@vellumai/assistant 0.6.5 → 0.7.0

package/src/cli/commands/image-generation.ts

@@ -5,16 +5,13 @@ import { join } from "node:path";
 import { Command } from "commander";
 
 import { getConfig } from "../../config/loader.js";
+import { resolveImageGenCredentials } from "../../media/image-credentials.js";
 import {
   generateImage,
   type ImageGenCredentials,
-  mapGeminiError,
-} from "../../media/gemini-image-service.js";
-import {
-  buildManagedBaseUrl,
-  resolveManagedProxyContext,
-} from "../../providers/managed-proxy/context.js";
-import { getProviderKeyAsync } from "../../security/secure-keys.js";
+  mapImageGenError,
+  providerForModel,
+} from "../../media/image-service.js";
 import { log } from "../logger.js";
 
 // ---------------------------------------------------------------------------
@@ -72,11 +69,12 @@ export function registerImageGenerationCommand(program: Command): void {
     `
 Modes:
   managed    — Uses platform-managed credentials (requires login to Vellum).
-  your-own   — Uses your own Gemini API key configured in settings.
+  your-own   — Uses your own Gemini or OpenAI API key depending on the configured model.
 
 Supported models:
   gemini-3.1-flash-image-preview (default)
   gemini-3-pro-image-preview
+  gpt-image-2
 
 Examples:
   $ assistant image-generation generate --prompt "A sunset over the ocean"
@@ -114,12 +112,14 @@ Notes:
   Edit mode (--mode edit) requires at least one --source image file.
   Output files are named image-1.png, image-2.png, etc. (extension matches MIME type).
   Default output directory is the system temp directory.
+  Uses your own Gemini or OpenAI API key depending on the configured model.
 
 Examples:
   $ assistant image-generation generate --prompt "A mountain landscape at dawn"
   $ assistant image-generation generate --prompt "Make it darker" --mode edit --source input.png
   $ assistant image-generation generate --prompt "Logo variations" --variants 4 --output-dir ./logos
-  $ assistant image-generation generate --prompt "A robot" --model gemini-3-pro-image-preview --json`,
+  $ assistant image-generation generate --prompt "A robot" --model gemini-3-pro-image-preview --json
+  $ assistant image-generation generate --prompt "A robot" --model gpt-image-2 --json`,
   );
 
   generate.action(async (opts) => {
@@ -149,33 +149,30 @@ Examples:
 
     // --- Resolve credentials ---
     const config = getConfig();
-    const imageGenMode = config.services["image-generation"].mode;
+    const svc = config.services["image-generation"];
 
-    let credentials: ImageGenCredentials | undefined;
+    // Derive provider from the explicit `--model` override when supplied so
+    // that `--model gpt-image-2` routes to OpenAI even when config.provider
+    // is `gemini` (and vice-versa). Without this, the Gemini service would
+    // silently downgrade unknown models to its default.
+    const provider = providerForModel(modelOverride, svc.provider);
 
-    if (imageGenMode === "managed") {
-      const managedBaseUrl = await buildManagedBaseUrl("gemini");
-      if (managedBaseUrl) {
-        const ctx = await resolveManagedProxyContext();
-        credentials = {
-          type: "managed-proxy",
-          assistantApiKey: ctx.assistantApiKey,
-          baseUrl: managedBaseUrl,
-        };
-      }
-    } else {
-      const apiKey = await getProviderKeyAsync("gemini");
-      if (apiKey) {
-        credentials = { type: "direct", apiKey };
-      }
-    }
+    const { credentials, errorHint } = await resolveImageGenCredentials({
+      provider,
+      mode: svc.mode,
+    });
 
     if (!credentials) {
+      const baseHint =
+        errorHint ?? "No credentials available for image generation.";
+      // The shared hint in image-credentials.ts is UI-neutral (it points at
+      // "Settings > Models & Services"), which is correct for tool surfaces
+      // but drops the actionable CLI command. Prepend the exact command for
+      // each mode so CLI users see a next-step they can copy-paste.
       const hint =
-        imageGenMode === "managed"
-          ? "Managed proxy is not available. Please log in to Vellum or switch to your-own mode:\n  Run 'assistant auth login' to authenticate, or set services.image-generation.mode to 'your-own' in config."
-          : "No Gemini API key configured. Add your Gemini API key:\n  Run 'assistant keys set gemini' or configure it in Settings > Models & Services.";
-
+        svc.mode === "managed"
+          ? `${baseHint}\n  Run 'assistant auth login' to authenticate, or set services.image-generation.mode to 'your-own' in config.`
+          : `Run: assistant keys set ${provider} <key>\n${baseHint}`;
       if (jsonOutput) {
         process.stdout.write(JSON.stringify({ ok: false, error: hint }) + "\n");
       } else {
@@ -185,6 +182,8 @@ Examples:
       return;
     }
 
