@vellumai/assistant 0.7.1 → 0.7.2

package/src/runtime/routes/memory-v2-routes.ts

@@ -8,6 +8,16 @@ import { z } from "zod";
 
 import { isAssistantFeatureFlagEnabled } from "../../config/assistant-feature-flags.js";
 import { loadConfig } from "../../config/loader.js";
+import {
+  applyCorrectionIfCalibrated,
+  explainedVarianceRatio,
+  fitAnisotropyCalibration,
+  saveCalibration,
+} from "../../memory/anisotropy.js";
+import {
+  embedWithBackend,
+  selectEmbeddingBackend,
+} from "../../memory/embedding-backend.js";
 import {
   enqueueMemoryJob,
   type MemoryJobType,
@@ -17,8 +27,22 @@ import {
   totalEdgeCount,
   validateEdgeTargets,
 } from "../../memory/v2/edge-index.js";
-import { listPages, readPage } from "../../memory/v2/page-store.js";
+import {
+  listPages,
+  readPage,
+  renderPageContent,
+} from "../../memory/v2/page-store.js";
+import {
+  hybridQueryConceptPages,
+  sampleConceptPageDenseVectors,
+} from "../../memory/v2/qdrant.js";
+import { effectiveWeights } from "../../memory/v2/sim.js";
 import { seedV2SkillEntries } from "../../memory/v2/skill-store.js";
+import {
+  generateBm25QueryEmbedding,
+  getConceptPageCorpusStats,
+  rebuildConceptPageCorpusStats,
+} from "../../memory/v2/sparse-bm25.js";
 import { getWorkspaceDir } from "../../util/platform.js";
 import { RouteError } from "./errors.js";
 import type { RouteDefinition } from "./types.js";
@@ -113,6 +137,79 @@ async function handleValidate({
   };
 }
 
+// ── Get concept page ────────────────────────────────────────────────────
+
+const MemoryV2GetConceptPageParams = z
+  .object({
+    slug: z.string().min(1),
+  })
+  .strict();
+
+export type MemoryV2GetConceptPageResult = {
+  slug: string;
+  /** Frontmatter + body, as produced by `renderPageContent`. */
+  rendered: string;
+};
+
+async function handleGetConceptPage({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2GetConceptPageResult> {
+  const { slug } = MemoryV2GetConceptPageParams.parse(body);
+  const workspaceDir = getWorkspaceDir();
+  let page;
+  try {
+    page = await readPage(workspaceDir, slug);
+  } catch (err) {
+    throw new RouteError(
+      `Failed to read concept page '${slug}': ${err instanceof Error ? err.message : String(err)}`,
+      "MEMORY_V2_PAGE_READ_FAILED",
+      400,
+    );
+  }
+  if (!page) {
+    throw new RouteError(
+      `Concept page '${slug}' not found on disk`,
+      "MEMORY_V2_PAGE_NOT_FOUND",
+      404,
+    );
+  }
+  return { slug, rendered: renderPageContent(page) };
+}
+
+// ── Rebuild BM25 corpus stats ───────────────────────────────────────────
+
+const MemoryV2RebuildCorpusStatsParams = z.object({}).strict();
+
+export interface MemoryV2RebuildCorpusStatsResult {
+  totalDocs: number;
+  avgDl: number;
+  /** Number of distinct hashed-token buckets that received any DF count. */
+  vocabularyBuckets: number;
+}
+
+async function handleRebuildCorpusStats({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2RebuildCorpusStatsResult> {
+  MemoryV2RebuildCorpusStatsParams.parse(body);
+  const workspaceDir = getWorkspaceDir();
+  await rebuildConceptPageCorpusStats(workspaceDir);
+  const stats = getConceptPageCorpusStats();
+  if (!stats) {
+    // The rebuild always swaps in a non-null table on success, so a missing
+    // value here means an unexpected reset between rebuild and read.
+    throw new RouteError(
+      "Corpus stats rebuild completed but no table is loaded",
+      "MEMORY_V2_CORPUS_STATS_MISSING",
+      500,
+    );
+  }
+  return {
+    totalDocs: stats.totalDocs,
+    avgDl: stats.avgDl,
+    vocabularyBuckets: stats.df.size,
+  };
+}
+
 // ── Reembed skills ──────────────────────────────────────────────────────
 
 const MemoryV2ReembedSkillsParams = z.object({}).strict();
@@ -149,6 +246,305 @@ async function handleReembedSkills({
   return { success: true };
 }
 
