@desplega.ai/agent-swarm 1.98.0 → 1.99.0

package/src/providers/claude-adapter.ts

@@ -248,6 +248,7 @@ export function mergeMcpConfig(
   baseConfig: { mcpServers?: Record<string, unknown> } | null,
   installedServers: Record<string, Record<string, unknown>> | null,
   taskId: string,
+  contextKey?: string,
 ): { mcpServers: Record<string, unknown> } {
   const config: { mcpServers: Record<string, unknown> } = {
     mcpServers: { ...(baseConfig?.mcpServers ?? {}) },
@@ -273,6 +274,9 @@ export function mergeMcpConfig(
     const server = config.mcpServers[serverKey] as Record<string, unknown>;
     if (!server.headers) server.headers = {};
     (server.headers as Record<string, string>)["X-Source-Task-Id"] = taskId;
+    if (contextKey) {
+      (server.headers as Record<string, string>)["X-Context-Key"] = contextKey;
+    }
   }
 
   return config;
@@ -291,6 +295,7 @@ export async function createSessionMcpConfig(
   cwd: string,
   taskId: string,
   installedServers?: Record<string, Record<string, unknown>> | null,
+  contextKey?: string,
 ): Promise<string | null> {
   // Collect every .mcp.json from cwd up to filesystem root. Stopping at the first
   // match silently drops the swarm-managed /workspace/.mcp.json when the cloned
@@ -341,7 +346,12 @@ export async function createSessionMcpConfig(
   }
 
   try {
-    const config = mergeMcpConfig({ mcpServers: mergedServers }, installedServers ?? null, taskId);
+    const config = mergeMcpConfig(
+      { mcpServers: mergedServers },
+      installedServers ?? null,
+      taskId,
+      contextKey,
+    );
     const sessionConfigPath = `/tmp/mcp-${taskId}.json`;
     await writeFile(sessionConfigPath, JSON.stringify(config, null, 2));
     return sessionConfigPath;
@@ -950,11 +960,12 @@ export class ClaudeAdapter implements ProviderAdapter {
       );
     }
 
-    // Create per-session MCP config with X-Source-Task-Id header + installed servers (no shared-file race condition)
+    // Create per-session MCP config with X-Source-Task-Id + X-Context-Key headers + installed servers (no shared-file race condition)
     const sessionMcpConfig = await createSessionMcpConfig(
       config.cwd,
       config.taskId,
       installedServers,
+      config.contextKey,
     );
 
     // Stage the system prompt on disk so it can be passed as a file path

package/src/providers/pricing-sources.md

@@ -1,16 +1,32 @@
 # Pricing sources
 
-This page lists the sources that feed the `pricing` table at server boot.
-Operators bumping a rate by hand should also update this file.
+This page lists the sources that feed the `pricing` table. Operators bumping a
+rate by hand should also update this file.
 
-## Primary: vendored models.dev snapshot
+## Primary pricing freshness: runtime models.dev refresh
 
-- **Source-of-truth path**: `src/be/modelsdev-cache.json`
+- **Runtime module**: `src/be/pricing-refresh.ts`
+- **Upstream**: `https://models.dev/api.json`, fetched with `If-None-Match`.
+- **Boot wiring**: after `seedPricingFromModelsDev()`, the API server starts one
+  non-blocking refresh and then repeats every 12 hours with `setInterval`.
+- **Update rule**: project upstream through `buildModelsDevSeedRows()` and insert
+  a new `effective_from=Date.now()` row only when the model/token class is new
+  or the active price changed. Identical prices are no-ops.
+- **Growth bound**: after each refresh, keep only the latest two rows per
+  `(provider, model, token_class)` triple.
+- **Pinned local entries**: safe by construction. The runtime refresh only adds
+  pricing rows; it does not rewrite or delete the committed snapshot.
+
+## Fallback/UI catalog: vendored models.dev snapshot
+
+- **Fallback path**: `src/be/modelsdev-cache.json`
 - **UI compatibility path**: `ui/src/lib/modelsdev-cache.json` symlinks to the
   backend snapshot so existing UI imports keep working.
 - **Loaded by**: `src/be/modelsdev-cache.ts` → `src/be/seed-pricing.ts` →
   `seedPricingFromModelsDev()`,
   called from `src/server.ts` after `initDb`.
+- **Role**: cold-start fallback seed for pricing when models.dev is unavailable,
+  plus the UI model-picker source for names, labels, and context windows.
 - **Projection rules** (see the same module for code-level detail):
   - Anthropic models → rows under `provider='claude'` AND `provider='claude-managed'`.
     Shortnames (`opus`, `sonnet`, `haiku`) ALSO get rows keyed by the current
@@ -22,12 +38,13 @@ Operators bumping a rate by hand should also update this file.
     stripped name and the full `google/...` id) so internal-ai callers find
     a hit either way.
 
