@desplega.ai/agent-swarm 1.149.1 → 1.151.0

package/dist/{cli-1ebpar1q.js → cli-4wppfkcm.js}

@@ -7,18 +7,16 @@ import {
   registerSessionResourceCleanup,
   uuidv7
 } from "./cli-bsakbbkm.js";
+import {
+  init_zod
+} from "./cli-sd0mxa3c.js";
+import {
+  exports_external
+} from "./cli-w5e1d1vv.js";
 import {
   bedrockConverseStreamApi,
   setBedrockProviderModule
 } from "./cli-c4eeb5q2.js";
-import {
-  getProviderEnvValue
-} from "./cli-btbzvq95.js";
-import {
-  parseJsonWithRepair,
-  parseStreamingJson,
-  repairJson
-} from "./cli-1m2e086m.js";
 import {
   InMemoryCredentialStore,
   InMemoryModelsStore,
@@ -35,6 +33,14 @@ import {
   modelsAreEqual,
   resolveProviderAuth
 } from "./cli-rgvp6r2x.js";
+import {
+  getProviderEnvValue
+} from "./cli-btbzvq95.js";
+import {
+  parseJsonWithRepair,
+  parseStreamingJson,
+  repairJson
+} from "./cli-1m2e086m.js";
 import {
   AssistantMessageEventStream,
   EventStream,
@@ -49,6 +55,330 @@ import {
   __require
 } from "./cli-bc05qbss.js";
 
+// src/be/memory/raters/llm.ts
+init_zod();
+
+// src/be/memory/raters/llm-client.ts
+var DEFAULT_TIMEOUT_MS = 30000;
+var PROMPT_TEMPLATE = `You are scoring the usefulness of one retrieved memory.
+
+Return ONLY a JSON object with these fields (no prose, no markdown):
+{
+  "score": number,        // 0 = misleading/unhelpful, 1 = highly useful
+  "reasoning": string     // 1..500 chars, why
+}
+
+QUERY:
+\${query}
+
+MEMORY:
+id: \${memoryId}
+name: \${memoryName}
+content: \${memoryContent}
+
+AGENT RESPONSE / SUMMARY:
+\${response}
+
+Score 0..1.`;
+var PLACEHOLDER_PREFIX = "$";
+var QUERY_PLACEHOLDER = `${PLACEHOLDER_PREFIX}{query}`;
+var MEMORY_ID_PLACEHOLDER = `${PLACEHOLDER_PREFIX}{memoryId}`;
+var MEMORY_NAME_PLACEHOLDER = `${PLACEHOLDER_PREFIX}{memoryName}`;
+var MEMORY_CONTENT_PLACEHOLDER = `${PLACEHOLDER_PREFIX}{memoryContent}`;
+var RESPONSE_PLACEHOLDER = `${PLACEHOLDER_PREFIX}{response}`;
+function buildPrompt(input) {
+  return PROMPT_TEMPLATE.replace(QUERY_PLACEHOLDER, input.query).replace(MEMORY_ID_PLACEHOLDER, input.memory.id).replace(MEMORY_NAME_PLACEHOLDER, input.memory.name).replace(MEMORY_CONTENT_PLACEHOLDER, input.memory.content).replace(RESPONSE_PLACEHOLDER, input.response);
+}
+function parseScoreAndReasoning(raw) {
+  if (typeof raw !== "string")
+    return null;
+  let parsed;
+  try {
+    parsed = JSON.parse(raw.trim());
+  } catch {
+    return null;
+  }
+  if (!parsed || typeof parsed !== "object")
+    return null;
+  const obj = parsed;
+  const score = typeof obj.score === "number" ? obj.score : null;
+  const reasoning = typeof obj.reasoning === "string" ? obj.reasoning : null;
+  if (score == null || reasoning == null)
+    return null;
+  if (!Number.isFinite(score) || score < 0 || score > 1)
+    return null;
+  if (reasoning.length === 0 || reasoning.length > 500)
+    return null;
+  return { score, reasoning };
+}
+
+class ClaudeCliLlmRaterClient {
+  model;
+  timeoutMs;
+  constructor(opts = {}) {
+    this.model = opts.model ?? process.env.MEMORY_LLM_RATER_MODEL ?? "haiku";
+    this.timeoutMs = opts.timeoutMs ?? DEFAULT_TIMEOUT_MS;
+  }
+  async rate(input) {
+    const prompt = buildPrompt(input);
+    const tmpFile = `/tmp/llm-rater-${Date.now()}-${Math.random().toString(36).slice(2)}.txt`;
