@assethub/cli 0.1.30 → 0.1.31

package/dist/index.js

@@ -1,24 +1,33 @@
 #!/usr/bin/env node
+import { childOperationId, expandProductionBatch, planProductionNodeMeshes, callApiOperation, describeApiOperation, discoverApiOperations, searchApiOperations, TERMINAL_WORKSPACE_SKILL_BUILD_STATUSES, WORKSPACE_SKILL_ENHANCE_SECTIONS, } from '@assethub/api-client';
 import { setTimeout as delay } from 'node:timers/promises';
 import { cliVersion, diagnose, mcpConfig } from './setup.js';
 import { AGENT_IDS, initAgents, syncInstalledSkill } from './agentSetup.js';
+import { confirmNo, confirmYes, defaultFileDeps, formatSetupSummary, pickFromList, promptHidden, runProcess, runSetup, } from './setupCommand.js';
+import { launchAgentProgram, loadKeyIntoLaunchd } from './appEnv.js';
+import { defaultInstallHooks } from './setupHooksBridge.js';
+import { canUploadSessions, installHooks as installSessionHooks, runHooksCommand } from './hooks/index.js';
+import { createNodeUpdateDeps, runUpdate } from './update.js';
 import { ingestProjectSources } from './projectSources.js';
 import { downloadCanvas } from './canvasDownload.js';
+import { downloadCanvasGraphZip } from './canvasGraphDownload.js';
 import { executeNodeMeshBatch, resumeNodeMeshBatch, } from './nodeMeshBatch.js';
 import { writeCanvasComparison } from './canvasComparison.js';
+import { productionBatchLabels, readProductionBatch, runProductionBatch, } from './productionBatch.js';
 import { importCanvasAssetWithState } from './canvasAssetImport.js';
 import { execFile, spawn } from 'node:child_process';
 import { createHash, randomUUID } from 'node:crypto';
 import { realpathSync } from 'node:fs';
-import { chmod, mkdir, readFile, rename, unlink, writeFile, } from 'node:fs/promises';
+import { chmod, mkdir, readdir, readFile, rename, unlink, writeFile, } from 'node:fs/promises';
 import { homedir, tmpdir } from 'node:os';
 import { basename, dirname, extname, join, resolve } from 'node:path';
 import { argv, env, exit, stdin, stderr, stdout } from 'node:process';
 import { fileURLToPath } from 'node:url';
 import { promisify } from 'node:util';
-import { AssetHubApiError, callApiOperation, describeApiOperation, discoverApiOperations, searchApiOperations, WorkspaceClientError, createWorkspaceClient, createAssetHubClient, } from '@assethub/api-client';
+import { AssetHubApiError, WorkspaceClientError, createWorkspaceClient, createAssetHubClient, } from '@assethub/api-client';
 import { resolveCanvasSelection, saveCanvasSelection } from './canvas.js';
-import { activeExecution, childOperationId, CliExecutionError, executeRecorded, hasRecordedOperation, openExecutionSession, resumeRecorded, waitForExecution, } from './execution.js';
+import { activeExecution, CliExecutionError, executeRecorded, hasRecordedOperation, openExecutionSession, resumeRecorded, waitForExecution, } from './execution.js';
+import { runProgressLine } from './runProgress.js';
 import { buildRunUploadPlan, formatBytes } from './runsUpload/buildRunUpload.js';
 import { readGraphFolder } from './runsUpload/graphFolder.js';
