@desplega.ai/agent-swarm 1.91.0 → 1.92.0

package/src/be/seed-scripts/catalog/task-context-gathering.ts

@@ -0,0 +1,92 @@
+import { z } from "zod";
+
+export const argsSchema = z.object({
+  taskId: z.string().describe("Task ID to fetch details for"),
+  queries: z
+    .array(z.string())
+    .min(1)
+    .describe("Search queries from the task description — 2-4 recommended"),
+  scope: z
+    .enum(["all", "agent", "swarm"])
+    .optional()
+    .describe("Memory scope filter (default all)"),
+  memoryLimit: z
+    .number()
+    .int()
+    .positive()
+    .optional()
+    .describe("Max memories per query before dedup (default 8)"),
+});
+
+function slimTask(task: any) {
+  if (!task || typeof task !== "object") return null;
+  return {
+    id: task.id,
+    status: task.status,
+    description: task.task,
+    dependsOn: task.dependsOn,
+    slackChannelId: task.slackChannelId,
+    slackThreadTs: task.slackThreadTs,
+    createdAt: task.createdAt,
+    finishedAt: task.finishedAt,
+    agentId: task.agentId,
+    output: task.output,
+    failureReason: task.failureReason,
+  };
+}
+
+/** Fetch slim task details plus deduped multi-query memories for task onboarding. */
+export default async function taskContextGathering(args: any, ctx: any) {
+  const parsed = argsSchema.safeParse(args);
+  if (!parsed.success) return { error: "invalid args: " + parsed.error.message };
+  const { taskId, queries, scope = "all", memoryLimit = 8 } = parsed.data;
+
+  const taskRes: any = await ctx.swarm.task_get({ taskId });
+  const taskPayload = taskRes?.data ?? taskRes;
+  if (taskPayload?.success === false) {
+    return { error: taskPayload.message ?? "task_get failed", taskId };
+  }
+
+  const allResults: any[] = [];
+  for (const query of queries) {
+    const res: any = await ctx.swarm.memory_search({ query, scope, limit: memoryLimit });
+    const payload = res?.data ?? res;
+    const results = payload?.results ?? [];
+    for (const memory of results) {
+      allResults.push({ ...memory, querySource: query });
+    }
+  }
+
+  const byId = new Map<string, any>();
+  const hitCounts = new Map<string, number>();
+  for (const memory of allResults) {
+    const id = typeof memory.id === "string" ? memory.id : JSON.stringify(memory);
+    hitCounts.set(id, (hitCounts.get(id) ?? 0) + 1);
+    const existing = byId.get(id);
+    const similarity = typeof memory.similarity === "number" ? memory.similarity : 0;
+    const existingSimilarity =
+      existing && typeof existing.similarity === "number" ? existing.similarity : -Infinity;
+    if (!existing || similarity > existingSimilarity) byId.set(id, memory);
+  }
+
+  const memories = Array.from(byId.entries()).map(([id, memory]) => {
+    const hits = hitCounts.get(id) ?? 1;
+    const similarity = typeof memory.similarity === "number" ? memory.similarity : 0;
+    return {
+      ...memory,
+      hits,
+      compositeScore: similarity + 0.05 * hits,
+    };
+  });
+  memories.sort((a: any, b: any) => b.compositeScore - a.compositeScore);
+
+  return {
+    task: slimTask(taskPayload?.task),
+    requestedBy: taskPayload?.requestedBy,
+    attachments: taskPayload?.attachments ?? [],
+    queriesRun: queries.length,
+    totalCandidates: allResults.length,
+    uniqueMemories: memories.length,
+    memories: memories.slice(0, memoryLimit * 2),
+  };
+}

package/src/be/seed-scripts/catalog/tool-usage.ts

@@ -25,7 +25,7 @@ export default async function toolUsage(args: any, ctx: any) {
   const agentId = parsed.data.agentId;
 
   const agentFilter = agentId ? `AND agent_id = '${agentId}'` : "";
-  const query = `
+  const sql = `
     SELECT tool_name, count(*) as calls
     FROM session_logs
     WHERE tool_name IS NOT NULL
@@ -36,9 +36,12 @@ export default async function toolUsage(args: any, ctx: any) {
     LIMIT ${limit}
   `;
 