+    let stdout = "";
+    try {
+      await Bun.write(tmpFile, prompt);
+      const proc = Bun.spawn(["bash", "-c", `cat "${tmpFile}" | claude -p --model ${this.model} --output-format json`], {
+        stdout: "pipe",
+        stderr: "pipe",
+        env: { ...process.env, SKIP_SESSION_SUMMARY: "1" }
+      });
+      const timeoutId = setTimeout(() => proc.kill(), this.timeoutMs);
+      stdout = await new Response(proc.stdout).text();
+      clearTimeout(timeoutId);
+    } catch (err) {
+      console.error("[memory-rater:llm] claude -p shell-out failed:", err.message);
+      return null;
+    } finally {
+      try {
+        await Bun.$`rm -f ${tmpFile}`.quiet();
+      } catch {}
+    }
+    let envelope;
+    try {
+      envelope = JSON.parse(stdout);
+    } catch {
+      return null;
+    }
+    return parseScoreAndReasoning(envelope.result);
+  }
+}
+
+// src/be/memory/raters/types.ts
+var REFERENCES_SOURCE_MAX_LENGTH = 512;
+var NUL_CHAR_CODE = 0;
+var DEL_CHAR_CODE = 127;
+var FIRST_PRINTABLE_ASCII = 32;
+function sanitizeReferencesSource(input) {
+  let stripped = "";
+  for (let i = 0;i < input.length; i++) {
+    const code = input.charCodeAt(i);
+    if (code === NUL_CHAR_CODE)
+      return null;
+    if (code < FIRST_PRINTABLE_ASCII || code === DEL_CHAR_CODE) {
+      continue;
+    }
+    stripped += input[i];
+  }
+  if (stripped.length === 0)
+    return null;
+  return stripped;
+}
+
+// src/be/memory/raters/llm.ts
+var LLM_RATER_WEIGHT = 0.8;
+var RatingSchema = exports_external.object({
+  id: exports_external.string().min(1),
+  score: exports_external.number().min(0).max(1),
+  reasoning: exports_external.string().min(1).max(500),
+  referencesSource: exports_external.string().min(1).max(REFERENCES_SOURCE_MAX_LENGTH).optional()
+});
+var SummaryWithRatingsSchema = exports_external.object({
+  summary: exports_external.string(),
+  ratings: exports_external.array(RatingSchema).default([])
+});
+var BASE_SUMMARIZE_PROMPT = `You are summarizing an AI agent's work session. Extract ONLY high-value learnings.
+
+DO NOT include:
+- Generic descriptions of what was done ("worked on task X")
+- Tool calls or file reads
+- Routine progress updates
+
+DO include (if present):
+- **Mistakes made and corrections** — what went wrong and what fixed it
+- **Discovered patterns** — reusable approaches, APIs, or codebase conventions
+- **Codebase knowledge** — important file paths, architecture decisions, gotchas
+- **Environment knowledge** — service URLs, config details, tool quirks
+- **Failed approaches** — what was tried and didn't work (and why)
+
+Format as a bulleted list of concrete, reusable facts. If the session was routine with no significant learnings, respond with exactly: "No significant learnings."`;
+
+class LlmRater {
+  client;
+  name = "llm";
+  constructor(client = new ClaudeCliLlmRaterClient) {
+    this.client = client;
+  }
+  async rate(ctx) {
+    const enriched = ctx;
+    if (enriched.retrievedMemoryIds.length === 0)
+      return [];
+    const memories = enriched.retrievedMemories;
+    if (!memories || memories.length === 0)
+      return [];
+    const events = [];
+    for (const memoryId of enriched.retrievedMemoryIds) {