+// ── Explain similarity ──────────────────────────────────────────────────
+
+const MemoryV2ExplainSimilarityParams = z
+  .object({
+    userText: z.string().min(1),
+    assistantText: z.string().optional(),
+    nowText: z.string().optional(),
+    top: z.number().int().min(1).default(25),
+  })
+  .strict();
+
+export interface MemoryV2ExplainSimilarityRow {
+  slug: string;
+  /** Raw dense cosine score, or null when the slug missed the dense channel. */
+  denseScore: number | null;
+  /** Raw sparse score (Qdrant scale), or null when the slug missed sparse. */
+  sparseRaw: number | null;
+  /** Sparse score divided by the per-batch max, in [0, 1]. */
+  sparseNorm: number | null;
+  /** `clamp01(dense_weight·dense + sparse_weight·sparseNorm)` — the simBatch fused value. */
+  fused: number;
+}
+
+export interface MemoryV2ExplainSimilarityStats {
+  count: number;
+  min: number;
+  max: number;
+  mean: number;
+  stddev: number;
+}
+
+export interface MemoryV2ExplainSimilarityChannel {
+  channel: "user" | "assistant" | "now";
+  textPreview: string;
+  maxSparse: number;
+  /**
+   * Spread (max - min) of normalized sparse scores across this channel's
+   * hits. Drives adaptive sparse weighting — low spread means the sparse
+   * channel can't discriminate, so its weight collapses for this query.
+   */
+  sparseSpread: number;
+  /** Sparse weight after adaptive collapse (≤ the configured base). */
+  effectiveSparseWeight: number;
+  /** Dense weight after adaptive compensation (≥ the configured base). */
+  effectiveDenseWeight: number;
+  rows: MemoryV2ExplainSimilarityRow[];
+  stats: {
+    dense: MemoryV2ExplainSimilarityStats;
+    sparseRaw: MemoryV2ExplainSimilarityStats;
+    sparseNorm: MemoryV2ExplainSimilarityStats;
+    fused: MemoryV2ExplainSimilarityStats;
+  };
+}
+
+export interface MemoryV2ExplainSimilarityResult {
+  config: {
+    dense_weight: number;
+    sparse_weight: number;
+  };
+  channels: MemoryV2ExplainSimilarityChannel[];
+}
+
+function summarizeStats(values: number[]): MemoryV2ExplainSimilarityStats {
+  if (values.length === 0) {
+    return { count: 0, min: 0, max: 0, mean: 0, stddev: 0 };
+  }
+  let min = Infinity;
+  let max = -Infinity;
+  let sum = 0;
+  for (const v of values) {
+    if (v < min) min = v;
+    if (v > max) max = v;
+    sum += v;
+  }
+  const mean = sum / values.length;
+  let sqDiff = 0;
+  for (const v of values) sqDiff += (v - mean) * (v - mean);
+  const stddev = Math.sqrt(sqDiff / values.length);
+  return { count: values.length, min, max, mean, stddev };
+}
+
+async function scoreChannel(
+  channel: "user" | "assistant" | "now",
+  text: string,
+  top: number,
+  denseWeight: number,
+  sparseWeight: number,
+  config: ReturnType<typeof loadConfig>,
+): Promise<MemoryV2ExplainSimilarityChannel> {
+  const denseResult = await embedWithBackend(config, [text]);
+  const denseVec = await applyCorrectionIfCalibrated(
+    denseResult.vectors[0],
+    denseResult.provider,
+    denseResult.model,
+  );
+  const sparseVec = generateBm25QueryEmbedding(text);
+
+  const hits = await hybridQueryConceptPages(denseVec, sparseVec, top);
+
+  let maxSparse = 0;
+  for (const hit of hits) {
+    if (hit.sparseScore !== undefined && hit.sparseScore > maxSparse) {
+      maxSparse = hit.sparseScore;
+    }
+  }
+
+  // Mirror simBatch's adaptive weighting so the printed `fused` matches what
+  // production retrieval would actually score for this query — otherwise
+  // operators staring at the diagnostic would see different numbers than
+  // the activation pipeline saw.
+  const {
+    dense: effDense,
+    sparse: effSparse,
+    spread: sparseSpread,
+  } = effectiveWeights(hits, maxSparse, denseWeight, sparseWeight, config);
+
+  const rows: MemoryV2ExplainSimilarityRow[] = hits.map((hit) => {
+    const dense = hit.denseScore ?? 0;
+    const sparseNorm =
+      hit.sparseScore !== undefined && maxSparse > 0
+        ? hit.sparseScore / maxSparse
+        : 0;
+    const fusedRaw = effDense * dense + effSparse * sparseNorm;
+    const fused = Math.max(0, Math.min(1, fusedRaw));
+    return {
+      slug: hit.slug,
+      denseScore: hit.denseScore ?? null,
+      sparseRaw: hit.sparseScore ?? null,
+      sparseNorm: hit.sparseScore !== undefined ? sparseNorm : null,
+      fused,
+    };
+  });
+
+  rows.sort((a, b) => b.fused - a.fused);
+
+  const denseValues: number[] = [];
+  const sparseRawValues: number[] = [];
+  const sparseNormValues: number[] = [];
+  const fusedValues: number[] = [];
+  for (const row of rows) {
+    if (row.denseScore !== null) denseValues.push(row.denseScore);
+    if (row.sparseRaw !== null) sparseRawValues.push(row.sparseRaw);
