@desplega.ai/agent-swarm 1.172.0 → 2.0.0

package/src/be/migrations/217_chat_sdk_state.sql

@@ -0,0 +1,104 @@
+-- Chat SDK (chat-sdk.dev) channel: StateAdapter tables + durable outbound relay.
+--
+-- The five chat_* state tables back `SqliteChatState` (src/integrations/chat-sdk/state.ts).
+-- Keys and thread ids written by the SDK already carry the adapter prefix
+-- (`telegram:...`, `slack:...`), so no adapter column is needed.
+-- Expiry columns are epoch milliseconds; NULL means no expiry.
+
+CREATE TABLE chat_state_kv (
+  key        TEXT PRIMARY KEY,
+  value      TEXT NOT NULL,
+  expires_at INTEGER
+);
+
+CREATE TABLE chat_subscriptions (
+  thread_id  TEXT PRIMARY KEY,
+  created_at TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ', 'now'))
+);
+
+CREATE TABLE chat_locks (
+  thread_id  TEXT PRIMARY KEY,
+  token      TEXT NOT NULL,
+  expires_at INTEGER NOT NULL
+);
+
+CREATE TABLE chat_lists (
+  key        TEXT NOT NULL,
+  seq        INTEGER NOT NULL,
+  value      TEXT NOT NULL,
+  expires_at INTEGER,
+  PRIMARY KEY (key, seq)
+);
+
+CREATE TABLE chat_queue (
+  id        INTEGER PRIMARY KEY AUTOINCREMENT,
+  thread_id TEXT NOT NULL,
+  entry     TEXT NOT NULL
+);
+
+CREATE INDEX idx_chat_queue_thread ON chat_queue(thread_id, id);
+
+-- Persist the obligation to publish a terminal result for every Chat SDK-rooted
+-- task. Mirrors 144_slack_relay_obligations.sql, plus an attempt cap: a poison
+-- row is marked given up (with the reason) instead of retrying forever.
+CREATE TABLE chat_relay_obligations (
+  task_id         TEXT PRIMARY KEY REFERENCES agent_tasks(id) ON DELETE CASCADE,
+  delivered_at    TEXT,
+  attempt_count   INTEGER NOT NULL DEFAULT 0,
+  last_attempt_at TEXT,
+  given_up_at     TEXT,
+  last_error      TEXT,
+  created_at      TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ', 'now')),
+  updated_at      TEXT NOT NULL DEFAULT (strftime('%Y-%m-%dT%H:%M:%fZ', 'now')),
+  created_by      TEXT,
+  updated_by      TEXT
+);
+
+CREATE INDEX idx_chat_relay_obligations_pending
+  ON chat_relay_obligations(last_attempt_at, created_at)
+  WHERE delivered_at IS NULL AND given_up_at IS NULL;
+
+-- Lead follow-ups (`source='system'`, `taskType='follow-up'`) inherit the chat
+-- contextKey and carry the lead's final answer, so they relay too. A resume
+-- (`source='system'`, `taskType='resume'`, from `createResumeFollowUp` after a
+-- graceful shutdown or crash recovery) inherits the contextKey too. It relays
+-- only when a lead owns it: a worker task (resumed or not) reports back through
+-- a lead follow-up, and the lead never gets one for its own tasks. Other system
+-- tasks that inherit the key stay silent. GLOB, not LIKE: LIKE is
+-- case-insensitive in SQLite.
+CREATE TRIGGER enqueue_chat_relay_after_terminal_update
+AFTER UPDATE OF status ON agent_tasks
+WHEN OLD.status NOT IN ('completed', 'failed', 'cancelled', 'superseded')
+  AND NEW.status IN ('completed', 'failed', 'cancelled')
+  AND NEW.contextKey GLOB 'task:chat-sdk:*'
+  AND (
+    NEW.source = 'chat'
+    OR (NEW.source = 'system' AND NEW.taskType = 'follow-up')
+    OR (
+      NEW.source = 'system'
+      AND NEW.taskType = 'resume'
+      AND EXISTS (SELECT 1 FROM agents WHERE agents.id = NEW.agentId AND agents.isLead = 1)
+    )
+  )
+BEGIN
+  INSERT OR IGNORE INTO chat_relay_obligations (task_id)
+  VALUES (NEW.id);
+END;
+
+CREATE TRIGGER enqueue_chat_relay_after_terminal_insert
+AFTER INSERT ON agent_tasks
+WHEN NEW.status IN ('completed', 'failed', 'cancelled')
+  AND NEW.contextKey GLOB 'task:chat-sdk:*'
+  AND (
+    NEW.source = 'chat'
+    OR (NEW.source = 'system' AND NEW.taskType = 'follow-up')
+    OR (
+      NEW.source = 'system'
+      AND NEW.taskType = 'resume'
+      AND EXISTS (SELECT 1 FROM agents WHERE agents.id = NEW.agentId AND agents.isLead = 1)
+    )
+  )
+BEGIN
+  INSERT OR IGNORE INTO chat_relay_obligations (task_id)
+  VALUES (NEW.id);
+END;

