@smithers-orchestrator/scheduler 0.23.0 → 0.24.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@smithers-orchestrator/scheduler",
-  "version": "0.23.0",
+  "version": "0.24.0",
   "description": "Pure decision engine: session, scheduler, and task state management for Smithers workflows",
   "type": "module",
   "sideEffects": false,
@@ -176,8 +176,8 @@
   ],
   "dependencies": {
     "effect": "^3.21.1",
-    "@smithers-orchestrator/errors": "0.23.0",
-    "@smithers-orchestrator/graph": "0.23.0"
+    "@smithers-orchestrator/errors": "0.24.0",
+    "@smithers-orchestrator/graph": "0.24.0"
   },
   "devDependencies": {
     "@types/bun": "latest",

package/src/makeWorkflowSession.js

@@ -4,6 +4,7 @@ import { toSmithersError } from "@smithers-orchestrator/errors/toSmithersError";
 import { buildPlanTree } from "./buildPlanTree.js";
 import { buildStateKey } from "./buildStateKey.js";
 import { cloneTaskStateMap } from "./cloneTaskStateMap.js";
+import { computeRetryDelayMs } from "./computeRetryDelayMs.js";
 import { parseStateKey } from "./parseStateKey.js";
 import { scheduleTasks } from "./scheduleTasks.js";
 /** @typedef {import("./ApprovalResolution.ts").ApprovalResolution} ApprovalResolution */
@@ -149,27 +150,6 @@ function parseDurationMs(value) {
             return amount;
     }
 }
-/**
- * @param {TaskDescriptor} descriptor
- * @param {number} failureCount
- * @returns {number}
- */
-function retryDelayMs(descriptor, failureCount) {
-    const policy = descriptor.retryPolicy;
-    if (!policy)
-        return 0;
-    const initial = policy.initialDelayMs ?? 0;
-    if (policy.backoff === "exponential") {
-        const multiplier = policy.multiplier ?? 2;
-        const computed = initial * Math.pow(multiplier, Math.max(0, failureCount - 1));
-        return Math.min(policy.maxDelayMs ?? computed, computed);
-    }
-    if (policy.backoff === "linear") {
-        const computed = initial * Math.max(1, failureCount);
-        return Math.min(policy.maxDelayMs ?? computed, computed);
-    }
-    return initial;
-}
 /**
  * @param {TaskDescriptor} descriptor
  * @param {unknown} error
@@ -199,6 +179,67 @@ function isRetryableFailure(descriptor, error) {
     }
     return true;
 }
+/**
+ * @param {unknown} error
+ * @returns {boolean}
+ */
+function isTransientSessionFailure(error) {
+    const normalized = toSmithersError(error);
+    const code = error && typeof error === "object" && typeof error.code === "string"
+        ? error.code
+        : normalized.code;
+    return code === "SESSION_ERROR" ||
+        code === "TASK_TIMEOUT" ||
+        code === "TASK_HEARTBEAT_TIMEOUT" ||
+        code === "TASK_ABORTED" ||
+        normalized.details?.failureRetryable === true;
+}
+/**
+ * Build a human-readable diagnostic for a dependency deadlock: pending tasks
+ * that can never run because their `dependsOn` edges point at tasks missing from
+ * the graph or themselves permanently blocked. The most common cause is a
+ * `deps`/`needs` mismatch — a `deps={{ key: ... }}` whose key is not the upstream
+ * task's id and was not remapped with `needs={{ key: '<id>' }}`, which the Task
+ * component (deriveDepNodeIds) turns into a dependency on a non-existent node id.
