@adhdev/daemon-core 0.9.82-rc.43 → 0.9.82-rc.430

package/dist/mesh/mesh-warmup-deadline.d.ts

@@ -0,0 +1,68 @@
+/**
+ * Await `work` under a warmup-aware deadline so a cold-open DataChannel handshake
+ * is NOT charged against the command response budget — the root cause of the
+ * "first mesh dispatch to a cold peer false-times-out, the warm retry succeeds"
+ * signature. Two budgets, switched by the live peer connection state:
+ *
+ *  - While `isConnected()` returns false the peer's channel is still opening; the
+ *    cold-open `connectTimeoutMs` budget applies. This phase is deliberately
+ *    generous because a TURN-relayed cross-machine handshake legitimately needs
+ *    many seconds — but a genuine connect *failure* is surfaced by `work`
+ *    rejecting on its own (the mesh manager fails the peer the instant its
+ *    PeerConnection state goes terminal), so a real failure is never masked for
+ *    the whole window.
+ *  - The first time `isConnected()` returns true the channel is warm; from that
+ *    instant the tight `responseTimeoutMs` governs how long the handler may take.
+ *    Warm-channel callers therefore see behavior identical to the old single
+ *    `Promise.race(work, responseTimeoutMs)`.
+ *
+ * Rejects with `Error('timeout')` when either budget is exhausted, mirroring the
+ * previous single-race contract. When no connection getter is wired callers must
+ * NOT pass `() => true` ("always warm") — that re-introduces the cold-open
+ * false-timeout. Use {@link resolveWarmupDeadlineOpts} which degrades conservatively.
+ */
+export declare function awaitWithWarmupDeadline<T>(work: Promise<T>, opts: {
+    isConnected: () => boolean;
+    connectTimeoutMs: number;
+    responseTimeoutMs: number;
+    pollIntervalMs?: number;
+}): Promise<T>;
+/** Minimal connection-state reader: a mesh peer snapshot stamps its live state on `.state`. */
+export declare function readWarmupConnectionState(connection: Record<string, unknown> | null | undefined): string | undefined;
+export interface ResolvedWarmupDeadlineOpts {
+    isConnected: () => boolean;
+    connectTimeoutMs: number;
+    responseTimeoutMs: number;
+}
+/**
+ * Build {@link awaitWithWarmupDeadline} opts from an OPTIONAL live peer-connection
+ * probe, handling the missing-getter case fail-loud instead of silently degrading.
+ *
+ * When `getConnection` is wired the normal cold-open/warm split applies: the probe
+ * is consulted live and the channel is "warm" only once it reports `connected`.
+ *
+ * When `getConnection` is ABSENT the old call sites fell back to `() => true`
+ * ("always warm"), which charges a still-opening cold channel against the response
+ * budget and silently re-introduces the exact cold-open false-timeout the warmup
+ * deadline exists to prevent. Instead we degrade CONSERVATIVELY and FAIL LOUD:
+ *   - `onMissingGetter` is invoked so the caller can warn (the degrade is visible,
+ *     never silent) — keyed/throttled by the caller as it sees fit.
+ *   - the channel is treated as NOT observably warm (`isConnected: () => false`),
+ *     so the response deadline never arms early on an unobservable channel.
+ *   - the cold peer is granted the COMBINED connect+response window as one deadline,
+ *     so a slow-but-live cold open is never false-timed at the shorter response
+ *     budget. A genuinely hung dispatch still rejects when the combined window
+ *     lapses, and `work` rejecting on its own (a real transport failure) still
+ *     settles immediately.
+ *
+ * Note: a present getter that returns a non-`connected` snapshot (or `null` for an
+ * unknown peer) correctly yields `isConnected() === false` — i.e. the generous
+ * connect budget, never "always warm". Only a wholly absent getter degrades.