+      const memory = memories.find((m) => m.id === memoryId);
+      if (!memory)
+        continue;
+      let result;
+      try {
+        result = await this.client.rate({
+          query: enriched.query ?? "",
+          memory,
+          response: enriched.response ?? enriched.evidence ?? ""
+        });
+      } catch (err) {
+        console.error(`[memory-rater:llm] client.rate threw for memoryId=${memoryId}:`, err.message);
+        continue;
+      }
+      if (!result)
+        continue;
+      events.push({
+        memoryId,
+        signal: 2 * result.score - 1,
+        weight: LLM_RATER_WEIGHT,
+        source: "",
+        reasoning: result.reasoning
+      });
+    }
+    return events;
+  }
+}
+function buildRatingsFromLlm(ratings, retrievals) {
+  const allowed = new Set(retrievals.map((r) => r.id));
+  const events = [];
+  for (const r of ratings) {
+    if (!allowed.has(r.id))
+      continue;
+    let cleanedReferencesSource;
+    if (r.referencesSource !== undefined) {
+      const cleaned = sanitizeReferencesSource(r.referencesSource);
+      if (cleaned !== null) {
+        cleanedReferencesSource = cleaned;
+      }
+    }
+    events.push({
+      memoryId: r.id,
+      signal: 2 * r.score - 1,
+      weight: LLM_RATER_WEIGHT,
+      source: "llm",
+      reasoning: r.reasoning,
+      ...cleanedReferencesSource !== undefined ? { referencesSource: cleanedReferencesSource } : {}
+    });
+  }
+  return events;
+}
+var RETRIEVAL_PROMPT_CONTENT_CAP = 600;
+function buildSummaryWithRatingsPrompt(basePrompt, retrievals) {
+  if (retrievals.length === 0)
+    return basePrompt;
+  const memoryBlock = retrievals.map((m, i) => {
+    const content = m.content.length > RETRIEVAL_PROMPT_CONTENT_CAP ? `${m.content.slice(0, RETRIEVAL_PROMPT_CONTENT_CAP)}…` : m.content;
+    return `Memory #${i + 1}
+  id: ${m.id}
+  name: ${m.name}
+  content: ${content}`;
+  }).join(`
+
+`);
+  return `${basePrompt}
+
+CRITICAL: Your entire response MUST be a single JSON object that conforms to the schema below. Do NOT wrap it in triple-backtick fences (no \`\`\`json or \`\`\`), do NOT add a prose preamble, do NOT add trailing commentary. Just the JSON object, nothing else.
+
+Schema:
+{
+  "summary": string,                        // your existing summary text
+  "ratings": [                              // one entry per memory you can score
+    {
+      "id": string,                         // memory id, copied from the list below
+      "score": number,                      // 0 = misleading/unhelpful, 1 = highly useful
+      "reasoning": string,                  // 1..500 chars, why
+      "referencesSource": string            // OPTIONAL — see note below
+    }
+  ]
+}
+
+Score ONLY memories present in the list below. Use the exact ids. Omit any you cannot evaluate.
+
+Optionally for each rating, if the memory clearly references a specific external source (a GitHub PR/issue, a Linear issue, a customer, a Slack thread, an AgentMail thread, etc.), include a \`referencesSource\` string using the convention "<source>:<identifier>" (e.g. "github:owner/repo#N", "linear:KEY-N", "customer:<slug>"). Any prefix is fine — pick what matches the source. Omit the field if no clear external source.