package/src/be/model-catalog-store.ts

@@ -13,6 +13,7 @@
  * See runbooks/model-catalog.md.
  */
 
+import { withAgentChatRegistrySections } from "@desplega/model-catalog";
 import { type RuntimeCatalog, setRuntimeModelCatalog } from "../utils/runtime-model-catalog";
 import {
   getActivePricingRow,
@@ -180,7 +181,10 @@ function pickNumber(value: unknown): number | undefined {
 }
 
 /** Project a models.dev payload into `model_catalog` rows for CATALOG_PROVIDER_IDS. */
-export function buildCatalogEntries(cache: ModelsDevCache, checkedAt: number): ModelCatalogEntry[] {
+export function buildCatalogEntries(raw: ModelsDevCache, checkedAt: number): ModelCatalogEntry[] {
+  // A live models.dev payload still carries the embeddings and image models of
+  // the registry gateways (`vercel`); pi and opencode cannot run them.
+  const cache = withAgentChatRegistrySections(raw);
   const entries: ModelCatalogEntry[] = [];
   for (const providerId of CATALOG_PROVIDER_IDS) {
     const provider = cache[providerId];

package/src/be/model-tier-resolution.ts

@@ -509,6 +509,8 @@ export type ClaimModelFields = Partial<TaskModelResolution> & {
 /**
  * Resolve and record the model for a task the given agent is claiming. Runs
  * inside the claim transaction. Returns the fields to merge into the trigger.
+ * `task.model` must be the task's own pin (`taskModelPin`), never the model a
+ * previous session reported into the column.
  */
 export async function recordClaimModelResolution(
   task: { id: string; model?: string | null; modelTier?: string | null },
@@ -521,7 +523,16 @@ export async function recordClaimModelResolution(
     harnessProvider,
     workerOverrides: await getAgentModelTierOverrides(agent.id),
   });
-  if (!resolved) return {};
+  if (!resolved) {
+    // A re-claimed row keeps the claim record of its previous run; drop it so
+    // the trigger and the row do not replay that run's model.
+    await getDbClient().run(
+      `UPDATE agent_tasks SET resolvedModel = NULL, modelSource = NULL, modelAlias = NULL
+         WHERE id = ? AND (resolvedModel IS NOT NULL OR modelSource IS NOT NULL OR modelAlias IS NOT NULL)`,
+      [task.id],
+    );
+    return {};
+  }
   if ("harnessMismatch" in resolved) return { modelUnsupported: resolved.harnessMismatch };
   const checked = await applyCliSupport(
     resolved,

package/src/be/model-validation.ts

@@ -29,6 +29,7 @@ import {
   type ModelsDevCatalog,
   parseAlias,
 } from "@desplega/model-catalog";
+import { type TaskModelPinFields, taskModelPin } from "../tasks/model-pin";
 import { ampModelError, ampPinCatalogError, ampPinProviderError } from "../utils/amp-models";
 import { getAgentById, getAllAgents } from "./db";
 import { loadModelsCatalog } from "./model-catalog-store";
@@ -99,10 +100,12 @@ export async function harnessModelErrorFor(
  * stays sync and catalog-free; this runs next to it.
  */
 export async function poolTaskRunsOnHarness(
-  task: { model?: string | null } | null | undefined,
+  task: TaskModelPinFields | null | undefined,
   harness: string | null | undefined,
 ): Promise<boolean> {
-  const model = task?.model?.trim();
+  // Only the task's own pin restricts harnesses; a model a previous run
+  // reported into `model` does not.
+  const model = taskModelPin(task ?? {})?.trim();
   if (!model) return true;
   return (await harnessModelErrorFor(model, harness)) === null;
 }

package/src/be/models-catalog.ts

@@ -1,3 +1,4 @@
+import { openAICompatibleProviderIds, registryCatalogSections } from "@desplega/model-catalog";
 import { loadModelsDevCache, type ModelsDevCache, type ModelsDevModel } from "./modelsdev-cache";
 
 /**
@@ -15,13 +16,18 @@ import { loadModelsDevCache, type ModelsDevCache, type ModelsDevModel } from "./
  * same snapshot as its own fallback while the request is in flight.
  */
 