+ */
+export declare function resolveWarmupDeadlineOpts(opts: {
+    getConnection?: ((daemonId: string) => Record<string, unknown> | null) | undefined;
+    daemonId: string;
+    connectTimeoutMs: number;
+    responseTimeoutMs: number;
+    onMissingGetter?: (daemonId: string) => void;
+}): ResolvedWarmupDeadlineOpts;

package/dist/mesh/mesh-work-queue.d.ts

@@ -1,21 +1,93 @@
+import type { RepoMeshDaemonRole } from '../repo-mesh-types.js';
+import { MeshRuntimeStore } from './mesh-runtime-store.js';
 export type MeshTaskStatus = 'pending' | 'assigned' | 'completed' | 'failed' | 'cancelled';
 export type MeshActiveTaskStatus = Extract<MeshTaskStatus, 'pending' | 'assigned'>;
 export type MeshHistoricalTaskStatus = Extract<MeshTaskStatus, 'completed' | 'failed' | 'cancelled'>;
+export type MeshTaskMode = 'code_change' | 'validation' | 'live_debug_readonly' | 'launch_app' | 'convergence';
 export declare const ACTIVE_MESH_QUEUE_STATUSES: MeshActiveTaskStatus[];
 export declare const HISTORICAL_MESH_QUEUE_STATUSES: MeshHistoricalTaskStatus[];
+export declare const MESH_TASK_MODES: MeshTaskMode[];
+/**
+ * QUEUE-NODE-SERIALIZATION: single source of truth for "is this task read-only?".
+ *
+ * Read-only classification used to be inlined as `task.taskMode === 'live_debug_readonly'`
+ * at every enforcement site (node-conflict claim gate, auto-launch isolation, the
+ * write/readonly cap counters, the write guardrail). That spread-out comparison is the
+ * exact recurring-defect class — one site drifting from the others silently makes the same
+ * task read-only at some gates and write at others, i.e. partial serialization. All sites
+ * MUST call this predicate so the classification is decided in exactly one place.
+ *
+ * Two orthogonal inputs feed the same boolean axis (kept backward-compatible):
+ *   • `readonly === true` — the explicit boolean axis (new API surface).
+ *   • `taskMode === 'live_debug_readonly'` — the original enum value, preserved as an
+ *     OR-fallback so existing live_debug_readonly tasks keep behaving identically.
+ *
+ * Accepts any task-like shape (full {@link MeshWorkQueueEntry} or a bare
+ * `{ readonly?, taskMode? }`) so the daemon-core and mcp-server boundaries can share it.
+ */
+export declare function isTaskReadonly(task: {
+    readonly?: boolean;
+    taskMode?: MeshTaskMode | string;
+} | null | undefined): boolean;
+export interface MeshTaskModeValidationResult {
+    valid: boolean;
+    taskMode?: MeshTaskMode;
+    violations: string[];
+    allowedOperations?: string[];
+}
+export declare function normalizeMeshTaskMode(value: unknown): MeshTaskMode | undefined;
+export declare function validateMeshTaskModeRequest(mode: unknown, message: string, readonly?: boolean): MeshTaskModeValidationResult;
 export interface MeshWorkQueueEntry {
     id: string;
     meshId: string;
     message: string;
     status: MeshTaskStatus;
+    taskMode?: MeshTaskMode;
+    /**
+     * QUEUE-NODE-SERIALIZATION: explicit read-only axis, orthogonal to taskMode. When
+     * true the task is treated as read-only by every scheduling gate (no node-busy
+     * isolation, counted under the read-only cap, write commands rejected) regardless of
+     * its taskMode. Decided exclusively through {@link isTaskReadonly}; `taskMode ===
+     * 'live_debug_readonly'` remains an OR-fallback so legacy rows behave unchanged.
+     */
+    readonly?: boolean;
     /** If specified, only this node can claim the task (used by legacy mesh_send_task) */
     targetNodeId?: string;
     /** If specified, only this runtime session can claim the task */
     targetSessionId?: string;
+    /** If specified, a node must expose all tags before it can claim the task. */
+    requiredTags?: string[];
+    /**
+     * M1: ids of tasks that must reach 'completed' before this task is claimable.
+     * Forward references (ids not yet enqueued) are allowed for batch flows and
+     * simply keep the task waiting until the referenced task exists and completes.
+     */
+    dependsOn?: string[];
+    /** M1/M3: mission this task belongs to (joins mesh_missions). */
+    missionId?: string;
+    /**
+     * MAGI: consensus group id shared by every replica of one mesh_magi_review
+     * fan-out. Marks the task as part of an INTENTIONAL same-prompt quorum so the
+     * completion-event dedup (mesh-events-pending) never collapses grouped
+     * replicas. Absent on ordinary tasks. Rides in the payload JSON (no column).