 import { uploadRun } from './runsUpload/uploadRun.js';
@@ -49,6 +58,81 @@ const partExtractorOptions = [
         apiValue: 'ah_part_extractor_v21_alpha',
         aliases: ['v2.1 alpha', 'v2.1-alpha', '2.1 alpha'],
     },
+    {
+        publicName: 'V2.1-beta',
+        apiValue: 'ah_part_extractor_v6b',
+        aliases: ['v2.1 beta', 'v2.1-beta', '2.1 beta', '2.1-beta'],
+    },
+    {
+        publicName: 'V1',
+        apiValue: 'ah_part_extractor_v5',
+        aliases: ['v1'],
+    },
+    {
+        publicName: 'V3.0.1',
+        apiValue: 'ah_agent_graph_v3_0_1',
+        aliases: ['v3.0.1', '3.0.1'],
+    },
+    {
+        publicName: 'V3.5.3 Complete Objects',
+        apiValue: 'ah_agent_graph_v3_5_3',
+        aliases: ['v3.5.3', '3.5.3'],
+    },
+    {
+        publicName: 'V3.5.3.1 Reliable Analysis',
+        apiValue: 'ah_agent_graph_v3_5_3_1',
+        aliases: ['v3.5.3.1', '3.5.3.1'],
+    },
+    {
+        publicName: 'V3.5.4 Evidence Review',
+        apiValue: 'ah_agent_graph_v3_5_4',
+        aliases: ['v3.5.4', '3.5.4'],
+    },
+    {
+        publicName: 'V3.5.5 Complete Part Review',
+        apiValue: 'ah_agent_graph_v3_5_5',
+        aliases: ['v3.5.5', '3.5.5'],
+    },
+    {
+        publicName: 'V3.5.6 Golden Recovery',
+        apiValue: 'ah_agent_graph_v3_5_6',
+        aliases: ['v3.5.6', '3.5.6'],
+    },
+    {
+        publicName: 'V3.5.7 Original Source First',
+        apiValue: 'ah_agent_graph_v3_5_7',
+        aliases: ['v3.5.7', '3.5.7'],
+    },
+    {
+        publicName: 'V3.5.8 Concise Inventory',
+        apiValue: 'ah_agent_graph_v3_5_8',
+        aliases: ['v3.5.8', '3.5.8'],
+    },
+    {
+        publicName: 'V3.5.9 Observed Parts',
+        apiValue: 'ah_agent_graph_v3_5_9',
+        aliases: ['v3.5.9', '3.5.9'],
+    },
+    {
+        publicName: 'V3.6 Multiview References',
+        apiValue: 'ah_agent_graph_v3_6',
+        aliases: ['v3.6', '3.6'],
+    },
+    {
+        publicName: 'Chibi Character (Pluffy)',
+        apiValue: 'ah_agent_graph_pluffy_v3',
+        aliases: ['pluffy v3', 'pluffy 3'],
+    },
+    {
+        publicName: 'V3-2.1',
+        apiValue: 'ah_agent_graph_v21',
+        aliases: ['v3-2.1', '3-2.1'],
+    },
+    {
+        publicName: 'V3.6.2 Single Analysis',
+        apiValue: 'ah_agent_graph_v3_6_2',
+        aliases: ['v3.6.2', '3.6.2'],
+    },
     {
         publicName: 'V3.6.1',
         apiValue: 'ah_agent_graph_v3_6_1',
@@ -69,21 +153,15 @@ const partExtractorOptions = [
         apiValue: 'ah_agent_graph_v3_6_5',
         aliases: ['v3.6.5', '3.6.5'],
     },
-    {
-        publicName: 'Chibi Character (Pluffy) v3.1',
-        apiValue: 'ah_agent_graph_pluffy_v3_1',
-        aliases: ['pluffy', 'pluffy v3.1', 'pluffy 3.1'],
-    },
     {
         publicName: 'V3.0.9 Garment Boundaries',
         apiValue: 'ah_agent_graph_v3_7_2',
         aliases: ['v3.0.9', '3.0.9'],
     },
     {
-        publicName: 'V3.7 Artist Skills',
-        apiValue: 'ah_agent_graph_v3_7',
-        // The web picker and `production agents` call it V3.0.7.
