO,CAAC,KAAgC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,EAAE,EAAE,CAAC;YAC/D,CAAC;YACD,cAAc,CAAC,CAAC,EAAE,GAAG,CAAC;SACvB,CAAC,CACH,CAAC;IACJ,OAAO,KAAK,CAAC;AACf,CAAC;AAED,MAAM,UAAU,cAAc,CAC5B,QAAgB,EAChB,MAAyB,OAAO,CAAC,GAAG;IAEpC,MAAM,GAAG,GAAG,EAAE,CAAC,YAAY,CAAC,QAAQ,EAAE,MAAM,CAAC,CAAC;IAC9C,OAAO,cAAc,CAAC,SAAS,CAAC,GAAG,CAAC,EAAE,GAAG,CAAe,CAAC;AAC3D,CAAC;AAED,MAAM,UAAU,sBAAsB,CAAC,UAAkB;IACvD,EAAE,CAAC,aAAa,CAAC,UAAU,EAAE,aAAa,EAAE,MAAM,CAAC,CAAC;AACtD,CAAC"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mcp-abap-adt/llm-agent-server-libs",
-  "version": "19.2.0",
+  "version": "20.4.0",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -56,15 +56,15 @@
     "test": "node --import tsx/esm --test --test-reporter=spec 'src/**/*.test.ts'"
   },
   "dependencies": {
-    "@mcp-abap-adt/llm-agent": "^19.2.0",
-    "@mcp-abap-adt/llm-agent-libs": "^19.2.0",
-    "@mcp-abap-adt/llm-agent-mcp": "^19.2.0",
-    "@mcp-abap-adt/llm-agent-rag": "^19.2.0",
+    "@mcp-abap-adt/llm-agent": "^20.4.0",
+    "@mcp-abap-adt/llm-agent-libs": "^20.4.0",
+    "@mcp-abap-adt/llm-agent-mcp": "^20.4.0",
+    "@mcp-abap-adt/llm-agent-rag": "^20.4.0",
     "pg": "^8.21.0",
     "yaml": "^2.9.0"
   },
   "peerDependencies": {
-    "@mcp-abap-adt/sap-aicore-llm": "^19.2.0"
+    "@mcp-abap-adt/sap-aicore-llm": "^20.4.0"
   },
   "peerDependenciesMeta": {
     "@mcp-abap-adt/sap-aicore-llm": {
@@ -72,7 +72,7 @@
     }
   },
   "devDependencies": {
-    "@mcp-abap-adt/sap-aicore-llm": "^19.2.0",
+    "@mcp-abap-adt/sap-aicore-llm": "^20.4.0",
     "@types/pg": "^8.11.0"
   },
   "publishConfig": {