-/** Mirrors the provider groups used by the direct harness and ACP model pickers. */
+/**
+ * Mirrors the provider groups used by the direct harness and ACP model
+ * pickers, plus every OpenAI-compatible registry provider
+ * (`@desplega/model-catalog`).
+ */
 export const CATALOG_PROVIDER_IDS = [
   "openrouter",
   "anthropic",
   "openai",
   "amazon-bedrock",
   "opencode",
+  ...openAICompatibleProviderIds(),
 ] as const;
 
 export type CatalogProviderId = (typeof CATALOG_PROVIDER_IDS)[number];
@@ -163,6 +169,24 @@ function mergePinnedEntries(catalog: ModelsCatalog): void {
   }
 }
 
+/**
+ * Registry-synthesized sections (`catalog: "registry"` entries such as
+ * `helmcode`) are never on models.dev, so the live catalog built from it
+ * lacks them. Fill them from the bundled snapshot, which carries them
+ * (`loadModelsDevCache`); models already present (an overlay row) win.
+ */
+function mergeRegistrySections(catalog: ModelsCatalog): void {
+  const snapshot = bundledSnapshotCatalog();
+  for (const providerId of Object.keys(registryCatalogSections()) as CatalogProviderId[]) {
+    const section = snapshot[providerId];
+    if (!section) continue;
+    catalog[providerId] ??= { ...section, models: {} };
+    for (const [modelId, model] of Object.entries(section.models)) {
+      catalog[providerId].models[modelId] ??= model;
+    }
+  }
+}
+
 /**
  * Install a freshly built live catalog (already merged with the persistent
  * table + overlay by `src/be/model-catalog-store.ts`). Pinned snapshot entries
@@ -170,6 +194,7 @@ function mergePinnedEntries(catalog: ModelsCatalog): void {
  */
 export function setLiveModelsCatalog(catalog: ModelsCatalog, updatedAt: number | null): void {
   mergePinnedEntries(catalog);
+  mergeRegistrySections(catalog);
   liveCatalog = catalog;
   liveUpdatedAt = updatedAt;
 }