-  const res: any = await ctx.swarm.db_query({ query });
+  const res: any = await ctx.swarm.db_query({ sql });
   const payload = res?.data ?? res;
-  const rows: any[] = payload?.rows ?? [];
+  const columns: string[] = payload?.columns ?? [];
+  const rows: any[] = (payload?.rows ?? []).map((row: any[]) =>
+    Object.fromEntries(columns.map((column, index) => [column, row[index]])),
+  );
 
   const totalCalls = rows.reduce((sum: number, r: any) => sum + (r.calls ?? 0), 0);
 

package/src/be/seed-scripts/index.ts

@@ -30,9 +30,11 @@ import groupCountSrc from "./catalog/group-count.ts" with { type: "text" };
 import jsonQuerySrc from "./catalog/json-query.ts" with { type: "text" };
 import linearIssueSrc from "./catalog/linear-issue.ts" with { type: "text" };
 import memoryDedupCheckSrc from "./catalog/memory-dedup-check.ts" with { type: "text" };
+import opsCatalogAuditSrc from "./catalog/ops-catalog-audit.ts" with { type: "text" };
 import scheduleHealthSrc from "./catalog/schedule-health.ts" with { type: "text" };
 import slackThreadFlattenSrc from "./catalog/slack-thread-flatten.ts" with { type: "text" };
 import smartRecallSrc from "./catalog/smart-recall.ts" with { type: "text" };
+import taskContextGatheringSrc from "./catalog/task-context-gathering.ts" with { type: "text" };
 import taskFailureAuditSrc from "./catalog/task-failure-audit.ts" with { type: "text" };
 import textDiffSrc from "./catalog/text-diff.ts" with { type: "text" };
 import toolUsageSrc from "./catalog/tool-usage.ts" with { type: "text" };
@@ -134,6 +136,14 @@ export const SEED_SCRIPTS: SeedScript[] = [
       "Recall relevant memories using multiple search angles — better coverage than a single query. Use for task onboarding, context gathering, or before writing new memories.",
     source: asText(smartRecallSrc),
   },
+  {
+    name: "task-context-gathering",
+    description:
+      "Get task details and recall relevant memories in one call — returns a slimmed task projection plus deduped and reranked memories from multi-query fan-out.",
+    intent:
+      "Task onboarding: one call instead of task_get plus multiple memory_search calls. Pass the task description split into 2-4 natural-language queries.",
+    source: asText(taskContextGatheringSrc),
+  },
   {
     name: "schedule-health",
     description:
@@ -153,11 +163,19 @@ export const SEED_SCRIPTS: SeedScript[] = [
   {
     name: "compound-insights",
     description:
-      "All-in-one swarm-wide daily ops snapshot: task completion/failure summary, real failure clusters (excludes superseded/cancelled bookkeeping), schedule health flags, tool usage top-25, memory health stats, and a per-agent breakdown. Aggregates across ALL agents via direct read-only SQL.",
+      "All-in-one swarm-wide daily ops snapshot: task completion/failure summary, real failure clusters (excludes superseded/cancelled bookkeeping), schedule health flags, tool usage top-25, memory health/pollution stats, seed-script candidate tool triplets, and a per-agent breakdown. Aggregates across ALL agents via direct read-only SQL.",
     intent:
-      "Single-call daily compounding Phase 0 helper — replaces ~25 raw tool roundtrips with one compressed JSON result covering every agent. For daily evolution, ops reviews, or heartbeat context.",
+      "Single-call daily compounding Phase 0 helper — replaces ~25 raw tool roundtrips with one compressed JSON result covering every agent. For daily evolution, self-scripting candidates, ops reviews, or heartbeat context.",
     source: asText(compoundInsightsSrc),
   },
+  {
+    name: "ops-catalog-audit",
+    description:
+      "Audit-as-code catalog check for schedules, workflows, and prompt/template drift. Clusters actionable findings by goal and can publish an authed HTML report page.",
+    intent:
+      "Re-run the ops inventory audit in one call: duplicate/dead schedules, code-work routing risks, enabled workflow fixtures, structured-output gate gaps, prompt registry drift, stale hosts, and systemDefault skill duplicates.",
+    source: asText(opsCatalogAuditSrc),
+  },
 ];
 