+ * @param {SessionState} state
+ * @returns {string}
+ */
+function describeDeadlock(state) {
+    const blocked = [];
+    let sawMissing = false;
+    for (const descriptor of state.descriptors.values()) {
+        const taskState = state.states.get(stateKeyFor(descriptor)) ?? "pending";
+        if (taskState !== "pending" && taskState !== "cancelled")
+            continue;
+        const unmet = [];
+        for (const depId of descriptor.dependsOn ?? []) {
+            const dep = state.descriptors.get(depId);
+            if (!dep) {
+                sawMissing = true;
+                unmet.push(`'${depId}' (no such task)`);
+            }
+            else {
+                const depState = state.states.get(stateKeyFor(dep)) ?? "pending";
+                unmet.push(`'${depId}' (${depState})`);
+            }
+        }
+        if (unmet.length > 0) {
+            blocked.push(`  - '${descriptor.nodeId}' is blocked on ${unmet.join(", ")}`);
+        }
+    }
+    const lines = [
+        "Workflow deadlocked: no task can run, and none is waiting on an approval, event, timer, or retry.",
+    ];
+    if (blocked.length > 0) {
+        lines.push("Pending tasks and their unsatisfied dependencies:", ...blocked);
+    }
+    if (sawMissing) {
+        lines.push("", "A dependency marked '(no such task)' references a node id that is not a mounted task. " +
+            "If it came from deps={{ <key>: ... }}, the key is treated as the upstream task's id unless you remap it: " +
+            "add needs={{ <key>: '<upstream task id>' }} (or rename the upstream task to match the key).");
+    }
+    return lines.join("\n");
+}
 /**
  * @param {unknown} error
  * @param {string} label
@@ -224,6 +265,7 @@ export function makeWorkflowSession(options = {}) {
         states: new Map(),
         outputs: new Map(),
         failures: new Map(),
+        failureDescriptors: new Map(),
         retryCounts: new Map(),
         retryWait: new Map(),
         approvals: new Set(),
@@ -231,6 +273,7 @@ export function makeWorkflowSession(options = {}) {
         schedule: null,
         cancelled: false,
         lastMountedSignature: null,
+        lastDeadlockSignature: null,
     };
     /**
    * @param {Pick<TaskOutput, "nodeId" | "iteration">} output
@@ -288,6 +331,7 @@ export function makeWorkflowSession(options = {}) {
                 state.retryWait.delete(key);
                 state.approvals.delete(key);
                 state.retryCounts.delete(key);
+                state.failureDescriptors.delete(key);
             }
         }
         for (const ralph of ralphs) {
@@ -317,6 +361,7 @@ export function makeWorkflowSession(options = {}) {
         state.states.set(key, "finished");
         state.outputs.set(key, output);
         state.retryWait.delete(key);
+        state.failureDescriptors.delete(key);
     }
     /**
    * @param {number} [iteration]
@@ -367,7 +412,7 @@ export function makeWorkflowSession(options = {}) {
         const canRetry = retryable &&
             (descriptor.retries === Infinity || failureCount <= descriptor.retries);
         if (canRetry) {
-            const delay = retryDelayMs(descriptor, failureCount);
+            const delay = computeRetryDelayMs(descriptor.retryPolicy, failureCount);
             state.states.set(key, "pending");
             if (delay > 0) {
                 state.retryWait.set(key, nowMs() + delay);
@@ -379,6 +424,7 @@ export function makeWorkflowSession(options = {}) {
         }
         state.states.set(key, "failed");
         state.failures.set(key, error);
+        state.failureDescriptors.set(key, descriptor);
         return decide();
     }
     /**
@@ -387,11 +433,15 @@ export function makeWorkflowSession(options = {}) {
     function unhandledFailureDecision(recoveryKeys = new Set()) {
         for (const [key, taskState] of state.states) {
             const parsed = parseStateKey(key);
-            const descriptor = findDescriptor(state, parsed.nodeId, parsed.iteration);
+            const descriptor = findDescriptor(state, parsed.nodeId, parsed.iteration) ??
+                state.failureDescriptors.get(key);
             if (taskState === "failed" && !descriptor?.continueOnFail) {
                 if (recoveryKeys.has(key)) {
                     continue;
                 }
+                if (descriptor?.agent && isTransientSessionFailure(state.failures.get(key))) {
+                    continue;
+                }
                 return {
                     _tag: "Failed",
                     error: new SmithersError("SESSION_ERROR", `Task failed: ${descriptor?.nodeId ?? key}`, { key }, state.failures.get(key)),
@@ -497,26 +547,17 @@ export function makeWorkflowSession(options = {}) {
         if (existingWait) {
             return { _tag: "Wait", reason: existingWait };
         }
-        if (schedule.pendingExists) {
-            if (schedule.nextRetryAtMs != null) {
-                return {
-                    _tag: "Wait",
-                    reason: {
-                        _tag: "RetryBackoff",
-                        waitMs: Math.max(0, schedule.nextRetryAtMs - nowMs()),
-                    },
-                };
-            }
-            return { _tag: "Wait", reason: { _tag: "ExternalTrigger" } };
-        }
-        if ([...state.states.values()].some((taskState) => taskState === "in-progress")) {
-            return { _tag: "Wait", reason: { _tag: "ExternalTrigger" } };
-        }
-        failure = unhandledFailureDecision(recoveryKeys);
-        if (failure) {
-            return failure;
-        }
-        if (schedule.readyRalphs.length > 0) {
+        if (schedule.readyRalphs.length > 0 && !unhandledFailureDecision(recoveryKeys)) {
+            // A ralph is ready only when every task in its own subtree is
+            // terminal, so pending or in-flight work elsewhere in the graph must
+            // not starve its next iteration (#267). Run-level continue-as-new
+            // handoffs stay quiescence-only: tearing down the run while sibling
+            // tasks are mid-flight is not safe, so those ralphs are deferred.