+    if (row.sparseNorm !== null) sparseNormValues.push(row.sparseNorm);
+    fusedValues.push(row.fused);
+  }
+
+  return {
+    channel,
+    textPreview: text.length > 120 ? `${text.slice(0, 120)}…` : text,
+    maxSparse,
+    sparseSpread,
+    effectiveSparseWeight: effSparse,
+    effectiveDenseWeight: effDense,
+    rows,
+    stats: {
+      dense: summarizeStats(denseValues),
+      sparseRaw: summarizeStats(sparseRawValues),
+      sparseNorm: summarizeStats(sparseNormValues),
+      fused: summarizeStats(fusedValues),
+    },
+  };
+}
+
+async function handleExplainSimilarity({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2ExplainSimilarityResult> {
+  const params = MemoryV2ExplainSimilarityParams.parse(body);
+  const config = loadConfig();
+  const { dense_weight: denseWeight, sparse_weight: sparseWeight } =
+    config.memory.v2;
+
+  const channels: MemoryV2ExplainSimilarityChannel[] = [];
+  channels.push(
+    await scoreChannel(
+      "user",
+      params.userText,
+      params.top,
+      denseWeight,
+      sparseWeight,
+      config,
+    ),
+  );
+  if (params.assistantText && params.assistantText.length > 0) {
+    channels.push(
+      await scoreChannel(
+        "assistant",
+        params.assistantText,
+        params.top,
+        denseWeight,
+        sparseWeight,
+        config,
+      ),
+    );
+  }
+  if (params.nowText && params.nowText.length > 0) {
+    channels.push(
+      await scoreChannel(
+        "now",
+        params.nowText,
+        params.top,
+        denseWeight,
+        sparseWeight,
+        config,
+      ),
+    );
+  }
+
+  return {
+    config: { dense_weight: denseWeight, sparse_weight: sparseWeight },
+    channels,
+  };
+}
+
+// ── Fit anisotropy calibration ──────────────────────────────────────────
+
+const MemoryV2FitAnisotropyParams = z
+  .object({
+    /**
+     * Number of leading principal components to project out at apply time.
+     * `1` is the canonical default for transformer embeddings; raise to 2-3
+     * only when the variance spectrum shows multiple dominant directions.
+     */
+    k: z.number().int().min(1).max(16).default(1),
+    /**
+     * Maximum number of stored vectors to pull from Qdrant for the fit.
+     * 5_000 is plenty for 3072-dim Gemini — power iteration converges fast
+     * and pulling the full corpus would just cost wall-clock time.
+     */
+    sample: z.number().int().min(1).max(100_000).default(5_000),
+  })
+  .strict();
+
+export interface MemoryV2FitAnisotropyResult {
+  provider: string;
+  model: string;
+  dim: number;
+  k: number;
+  sampleCount: number;
+  totalVariance: number;
+  componentVariance: number[];
+  /** `componentVariance[i] / totalVariance` for each component. */
+  explainedVarianceRatio: number[];
+  /** Absolute path the calibration was written to. */
+  path: string;
+}
+
+async function handleFitAnisotropy({
+  body = {},
+}: RouteHandlerArgs): Promise<MemoryV2FitAnisotropyResult> {
+  const { k, sample } = MemoryV2FitAnisotropyParams.parse(body);
+  const config = loadConfig();
+
+  const selection = await selectEmbeddingBackend(config);
+  if (!selection.backend) {
+    throw new RouteError(
+      `Cannot fit anisotropy calibration: ${selection.reason ?? "no embedding backend configured"}`,
+      "MEMORY_V2_NO_EMBEDDING_BACKEND",
+      409,
+    );
+  }
+
+  const vectors = await sampleConceptPageDenseVectors(sample);
+  if (vectors.length === 0) {
+    throw new RouteError(
+      "Cannot fit anisotropy calibration: the v2 concept-page collection is empty. " +
+        "Embed some concept pages first (run `assistant memory v2 reembed`), then retry.",
+      "MEMORY_V2_NO_VECTORS",
+      409,
+    );
+  }
+  if (vectors.length < k * 4) {
+    // PCA on too-few samples is unstable — refuse rather than hand back
+    // overfit components. The 4× heuristic is conservative; in practice
+    // anisotropy fits stabilise at a few hundred samples per component.
+    throw new RouteError(
+      `Cannot fit k=${k} components from only ${vectors.length} vectors — need at least ${k * 4}. ` +
+        "Embed more concept pages or fit a smaller k.",
+      "MEMORY_V2_INSUFFICIENT_VECTORS",
+      409,
+    );
+  }
+
+  const { provider, model } = selection.backend;
+  const calib = fitAnisotropyCalibration(vectors, k, { provider, model });
+  const path = await saveCalibration(calib);
+
+  return {
+    provider,
+    model,
+    dim: calib.dim,
+    k,
+    sampleCount: calib.sampleCount,
+    totalVariance: calib.totalVariance,
+    componentVariance: calib.componentVariance,
+    explainedVarianceRatio: explainedVarianceRatio(calib),
+    path,
+  };
+}
+
 // ── Route definitions ───────────────────────────────────────────────────
 