-        aliases: ['v3.7', '3.7', 'v3.0.7 artist skills', 'v3.0.7', '3.0.7'],
+        publicName: 'V3.0.10-dev.1 Body Fit',
+        apiValue: 'ah_agent_graph_v3_7_3',
+        aliases: ['v3.0.10-dev.1', '3.0.10-dev.1'],
     },
     {
         publicName: 'V3.7.1 Building Modules',
@@ -92,11 +170,31 @@ const partExtractorOptions = [
         aliases: ['v3.7.1', '3.7.1', 'v3.0.8 building modules', 'v3.0.8', '3.0.8'],
     },
     {
-        publicName: 'Humanoid Assembly (internal)',
+        publicName: 'Chibi Character (Pluffy) v3.1',
+        apiValue: 'ah_agent_graph_pluffy_v3_1',
+        aliases: ['pluffy', 'pluffy v3.1', 'pluffy 3.1'],
+    },
+    {
+        publicName: 'V3.7 Artist Skills',
+        apiValue: 'ah_agent_graph_v3_7',
+        // The web picker and `production agents` call it V3.0.7.
+        aliases: ['v3.7', '3.7', 'v3.0.7 artist skills', 'v3.0.7', '3.0.7'],
+    },
+    {
+        publicName: 'Humanoid Assembly',
         apiValue: 'ah_agent_graph_humanoid_assembly_auto',
-        aliases: ['humanoid assembly', 'humanoid-assembly', 'humanoid_assembly_auto'],
+        // The name CLI 0.1.26-0.1.31 used stays accepted.
+        aliases: ['humanoid assembly (internal)', 'humanoid-assembly', 'humanoid_assembly_auto'],
+    },
+    {
+        publicName: 'V4 Character Assembly',
+        apiValue: 'ah_agent_graph_harpy_assembly_v2',
+        aliases: ['v4', '4', 'v4 character assembly', 'character assembly'],
     },
 ];
+// A V4 Character Assembly run (image → parts → meshes → assembled character) takes
+// about an hour, so `--wait` defaults past the generic 15 minute limit.
+const harpyAssemblyV2WaitMs = 2 * 60 * 60 * 1000;
 // Graph-engine extractors run through the artifact-graph endpoint. Derived from
 // the option list so a newly listed graph agent cannot miss the graph path.
 const artifactGraphPartExtractorApiValues = new Set(partExtractorOptions
@@ -124,6 +222,9 @@ CLI and MCP are available to all AssetHub users. Workspace permissions and featu
 
 Usage:
   assethub --version
+  assethub setup [--client claude|codex|both] [--profile <name>] [--workspace <id>] [--api-key-stdin] [--base-url <url>] [--no-app-env] [--dry-run] [--yes] [--print-env] [--json]
+  assethub env load [--profile <name>]
+  assethub update [--check] [--dry-run] [--yes]
   assethub doctor [--mcp] [--profile <name>] [--timeout-ms <n>]
   assethub init [--agent claude-code|codex|cursor...] [--project] [--dry-run] [--base-url <url>] [--workspace <id>] [--profile <name>]
   assethub mcp tools [tool-name] [--account] [--profile <name>] [--timeout-ms <n>]
@@ -167,31 +268,30 @@ Usage:
   assethub api search [query]
   assethub api describe "<METHOD /path>"
   assethub api call "<METHOD /path>" [--path-json <json|@file>] [--query-json <json|@file>] [--input-json <json|@file|@->] [--operation-id <uuid>]
-  assethub skills list [--cursor <cursor>]
-  assethub skills get <skill-id> [--revision <n>]
-  assethub skills learn --input-json <json|@file|@-> --operation-id <uuid>
-  assethub skills update <skill-id> --input-json <json|@file|@->
-  assethub skills controls <skill-id> --revision <n> --grant-revision <n> --mode automatic|manual|off [--share]
-  assethub skills prepare --input-json <json|@file|@->
-  assethub skills proposal <proposal-id>
-  assethub skills accept <proposal-id> --input-json <json|@file|@->
   assethub composer models