+            // An unhandled task failure keeps its precedence over further loop
+            // iterations (decide() already returns it at the top; this guard
+            // makes the ordering explicit).
+            const hasInProgress = [...state.states.values()].some((taskState) => taskState === "in-progress");
+            let advanced = false;
             for (const ralph of schedule.readyRalphs) {
                 const current = state.ralphState.get(ralph.id) ?? {
                     iteration: 0,
@@ -524,6 +565,7 @@ export function makeWorkflowSession(options = {}) {
                 };
                 if (ralph.until) {
                     state.ralphState.set(ralph.id, { ...current, done: true });
+                    advanced = true;
                     continue;
                 }
                 const nextIteration = current.iteration + 1;
@@ -535,13 +577,18 @@ export function makeWorkflowSession(options = {}) {
                         };
                     }
                     state.ralphState.set(ralph.id, { iteration: current.iteration, done: true });
+                    advanced = true;
                     continue;
                 }
-                state.ralphState.set(ralph.id, { iteration: nextIteration, done: false });
-                if (ralph.continueAsNewEvery != null &&
+                const wantsContinueAsNew = ralph.continueAsNewEvery != null &&
                     ralph.continueAsNewEvery > 0 &&
                     nextIteration > 0 &&
-                    nextIteration % ralph.continueAsNewEvery === 0) {
+                    nextIteration % ralph.continueAsNewEvery === 0;
+                if (wantsContinueAsNew && (hasInProgress || schedule.pendingExists)) {
+                    continue;
+                }
+                state.ralphState.set(ralph.id, { iteration: nextIteration, done: false });
+                if (wantsContinueAsNew) {
                     return {
                         _tag: "ContinueAsNew",
                         transition: {
@@ -551,8 +598,51 @@ export function makeWorkflowSession(options = {}) {
                         },
                     };
                 }
+                advanced = true;
+            }
+            if (advanced) {
+                return { _tag: "ReRender", context: renderContext(state) };
+            }
+        }
+        if (schedule.pendingExists) {
+            if (schedule.nextRetryAtMs != null) {
+                return {
+                    _tag: "Wait",
+                    reason: {
+                        _tag: "RetryBackoff",
+                        waitMs: Math.max(0, schedule.nextRetryAtMs - nowMs()),
+                    },
+                };
+            }
+            // Nothing is runnable, in flight, or waiting on an approval, event, or
+            // timer, yet tasks remain pending. They are blocked on dependencies
+            // nothing will ever satisfy — most often a deps/needs key that maps to
+            // a node id no task produces, which becomes a dependsOn on a missing
+            // node. Returning Wait here suspends the run forever with no error.
+            // Give a reactive re-render one chance to mount a producer (the mounted
+            // signature changes), then fail loudly with a diagnostic.
+            const noInProgress = ![...state.states.values()].some((taskState) => taskState === "in-progress");
+            if (noInProgress) {
+                if (options.requireStableFinish && state.graph) {
+                    const signature = mountedSignature(state.graph);
+                    if (state.lastDeadlockSignature !== signature) {
+                        state.lastDeadlockSignature = signature;
+                        return { _tag: "ReRender", context: renderContext(state) };
+                    }
+                }
+                return {
+                    _tag: "Failed",
+                    error: new SmithersError("DEPENDENCY_DEADLOCK", describeDeadlock(state)),
+                };
             }
-            return { _tag: "ReRender", context: renderContext(state) };
+            return { _tag: "Wait", reason: { _tag: "ExternalTrigger" } };
+        }
+        if ([...state.states.values()].some((taskState) => taskState === "in-progress")) {
+            return { _tag: "Wait", reason: { _tag: "ExternalTrigger" } };
+        }
+        failure = unhandledFailureDecision(recoveryKeys);
+        if (failure) {
+            return failure;
         }
         if (options.requireStableFinish && state.graph) {
             const signature = mountedSignature(state.graph);