-- **Refresh procedure** (the only place to update the snapshot):
+- **Snapshot refresh procedure**:
   - Run `bun run scripts/refresh-modelsdev-pricing.ts` (Phase 2 — adds the
     script). It fetches the latest snapshot from models.dev, diffs against
     the vendored copy, prints a summary, and writes the new file.
   - Commit the regenerated `src/be/modelsdev-cache.json` together with a bump
-    note in the PR description.
+    note in the PR description. This is no longer the pricing freshness path;
+    use it when the fallback/UI catalog needs new labels or context-window data.
 
 ## Manual overrides
 
@@ -50,6 +67,7 @@ no input/output pricing rows at the lookup time, the row is persisted with
 `costSource='unpriced'` (rather than 'harness'). The UI surfaces this as a
 yellow badge.
 
-To fix: either add the model to `src/be/modelsdev-cache.json` (preferred — the
-upstream snapshot probably needs refreshing) or add a manual override row via
-the existing admin route `POST /api/pricing`.
+To fix: first check whether the runtime refresh is failing. If the model must
+also appear in the UI picker or cold-start fallback, add it to
+`src/be/modelsdev-cache.json`; otherwise add a manual override row via the
+existing admin route `POST /api/pricing`.

package/src/providers/types.ts

@@ -92,6 +92,7 @@ export interface ProviderSessionConfig {
   apiKey: string;
   cwd: string;
   vcsRepo?: string;
+  contextKey?: string;
   /**
    * @deprecated Never set by the runner — native session resume was removed in
    * the 2026-05-28 plan. Adapters log + ignore any stray value. Follow-up

package/src/scripts-runtime/swarm-sdk.ts

@@ -56,7 +56,11 @@ function bridgeRequestFor(name: string, args: unknown): BridgeRequest | null {
     case "memory_get": {
       const memoryId = typeof body.memoryId === "string" ? body.memoryId : undefined;
       if (!memoryId) throw new Error("memory_get requires string `memoryId`");
-      return { method: "GET", path: `/api/memory/${encodeURIComponent(memoryId)}` };
+      const getIntent = typeof body.intent === "string" ? body.intent : "script-sdk";
+      return {
+        method: "GET",
+        path: `/api/memory/${encodeURIComponent(memoryId)}?intent=${encodeURIComponent(getIntent)}`,
+      };
     }
     case "memory_rate": {
       const event = {

package/src/scripts-runtime/types/stdlib.d.ts

@@ -50,11 +50,12 @@ declare module "swarm-sdk" {
     // --- memory ---
     memory_search(args: {
       query: string;
+      intent: string;
       scope?: "all" | "agent" | "swarm";
       limit?: number;
       source?: string;
     }): Promise<unknown>;
-    memory_get(args: { memoryId: string }): Promise<unknown>;
+    memory_get(args: { memoryId: string; intent: string }): Promise<unknown>;
     memory_rate(args: { id: string; useful: boolean; note?: string }): Promise<unknown>;
     // --- tasks ---
     task_list(args?: Record<string, unknown>): Promise<unknown>;

package/src/scripts-runtime/types/swarm-sdk.d.ts

@@ -32,11 +32,12 @@ declare module "swarm-sdk" {
     // --- memory ---
     memory_search(args: {
       query: string;
+      intent: string;
       scope?: "all" | "agent" | "swarm";
       limit?: number;
       source?: string;
     }): Promise<unknown>;
-    memory_get(args: { memoryId: string }): Promise<unknown>;
+    memory_get(args: { memoryId: string; intent: string }): Promise<unknown>;
     memory_rate(args: { id: string; useful: boolean; note?: string }): Promise<unknown>;
     // --- tasks ---
     task_list(args?: Record<string, unknown>): Promise<unknown>;