   assethub composer refine --list-modes
-  assethub composer refine --from-run <run-id> --instruction <text> [--mode standard|thorough|placement|workshop|blender|codex] [--effort light|standard|thorough] [--max-rounds <n>] [--transforms-json <json|@file>] [--wait] [--download --out-dir <dir>]
+  assethub composer refine --from-run <run-id> --instruction <text> [--mode standard|thorough|placement|workshop|blender|codex] [--effort light|standard|thorough] [--max-rounds <n>] [--agent-model <provider:id|unambiguous-id>] [--transforms-json <json|@file>] [--wait] [--download --out-dir <dir>]
   assethub composer refine --input-json <json|@file> [--canvas <id>] [--operation-id <uuid>] [--wait] [--download --out-dir <dir>]
-  assethub composer run --part <mesh-asset-id>[:<part-image-asset-id>] [--part <mesh-asset-id>[:<part-image-asset-id>]...] --reference <image-asset-id> [--model <id>] [--mode quick|quality] [--transforms-json <json|@file>] [--canvas <id>] [--wait]
+  assethub composer refine --from-run <run-id> --skill-assembly [--reasoning low|medium|high] [--agent-model <agents-api model>] [--wait]
+  assethub composer run --part <mesh-asset-id>[=<part-image-id>] [--part <mesh-asset-id>[=<part-image-id>]...] --reference <image-asset-id> [--model <id>] [--agent-model <provider:id|unambiguous-id>] [--mode quick|quality] [--optimize off|light|medium|heavy] [--transforms-json <json|@file>] [--canvas <id>] [--wait]
+  assethub composer run --part <mesh-asset-id>=<part-image-id>... --reference <image-asset-id> --model V6 [--optimize off|light|medium|heavy] [--canvas <id>] [--wait]
+  assethub composer run --part <mesh-asset-id>... --reference <image-asset-id> --model V5.1 [--reasoning low|medium|high] [--agent-model <agents-api model>] [--transforms-json <json|@file>] [--canvas <id>] [--wait]
   assethub composer run --input-json <json|@file> [--canvas <id>] [--operation-id <uuid>] [--wait] [--download --out-dir <dir>]
   assethub composer run --from-run <run-id> [--transforms-json <json|@file>] [--mode quick|quality] [--wait]
   assethub composer run --node <shape:id> --canvas <id> [--model <id>] [--mode quick|quality] [--operation-id <uuid>] [--wait]
   assethub canvas create [--name <name>] [--operation-id <uuid>]
   assethub canvas list [--cursor <cursor>] [--limit <n>]
   assethub canvas get|use|open [<id>]
+  assethub canvas graph-id <canvas-id>
   assethub runs list [--canvas <id>] [--cursor <cursor>] [--limit <n>]
-  assethub runs get|watch|wait <run-id> [--interval-ms <ms>] [--timeout-ms <n>] [--summary]
+  assethub runs get|watch|wait <run-id> [--interval-ms <ms>] [--timeout-ms <n>] [--summary] [--download --out-dir <dir>]
   assethub runs resume <operation-id> [--wait]
   assethub graph list [--cursor <cursor>] [--limit <n>]
   assethub graph show|export [--canvas <id> | --graph <id>] [--source generated|upload] [--artifact <node-id>] [--out <path>] [--allow-truncated]
   assethub graph lineage --artifact <node-id> [--canvas <id> | --graph <id>] [--source generated|upload] [--direction ancestors|descendants|both] [--depth <n>] [--out <path>]
+  assethub graph snapshot <graph-id> [--offset <n>] [--limit <n>]
+  assethub graph node <graph-id> <node-id>
+  assethub graph image <graph-id> <node-id> [--out <path>]
   assethub evaluate list