+    const resolvedCredentials: ImageGenCredentials = credentials;
+
     // --- Read source images for edit mode ---
     let sourceImages:
       | Array<{ mimeType: string; dataBase64: string }>
@@ -231,11 +230,11 @@ Examples:
     }
 
     // --- Resolve model ---
-    const model = modelOverride ?? config.services["image-generation"].model;
+    const model = modelOverride ?? svc.model;
 
     // --- Generate image ---
     try {
-      const result = await generateImage(credentials, {
+      const result = await generateImage(provider, resolvedCredentials, {
         prompt,
         mode,
         sourceImages,
@@ -286,7 +285,7 @@ Examples:
         }
       }
     } catch (error) {
-      const errorMsg = mapGeminiError(error);
+      const errorMsg = mapImageGenError(provider, error);
       if (jsonOutput) {
         process.stdout.write(
           JSON.stringify({ ok: false, error: errorMsg }) + "\n",

package/src/cli/commands/keys.ts

@@ -1,9 +1,9 @@
 import type { Command } from "commander";
 
 import { API_KEY_PROVIDERS } from "../../config/loader.js";
+import { getSecureKeyAsync } from "../../security/secure-keys.js";
 import {
   deleteSecureKeyViaDaemon,
-  getSecureKeyViaDaemon,
   setSecureKeyViaDaemon,
 } from "../lib/daemon-credential-client.js";
 import { log } from "../logger.js";
@@ -61,7 +61,7 @@ Examples:
     .action(async () => {
       const stored: string[] = [];
       for (const provider of API_KEY_PROVIDERS) {
-        const value = await getSecureKeyViaDaemon(provider);
+        const value = await getSecureKeyAsync(provider);
         if (value) stored.push(provider);
       }
       if (stored.length === 0) {

package/src/cli/commands/mcp.ts

@@ -17,7 +17,7 @@ import {
 import { getSignalsDir } from "../../util/platform.js";
 import { log } from "../logger.js";
 
-export const HEALTH_CHECK_TIMEOUT_MS = 10_000;
+const HEALTH_CHECK_TIMEOUT_MS = 10_000;
 