+
+Memories retrieved during this session:
+
+${memoryBlock}`;
+}
+function isLlmRaterEnabled() {
+  const raw = process.env.MEMORY_RATERS;
+  if (!raw || raw.trim() === "")
+    return false;
+  return raw.split(",").map((s) => s.trim()).includes("llm");
+}
+function dedupeRetrievalsForRater(rows) {
+  const seenSchedules = new Set;
+  const out = [];
+  for (const row of rows) {
+    const scheduleId = row.scheduleId;
+    if (typeof scheduleId === "string" && scheduleId.length > 0) {
+      if (seenSchedules.has(scheduleId))
+        continue;
+      seenSchedules.add(scheduleId);
+    }
+    out.push(row);
+  }
+  return out;
+}
+async function fetchRetrievalsForTask(opts) {
+  const fetchFn = opts.fetchImpl ?? fetch;
+  try {
+    const url = `${opts.apiUrl}/api/memory/retrievals?taskId=${encodeURIComponent(opts.taskId)}`;
+    const res = await fetchFn(url, {
+      headers: {
+        "X-Agent-ID": opts.agentId,
+        ...opts.apiKey ? { Authorization: `Bearer ${opts.apiKey}` } : {}
+      }
+    });
+    if (!res.ok) {
+      console.error(`[memory-rater:llm] GET /api/memory/retrievals failed: ${res.status} ${res.statusText}`);
+      return [];
+    }
+    const body = await res.json();
+    return body.results ?? [];
+  } catch (err) {
+    console.error("[memory-rater:llm] fetchRetrievalsForTask threw:", err.message);
+    return [];
+  }
+}
+async function postRatings(opts) {
+  if (opts.events.length === 0)
+    return { ok: true, status: 0 };
+  const fetchFn = opts.fetchImpl ?? fetch;
+  const events = opts.events.map((e) => ({
+    memoryId: e.memoryId,
+    signal: e.signal,
+    weight: e.weight,
+    source: e.source,
+    ...e.reasoning !== undefined ? { reasoning: e.reasoning } : {},
+    ...e.referencesSource !== undefined ? { referencesSource: e.referencesSource } : {},
+    ...opts.taskId ? { taskId: opts.taskId } : {}
+  }));
+  try {
+    const res = await fetchFn(`${opts.apiUrl}/api/memory/rate`, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        "X-Agent-ID": opts.agentId,
+        ...opts.apiKey ? { Authorization: `Bearer ${opts.apiKey}` } : {}
+      },
+      body: JSON.stringify({ events })
+    });
+    if (!res.ok) {
+      const text = await res.text().catch(() => "");
+      console.error(`[memory-rater:llm] POST /api/memory/rate failed: ${res.status} ${res.statusText} ${text.slice(0, 200)}`);
+    }
+    return { ok: res.ok, status: res.status };
+  } catch (err) {
+    console.error("[memory-rater:llm] postRatings threw:", err.message);
+    return { ok: false, status: 0 };
+  }
+}
+
 // node_modules/typebox/build/typebox.mjs
 var exports_typebox = {};
 __export(exports_typebox, {
@@ -5592,7 +5922,7 @@ function amazonBedrockProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/openai-completions.lazy.js
-var openAICompletionsApi = () => lazyApi(() => import("./openai-completions-0bqmsajw.js"));
+var openAICompletionsApi = () => lazyApi(() => import("./openai-completions-32rdwmze.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/auth/helpers.js
 function envApiKeyAuth(name, envVars) {
@@ -5648,7 +5978,7 @@ function antLingProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/anthropic-messages.lazy.js
-var anthropicMessagesApi = () => lazyApi(() => import("./anthropic-messages-vs2rabzp.js"));
+var anthropicMessagesApi = () => lazyApi(() => import("./anthropic-messages-2y8nktar.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/auth/oauth/load.js
 var __rewriteRelativeImportExtension = function(path, preserveJsx) {
@@ -5886,7 +6216,7 @@ function anthropicProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/azure-openai-responses.lazy.js
-var azureOpenAIResponsesApi = () => lazyApi(() => import("./azure-openai-responses-cqar64rc.js"));
+var azureOpenAIResponsesApi = () => lazyApi(() => import("./azure-openai-responses-f38x2avq.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/azure-openai-responses.js
 function azureOpenAIResponsesProvider() {
@@ -5924,7 +6254,7 @@ function cerebrasProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/openai-responses.lazy.js
-var openAIResponsesApi = () => lazyApi(() => import("./openai-responses-nmeyptke.js"));
+var openAIResponsesApi = () => lazyApi(() => import("./openai-responses-6xwf7bzg.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/cloudflare-auth.js
 var CLOUDFLARE_API_KEY = "CLOUDFLARE_API_KEY";
@@ -6112,7 +6442,7 @@ function githubCopilotProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/google-generative-ai.lazy.js