   assethub evaluations submit --artifact <asset-id> --report <json|file|@file|@-> [--agent <name>] [--canvas <id>] [--reference <asset-id>...] [--run <run-id>] [--operation-id <uuid>] [--require-pass]
   assethub evaluations list [--canvas <id>] [--artifact <asset-id>]
@@ -212,7 +312,10 @@ Usage:
   assethub animate retarget --resource-id <id> --animation <preset-id> [--animation <preset-id>... up to 5 total] [--out-format glb|fbx] [--bake-animation true|false] [--export-with-geometry true|false] [--animate-in-place true|false] [--name <name>] [--wait] [--download --out-dir <dir>]
   assethub jobs get <job-id> [--download --out-dir <dir>]
   assethub jobs watch <job-id> [--interval-ms <ms>] [--timeout-ms <ms>] [--download --out-dir <dir>]
-  assethub production analyze (--file <path> | --source-url <url> | --source-id <id> | --image-url <url> | --file-ref-json <json> | --stdin | --stdin-base64 | --stdin-data-uri | --stdin-json | --source-json <json|@file|@-> | --data-uri <uri> | --clipboard) [--file-name <name>] [--content-type <type>] [--part-extractor <name>] [--name <name>] [--skill-mode auto|manual|off] [--skill-id <id>...] [--context <canvas-id> --context-version <n>] [--wait] [--download --out-dir <dir>]
+  assethub production analyze (--file <path> | --source-url <url> | --source-id <id> | --image-url <url> | --file-ref-json <json> | --stdin | --stdin-base64 | --stdin-data-uri | --stdin-json | --source-json <json|@file|@-> | --data-uri <uri> | --clipboard) [--file-name <name>] [--content-type <type>] [--part-extractor <name>] [--name <name>] [--base-body <mesh-asset-id>] [--skill-planner-model <model>] [--auto-repair true|false] [--skill-mode auto|manual|off] [--skill-id <id>...] [--context <canvas-id> --context-version <n>] [--wait] [--download --out-dir <dir>]
+    Production graph options: --pipeline-depth parts|mesh|composition [--assembly-policy concept-to-character-v1] [--mesh-generation-json <json|@file>]
+  assethub production batch (--file <path>... | --files-dir <dir> | --source-id <id>... | --source-url <url>...) --part-extractor <name> [--repeat <n>] [--yes] [--canvas <id>] [--operation-id <uuid>] [--wait [--concurrency <n>]] [--download --out-dir <dir>]
+    One production analyze per image × --repeat, on one canvas. Takes the production analyze options. Rerun with the printed --operation-id to resume; nothing is started twice.
   assethub production agents
   assethub production automation --input-json <json|@file|@-> [--canvas <id>] [--operation-id <uuid>] [--wait]
   assethub production status <order-id> [--wait] [--download --out-dir <dir>]
@@ -222,10 +325,32 @@ Usage:
   assethub production watch <order-id> [--interval-ms <ms>] [--timeout-ms <ms>] [--download --out-dir <dir>]
   assethub production intervene <order-id> (--exclude <target-id> | --include <target-id> | --add-part <name> | --rename <target-id>=<name> | --reject <target-id>[=<reason>] | --regenerate <target-id>=<mode> | --set-param <key>=<value> | --ops-json <json|@file|@->)... [--idempotency-key <key>]
   assethub production interventions <order-id>
+  assethub production resume <run-id|order-id> [--expected-resume-count <n>] [--wait] [--timeout-ms <ms>] [--download --out-dir <dir>]   (after a fix, pick a stopped or stalled graph run back up; no new charge)
   assethub runs upload <path> [--graph-id <id>] [--stream-id <id>] [--rev <n>] [--description <text>] [--tag <tag>...] [--skip-register] [--dry-run]