package/src/server.ts

@@ -1,6 +1,7 @@
 import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import pkg from "../package.json";
 import { initDb } from "./be/db";
+import { startPricingRefreshLoop } from "./be/pricing-refresh";
 import { seedPricingFromModelsDev } from "./be/seed-pricing";
 import { registerCancelTaskTool } from "./tools/cancel-task";
 import { registerContextDiffTool } from "./tools/context-diff";
@@ -172,6 +173,7 @@ export function createServer() {
   // call on every boot. See src/be/seed-pricing.ts for the projection logic
   // and the manual-override constants for runtime-fee / ACU pricing.
   seedPricingFromModelsDev();
+  startPricingRefreshLoop();
 
   const server = new McpServer(
     {

package/src/slack/blocks.ts

@@ -8,8 +8,9 @@
 import type { AgentTaskStatus, TaskAttachment } from "../types";
 import { buildAgentFsLiveUrl, getAppUrl } from "../utils/constants";
 
-// Slack limits section text to 3000 chars; we use 2900 for safety
-const MAX_SECTION_LENGTH = 2900;
+// Slack limits section text to 3000 chars; we use 2900 for safety.
+export const MAX_SECTION_LENGTH = 2900;
+export const MAX_BLOCKS_PER_COMPLETION_MESSAGE = 45;
 
 // biome-ignore lint/suspicious/noExplicitAny: Slack block types are complex unions; we build plain objects
 type SlackBlock = any;
@@ -38,24 +39,33 @@ export function getTaskUrl(taskId: string): string {
  * Convert GitHub-flavored markdown to Slack mrkdwn format.
  *
  * Key differences:
- * - GitHub: **bold**, *italic*, ~~strike~~, [text](url)
- * - Slack:  *bold*,  _italic_, ~strike~,   <url|text>
+ * - GitHub: **bold**, __bold__, *italic*, ~~strike~~, ### Header, [text](url)
+ * - Slack:  *bold*,  *bold*,   _italic_, ~strike~,   *Header*, text (url)
  */
 export function markdownToSlack(text: string): string {
   return (
     text
+      // Images: keep alt text and expose the URL plainly.
+      .replace(/!\[([^\]]*)\]\(([^)\s]+)(?:\s+"[^"]*")?\)/g, (_match, alt, url) =>
+        alt ? `${alt} (${url})` : url,
+      )
+      // Links: keep a plain URL fallback instead of Slack's <url|text> shortcut.
+      // Slack block auto-promotion has historically rejected that shortcut with
+      // invalid_blocks, while plain URLs remain copyable and auto-unfurlable.
+      .replace(/\[([^\]]+)\]\(([^)\s]+)(?:\s+"[^"]*")?\)/g, "$1 ($2)")
       // Headers to bold placeholder (# Header -> bold, protected from italic)
       .replace(/^#{1,6}\s+(.+)$/gm, "\uE000$1\uE001")
       // Bold **text** -> placeholder (to avoid italic chain converting *bold* to _italic_)
       .replace(/\*\*(.+?)\*\*/g, "\uE000$1\uE001")
+      // Bold __text__ -> placeholder
+      .replace(/__(.+?)__/g, "\uE000$1\uE001")
       // Italic *text* -> _text_ (single asterisks, now safe from bold placeholders)
       .replace(/(?<!\*)\*([^*]+)\*(?!\*)/g, "_$1_")
+      // Italic _text_ already matches Slack mrkdwn; leave it alone.
       // Restore bold from placeholder -> *text*
       .replace(/\uE000(.+?)\uE001/g, "*$1*")
       // Strikethrough ~~text~~ -> ~text~
       .replace(/~~(.+?)~~/g, "~$1~")
-      // Links [text](url) -> <url|text>
-      .replace(/\[([^\]]+)\]\(([^)]+)\)/g, "<$2|$1>")
       // Inline code already works the same
       // Bullet points already work the same
       // Remove excessive blank lines
@@ -66,7 +76,7 @@ export function markdownToSlack(text: string): string {
 /**
  * Split text into chunks that fit within Slack's section text limit.
  */
-function splitText(text: string): string[] {
+export function splitSlackSectionText(text: string): string[] {
   if (text.length <= MAX_SECTION_LENGTH) return [text];
 
   const chunks: string[] = [];
@@ -248,17 +258,50 @@ export function buildCompletedBlocks(opts: {
 