-var googleGenerativeAIApi = () => lazyApi(() => import("./google-generative-ai-tsskx7te.js"));
+var googleGenerativeAIApi = () => lazyApi(() => import("./google-generative-ai-ndjmj5fa.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/google.js
 function googleProvider() {
@@ -6127,7 +6457,7 @@ function googleProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/google-vertex.lazy.js
-var googleVertexApi = () => lazyApi(() => import("./google-vertex-ppw8ydpa.js"));
+var googleVertexApi = () => lazyApi(() => import("./google-vertex-fr665tqs.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/google-vertex.js
 var VERTEX_ADC_PATH = "~/.config/gcloud/application_default_credentials.json";
@@ -6281,7 +6611,7 @@ function minimaxCnProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/mistral-conversations.lazy.js
-var mistralConversationsApi = () => lazyApi(() => import("./mistral-conversations-4ryk7h83.js"));
+var mistralConversationsApi = () => lazyApi(() => import("./mistral-conversations-qhr84ph3.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/mistral.js
 function mistralProvider() {
@@ -6344,7 +6674,7 @@ function openaiProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/openai-codex-responses.lazy.js
-var openAICodexResponsesApi = () => lazyApi(() => import("./openai-codex-responses-g7w48y30.js"));
+var openAICodexResponsesApi = () => lazyApi(() => import("./openai-codex-responses-wy6a43wd.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/openai-codex.js
 function openaiCodexProvider() {
@@ -6419,7 +6749,7 @@ function openrouterProvider() {
 
 // node_modules/@earendil-works/pi-ai/dist/api/openrouter-images.lazy.js
 var openrouterImagesApi = () => ({
-  generateImages: async (model, context, options) => (await import("./openrouter-images-qx18fyp8.js")).generateImages(model, context, options)
+  generateImages: async (model, context, options) => (await import("./openrouter-images-9f9ek4rw.js")).generateImages(model, context, options)
 });
 
 // node_modules/@earendil-works/pi-ai/dist/image-models.generated.js
@@ -7263,7 +7593,7 @@ function qwenTokenPlanIndividualProvider() {
 }
 
 // node_modules/@earendil-works/pi-ai/dist/api/pi-messages.lazy.js
-var piMessagesApi = () => lazyApi(() => import("./pi-messages-79jvzz1d.js"));
+var piMessagesApi = () => lazyApi(() => import("./pi-messages-b1mbr6gp.js"));
 
 // node_modules/@earendil-works/pi-ai/dist/providers/radius-config.js
 var DEFAULT_RADIUS_GATEWAY = "https://radius.pi.dev";
@@ -13538,7 +13868,7 @@ function createLazyLoadErrorImages(model, error2) {
   };
 }
 function loadOpenRouterImagesProviderModule() {
-  openRouterImagesProviderModulePromise ||= import("./openrouter-images-qx18fyp8.js").then((module3) => module3);
+  openRouterImagesProviderModulePromise ||= import("./openrouter-images-9f9ek4rw.js").then((module3) => module3);
   return openRouterImagesProviderModulePromise;
 }
 var generateImagesOpenRouter = async (model, context2, options) => {
@@ -16051,4 +16381,4 @@ async function resolveCredential(opts = {}) {
   return null;
 }
 
-export { exports_typebox, exports_build, AssistantMessageFrameEncoder, reduceAssistantMessageFrames, isContextOverflow, isRecoverableLength, retryAssistantCall, isRetryableAssistantError, contentText, exports_value2 as exports_value, Compile, exports_compile, validateToolArguments, DEFAULT_RADIUS_GATEWAY, radiusProvider, getBuiltinModel, getBuiltinModelDataGeneratedAt, getBuiltinModels, builtinProviders, exports_all, getApiProvider, resetApiProviders, stream, complete, streamSimple, completeSimple, exports_compat, DEFAULT_MODEL, parseModelStr, resolveCredential };
+export { REFERENCES_SOURCE_MAX_LENGTH, sanitizeReferencesSource, SummaryWithRatingsSchema, BASE_SUMMARIZE_PROMPT, LlmRater, buildRatingsFromLlm, buildSummaryWithRatingsPrompt, isLlmRaterEnabled, dedupeRetrievalsForRater, fetchRetrievalsForTask, postRatings, exports_typebox, exports_build, AssistantMessageFrameEncoder, reduceAssistantMessageFrames, isContextOverflow, isRecoverableLength, retryAssistantCall, isRetryableAssistantError, contentText, exports_value2 as exports_value, Compile, exports_compile, validateToolArguments, DEFAULT_RADIUS_GATEWAY, radiusProvider, getBuiltinModel, getBuiltinModelDataGeneratedAt, getBuiltinModels, builtinProviders, exports_all, getApiProvider, resetApiProviders, stream, complete, streamSimple, completeSimple, exports_compat, DEFAULT_MODEL, parseModelStr, resolveCredential };