 /** A catalog entry resolved into a generic {@link SeedItem}. */

package/src/be/seed-skills/index.ts

@@ -31,6 +31,12 @@ import scheduledTaskResilienceConfig from "../../../templates/skills/scheduled-t
 import scheduledTaskResilienceContent from "../../../templates/skills/scheduled-task-resilience/content.md" with {
   type: "text",
 };
+import scriptWorkflowsConfig from "../../../templates/skills/script-workflows/config.json" with {
+  type: "text",
+};
+import scriptWorkflowsContent from "../../../templates/skills/script-workflows/content.md" with {
+  type: "text",
+};
 import swarmScriptsConfig from "../../../templates/skills/swarm-scripts/config.json" with {
   type: "text",
 };
@@ -72,6 +78,7 @@ const BUILT_IN_SKILL_SOURCES = [
   { config: kvStorageConfig, body: kvStorageContent },
   { config: pagesConfig, body: pagesContent },
   { config: scheduledTaskResilienceConfig, body: scheduledTaskResilienceContent },
+  { config: scriptWorkflowsConfig, body: scriptWorkflowsContent },
   { config: swarmScriptsConfig, body: swarmScriptsContent },
   { config: workflowIterateConfig, body: workflowIterateContent },
   { config: workflowStructuredOutputConfig, body: workflowStructuredOutputContent },

package/src/be/swarm-config-guard.ts

@@ -41,6 +41,23 @@ const VALIDATED_KEYS: Record<string, (value: unknown) => string | null> = {
     if (parsed.success) return null;
     return `Invalid HARNESS_PROVIDER value (must be one of: ${ProviderNameSchema.options.join(", ")})`;
   },
+  // Codex credits-exhausted cooldown (ms). Permissive on range here (positive
+  // integer) — the worker clamps to [5m, 7d] via resolveCodexCreditsExhaustedCooldownMs.
+  CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS: (value) => {
+    const str = String(value).trim();
+    if (!/^\d+$/.test(str) || Number(str) <= 0) {
+      return "Invalid CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS (must be a positive integer of milliseconds)";
+    }
+    return null;
+  },
+  SWARM_USE_CLAUDE_BRIDGE: (value) => {
+    if (typeof value !== "string") {
+      return "Invalid SWARM_USE_CLAUDE_BRIDGE value (must be one of: true, false, 1, 0)";
+    }
+    const normalized = value.trim().toLowerCase();
+    if (["true", "false", "1", "0"].includes(normalized)) return null;
+    return "Invalid SWARM_USE_CLAUDE_BRIDGE value (must be one of: true, false, 1, 0)";
+  },
 };
 
 export function validateConfigValue(key: string, value: unknown): string | null {

package/src/commands/runner.ts

@@ -40,9 +40,11 @@ import { getMcpBaseUrl } from "../utils/constants.ts";
 import { getContextWindowSize } from "../utils/context-window.ts";
 import { type CredentialSelection, resolveCredentialPools } from "../utils/credentials.ts";
 import {
+  isCodexCreditsExhaustedMessage,
   isRateLimitMessage,
   MAX_RATE_LIMIT_RESET_MS,
   parseRateLimitResetTime,
+  resolveCodexCreditsExhaustedCooldownMs,
 } from "../utils/error-tracker.ts";
 import { resolveHarnessProvider } from "../utils/harness-provider.ts";
 import { prettyPrintLine, prettyPrintStderr } from "../utils/pretty-print.ts";
@@ -424,6 +426,7 @@ async function fetchResolvedEnv(
 const RELOADABLE_ENV_KEYS: ReadonlySet<string> = new Set([
   "MODEL_OVERRIDE",
   "AGENT_FS_SHARED_ORG_ID",
+  "SWARM_USE_CLAUDE_BRIDGE",
 ]);
 