   // Only include body if not minimal (agent didn't reply via slack-reply)
   if (!opts.minimal) {
-    for (const chunk of splitText(opts.body)) {
+    for (const chunk of splitSlackSectionText(opts.body)) {
       blocks.push(sectionBlock(chunk));
     }
   } else if (opts.trailer && opts.trailer.length > 0) {
-    for (const chunk of splitText(opts.trailer)) {
+    for (const chunk of splitSlackSectionText(opts.trailer)) {
       blocks.push(sectionBlock(chunk));
     }
   }
   return blocks;
 }
 
+/**
+ * Build one or more completed-task block payloads. The first payload carries
+ * the normal completion header; continuation payloads carry a compact part
+ * header. This keeps long completion summaries inside Slack's block limits
+ * without dropping body text.
+ */
+export function buildCompletedBlockBatches(opts: Parameters<typeof buildCompletedBlocks>[0]) {
+  const allBlocks = buildCompletedBlocks(opts);
+  if (allBlocks.length <= MAX_BLOCKS_PER_COMPLETION_MESSAGE) return [allBlocks];
+
+  const header = allBlocks[0];
+  const bodyBlocks = allBlocks.slice(1);
+  const bodyLimit = MAX_BLOCKS_PER_COMPLETION_MESSAGE - 1;
+  const batches: SlackBlock[][] = [];
+
+  for (let start = 0; start < bodyBlocks.length; start += bodyLimit) {
+    const partBlocks = bodyBlocks.slice(start, start + bodyLimit);
+    if (start === 0) {
+      batches.push([header, ...partBlocks]);
+    } else {
+      const part = batches.length + 1;
+      batches.push([
+        sectionBlock(
+          `↳ *${opts.agentName}* (${getTaskLink(opts.taskId)}) continued · part ${part}`,
+        ),
+        ...partBlocks,
+      ]);
+    }
+  }
+
+  return batches;
+}
+
 /**
  * Build blocks for a failed task response.
  * Single-line header, then error in code block.
@@ -369,7 +412,10 @@ function truncateOutput(text: string): string {
   const sentenceEnd = text.search(/\.\s/);
   const firstSentence = sentenceEnd !== -1 ? text.slice(0, sentenceEnd + 1) : text;
   if (firstSentence.length <= MAX_OUTPUT_LENGTH) return firstSentence;
-  return `${text.slice(0, MAX_OUTPUT_LENGTH)}…`;
+  const boundary = text.lastIndexOf(" ", MAX_OUTPUT_LENGTH);
+  const cut = boundary >= MAX_OUTPUT_LENGTH / 2 ? boundary : MAX_OUTPUT_LENGTH;
+  const omitted = text.length - cut;
+  return `${text.slice(0, cut).trimEnd()}… (${omitted} more chars; full output in thread)`;
 }
 
 /**
@@ -399,7 +445,7 @@ function renderChildDetail(node: TreeNode, indent: string): string[] {
   }
 
   if (node.status === "completed" && !node.slackReplySent && node.output) {
-    lines.push(`${indent}${truncateOutput(node.output)}`);
+    lines.push(`${indent}${truncateOutput(markdownToSlack(node.output))}`);
   }
 
   return lines;
@@ -423,7 +469,7 @@ function renderTree(root: TreeNode): string {
       lines.push(`    Error: ${root.failureReason}`);
     }
     if (root.status === "completed" && !root.slackReplySent && root.output) {
-      lines.push(`    ${truncateOutput(root.output)}`);
+      lines.push(`    ${truncateOutput(markdownToSlack(root.output))}`);
     }
     return lines.join("\n");
   }