package/src/be/modelsdev-cache.json

@@ -147258,7 +147258,7 @@
         "cost": {
           "input": 2,
           "output": 10,
-          "cache_read": 0.2,
+          "cache_read": 0.1,
           "cache_write": 2.5
         },
         "limit": {
@@ -314158,23 +314158,59 @@
     "name": "NaN",
     "doc": "https://nan.builders/docs/models",
     "models": {
-      "qwen3.6": {
-        "id": "qwen3.6",
-        "name": "Qwen3.6 35B-A3B",
-        "description": "Open multimodal Qwen MoE for local agents that need vision, audio, and code",
+      "qwen-image-2.1": {
+        "id": "qwen-image-2.1",
+        "name": "Qwen-Image-2.1",
+        "description": "Unified text-to-image generation and image editing model with native RGBA transparency and reference-image editing",
         "family": "qwen",
+        "attachment": false,
+        "reasoning": false,
+        "tool_call": false,
+        "temperature": true,
+        "release_date": "2026-09-14",
+        "last_updated": "2026-09-21",
+        "modalities": {
+          "input": [
+            "text"
+          ],
+          "output": [
+            "image"
+          ]
+        },
+        "open_weights": true,
+        "limit": {
+          "context": 8192,
+          "output": 0
+        },
+        "cost": {
+          "input": 0,
+          "output": 0
+        },
+        "canonical_model_id": "alibaba/qwen-image-2.1"
+      },
+      "glm5.3-flash": {
+        "id": "glm5.3-flash",
+        "name": "GLM-5.3-Flash",
+        "description": "Native multimodal GLM model for efficient coding and long-horizon agent tasks",
+        "family": "glm-flash",
         "attachment": true,
         "reasoning": true,
         "reasoning_options": [
           {
-            "type": "toggle"
+            "type": "effort",
+            "values": [
+              "low",
+              "medium",
+              "high",
+              "max"
+            ]
           }
         ],
         "tool_call": true,
         "structured_output": true,
         "temperature": true,
-        "release_date": "2026-04-17",
-        "last_updated": "2026-04-17",
+        "release_date": "2026-08-26",
+        "last_updated": "2026-08-26",
         "modalities": {
           "input": [
             "text",
@@ -314186,31 +314222,43 @@
         },
         "open_weights": true,
         "limit": {
-          "context": 262144,
-          "output": 65536
+          "context": 1000000,
+          "output": 131072
         },
         "cost": {
           "input": 0,
           "output": 0
-        }
+        },
+        "canonical_model_id": "zhipuai/glm-5.3-flash"
       },
-      "deepseek-v4-flash": {
-        "id": "deepseek-v4-flash",
-        "name": "DeepSeek V4 Flash",
-        "description": "Fast DeepSeek V4 lane for economical reasoning, coding, and long-context work",
-        "family": "deepseek-flash",
-        "attachment": false,
+      "gemma4": {
+        "id": "gemma4",
+        "name": "Gemma 4 26B A4B IT",
+        "description": "Open Gemma instruction model for efficient chat and self-hosted deployments",
+        "family": "gemma",
+        "attachment": true,
         "reasoning": true,
-        "reasoning_options": [],
+        "reasoning_options": [
+          {
+            "type": "effort",
+            "values": [
+              "none",
+              "low",
+              "medium",
+              "high",
+              "max"
+            ]
+          }
+        ],
         "tool_call": true,
         "structured_output": true,
         "temperature": true,
-        "knowledge": "2025-05",
-        "release_date": "2026-04-24",
-        "last_updated": "2026-04-24",
+        "release_date": "2026-04-02",
+        "last_updated": "2026-04-02",
         "modalities": {
           "input": [
-            "text"
+            "text",
+            "image"
           ],
           "output": [
             "text"
@@ -314218,30 +314266,43 @@
         },
         "open_weights": true,
         "limit": {
-          "context": 1000000,
-          "output": 384000
+          "context": 262144,
+          "output": 32768
         },
         "cost": {
           "input": 0,
           "output": 0
-        }
+        },
+        "canonical_model_id": "google/gemma-4-26b-a4b-it"
       },
-      "glm5.2": {
-        "id": "glm5.2",
-        "name": "GLM-5.2",
-        "description": "Open flagship GLM for long-horizon coding agents, served at 500K context on the NaN premium tier",
-        "family": "glm",
-        "attachment": false,
+      "qwen3.6": {
+        "id": "qwen3.6",
+        "name": "Qwen3.6 35B-A3B",
+        "description": "Open multimodal Qwen MoE for local agents that need vision, audio, and code",
+        "family": "qwen",
+        "attachment": true,
         "reasoning": true,
-        "reasoning_options": [],
+        "reasoning_options": [
+          {
+            "type": "effort",
+            "values": [
+              "none",
+              "low",
+              "medium",
+              "high",
+              "max"
+            ]
+          }
+        ],
         "tool_call": true,
         "structured_output": true,
         "temperature": true,
-        "release_date": "2026-06-13",
-        "last_updated": "2026-06-13",
+        "release_date": "2026-04-17",
+        "last_updated": "2026-04-17",
         "modalities": {
           "input": [
-            "text"
+            "text",
+            "image"
           ],
           "output": [
             "text"
@@ -314249,35 +314310,32 @@
         },
         "open_weights": true,
         "limit": {
-          "context": 500000,
-          "output": 131072
+          "context": 262144,
+          "output": 65536
         },
         "cost": {
           "input": 0,
           "output": 0
-        }
+        },
+        "canonical_model_id": "alibaba/qwen3.6-35b-a3b"
       },
-      "gemma4": {
-        "id": "gemma4",
-        "name": "Gemma 4 26B A4B IT",
-        "description": "Open Gemma instruction model for efficient chat and self-hosted deployments",
-        "family": "gemma",
+      "mimo-v2.6-flash": {
+        "id": "mimo-v2.6-flash",
+        "name": "MiMo-V2.6-Flash",
+        "description": "MiMo Flash model for multimodal coding agents and long-context automation",
+        "family": "mimo",
         "attachment": true,
         "reasoning": true,
-        "reasoning_options": [
-          {
-            "type": "toggle"
-          }
-        ],
+        "reasoning_options": [],
         "tool_call": true,
-        "structured_output": true,
         "temperature": true,
-        "release_date": "2026-04-02",
-        "last_updated": "2026-04-02",
+        "release_date": "2026-09-22",
+        "last_updated": "2026-09-22",
         "modalities": {
           "input": [
             "text",
-            "image"
+            "image",
+            "audio"
           ],
           "output": [
             "text"
@@ -314285,13 +314343,14 @@
         },
         "open_weights": true,
         "limit": {
-          "context": 262144,
-          "output": 32768
+          "context": 1048576,
+          "output": 131072
         },
         "cost": {
           "input": 0,
           "output": 0
-        }
+        },
+        "canonical_model_id": "xiaomi/mimo-v2.6-flash"
       },
       "qwen3.8-flash": {
         "id": "qwen3.8-flash",
@@ -314322,21 +314381,65 @@
         "cost": {
           "input": 0,
           "output": 0
-        }
+        },
+        "canonical_model_id": "alibaba/qwen3.8-flash"
       },
-      "glm5.3-flash": {
-        "id": "glm5.3-flash",
-        "name": "GLM-5.3-Flash",
-        "description": "Native multimodal GLM model for efficient coding and long-horizon agent tasks",
+      "glm5.3": {
+        "id": "glm5.3",
+        "name": "GLM-5.3",
+        "description": "Flagship GLM model for long-horizon coding, agents, and complex project delivery",
         "family": "glm",
+        "attachment": false,
+        "reasoning": true,
+        "reasoning_options": [
+          {
+            "type": "effort",
+            "values": [
+              "low",
+              "medium",
+              "high",
+              "max"
+            ]
+          }
+        ],
+        "tool_call": true,
+        "structured_output": true,
+        "temperature": true,
+        "release_date": "2026-08-14",
+        "last_updated": "2026-08-14",
+        "modalities": {
+          "input": [
+            "text"
+          ],
+          "output": [
+            "text"
+          ]
+        },
+        "open_weights": true,
+        "limit": {
+          "context": 1000000,
+          "output": 131072
+        },
+        "cost": {
+          "input": 0,
+          "output": 0
+        },
+        "canonical_model_id": "zhipuai/glm-5.3"
+      },
+      "deepseek-v4-flash": {
+        "id": "deepseek-v4-flash",
+        "name": "DeepSeek V4.1 Flash",
+        "description": "DeepSeek V4.1 Flash model for reasoning and agentic coding",
+        "family": "deepseek-flash",
         "attachment": true,
         "reasoning": true,
         "reasoning_options": [],
         "tool_call": true,
         "structured_output": true,
         "temperature": true,
-        "release_date": "2026-08-26",
-        "last_updated": "2026-08-26",
+        "knowledge": "2025-05",
+        "release_date": "2026-09-10",
+        "last_updated": "2026-09-10",
         "modalities": {
           "input": [
             "text",
@@ -314346,9 +314449,41 @@
             "text"
           ]
         },
-        "open_weights": false,
+        "open_weights": true,
         "limit": {
           "context": 1000000,
+          "output": 384000
+        },
+        "cost": {
+          "input": 0,
+          "output": 0
+        },
+        "canonical_model_id": "deepseek/deepseek-v4.1-flash"
+      },
+      "glm5.2": {
+        "id": "glm5.2",
+        "name": "GLM-5.2",
+        "description": "Open flagship GLM for long-horizon coding agents, served at 500K context on the NaN premium tier",
+        "family": "glm",
+        "attachment": false,
+        "reasoning": true,
+        "reasoning_options": [],
+        "tool_call": true,
+        "structured_output": true,
+        "temperature": true,
+        "release_date": "2026-06-13",
+        "last_updated": "2026-06-13",
+        "modalities": {
+          "input": [
+            "text"
+          ],
+          "output": [
+            "text"
+          ]
+        },
+        "open_weights": true,
+        "limit": {
+          "context": 500000,
           "output": 131072
         },
         "cost": {