 /**
@@ -1483,6 +1486,13 @@ interface RunnerState {
    * (per-task live re-resolution) will mutate this between tasks.
    */
   harnessProvider: ProviderName;
+  /**
+   * Effective Codex credits-exhausted cooldown (ms), resolved from `swarm_config`
+   * (key `CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS`) > default 2h constant, clamped to
+   * [5m, 7d]. Reconciled live by `applySwarmConfigDrift` — read at the cooldown
+   * application site so a fresh value applies to the next credits-exhausted failure.
+   */
+  codexCreditsExhaustedCooldownMs: number;
 }
 
 /** Buffer for session logs */
@@ -3209,6 +3219,12 @@ async function checkCompletedProcesses(
             console.log(
               `[credentials] Parsed rate limit reset time from error: ${rateLimitedUntil}`,
             );
+          } else if (isCodexCreditsExhaustedMessage(failureReason)) {
+            const cooldownMs = state.codexCreditsExhaustedCooldownMs;
+            rateLimitedUntil = new Date(Date.now() + cooldownMs).toISOString();
+            console.log(
+              `[credentials] Codex credits exhausted — applying cooldown (${cooldownMs}ms): ${rateLimitedUntil}`,
+            );
           } else {
             rateLimitedUntil = new Date(Date.now() + 5 * 60 * 1000).toISOString();
           }
@@ -3403,10 +3419,22 @@ export async function runAgent(config: RunnerConfig, opts: RunnerOptions) {
   // Failures (network, API down, malformed value) fall back to env then "claude"
   // so a swarm_config outage cannot wedge boot.
   let bootProvider: ProviderName;
+  // Codex credits-exhausted cooldown is sourced solely from the global swarm_config
+  // (key `CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS`). Initialize to the default constant
+  // for the case where it is unset, then apply the swarm_config value below; on a
+  // boot-fetch failure it stays at the default. Reconciled live thereafter by
+  // `applySwarmConfigDrift`.
+  let bootCooldownMs = resolveCodexCreditsExhaustedCooldownMs(undefined);
   try {
-    bootProvider = (await fetchResolvedEnv(apiUrl, apiKey, agentId)).resolvedProvider;
+    const bootEnv = await fetchResolvedEnv(apiUrl, apiKey, agentId);
+    bootProvider = bootEnv.resolvedProvider;
+    bootCooldownMs = resolveCodexCreditsExhaustedCooldownMs(
+      bootEnv.env.CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS,
+    );
   } catch (err) {
-    console.warn(`[runner] fetchResolvedEnv failed at boot, falling back to env: ${err}`);
+    console.warn(
+      `[runner] fetchResolvedEnv failed at boot, falling back to env for provider and the default cooldown: ${err}`,
+    );
     bootProvider = resolveHarnessProvider({}, process.env);
   }
   console.log(`[runner] Resolved HARNESS_PROVIDER: ${bootProvider}`);
@@ -3610,6 +3638,7 @@ export async function runAgent(config: RunnerConfig, opts: RunnerOptions) {
     activeTasks: new Map(),
     maxConcurrent,
     harnessProvider: bootProvider,
+    codexCreditsExhaustedCooldownMs: bootCooldownMs,
   };
 
   // Track tasks already signaled for cancellation to avoid repeated SIGTERM
@@ -3690,6 +3719,18 @@ export async function runAgent(config: RunnerConfig, opts: RunnerOptions) {
       agentVisibleChanged = true;
     }
 
+    // (2b) Codex credits-exhausted cooldown — operator-tunable live. Not
+    // agent-visible (doesn't change provider/maxConcurrent → no re-register).
+    const nextCooldown = resolveCodexCreditsExhaustedCooldownMs(
+      freshEnv.CODEX_CREDITS_EXHAUSTED_COOLDOWN_MS,
+    );
+    if (nextCooldown !== state.codexCreditsExhaustedCooldownMs) {
+      console.log(
+        `[${role}] [config] codexCreditsExhaustedCooldownMs: ${state.codexCreditsExhaustedCooldownMs} → ${nextCooldown}`,
+      );
+      state.codexCreditsExhaustedCooldownMs = nextCooldown;
+    }
+
     // (3) Apply the small allowlist of safe-to-mutate env keys to process.env.
     const changedKeys = applyResolvedEnvToProcessEnv(freshEnv);
     if (changedKeys.length > 0) {