 export const ROUTES: RouteDefinition[] = [
@@ -174,6 +570,17 @@ export const ROUTES: RouteDefinition[] = [
     tags: ["memory"],
     requestBody: MemoryV2ValidateParams,
   },
+  {
+    operationId: "memory_v2_get_concept_page",
+    method: "POST",
+    endpoint: "memory/v2/concept-page",
+    handler: handleGetConceptPage,
+    summary: "Read a single memory v2 concept page",
+    description:
+      "Returns the rendered (frontmatter + body) markdown for a slug. 404 when the slug has no on-disk page — the activation log inspector uses this to show what got injected.",
+    tags: ["memory"],
+    requestBody: MemoryV2GetConceptPageParams,
+  },
   {
     operationId: "memory_v2_reembed_skills",
     method: "POST",
@@ -185,4 +592,37 @@ export const ROUTES: RouteDefinition[] = [
     tags: ["memory"],
     requestBody: MemoryV2ReembedSkillsParams,
   },
+  {
+    operationId: "memory_v2_explain_similarity",
+    method: "POST",
+    endpoint: "memory/v2/explain-similarity",
+    handler: handleExplainSimilarity,
+    summary: "Diagnose dense vs sparse similarity score distributions",
+    description:
+      "Read-only diagnostic. Embeds the supplied text(s), runs hybrid dense + sparse queries against the concept-page collection, and returns per-slug raw dense, raw sparse, normalized sparse, and fused scores plus per-channel summary stats. Used to investigate score-compression at the head of the activation distribution.",
+    tags: ["memory"],
+    requestBody: MemoryV2ExplainSimilarityParams,
+  },
+  {
+    operationId: "memory_v2_rebuild_corpus_stats",
+    method: "POST",
+    endpoint: "memory/v2/rebuild-corpus-stats",
+    handler: handleRebuildCorpusStats,
+    summary: "Rebuild the BM25 corpus statistics for memory v2",
+    description:
+      "Walks every concept page on disk, recomputes the document-frequency table and average document length used by the BM25 sparse channel, and atomically swaps the in-memory stats. Run after bulk content imports or to recover from a rebuild that errored at startup. Does not reembed individual page sparse vectors — pair with `assistant memory v2 reembed` when document-side weights need refreshing.",
+    tags: ["memory"],
+    requestBody: MemoryV2RebuildCorpusStatsParams,
+  },
+  {
+    operationId: "memory_v2_fit_anisotropy",
+    method: "POST",
+    endpoint: "memory/v2/fit-anisotropy",
+    handler: handleFitAnisotropy,
+    summary: "Fit the embedding anisotropy correction for memory v2",
+    description:
+      "Samples stored dense vectors from the concept-page Qdrant collection, fits a corpus mean + top-k principal components (Mu & Viswanath 'all-but-the-top'), and persists the calibration so subsequent embeds and queries apply the correction. Run `assistant memory v2 reembed` after fitting so stored vectors are written under the new calibration — until then, queries (corrected) and stored vectors (uncorrected) live in different spaces.",
+    tags: ["memory"],
+    requestBody: MemoryV2FitAnisotropyParams,
+  },
 ];

