@sema-agent/core 7.32.0 → 7.33.0

package/dist/orchestration/workflow.d.ts

@@ -588,8 +588,8 @@ export declare const MAX_WORKFLOW_ITEMS = 4096;
  * own. This watchdog does not re-send a model call — it aborts the attempt and restarts the whole agent
  * on a FRESH session, so tool effects from before the stall can be repeated. The engine, one layer down,
  * already answers a silent provider call the cheap way: it re-sends that ONE call. So the floor here is
- * what a single provider call may legitimately occupy while producing NOTHING — the connect wait plus the
- * first-token wait ({@link STALL_CONNECT_MS} + {@link STALL_FIRST_TOKEN_MS}). It was a hand-written
+ * what a single provider call may legitimately occupy while producing NOTHING — the whole first-byte lane at the default knob
+ * ({@link firstByteLaneKnobMs}) plus the first-token wait; a floor (upload allowance and backoff between attempts not counted). It was a hand-written
  * 180_000 and CC-same until the first-token bound rose to 600s past it (measured: the same silent call
  * then produced two agent runs on two sessions instead of one run with a re-sent call), which is the
  * systematic inversion a derived value cannot re-open: raise or lower the engine's bound and this follows.
@@ -599,8 +599,8 @@ export declare const MAX_WORKFLOW_ITEMS = 4096;
  * this timer starts before the Runner prepares the attempt (an execution-env factory, a hook, a store
  * read all run inside this window and outside the engine's), and the first-token timer only starts once
  * headers arrive — so the margin by which the inner deadline can land AFTER this one is preparation plus
- * the connect time actually spent, less the {@link STALL_CONNECT_MS} this window already budgets for it
- * (40s of preparation and instant headers puts the first inner deadline 10s past this timer). An engine retry does not re-arm this timer either, so two silent provider attempts
+ * the header wait actually spent, less the {@link STALL_FIRST_BYTE_MS} this window already budgets for it
+ * (instant headers and more than 300s of preparation put the first inner deadline past this timer). An engine retry does not re-arm this timer either, so two silent provider attempts
  * exceed it by construction. The derivation removes the case that was guaranteed to invert; it does not
  * make inversion impossible. A deployment that wants the old aggression passes `stallMs` explicitly, and
  * a deployment that wants a real guarantee must set it from its own preparation budget.

package/dist/orchestration/workflow.js

@@ -22,7 +22,7 @@ import { OUTPUT_TOOL_NAME } from "../core/runner/synthetic-tools.js";
 import { compileOutputSchema } from "../core/runner/strict-output-schema.js";
 import { WorkflowBudgetExceededError, WorkflowNestingError, WorkflowAgentBlockedError, WorkflowAgentParkedError, WorkflowAgentSchemaError, WorkflowAgentStalledError, WorkflowMaxAgentsError, WORKFLOW_SPAWN_BLOCKED_ERROR_CODE } from "./workflow-types.js";
 import { WorkflowScriptError } from "./workflow-meta.js";
-import { STALL_CONNECT_MS, STALL_FIRST_TOKEN_MS } from "../brain/timeout.js";
+import { firstByteLaneKnobMs, STALL_FIRST_BYTE_MS, STALL_FIRST_TOKEN_MS } from "../brain/timeout.js";
 import { asThrownError, isErrorLike, readMember, thrownIs, thrownName, thrownText } from "../core/thrown-value.js";
 export * from "./workflow-types.js";
 const MAX_TRANSCRIPT_CHARS = 4000;
@@ -310,7 +310,7 @@ function assertFanOutOptionsShape(v, position) {
             `Nothing was run.`);
     }
 }
-export const WORKFLOW_AGENT_STALL_MS = STALL_CONNECT_MS + STALL_FIRST_TOKEN_MS;
+export const WORKFLOW_AGENT_STALL_MS = firstByteLaneKnobMs(STALL_FIRST_BYTE_MS) + STALL_FIRST_TOKEN_MS;
 export const WORKFLOW_THROTTLE_SLOW_MS = 90_000;
 export const WORKFLOW_AGENT_MAX_RETRIES = 5;
 export const WORKFLOW_AGENT_THROTTLE_BACKOFF_MS = 45_000;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sema-agent/core",
-  "version": "7.32.0",
+  "version": "7.33.0",
   "description": "Stateless, task-oriented AI agent core",
   "type": "module",
   "license": "BUSL-1.1",
@@ -44,7 +44,7 @@
     "wire-changes/"
   ],
   "engines": {
-    "node": ">=20.3.0"
+    "node": ">=20.18.1"
   },
   "scripts": {
     "build": "rm -rf dist && tsc -p tsconfig.build.json && tsc -p tsconfig.build.json --emitDeclarationOnly --removeComments false",
@@ -95,7 +95,8 @@
   },
   "dependencies": {
     "@modelcontextprotocol/sdk": "1.30.0",
-    "typebox": "1.1.39"
+    "typebox": "1.1.39",
+    "undici": "7.30.0"
   },
   "optionalDependencies": {
     "sharp": "^0.35.0"
@@ -107,6 +108,8 @@
     "pg-mem": "^3.0.14",
     "tsx": "4.19.2",
     "typescript": "5.7.2",
+    "undici6": "npm:undici@^6.29.0",
+    "undici8": "npm:undici@^8.11.2",
     "vitest": "^2.1.8"
   }
 }