 export async function checkServerHealth(
   serverId: string,

package/src/cli/commands/memory-v2.ts

@@ -0,0 +1,343 @@
+/**
+ * `assistant memory v2` CLI subgroup.
+ *
+ * Operator-facing subcommands for the v2 memory subsystem (concept-page
+ * activation model). All commands are thin wrappers over two IPC routes:
+ *
+ *   - `memory_v2/backfill` — enqueues one of four mutating jobs against the
+ *     memory job queue (`migrate`, `rebuild-edges`, `reembed`,
+ *     `activation-recompute`). Returns a `jobId` so the operator can poll
+ *     status from the regular memory subsystem.
+ *   - `memory_v2/validate` — read-only structural validation of the v2
+ *     workspace state (concept pages + `memory/edges.json`). Returns an
+ *     aggregate report of orphan edges, oversized pages, and parse failures.
+ *
+ * Subcommands:
+ *
+ *   - `migrate [--force]` — one-shot v1->v2 synthesis. Refuses to run a
+ *     second time unless `--force` is passed (the migration handler writes a
+ *     sentinel after a successful first run).
+ *   - `rebuild-edges` — recompute every concept page's `edges:` frontmatter
+ *     from the canonical `memory/edges.json` index.
+ *   - `reembed` — fan out an `embed_concept_page` job per page slug to
+ *     refresh dense + sparse vectors in Qdrant.
+ *   - `activation` — refresh persisted activation state for every
+ *     conversation that has a stored row.
+ *   - `validate` — print a diagnostic report (page count, edge count, and
+ *     violation lists). Does not mutate the workspace.
+ *
+ * Lives alongside the existing v1 `memory` command rather than replacing it
+ * because v1 graph + PKB stays write-active until the cutover PR. Until the
+ * `memory-v2-enabled` feature flag flips on, the workspace keeps both v1 and
+ * v2 state side-by-side.
+ */
+
+import type { Command } from "commander";
+
+import { cliIpcCall } from "../../ipc/cli-client.js";
+import type {
+  MemoryV2BackfillOp,
+  MemoryV2BackfillResult,
+  MemoryV2ReembedSkillsResult,
+  MemoryV2ValidateResult,
+} from "../../runtime/routes/memory-v2-routes.js";
+import { log } from "../logger.js";
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+/**
+ * Issue a backfill IPC call, log the resulting `jobId`, and set a non-zero
+ * exit code on failure. Centralises the error-handling boilerplate for the
+ * four mutating subcommands.
+ */
+async function runBackfillOp(
+  op: MemoryV2BackfillOp,
+  opts: { force?: boolean } = {},
+): Promise<void> {
+  // Only forward `force: true` — the route handler already defaults missing
+  // to false, so omitting the field keeps the queued JSON minimal.
+  const params: Record<string, unknown> = { op };
+  if (opts.force === true) params.force = true;
+
+  const result = await cliIpcCall<MemoryV2BackfillResult>(
+    "memory_v2_backfill",
+    { body: params },
+  );
+
+  if (!result.ok) {
+    log.error(result.error ?? `Failed to enqueue ${op} job`);
+    process.exitCode = 1;
+    return;
+  }
+
+  log.info(`Queued ${op} job: ${result.result!.jobId}`);
+}
+
+// ---------------------------------------------------------------------------
+// Registration
+// ---------------------------------------------------------------------------
+
+export function registerMemoryV2Command(program: Command): void {
+  // Attach the `v2` subgroup to the existing top-level `memory` command so
+  // operators invoke it as `assistant memory v2 <op>`. The plan deliberately
+  // namespaces v2 under v1 rather than promoting it to a top-level command,
+  // so `assistant memory --help` surfaces both eras side-by-side until the
+  // post-soak cutover removes v1.
+  const memory = program.commands.find((c) => c.name() === "memory");
+  if (!memory) {
+    throw new Error(
+      "registerMemoryV2Command: parent `memory` command not found. " +
+        "Call registerMemoryCommand(program) before registerMemoryV2Command(program).",
+    );
+  }
+
+  const v2 = memory
+    .command("v2")
+    .description("Memory v2 subsystem operations (concept-page model)");
+
+  v2.addHelpText(
+    "after",
+    `
+The v2 subsystem replaces the v1 graph + PKB with prose concept pages,
+bidirectional edges in memory/edges.json, and activation-based retrieval.
+v2 stays gated behind the memory-v2-enabled feature flag — these subcommands
+remain useful operator tools regardless of whether the flag is on.
+
+Subcommands fall into three groups:
+
+  Mutating (return a jobId enqueued on the memory job queue):
+    migrate          One-shot v1->v2 synthesis. Refuses to overwrite an
+                     existing v2 state without --force.
+    rebuild-edges    Regenerate every concept page's edges: frontmatter from
+                     memory/edges.json.
+    reembed          Refresh dense + sparse vectors for every concept page.
+    activation       Refresh persisted activation state for every conversation.
+
+  Mutating (synchronous — runs inside the daemon and returns when done):
+    reembed-skills   Re-seed v2 skill entries from the current skill catalog.
+
+  Read-only:
+    validate         Print a diagnostic report of orphan edges, oversized
+                     pages, and parse failures.
+
+Examples:
+  $ assistant memory v2 validate
+  $ assistant memory v2 migrate
+  $ assistant memory v2 migrate --force
+  $ assistant memory v2 rebuild-edges
+  $ assistant memory v2 reembed
+  $ assistant memory v2 reembed-skills
+  $ assistant memory v2 activation`,
+  );
+
+  // ── migrate ───────────────────────────────────────────────────────────
+
+  v2.command("migrate")
+    .description("Run the v1->v2 migration (one-shot synthesis)")
+    .option(
+      "-f, --force",
+      "Re-run the migration even if the v2 sentinel is already present",
+    )
+    .addHelpText(
+      "after",
+      `
+Synthesises v2 concept pages from the v1 graph + PKB and writes
+memory/edges.json. The migration handler stores a sentinel at
+memory/.v2-state/.migration-complete-v1-to-v2 after a successful run and
+refuses to run again unless --force is passed.
+
+The job runs on the background memory worker — this command returns once
+the job is enqueued. Track progress via 'assistant memory status' or the
+assistant logs.
+
+Examples:
+  $ assistant memory v2 migrate
+  $ assistant memory v2 migrate --force`,
+    )
+    .action(async (opts: { force?: boolean }) => {
+      await runBackfillOp("migrate", { force: opts.force });
+    });
+
+  // ── rebuild-edges ─────────────────────────────────────────────────────
+
+  v2.command("rebuild-edges")
+    .description(
+      "Regenerate every concept page's edges: frontmatter from memory/edges.json",
+    )
+    .addHelpText(
+      "after",
+      `
+Walks every concept page and rewrites the edges: frontmatter list to match
+the canonical memory/edges.json index. Useful after a manual edit to
+edges.json or to recover from a partially-written page that drifted from
+the index.
+
+The job runs on the background memory worker — this command returns once
+the job is enqueued.
+
+Examples:
+  $ assistant memory v2 rebuild-edges`,
+    )
+    .action(async () => {
+      await runBackfillOp("rebuild-edges");
+    });
+
+  // ── reembed ───────────────────────────────────────────────────────────
+
+  v2.command("reembed")
+    .description(
+      "Refresh dense + sparse vectors for every concept page in Qdrant",
+    )
+    .addHelpText(
+      "after",
+      `
+Fans out an embed_concept_page job per concept page slug (plus the four
+reserved meta-file slugs) so each page's dense and sparse vectors get
+recomputed against the current embedding backend. Useful after upgrading
+the embedding model or recovering a corrupted Qdrant collection.
+
+The fan-out runs on the background memory worker — this command returns
+once the parent job is enqueued.
+
+Examples:
+  $ assistant memory v2 reembed`,
+    )
+    .action(async () => {
+      await runBackfillOp("reembed");
+    });
+
+  // ── reembed-skills ────────────────────────────────────────────────────
+
+  v2.command("reembed-skills")
+    .description(
+      "Re-seed v2 skill entries from the current skill catalog (synchronous)",
+    )
+    .addHelpText(
+      "after",
+      `
+Re-runs the v2 skill catalog seed against the current skill set, replacing
+both the in-process skill cache and the memory_v2_skills Qdrant collection.
+Useful after editing a skill's SKILL.md, after a feature-flag flip changes
+the enabled-skill set, or to recover a corrupted skills collection.
+
+Unlike 'reembed' (concept pages), this runs synchronously inside the
+daemon — the command returns only once the seed completes. Requires both
+the memory-v2-enabled feature flag and memory.v2.enabled to be on.
+
+Examples:
+  $ assistant memory v2 reembed-skills`,
+    )
+    .action(async () => {
+      const result = await cliIpcCall<MemoryV2ReembedSkillsResult>(
+        "memory_v2_reembed_skills",
+        { body: {} },
+      );
+
+      if (!result.ok) {
+        log.error(result.error ?? "Failed to re-seed v2 skill entries");
+        process.exitCode = 1;
+        return;
+      }
+
+      log.info("Skill re-seed complete.");
+    });
+
+  // ── activation ────────────────────────────────────────────────────────
+
+  v2.command("activation")
+    .description(
+      "Refresh persisted activation state for every active conversation",
+    )
+    .addHelpText(
+      "after",
+      `
+Walks every conversation row in the activation_state table and
+recomputes the persisted state without rendering or injecting a memory
+block. Useful after tuning the activation params (d, c_user, c_assistant,
+c_now, k, hops) so subsequent retrievals reflect the new weights without
+waiting for organic per-turn updates.
+
+The job runs on the background memory worker — this command returns once
+the job is enqueued.
+
+Examples:
+  $ assistant memory v2 activation`,
+    )
+    .action(async () => {
+      await runBackfillOp("activation-recompute");
+    });
+
+  // ── validate ──────────────────────────────────────────────────────────
+
+  v2.command("validate")
+    .description("Print a diagnostic report of v2 workspace state (read-only)")
+    .addHelpText(
+      "after",
+      `
+Walks every concept page and the memory/edges.json index, returning an
+aggregate report of:
+
+  pageCount             Number of concept pages successfully parsed.
+  edgeCount             Number of edges in memory/edges.json.
+  missingEdgeEndpoints  Edges whose endpoints reference a slug that no
+                        concept page exists for (orphan endpoints).
+  oversizedPages        Pages whose body exceeds memory.v2.max_page_chars
+                        (a soft cap; consolidation will eventually split).
+  parseFailures         Pages whose YAML frontmatter or schema validation
+                        failed during read.
+
+This is purely diagnostic — the command never mutates the workspace.
+
+Examples:
+  $ assistant memory v2 validate`,
+    )
+    .action(async () => {
+      const result = await cliIpcCall<MemoryV2ValidateResult>(
+        "memory_v2_validate",
+        {
+          body: {},
+        },
+      );
+
+      if (!result.ok) {
+        log.error(result.error ?? "Failed to validate memory v2 state");
+        process.exitCode = 1;
+        return;
+      }
+
+      const report = result.result!;
+      log.info(`Pages: ${report.pageCount}`);
+      log.info(`Edges: ${report.edgeCount}`);
+
+      if (report.missingEdgeEndpoints.length === 0) {
+        log.info("Missing edge endpoints: none");
+      } else {
+        log.info(
+          `Missing edge endpoints: ${report.missingEdgeEndpoints.length}`,
+        );
+        for (const { from, to } of report.missingEdgeEndpoints) {
+          log.info(`  ${from} <-> ${to}`);
+        }
+      }
+
+      if (report.oversizedPages.length === 0) {
+        log.info("Oversized pages: none");
+      } else {
+        log.info(`Oversized pages: ${report.oversizedPages.length}`);
+        for (const { slug, chars } of report.oversizedPages) {
+          log.info(`  ${slug} (${chars.toLocaleString()} chars)`);
+        }
+      }
+
+      if (report.parseFailures.length === 0) {
+        log.info("Parse failures: none");
+      } else {
+        log.info(`Parse failures: ${report.parseFailures.length}`);
+        for (const { slug, error } of report.parseFailures) {
+          log.info(`  ${slug}: ${error}`);
+        }
+      }
+    });
+}