package/src/runtime/routes/migration-routes.ts

@@ -65,6 +65,11 @@ import {
   analyzeImport,
   DefaultPathResolver,
 } from "../migrations/vbundle-import-analyzer.js";
+import {
+  evaluateRuntimeCompatibility,
+  formatRuntimeCompatibilityMessage,
+  type RuntimeCompatibility,
+} from "../migrations/vbundle-import-policy.js";
 import {
   commitImport,
   extractCredentialsFromBundle,
@@ -79,6 +84,7 @@ import {
   InternalError,
   NotFoundError,
   RouteError,
+  UnprocessableEntityError,
 } from "./errors.js";
 import type { RouteDefinition, RouteHandlerArgs } from "./types.js";
 import { RouteResponse } from "./types.js";
@@ -866,6 +872,23 @@ export async function handleMigrationImport(
       };
     }
 
+    // Pre-check runtime-version compat before the DB close/reopen cycle.
+    // commitImport runs the same gate as defense-in-depth for callers that
+    // don't pre-check; we run it here too so an incompatible bundle short-
+    // circuits before resetDb().
+    const compatResult = evaluateRuntimeCompatibility(
+      validation.manifest!.compatibility,
+      APP_VERSION,
+    );
+    if (!compatResult.ok) {
+      throwImportCommitFailure({
+        ok: false,
+        reason: "version_incompatible",
+        bundle_compat: compatResult.bundle_compat,
+        runtime_version: compatResult.runtime_version,
+      });
+    }
+
     const pathResolver = new DefaultPathResolver(
       getWorkspaceDir(),
       getWorkspaceHooksDir(),
@@ -918,6 +941,12 @@ export async function handleMigrationImport(
 
     return importCommitSuccessResult(result.report, credentialsImported);
   } catch (err) {
+    // Preserve typed RouteError instances (e.g. UnprocessableEntityError for
+    // version_incompatible, BadRequestError for validation_failed) — only
+    // wrap genuinely unexpected errors as 500 InternalError.
+    if (err instanceof RouteError) {
+      throw err;
+    }
     log.error({ err }, "Unexpected error during import commit");
     throw new InternalError(
       err instanceof Error ? err.message : "Unexpected import error",
@@ -1014,12 +1043,18 @@ interface GcsImportErrorInit {
     | "fetch_failed"
     | "validation_failed"
     | "extraction_failed"
+    | "version_incompatible"
     | "write_failed";
   message: string;
   upstreamStatus?: number;
   reason?: string;
   errors?: Array<{ code: string; message: string; path?: string }>;
   partial_report?: ImportCommitReport;
+  /** Populated for `version_incompatible` — mirrors the platform's PR #5470
+   *  response shape so the URL-body endpoint can return the same body. */
+  bundle_compat?: RuntimeCompatibility;
+  /** Populated for `version_incompatible`. */
+  runtime_version?: string;
 }
 
 class GcsImportError extends Error {
@@ -1028,6 +1063,8 @@ class GcsImportError extends Error {
   public readonly reason?: string;
   public readonly errors?: GcsImportErrorInit["errors"];
   public readonly partial_report?: ImportCommitReport;
+  public readonly bundle_compat?: RuntimeCompatibility;
+  public readonly runtime_version?: string;
 
   constructor(init: GcsImportErrorInit) {
     super(init.message);
@@ -1045,6 +1082,12 @@ class GcsImportError extends Error {
     if (init.partial_report !== undefined) {
       this.partial_report = init.partial_report;
     }
+    if (init.bundle_compat !== undefined) {
+      this.bundle_compat = init.bundle_compat;
+    }
+    if (init.runtime_version !== undefined) {
+      this.runtime_version = init.runtime_version;
+    }
   }
 }
 