package/wire-changes/core@7.32.1.json

@@ -0,0 +1,78 @@
+{
+  "repo": "core",
+  "version": "7.32.1",
+  "from": "5aeef6b3eb1ec7dfc7aef1dd0c9f6a7aebfe3cfd",
+  "to": "v7.32.1",
+  "added": [],
+  "removed": [],
+  "changed": [],
+  "narrowed": [],
+  "closedSetMembers": [],
+  "closedSetsUnreadable": [],
+  "typeKeys": [],
+  "typeKeysUnreadableArms": [],
+  "typeKeysFlattenedCount": 217,
+  "declarationMoves": [],
+  "mintMoves": [],
+  "criteriaObserves": [
+    {
+      "kind": "taskEventField",
+      "name": "tool_end.output",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "BackgroundPoll.status",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "BackgroundShellHardKillReceipt.outcome",
+      "assertions": [
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskRegistry.list",
+      "assertions": [
+        "G1.a",
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskRegistry.reapSessionBackground",
+      "assertions": [
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskRegistry.runningShellIds",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskSpec.retainBackgroundProcesses",
+      "assertions": [
+        "G1.a"
+      ]
+    }
+  ],
+  "consumersHint": {
+    "cli": [],
+    "server": [],
+    "sdk": [],
+    "note": "scanned cli@4e593722 server@582dee2f sdk@9b5e65b — a name listed is MENTIONED there today, which is evidence of consumption, not proof of it"
+  },
+  "breakingCount": 0,
+  "breakingItems": [],
+  "toResolvedAt": "b09a04bd1f9e09fd7e28914d488f12f517dc8d4c"
+}