+     */
+    consensusGroupId?: string;
+    /**
+     * M1: why this task is held back (e.g. "dependency_failed:<taskId>").
+     * Only set by the system on dependency failure under the 'block' policy;
+     * waiting-on-dependency state is computed at view time, not stored.
+     */
+    blockedReason?: string;
     /** The node that actually claimed and is executing the task */
     assignedNodeId?: string;
     /** The session currently executing the task */
     assignedSessionId?: string;
+    /**
+     * Provider type of the session that claimed the task. Recorded so the queue
+     * can enforce per-(node, provider) maxParallel caps (RepoMeshNodePolicy
+     * providerRoles) by counting active assignments grouped by node + provider.
+     */
+    assignedProviderType?: string;
     /** Human/operator reason for terminal cancellation. */
     cancelReason?: string;
     cancelledAt?: string;
@@ -23,6 +95,16 @@ export interface MeshWorkQueueEntry {
     requeueReason?: string;
     requeuedAt?: string;
     requeueCount?: number;
+    /** Max automatic requeue attempts. When requeueCount reaches this, task is auto-failed. */
+    maxRetries?: number;
+    /**
+     * Bug B: number of times the reconcile assigned-stranded watchdog has reclaimed this
+     * row from 'assigned' back to 'pending' because its dispatch was never confirmed
+     * delivered. Separate from requeueCount (operator/execution retries) and bounded by
+     * MAX_STRANDED_RECLAIMS so a permanently-undeliverable target auto-fails rather than
+     * cycling reclaim→re-dispatch→strand forever.
+     */
+    strandedReclaimCount?: number;
     /** Last automatic queue session spin-up attempt, for mesh_view_queue/debug visibility. */
     autoLaunch?: {
         status: 'skipped' | 'started' | 'failed' | 'completed';
@@ -34,55 +116,213 @@ export interface MeshWorkQueueEntry {
     };
     /** ISO timestamp when the task was dispatched (assigned) to a node/session. Used for precise matching on completion. */
     dispatchTimestamp?: string;
+    /**
+     * (3) The ORIGINATING coordinator session that enqueued this task. Stamped onto the
+     * worker at dispatch (meshCoordinatorSessionId) so the task's completion routes back to
+     * the exact coordinator session — even when several coordinator sessions share one
+     * daemon. Rides in the queue payload JSON (no column migration); absent on legacy rows
+     * → daemon-level routing fallback (backward + version-skew safe).
+     */
+    sourceCoordinatorSessionId?: string;
     createdAt: string;
     updatedAt: string;
 }
+export interface MeshQueueMutationOptions {
+    ownerRole?: RepoMeshDaemonRole;
+}
+export declare function normalizeMeshCapabilityTags(value: unknown): string[];
+export declare function buildMeshNodeCapabilityTags(node: {
+    capabilities?: unknown;
+    policy?: unknown;
+    isLocalWorktree?: unknown;
+    worktreeBranch?: unknown;
+    userOverrides?: unknown;
+    reportedPlatform?: unknown;
+    reportedArch?: unknown;
+} | undefined, providerType?: string): string[];
+export declare function nodeSatisfiesRequiredTags(requiredTags: unknown, capabilityTags: unknown): boolean;
+/**
+ * Convergence-aware required-tags resolution (load-balancing scheduler, opt-in).
+ *
+ * When the mesh enables policy.autoConvergeCodeChange, a `converge=refine` required
+ * tag is merged into a code_change task's required tags at enqueue time, so the
+ * scheduler hard-filters the task onto refine-capable worktree nodes only (on any
+ * machine — refine_mesh_node forwards to the owning daemon). Because the tag is
+ * persisted on the queue entry, BOTH the eligibility scan (maybeAutoLaunchOneQueueSession)
+ * and the claim transaction (claimNextQueueTask → nodeSatisfiesRequiredTags) enforce
+ * it consistently.