package/src/slack/responses.ts

@@ -4,6 +4,7 @@ import type { Agent, AgentTask } from "../types";
 import { getSlackApp } from "./app";
 import {
   buildCancelledBlocks,
+  buildCompletedBlockBatches,
   buildCompletedBlocks,
   buildFailedBlocks,
   buildProgressBlocks,
@@ -74,7 +75,7 @@ export async function sendTaskResponse(task: AgentTask): Promise<boolean> {
       console.log(
         `[Slack] sendTaskResponse: task=${task.id} slackReplySent=${!!task.slackReplySent} minimal=${!!task.slackReplySent}`,
       );
-      const blocks = buildCompletedBlocks({
+      const completionOpts = {
         agentName,
         taskId: task.id,
         body,
@@ -84,15 +85,21 @@ export async function sendTaskResponse(task: AgentTask): Promise<boolean> {
         // trailing addendum so links are visible without expanding the card.
         minimal: !!task.slackReplySent,
         trailer: task.slackReplySent ? attachmentsBlock : undefined,
-      });
-      await sendWithPersona(client, {
-        channel: task.slackChannelId,
-        thread_ts: task.slackThreadTs,
-        text: task.slackReplySent ? `✅ ${agentName} completed` : body,
-        username: getAgentDisplayName(agent),
-        icon_emoji: getAgentEmoji(agent),
-        blocks,
-      });
+      };
+      const blockBatches = buildCompletedBlockBatches(completionOpts);
+      for (let i = 0; i < blockBatches.length; i++) {
+        await sendWithPersona(client, {
+          channel: task.slackChannelId,
+          thread_ts: task.slackThreadTs,
+          text:
+            task.slackReplySent || i > 0
+              ? `✅ ${agentName} completed${i > 0 ? ` (continued ${i + 1}/${blockBatches.length})` : ""}`
+              : body,
+          username: getAgentDisplayName(agent),
+          icon_emoji: getAgentEmoji(agent),
+          blocks: blockBatches[i],
+        });
+      }
     } else if (task.status === "failed") {
       const reason = task.failureReason || "Unknown error";
       const blocks = buildFailedBlocks({ agentName, taskId: task.id, reason });
@@ -199,6 +206,7 @@ export async function updateToFinal(task: AgentTask, messageTs: string): Promise
   const agentName = agent.name;
   let blocks: unknown[];
   let text: string;
+  let completionBlockBatches: unknown[][] | undefined;
 
   if (task.status === "completed") {
     const output = task.output || "Task completed.";
@@ -212,14 +220,16 @@ export async function updateToFinal(task: AgentTask, messageTs: string): Promise
     console.log(
       `[Slack] updateToFinal: task=${task.id} slackReplySent=${!!task.slackReplySent} minimal=${!!task.slackReplySent}`,
     );
-    blocks = buildCompletedBlocks({
+    const completionOpts = {
       agentName,
       taskId: task.id,
       body,
       duration,
       minimal: !!task.slackReplySent,
       trailer: task.slackReplySent ? attachmentsBlock : undefined,
-    });
+    };
+    completionBlockBatches = buildCompletedBlockBatches(completionOpts);
+    blocks = completionBlockBatches[0] ?? buildCompletedBlocks(completionOpts);
     text = task.slackReplySent ? `✅ ${agentName} completed` : body;
   } else if (task.status === "cancelled") {
     blocks = buildCancelledBlocks({ agentName, taskId: task.id });
@@ -238,6 +248,19 @@ export async function updateToFinal(task: AgentTask, messageTs: string): Promise
       // biome-ignore lint/suspicious/noExplicitAny: Block Kit objects
       blocks: blocks as any,
     });
+
+    if (completionBlockBatches) {
+      for (let i = 1; i < completionBlockBatches.length; i++) {
+        await sendWithPersona(app.client, {
+          channel: task.slackChannelId,
+          thread_ts: task.slackThreadTs ?? messageTs,
+          text: `✅ ${agentName} completed (continued ${i + 1}/${completionBlockBatches.length})`,
+          username: getAgentDisplayName(agent),
+          icon_emoji: getAgentEmoji(agent),
+          blocks: completionBlockBatches[i],
+        });
+      }
+    }
     return true;
   } catch (error) {
     console.error(`[Slack] Failed to update task message to final state:`, error);

package/src/slack/watcher.ts

@@ -696,13 +696,6 @@ export function startTaskWatcher(intervalMs = 3000): void {
           }
         }
 
-        // Skip tasks tracked in a tree — they're rendered by processTreeMessages()
-        // But mark as notified to prevent re-processing if tree is cleaned up
-        if (taskToTree.has(task.id)) {
-          notifiedCompletions.set(task.id, now);
-          continue;
-        }
-
         const completionKey = `completion:${task.id}`;
 