package/wire-changes/core@7.33.0.json

@@ -0,0 +1,351 @@
+{
+  "repo": "core",
+  "version": "7.33.0",
+  "from": "a36eb21ebe39f5191f30102623e1570c3eb77f41",
+  "to": "v7.33.0",
+  "added": [],
+  "removed": [],
+  "changed": [],
+  "narrowed": [
+    {
+      "id": "F4",
+      "source": "behaviorFlips.summary",
+      "text": "Behaviour change with a NARROWING (#1121; +0 exports): ONE resolution chain for every model-call watchdog (`resolveWatchdogMs`): construction config (an explicit `0` is OFF) > per-call advisory > the engine default constant (`STALL_FIRST_BYTE_MS` 300 000 + the upload allowance / `STALL_FIRST_TOKEN_MS` 600 000 / `STALL_IDLE_MS` 300 000, F5). The old third level \"both absent ⇒ off\" is retired: a first-party brain built with no timeout fields and driven without the Runner's advisory (a bare SDK brain; compaction and side-query calls) now arms all three — a call that never gets headers ends after 301 s + a widened 601 s retry, one with no delta after 600 s, silence after a delta after 300 s (the same 300 s the transport's body clock used to give such a call, now the engine's own watchdog announcing a `stall` instead of a transport failure; an Anthropic `ping` is not a content delta, so with text already streamed the call is committed by tier A, \"Response stalled mid-stream. The response above may be incomplete.\"). `0` on a watchdog now means no bound at all on that segment, the transport's included (F2). NARROWING: such a brain whose fetch keeps no clock of its own (a `fetchImpl` that is not undici's, a test double) waited indefinitely on every segment and is now cut at those defaults; on Node's fetch it was bounded by the transport's 300 s clocks — the header cut reported `transport` and re-sent through the shared ladder up to `maxRetries` times — and is now bounded by the engine's (headers 301 s + one widened 601 s retry; first token 600 s, wider; idle 300 s, the same bound under a new announcer). WHO IS HURT: a deployment that constructed a first-party brain with no timeout fields and relied on it having none. COMPENSATION: `0` on the field it wants off, or a larger value."
+    }
+  ],
+  "closedSetMembers": [],
+  "closedSetsUnreadable": [],
+  "typeKeys": [
+    {
+      "type": "AnthropicBrainConfig",
+      "at": [
+        "src/brain/anthropic.ts:25"
+      ],
+      "added": [
+        "firstByteTimeoutMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteTimeoutMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectTimeoutMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": true
+    },
+    {
+      "type": "BrainTimeoutConfig",
+      "at": [
+        "src/brain/timeout.ts:166"
+      ],
+      "added": [
+        "firstByteTimeoutMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteTimeoutMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectTimeoutMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": true
+    },
+    {
+      "type": "OpenAIBrainConfig",
+      "at": [
+        "src/brain/openai.ts:27"
+      ],
+      "added": [
+        "firstByteTimeoutMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteTimeoutMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectTimeoutMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": true
+    },
+    {
+      "type": "OpenResponsesBrainConfig",
+      "at": [
+        "src/brain/open-responses.ts:101"
+      ],
+      "added": [
+        "firstByteTimeoutMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteTimeoutMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectTimeoutMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": true
+    },
+    {
+      "type": "StallTimeouts",
+      "at": [
+        "src/engine/llm/types.ts:70"
+      ],
+      "via": "CompleteSimpleFn.SimpleStreamOptions.StreamOptions.stallTimeouts",
+      "added": [
+        "firstByteMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": false
+    },
+    {
+      "type": "StreamEngineConfig",
+      "at": [
+        "src/brain/stream-engine.ts:45"
+      ],
+      "via": "AnthropicBrainConfig.StreamEngineConfig",
+      "added": [
+        "firstByteTimeoutMs"
+      ],
+      "addedDetail": [
+        {
+          "key": "firstByteTimeoutMs",
+          "optional": true
+        }
+      ],
+      "addedRequired": [],
+      "removed": [
+        "connectTimeoutMs"
+      ],
+      "breaking": true,
+      "criteriaCovered": false
+    }
+  ],
+  "typeKeysUnreadableArms": [],
+  "typeKeysFlattenedCount": 217,
+  "declarationMoves": [],
+  "mintMoves": [],
+  "criteriaObserves": [
+    {
+      "kind": "taskEventField",
+      "name": "status.attempt",
+      "assertions": [
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "taskEventField",
+      "name": "status.detail",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "taskEventField",
+      "name": "status.errClass",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "taskEventField",
+      "name": "status.maxRetries",
+      "assertions": [
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "taskEventField",
+      "name": "status.timeoutMs",
+      "assertions": [
+        "G1.a",
+        "G2.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "AnthropicBrainConfig.firstByteTimeoutMs",
+      "assertions": [
+        "G3.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "AssistantMessage.errorMessage",
+      "assertions": [
+        "G3.a",
+        "G4.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "AssistantMessage.partialFinalized",
+      "assertions": [
+        "G4.c"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "AssistantMessage.stopReason",
+      "assertions": [
+        "G3.a",
+        "G4.a",
+        "G4.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "Brain.stream",
+      "assertions": [
+        "G3.a",
+        "G4.a",
+        "G4.c"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "BrainTimeoutConfig.firstByteTimeoutMs",
+      "assertions": [
+        "G1.a",
+        "G2.a",
+        "G3.a",
+        "G4.a",
+        "G4.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "BrainTimeoutConfig.firstTokenTimeoutMs",
+      "assertions": [
+        "G2.a",
+        "G4.a",
+        "G4.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "BrainTimeoutConfig.idleTimeoutMs",
+      "assertions": [
+        "G4.b",
+        "G4.c"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "OpenAIBrainConfig.fetchImpl",
+      "assertions": [
+        "G1.b",
+        "G2.a",
+        "G2.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "OpenAIBrainConfig.firstByteTimeoutMs",
+      "assertions": [
+        "G1.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "OpenAIBrainConfig.maxRetries",
+      "assertions": [
+        "G1.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "OpenResponsesBrainConfig.firstByteTimeoutMs",
+      "assertions": [
+        "G3.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "RunWorkflowOptions.stallMs",
+      "assertions": [
+        "G5.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "RunWorkflowOptions.timers",
+      "assertions": [
+        "G5.a"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskRegistry.list",
+      "assertions": [
+        "G3.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskRegistry.runningShellIds",
+      "assertions": [
+        "G3.b"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskResult.result",
+      "assertions": [
+        "G2.b",
+        "G4.c"
+      ]
+    },
+    {
+      "kind": "typeMember",
+      "name": "TaskResult.terminal",
+      "assertions": [
+        "G1.a",
+        "G1.b",
+        "G2.a",
+        "G2.b",
+        "G4.c"
+      ]
+    }
+  ],
+  "consumersHint": {
+    "cli": [],
+    "server": [],
+    "sdk": [],
+    "note": "scanned cli@4e593722 server@907e1c76 sdk@4fe310e — a name listed is MENTIONED there today, which is evidence of consumption, not proof of it. Ledger files (**/docs/core-consumption/**, **/docs/wire-consumption/**, **/*LEDGER*.json) are not evidence of use — server: 4 name(s) matched ONLY in bookkeeping files and were set aside (AnthropicBrainConfig, BrainTimeoutConfig, OpenAIBrainConfig, OpenResponsesBrainConfig)"
+  },
+  "breakingCount": 6,
+  "breakingItems": [
+    "type lost keys: AnthropicBrainConfig -connectTimeoutMs",
+    "type lost keys: BrainTimeoutConfig -connectTimeoutMs",
+    "type lost keys: OpenAIBrainConfig -connectTimeoutMs",
+    "type lost keys: OpenResponsesBrainConfig -connectTimeoutMs",
+    "type lost keys: StallTimeouts -connectMs",
+    "type lost keys: StreamEngineConfig -connectTimeoutMs"
+  ],
+  "toResolvedAt": "0ce5bb4b0d729b70f84bdd3ed2ea93bbdaedb68a"
+}