+ *
+ * Strict backward compatibility — the injection is skipped (returns the explicit tags
+ * unchanged) when ANY of:
+ *   - the mesh does not opt in (autoConvergeCodeChange !== true), or
+ *   - the task is not code_change (validation / live_debug_readonly / launch_app /
+ *     convergence carry no merge cost and may run anywhere), or
+ *   - the task is explicitly targeted (targetNodeId): the operator chose the node, so
+ *     we do not second-guess it by filtering on convergence capability.
+ * Idempotent: normalizeMeshCapabilityTags dedupes, so re-injection is a no-op.
+ */
+export declare function resolveConvergeRequiredTags(meshId: string, taskMode: MeshTaskMode | undefined, explicitRequiredTags: string[], opts?: {
+    targetNodeId?: string;
+}): string[];
+/**
+ * M1: detect dependency cycles before enqueue. Walks the dependency graph of
+ * existing queue entries plus the new task's edges. Fail-closed: a cycle
+ * rejects the enqueue entirely. Synchronous and bounded by queue size.
+ */
+export declare function assertNoDependencyCycle(meshId: string, newTaskId: string, dependsOn: string[]): void;
 /**
  * Add a new task to the mesh queue.
  */
 export declare function enqueueTask(meshId: string, message: string, opts?: {
     targetNodeId?: string;
     targetSessionId?: string;
-}): MeshWorkQueueEntry;
+    taskMode?: MeshTaskMode | string;
+    /** QUEUE-NODE-SERIALIZATION: explicit read-only axis (orthogonal to taskMode). */
+    readonly?: boolean;
+    requiredTags?: string[];
+    /** M1: tasks that must complete before this one is claimable. */
+    dependsOn?: string[];
+    /** M1/M3: mission this task belongs to. */
+    missionId?: string;
+    /** MAGI: consensus group id shared by every replica of a mesh_magi_review fan-out. */
+    consensusGroupId?: string;
+    /** Explicit task id for batch/template flows (M5). Random UUID when omitted. */
+    id?: string;
+    /** (3) Originating coordinator session id (for session-anchored completion routing). */
+    sourceCoordinatorSessionId?: string;
+} & MeshQueueMutationOptions): MeshWorkQueueEntry;
+/**
+ * Record a direct-dispatch task (mesh_send_task) as an already-assigned queue
+ * entry so it is attributable to a mission.
+ *
+ * Direct dispatch normally bypasses the queue entirely — the task lives only in
+ * the ledger + mesh_direct_dispatches table, neither of which carries a
+ * missionId, so {@link summarizeMissionTasks}/{@link computeMeshTaskStats}
+ * (which both scan the queue for `task.missionId`) count it as 0. When a
+ * mission is attached, we materialise the same queue entry shape an enqueued
+ * task would have, but pre-assigned to the dispatched node/session and stamped
+ * with the dispatch timestamp. The terminal event path (updateSessionTaskStatus
+ * → findAssignedBySession) then flips it to completed/failed exactly like a
+ * pulled task, so mission total + completed aggregates work with no extra wiring.
+ *
+ * Intentionally separate from {@link enqueueTask}: enqueue creates `pending`
+ * work for the queue to assign, whereas this records work already dispatched
+ * out-of-band. They share the missionId stamping rule and mode validation.
+ */
+export declare function recordDirectDispatchTask(meshId: string, message: string, opts: {
+    id: string;
+    missionId: string;
+    assignedNodeId?: string;
+    assignedSessionId?: string;
+    taskMode?: MeshTaskMode | string;
+    /** QUEUE-NODE-SERIALIZATION: explicit read-only axis (orthogonal to taskMode). */
+    readonly?: boolean;
+    dispatchedAt?: string;
+}): MeshWorkQueueEntry | null;
 /**
  * Get all tasks in the queue, optionally filtered by status.
  */
 export declare function getQueue(meshId: string, opts?: {
     status?: MeshTaskStatus[];
 }): MeshWorkQueueEntry[];
+export declare function getMeshQueueRevision(meshId: string): string;
 /**
  * Find the next pending task that this node is allowed to claim, and mark it as assigned.
+ *
+ * `opts.providerType` is stamped onto the claimed entry (assignedProviderType) so
+ * per-(node, provider) caps can be counted. `opts.providerMaxParallel`, when set,
+ * is the enforced per-(node, provider) cap from RepoMeshNodePolicy.providerRoles:
+ * a task is not assigned to this (node, provider) once it already has that many
+ * active assignments. This composes with the global/taskMode caps (stricter wins).