-  assethub parts split (--file <path> | --source-url <url> | --source-id <id> | --image-url <url> | --file-ref-json <json> | --stdin | --stdin-base64 | --stdin-data-uri | --stdin-json | --source-json <json|@file|@-> | --data-uri <uri> | --clipboard | --order-id <id>) [--file-name <name>] [--content-type <type>] [--part-extractor <name>] [--skill-mode auto|manual|off] [--skill-id <id>...] [--wait] [--task-id <id>...] [--mission-id <id>] [--all-ready] [--download --out-dir <dir>]
+  assethub parts split (--file <path> | --source-url <url> | --source-id <id> | --image-url <url> | --file-ref-json <json> | --stdin | --stdin-base64 | --stdin-data-uri | --stdin-json | --source-json <json|@file|@-> | --data-uri <uri> | --clipboard | --order-id <id>) [--file-name <name>] [--content-type <type>] [--part-extractor <name>] [--base-body <mesh-asset-id>] [--skill-planner-model <model>] [--auto-repair true|false] [--skill-mode auto|manual|off] [--skill-id <id>...] [--wait] [--task-id <id>...] [--mission-id <id>] [--all-ready] [--download --out-dir <dir>]
   assethub parts compare (--file <path> | --source-url <url> | --source-id <id> | --image-url <url> | --file-ref-json <json> | --stdin | --stdin-base64 | --stdin-data-uri | --stdin-json | --source-json <json|@file|@-> | --data-uri <uri> | --clipboard) [--file-name <name>] [--content-type <type>] [--preprocess-prompt <text>] [--preprocess-model-id <id>] [--fail-on-preprocess-error] [--part-extractor <name>] [--wait] [--task-id <id>...] [--mission-id <id>] [--all-ready] [--download --out-dir <dir>]
   assethub autopilot [<image>] --demo [--yolo] [--json] [--max-regen <n>] [--stall-rounds <n>] [--tie-epsilon <x>] [--multiview 2-view|4-view|6-view] [--ab-mode serial|cross] [--credit-budget <n>]
+  assethub memory memorize --canvas <id> [--nodes <shape-id>,<shape-id>,...] [--wait] [--timeout <s>]
+  assethub memory replay <memory-id> --source <asset-id> [--wait] [--timeout <s>]
+  assethub skills list [--cursor <cursor>]
+  assethub skills get <skill-id> [--revision <n>]
+  assethub skills memory <skill-id> [--revision <n>] [--images <dir>]
+  assethub skills learn --input-json <json|@file|@-> --operation-id <uuid>
+  assethub skills update <skill-id> --input-json <json|@file|@->
+  assethub skills controls <skill-id> --revision <n> --grant-revision <n> --mode automatic|manual|off [--share]
+  assethub skills prepare --input-json <json|@file|@->
+  assethub skills proposal <proposal-id>
+  assethub skills accept <proposal-id> --input-json <json|@file|@->
+  assethub skills validate --input-json <json|@file|@-> [--build <build-id>]
+  assethub skills schema
+  assethub skills build --goal <text> (--canvas <id>... | --memory <id>...) [--instructions <text>] [--operation-id <uuid>] [--wait] [--dry-run]
+  assethub skills official list
+  assethub skills official install <skill-id> --revision <n> --content-sha256 <sha256> [--expected-revision <n>] [--expected-grant-revision <n>]
+  assethub skills official prepare <skill-id> --input-json <json|@file|@->
+  assethub skills official customize <skill-id> --input-json <json|@file|@->
+  assethub org search <query>
+  assethub org canvases --org <orgId>
+  assethub graph export-canvas --org <orgId> --canvas <id>[,<id>...] [--wait] [--timeout <s>] [--out-dir <dir>]
 
 Global options:
   --api-key <key>       Overrides saved auth and ASSETHUB_API_KEY.