package/src/http/db-query.ts

@@ -11,6 +11,24 @@ export interface DbQueryResult {
   total: number;
 }
 
+export const DbQueryInputShape = {
+  sql: z.string().min(1).max(10_000).optional(),
+  query: z.string().min(1).max(10_000).optional().describe("Deprecated runtime alias for sql."),
+  params: z.array(z.any()).optional().default([]),
+};
+
+export const DbQueryInputSchema = z
+  .object(DbQueryInputShape)
+  .refine((body) => body.sql !== undefined || body.query !== undefined, {
+    message: "Either sql or query is required",
+  });
+
+export type DbQueryInput = z.infer<typeof DbQueryInputSchema>;
+
+export function resolveDbQuerySql(input: Pick<DbQueryInput, "sql" | "query">): string {
+  return input.sql ?? input.query ?? "";
+}
+
 function stripTrailingSemicolon(sql: string): string {
   return sql.trim().replace(/;\s*$/, "").trim();
 }
@@ -67,10 +85,7 @@ const dbQueryRoute = route({
   pattern: ["api", "db-query"],
   summary: "Execute a read-only SQL query",
   tags: ["Debug"],
-  body: z.object({
-    sql: z.string().min(1).max(10_000),
-    params: z.array(z.any()).optional().default([]),
-  }),
+  body: DbQueryInputSchema,
   responses: {
     200: {
       description: "Query results",
@@ -100,7 +115,7 @@ export async function handleDbQuery(
   if (!parsed) return true;
 
   try {
-    const result = executeReadOnlyQuery(parsed.body.sql, parsed.body.params);
+    const result = executeReadOnlyQuery(resolveDbQuerySql(parsed.body), parsed.body.params);
     json(res, result);
   } catch (err: unknown) {
     const message = err instanceof Error ? err.message : String(err);

package/src/http/index.ts

@@ -21,9 +21,11 @@ import {
   withRemoteContext,
   withSpanContext,
 } from "../otel";
+import { startScriptRunSupervisor, stopScriptRunSupervisor } from "../script-workflows/supervisor";
 import { startSlackApp, stopSlackApp } from "../slack";
 import { initTelemetry, telemetry } from "../telemetry";
 import { getApiKey } from "../utils/api-key";
+import { getMcpBaseUrl } from "../utils/constants";
 import { scrubSecrets } from "../utils/secret-scrubber";
 import { initWorkflows } from "../workflows";
 import { handleActiveSessions } from "./active-sessions";
@@ -57,6 +59,7 @@ import { handlePromptTemplates } from "./prompt-templates";
 import { handleRepos } from "./repos";
 import { describeRequestRoute } from "./route-def";
 import { handleSchedules } from "./schedules";
+import { handleScriptRuns } from "./script-runs";
 import { handleScripts } from "./scripts";
 import { handleSessionData } from "./session-data";
 import { handleSessions } from "./sessions";
@@ -272,6 +275,7 @@ const httpServer = createHttpServer(async (req, res) => {
         () => handleMetrics(req, res, pathSegments, queryParams, myAgentId),
         () => handleRepos(req, res, pathSegments, queryParams),
         () => handleSkills(req, res, pathSegments, queryParams, myAgentId),
+        () => handleScriptRuns(req, res, pathSegments, queryParams, myAgentId),
         () => handleScripts(req, res, pathSegments, queryParams, myAgentId),
         () => handleMcpBridge(req, res, pathSegments, queryParams, myAgentId),
         () => handleMcpServers(req, res, pathSegments, queryParams),
@@ -343,6 +347,9 @@ async function shutdown() {
   // Stop heartbeat triage
   stopHeartbeat();
 
+  // Stop durable script workflow subprocesses
+  stopScriptRunSupervisor();
+
   // Stop Slack bot
   await stopSlackApp();
 
@@ -487,6 +494,9 @@ httpServer
     // Initialize workflow engine (trigger subscriptions + resume listener)
     initWorkflows();
 
+    // Reconcile durable script workflow subprocesses
+    startScriptRunSupervisor(getMcpBaseUrl());
+
     // Start scheduler (if enabled)
     if (hasCapability("scheduling")) {
       const { startScheduler } = await import("../scheduler");