package/src/cli/commands/memory.ts

@@ -12,7 +12,7 @@ import {
   requestReextract,
 } from "../../memory/admin.js";
 import { listConversations } from "../../memory/conversation-queries.js";
-import { initializeDb } from "../db.js";
+import { getDb } from "../../memory/db-connection.js";
 import { log } from "../logger.js";
 
 export function registerMemoryCommand(program: Command): void {
@@ -61,7 +61,7 @@ Examples:
   $ assistant memory status`,
     )
     .action(async () => {
-      initializeDb();
+      getDb();
       const status = await getMemorySystemStatus();
       log.info(`Memory enabled: ${status.enabled ? "yes" : "no"}`);
       log.info(`Memory degraded: ${status.degraded ? "yes" : "no"}`);
@@ -101,7 +101,7 @@ Examples:
   $ assistant memory backfill --force`,
     )
     .action((opts: { force?: boolean }) => {
-      initializeDb();
+      getDb();
       const jobId = requestMemoryBackfill(Boolean(opts?.force));
       log.info(`Queued backfill job: ${jobId}`);
     });
@@ -125,7 +125,7 @@ Examples:
   $ assistant memory cleanup-segments --dry-run`,
     )
     .action(async (opts: { dryRun?: boolean }) => {
-      initializeDb();
+      getDb();
       const result = await cleanupShortSegments({ dryRun: opts.dryRun });
       if (opts.dryRun) {
         log.info(
@@ -168,7 +168,7 @@ Examples:
   $ assistant memory query "API rate limits"`,
     )
     .action(async (text: string, opts?: { conversation?: string }) => {
-      initializeDb();
+      getDb();
       let conversationId = opts?.conversation;
       if (!conversationId) {
         const latest = listConversations(1)[0];
@@ -212,7 +212,7 @@ Examples:
   $ assistant memory status`,
     )
     .action(() => {
-      initializeDb();
+      getDb();
       const jobId = requestMemoryRebuildIndex();
       log.info(`Queued rebuild-index job: ${jobId}`);
     });
@@ -253,7 +253,7 @@ Examples:
     )
     .action(
       (opts: { conversation?: string[]; top?: string; dryRun?: boolean }) => {
-        initializeDb();
+        getDb();
 
         const targets = [];
 
@@ -355,7 +355,7 @@ Examples:
     )
     .action(
       async (opts: { threshold: number; limit?: number; apply?: boolean }) => {
-        initializeDb();
+        getDb();
         const apply = Boolean(opts.apply);
 
         if (!Number.isFinite(opts.threshold) || opts.threshold <= 0) {