@@ -254,6 +379,8 @@ Examples:
   assethub rig create --source-resource-id mesh_x --rig-type biped --wait
   assethub animate retarget --resource-id rig_x --animation preset:idle --wait
   assethub production analyze --file ./input.png --part-extractor "V1.5" --wait
+  assethub production analyze --file ./concept.png --part-extractor v4 --wait --download --out-dir ./character
+  assethub production batch --file ./harpy.png --file ./satyr.png --repeat 2 --part-extractor v4 --yes --wait --download --out-dir ./batch
   assethub production run --order-id ord_x --mission-id ms_x --run-mode full_auto --max-cost-credits 200 --wait
   assethub production run --image ./character.png --estimate
   assethub production run --image ./character.png --compose v6 --max-cost 500 --wait --download --out-dir ./out/asset
@@ -262,12 +389,17 @@ Examples:
   assethub runs wait run_x --timeout-ms 3600000
   assethub production intervene ord_x --exclude 7 --rename 3="left wing" --idempotency-key review-2026-07-29
   assethub production interventions ord_x
+  assethub production resume ord_x
+  assethub production resume <run-id> --wait --download --out-dir ./character
   assethub runs upload ./out/pluffy-run-2026-07-30 --dry-run
   assethub runs upload ./out/pluffy-run-2026-07-30 --tag pluffy --description "V3 humanoid, 18 parts"
   assethub parts split --file ./input.png --part-extractor "V1.5" --wait --all-ready
   assethub parts compare --file ./input.png --preprocess-prompt "clean white background, centered product photo" --part-extractor "V1.5" --wait --all-ready
 `;
+let printSuppressed = false;
 const print = (value) => {
+    if (printSuppressed)
+        return;
     stdout.write(`${JSON.stringify(value, null, 2)}\n`);
 };
 const parseArgs = (args) => {
@@ -369,8 +501,8 @@ const resolveProductionSkillSelection = (flags, agentVersion) => {
     }
     if (mode !== 'auto' && mode !== 'manual' && mode !== 'off')
         throw new Error('--skill-mode must be "auto", "manual", or "off"');
-    if (!['ah_agent_graph_v3_7', 'ah_agent_graph_v3_7_1'].includes(agentVersion))
-        throw new Error('--skill-mode requires V3.7 Artist Skills or V3.7.1 Building Modules');
+    if (agentVersion !== 'ah_agent_graph_v3_7' && agentVersion !== 'ah_agent_graph_v3_7_1' && agentVersion !== 'ah_agent_graph_v3_7_2' && agentVersion !== 'ah_agent_graph_v3_7_3')
+        throw new Error('--skill-mode is supported only with V3 Artist Skills');
     if (skillIds.length > 3)
         throw new Error('--skill-id accepts at most 3 values');
     if (new Set(skillIds).size !== skillIds.length)
@@ -383,6 +515,14 @@ const resolveProductionSkillSelection = (flags, agentVersion) => {
         throw new Error('--skill-mode off does not accept --skill-id');
     return { mode, skillIds };
 };
+const resolveProductionBaseBody = (flags) => {
+    const value = getFlag(flags, 'base-body');
+    return value == null ? undefined : assertFlagHasValue(value, '--base-body <mesh-asset-id>');
+};
+const resolveSkillPlannerModel = (flags) => {
+    const value = getFlag(flags, 'skill-planner-model');
+    return value == null ? undefined : assertFlagHasValue(value, '--skill-planner-model <model>');
+};
 const publicPartExtractorName = (value) => {
     if (value == null) {
         return null;
@@ -759,17 +899,6 @@ const assertFlagHasValue = (value, flagDescription) => {
  * attaches the image the part's mesh was generated from, matching what the
  * canvas node graph supplies automatically for UI-driven runs.