  */
-export declare function claimNextTask(meshId: string, nodeId: string, sessionId: string): MeshWorkQueueEntry | null;
+export declare function claimNextTask(meshId: string, nodeId: string, sessionId: string, capabilityTags?: string[], opts?: {
+    providerType?: string;
+    providerMaxParallel?: number;
+    nodeIsWorktree?: boolean;
+}): MeshWorkQueueEntry | null;
+export type DependencyFailurePolicy = 'block' | 'cancel';
 /**
  * Update the status of a specific task.
  * Used when a session completes, fails, or stalls.
  */
-export declare function updateTaskStatus(meshId: string, taskId: string, status: MeshTaskStatus): MeshWorkQueueEntry | null;
+export declare function updateTaskStatus(meshId: string, taskId: string, status: MeshTaskStatus, opts?: MeshQueueMutationOptions): MeshWorkQueueEntry | null;
 export declare function recordTaskAutoLaunch(meshId: string, taskId: string, autoLaunch: Omit<NonNullable<MeshWorkQueueEntry['autoLaunch']>, 'updatedAt'>): MeshWorkQueueEntry | null;
 /**
  * Mark a queue task as manually cancelled without deleting audit history.
  */
 export declare function cancelTask(meshId: string, taskId: string, opts?: {
     reason?: string;
-}): MeshWorkQueueEntry | null;
+} & MeshQueueMutationOptions): MeshWorkQueueEntry | null;
 /**
  * Return a queue task to pending for retry. By default, dead session targeting
  * and assigned ownership are cleared so stale assignments do not strand again.
  */
+export type RequeueResult = {
+    status: 'requeued';
+    entry: MeshWorkQueueEntry;
+} | {
+    status: 'failed_max_retries';
+    entry: MeshWorkQueueEntry;
+    maxRetries: number;
+    requeueCount: number;
+} | {
+    status: 'not_found';
+};
 export declare function requeueTask(meshId: string, taskId: string, opts?: {
     reason?: string;
     targetNodeId?: string;
     targetSessionId?: string;
     clearTargetNode?: boolean;
     clearTargetSession?: boolean;
-}): MeshWorkQueueEntry | null;
+    /**
+     * Override the retry cap for this call. Use only for explicit operator actions.
+     * If true, the task is requeued even when requeueCount >= maxRetries.
+     */
+    force?: boolean;
+    /** Per-task retry cap override. Falls back to mesh policy maxTaskRetries (default 1). */
+    maxRetries?: number;
+} & MeshQueueMutationOptions): MeshWorkQueueEntry | null;
+/**
+ * Bug B: reclaim a task stuck in 'assigned' because its dispatch was never confirmed.
+ *
+ * claimNextTask atomically marks a row 'assigned' BEFORE the fire-and-forget dispatch
+ * runs. If that dispatch neither rejects (→ no .catch requeue) nor is confirmed
+ * delivered — a relay that hangs without acking, or a confirm timer lost across a
+ * daemon restart — the row stays 'assigned' forever, contributing 0 pending so PHASE 3
+ * reconcile never re-examines it. This returns such a row to 'pending' and clears its
+ * dead assignment ownership (node / session / provider / dispatchTimestamp) — the same
+ * ownership-clear requeueTask applies — so PHASE 3 can re-dispatch it onto a fresh idle
+ * session.
+ *
+ * Guarded to 'assigned' rows only (a completion/cancel that already moved the row off
+ * 'assigned' must never be resurrected) and bounded by MAX_STRANDED_RECLAIMS (beyond
+ * which the task is failed so dependents unblock).
+ */
+export declare function reclaimStrandedAssignedTask(meshId: string, taskId: string, opts?: {
+    reason?: string;
+    ageMs?: number;
+} & MeshQueueMutationOptions): MeshWorkQueueEntry | null;
 /**
  * Update the status of the task currently assigned to a specific session.
  */
 export declare function updateSessionTaskStatus(meshId: string, sessionId: string, status: MeshTaskStatus, opts?: {
     occurredAt?: string;
+    taskId?: string;
 }): MeshWorkQueueEntry | null;
+/**
+ * M1-3: true when at least one pending task is waiting on the given task.