@@ -1343,6 +1386,22 @@ async function runGcsImport(
         reason: result.reason,
       });
     }
+    if (result.reason === "version_incompatible") {
+      // Returned by commitImport / streamCommitImport when the runtime falls
+      // outside the bundle's compat range. The platform-side gate is the
+      // primary check; this catches legacy bundles whose ExportJob row
+      // predates PR #5470 (compat columns NULL → platform gate skipped).
+      throw new GcsImportError({
+        code: "version_incompatible",
+        message: formatRuntimeCompatibilityMessage(
+          result.bundle_compat,
+          result.runtime_version,
+        ),
+        reason: result.reason,
+        bundle_compat: result.bundle_compat,
+        runtime_version: result.runtime_version,
+      });
+    }
     // write_failed
     throw new GcsImportError({
       code: "write_failed",
@@ -1442,6 +1501,20 @@ function throwGcsImportError(err: unknown): never {
         }),
       );
     }
+    if (err.code === "version_incompatible") {
+      // 422 (not 500) — the bundle is structurally valid but cannot be
+      // imported on this runtime. Body mirrors the platform's PR #5470
+      // response shape.
+      throw new UnprocessableEntityError(err.message, {
+        reason: "version_incompatible" as const,
+        ...(err.bundle_compat !== undefined && {
+          bundle_compat: err.bundle_compat,
+        }),
+        ...(err.runtime_version !== undefined && {
+          runtime_version: err.runtime_version,
+        }),
+      });
+    }
     if (err.code === "extraction_failed") {
       throw new InternalError(err.message);
     }
@@ -1674,6 +1747,29 @@ function throwImportCommitFailure(
     throw new InternalError(result.message);
   }
 
+  if (result.reason === "version_incompatible") {
+    // Returned by commitImport / streamCommitImport when the runtime falls
+    // outside the bundle's compat range. The platform-side gate is the
+    // primary check; this catches legacy bundles whose ExportJob row
+    // predates PR #5470 (compat columns NULL → platform gate skipped).
+    //
+    // 422 (not 500) — the bundle is structurally valid but cannot be
+    // imported on this runtime; the caller can act on it (upgrade the
+    // runtime, choose a different bundle). Body mirrors the platform's
+    // PR #5470 response shape.
+    throw new UnprocessableEntityError(
+      formatRuntimeCompatibilityMessage(
+        result.bundle_compat,
+        result.runtime_version,
+      ),
+      {
+        reason: "version_incompatible" as const,
+        bundle_compat: result.bundle_compat,
+        runtime_version: result.runtime_version,
+      },
+    );
+  }
+
   // write_failed
   throw new InternalError(result.message);
 }

package/src/runtime/routes/schedule-routes.ts

@@ -54,6 +54,9 @@ function handleListSchedules(queryParams: Record<string, string>) {
       nextRunAt: j.nextRunAt,
       lastRunAt: j.lastRunAt,
       lastStatus: j.lastStatus,
+      retryCount: j.retryCount,
+      maxRetries: j.maxRetries,
+      retryBackoffMs: j.retryBackoffMs,
       description:
         j.syntax === "cron"
           ? describeCronExpression(j.cronExpression)
@@ -139,6 +142,8 @@ function handleUpdateSchedule(id: string, body: Record<string, unknown>) {
     "quiet",
     "reuseConversation",
     "wakeConversationId",
+    "maxRetries",
+    "retryBackoffMs",
   ] as const) {
     if (key in body) {
       updates[key] = body[key];
@@ -294,6 +299,8 @@ export const ROUTES: RouteDefinition[] = [
         .describe("single_channel, multi_channel, or all_channels"),
       quiet: z.boolean(),
       reuseConversation: z.boolean(),
+      maxRetries: z.number().describe("Maximum retry attempts"),
+      retryBackoffMs: z.number().describe("Retry backoff in milliseconds"),
     }),
     responseBody: z.object({
       schedules: z.array(z.unknown()).describe("Updated schedule list"),