         // Skip if already notified or currently sending or sent recently
@@ -710,6 +703,34 @@ export function startTaskWatcher(intervalMs = 3000): void {
         const lastSent = lastSendTime.get(completionKey);
         if (lastSent && now - lastSent < MIN_SEND_INTERVAL) continue;
 
+        // Tasks tracked in a tree are still rendered by processTreeMessages().
+        // Successful tasks without their own slack-reply also get a full
+        // threaded completion message so the tree's compact preview is not the
+        // only place their output appears.
+        if (taskToTree.has(task.id)) {
+          if (task.status !== "completed" || task.slackReplySent) {
+            notifiedCompletions.set(task.id, now);
+            continue;
+          }
+
+          pendingSends.add(completionKey);
+          notifiedCompletions.set(task.id, now);
+          lastSendTime.set(completionKey, now);
+          try {
+            await sendTaskResponse(task);
+            console.log(
+              `[Slack] Sent full tree-tracked completion for task ${task.id.slice(0, 8)}`,
+            );
+          } catch (error) {
+            notifiedCompletions.delete(task.id);
+            lastSendTime.delete(completionKey);
+            console.error(`[Slack] Failed to send tree-tracked completion:`, error);
+          } finally {
+            pendingSends.delete(completionKey);
+          }
+          continue;
+        }
+
         // Mark as pending and notified BEFORE sending
         pendingSends.add(completionKey);
         notifiedCompletions.set(task.id, now);

package/src/tests/internal-ai/complete-structured.test.ts

@@ -1,7 +1,10 @@
 import { describe, expect, test } from "bun:test";
 import { Type } from "typebox";
 import { z } from "zod";
-import { completeStructured } from "../../utils/internal-ai/complete-structured.js";
+import {
+  completeStructured,
+  defaultSpawnClaudeCli,
+} from "../../utils/internal-ai/complete-structured.js";
 import type { ResolvedCredential } from "../../utils/internal-ai/credentials.js";
 
 const ResultZodSchema = z.object({
@@ -274,3 +277,33 @@ describe("completeStructured", () => {
     expect(match).toBeDefined();
   });
 });
+
+describe("defaultSpawnClaudeCli", () => {
+  test("sets SKIP_SESSION_SUMMARY=1 in the child env (Stop-hook recursion guard)", async () => {
+    // Without this guard, the spawned `claude -p` summarizer session fires the
+    // same global Stop hook on exit, which spawns another summarizer claude,
+    // recursively — observed OOM-wedging 8GB E2B worker sandboxes.
+    const fakeBinary = `/tmp/fake-claude-${Date.now()}-${Math.random().toString(36).slice(2)}.sh`;
+    await Bun.write(
+      fakeBinary,
+      '#!/usr/bin/env bash\ncat >/dev/null\nprintf \'{"result":"SKIP_SESSION_SUMMARY=%s"}\' "$SKIP_SESSION_SUMMARY"\n',
+    );
+    const savedBinary = process.env.CLAUDE_BINARY;
+    const savedSkip = process.env.SKIP_SESSION_SUMMARY;
+    const savedBridge = process.env.SWARM_USE_CLAUDE_BRIDGE;
+    process.env.CLAUDE_BINARY = `bash ${fakeBinary}`;
+    // Ensure a false-pass via inheritance is impossible.
+    delete process.env.SKIP_SESSION_SUMMARY;
+    delete process.env.SWARM_USE_CLAUDE_BRIDGE;
+    try {
+      const out = await defaultSpawnClaudeCli("prompt", "haiku");
+      expect(out).toBe("SKIP_SESSION_SUMMARY=1");
+    } finally {
+      if (savedBinary === undefined) delete process.env.CLAUDE_BINARY;
+      else process.env.CLAUDE_BINARY = savedBinary;
+      if (savedSkip !== undefined) process.env.SKIP_SESSION_SUMMARY = savedSkip;
+      if (savedBridge !== undefined) process.env.SWARM_USE_CLAUDE_BRIDGE = savedBridge;
+      await Bun.$`rm -f ${fakeBinary}`.quiet();
+    }
+  });
+});