+ * Used by the completion event path to decide whether to wake the queue.
+ */
+export declare function hasPendingDependents(meshId: string, taskId: string): boolean;
+/**
+ * M1-4: view-time dependency state for a task — unmet dependency ids and
+ * whether the task is currently claimable from a dependency standpoint.
+ * Not stored (truth stays in task statuses).
+ */
+export declare function describeTaskDependencyState(entry: Pick<MeshWorkQueueEntry, 'dependsOn' | 'blockedReason'>, statusById: Map<string, MeshTaskStatus | string>): {
+    waitingOn: string[];
+    dependenciesSatisfied: boolean;
+};
 export interface MeshWorkQueueStats {
     total: number;
     active: number;
@@ -107,3 +347,48 @@ export interface MeshWorkQueueStats {
  * Return aggregate queue statistics for the given mesh.
  */
 export declare function getMeshQueueStats(meshId: string): MeshWorkQueueStats;
+export declare function __replaceMeshQueueForTests(meshId: string, queue: MeshWorkQueueEntry[]): void;
+export declare function __clearMeshQueueForTests(meshId: string): void;
+export declare function __clearDirectDispatchesForTests(meshId: string): void;
+export declare function __resetMeshRuntimeStoreForTests(): void;
+export type DirectDispatchRecord = ReturnType<MeshRuntimeStore['getActiveDirectDispatches']>[number];
+export declare function insertDirectDispatch(meshId: string, data: {
+    taskId: string;
+    nodeId?: string;
+    sessionId?: string;
+    providerType?: string;
+    message: string;
+    taskMode?: string;
+    via: string;
+    dispatchedToIdleSession?: boolean;
+    dispatchedAt: string;
+}): void;
+export declare function getActiveDirectDispatches(meshId: string): DirectDispatchRecord[];
+export declare function updateDirectDispatchStatus(meshId: string, sessionId: string, status: 'acked' | 'completed' | 'failed' | 'stale', taskId?: string): void;
+export declare function cleanupTerminalDirectDispatches(olderThanMs?: number): void;
+export declare function markStaleDirectDispatches(meshId: string, olderThanMs?: number): void;
+/**
+ * Delete specific direct dispatch rows by taskId. Returns the number of rows deleted.
+ * Used by the staleDirect prune path to evict orphaned/terminal dispatch records from the
+ * active staleDirect surface while leaving the append-only mesh ledger (audit history) intact.
+ */
+export declare function deleteDirectDispatchesByTaskId(meshId: string, taskIds: string[]): number;
+export type MeshToolCallRateResult = {
+    rateLimitExceeded: boolean;
+    callsInWindow: number;
+    advisory: string | null;
+};
+/**
+ * Record a coordinator tool call and return a rate-limit advisory when the
+ * call rate for that tool exceeds the allowed threshold.
+ *
+ * Defaults: 10-second sliding window, max 5 calls before advisory is raised.
+ * Returns { rateLimitExceeded: false } on any store error so callers are not blocked.
+ */
+export declare function recordMeshToolCall(opts: {
+    meshId: string;
+    tool: string;
+    sessionId?: string | null;
+    windowMs?: number;
+    maxCalls?: number;
+}): MeshToolCallRateResult;

package/dist/mesh/preview-freshness.d.ts

@@ -0,0 +1,18 @@
+export type PreviewFreshnessStatus = 'fresh' | 'stale' | 'unknown' | 'not_configured';
+export interface PreviewFreshness {
+    status: PreviewFreshnessStatus;
+    lastPreviewCommit: string | null;
+    currentMainCommit: string | null;
+    currentMainCommitSource: 'origin/main' | 'HEAD' | 'unknown';
+    recordPath: string;
+    lastDeployedAt?: string;
+    lastTarget?: string;
+    previewVersion?: string;
+    targets: Record<'npm' | 'server' | 'web', {
+        commit: string | null;
+        deployedAt?: string;
+        status: PreviewFreshnessStatus;
+    }>;
+    nextAction: string;
+}
+export declare function buildPreviewFreshness(repoRoot: string): PreviewFreshness;