package/src/be/modelsdev-cache.ts

@@ -1,5 +1,6 @@
 import { readFileSync } from "node:fs";
 import path from "node:path";
+import { withRegistryCatalogSections } from "@desplega/model-catalog";
 
 export interface ModelsDevCostBlock {
   input?: number;
@@ -53,6 +54,11 @@ const parsedCaches = new Map<string, WeakRef<ModelsDevCache>>();
  * fetches the live catalog from `GET /api/models-catalog`
  * (`src/be/models-catalog.ts`) and only falls back to its bundled copy of
  * this snapshot when that request hasn't resolved.
+ *
+ * The result also carries the sections of the OpenAI-compatible registry
+ * providers models.dev does not list (`helmcode`), synthesized from the
+ * registry (`withRegistryCatalogSections`), so every snapshot consumer
+ * (catalog, pricing seed, tier resolution) sees them like any other section.
  */
 export function loadModelsDevCache(): ModelsDevCache | null {
   const explicitPath = process.env.MODELSDEV_CACHE_PATH;
@@ -67,7 +73,9 @@ export function loadModelsDevCache(): ModelsDevCache | null {
     const memoized = parsedCaches.get(candidate)?.deref();
     if (memoized) return memoized;
     try {
-      const parsed = JSON.parse(readFileSync(candidate, "utf-8")) as ModelsDevCache;
+      const parsed = withRegistryCatalogSections(
+        JSON.parse(readFileSync(candidate, "utf-8")) as ModelsDevCache,
+      );
       parsedCaches.set(candidate, new WeakRef(parsed));
       return parsed;
     } catch {