  */
-const parsePartFlagValue = (raw) => {
-    const value = assertFlagHasValue(raw, '--part <mesh-asset-id>[:<part-image-asset-id>]');
-    const separatorIndex = value.indexOf(':');
-    if (separatorIndex === -1)
-        return { assetId: value };
-    const assetId = value.slice(0, separatorIndex);
-    const partImageAssetId = value.slice(separatorIndex + 1);
-    if (!assetId || !partImageAssetId)
-        throw new Error('--part must be <mesh-asset-id> or <mesh-asset-id>:<part-image-asset-id>');
-    return { assetId, partImageAssetId };
-};
 const sourceInputCount = (flags) => [
     getSourceResourceIdFlag(flags) != null,
     getSourceUrlFlag(flags) != null,
@@ -1408,68 +1537,6 @@ const commandApi = async (subcommand, positionals, ctx) => {
         throw error;
     }
 };
-export const commandSkills = async (subcommand, positionals, ctx) => {
-    if (subcommand === 'list') {
-        print(await ctx.client.v2.listWorkspaceSkills({
-            cursor: getFlag(ctx.flags, 'cursor'),
-        }));
-        return;
-    }
-    if (subcommand === 'get') {
-        print(await ctx.client.v2.getWorkspaceSkill(requirePositional(positionals, 2, 'skill-id'), {
-            revision: hasFlag(ctx.flags, 'revision')
-                ? parsePositiveIntegerFlag(ctx.flags, 'revision', 1)
-                : undefined,
-        }));
-        return;
-    }
-    if (subcommand === 'controls') {
-        const mode = parseEnumFlag(ctx.flags, 'mode', [
-            'automatic',
-            'manual',
-            'off',
-        ]);
-        if (!mode)
-            throw new Error('Missing required flag: --mode');
-        requireFlag(ctx.flags, 'revision');
-        requireFlag(ctx.flags, 'grant-revision');
-        print(await ctx.client.v2.setWorkspaceSkillControls(requirePositional(positionals, 2, 'skill-id'), {
-            expectedRevision: parsePositiveIntegerFlag(ctx.flags, 'revision', 1),
-            expectedGrantRevision: parsePositiveIntegerFlag(ctx.flags, 'grant-revision', 1),
-            mode,
-            ...(hasFlag(ctx.flags, 'share') ? { share: true } : {}),
-        }));
-        return;
-    }
-    if (subcommand === 'proposal') {
-        print(await ctx.client.v2.getWorkspaceSkillProposal(requirePositional(positionals, 2, 'proposal-id')));
-        return;
-    }
-    if (['learn', 'prepare'].includes(subcommand ?? '')) {
-        const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
-        const operationId = subcommand === 'learn'
-            ? requireFlag(ctx.flags, 'operation-id')
-            : undefined;
-        if (operationId != null &&
-            !/^[0-9a-f]{8}-[0-9a-f]{4}-[1-8][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i.test(operationId))
-            throw new Error('--operation-id must be a UUID; reuse the same ID and input after an uncertain response');
-        if (operationId != null)
-            activeExecution.operationId = operationId;
-        print(subcommand === 'learn'
-            ? await ctx.client.v2.learnWorkspaceSkill(input, { idempotencyKey: operationId })
-            : await ctx.client.v2.prepareWorkspaceSkillProposal(input));
-        return;
-    }
-    if (['update', 'accept'].includes(subcommand ?? '')) {
-        const id = requirePositional(positionals, 2, subcommand === 'update' ? 'skill-id' : 'proposal-id');
-        const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
-        print(subcommand === 'update'
-            ? await ctx.client.v2.updateWorkspaceSkill(id, input)
-            : await ctx.client.v2.acceptWorkspaceSkillProposal(id, input));
-        return;
-    }
-    throw new Error('Use skills list|get|learn|update|controls|prepare|proposal|accept');
-};
 const commandModels = async (subcommand, positionals, ctx) => {
     if (subcommand === 'list') {
         const domain = getFlag(ctx.flags, 'domain');
@@ -1837,6 +1904,133 @@ const commandAuth = async (subcommand, flags) => {
     }
     throw new Error('Unknown auth command. Use "auth login", "auth status", or "auth logout".');
 };
+// One-command onboarding. Login and workspace selection reuse commandAuth and
+// commandWorkspace (their JSON output is suppressed); setupCommand.ts owns the
+// client configuration and the summary.
+export const commandSetup = async (flags, installHooks = defaultInstallHooks) => {
+    const client = getFlag(flags, 'client') ?? 'both';
+    if (client !== 'claude' && client !== 'codex' && client !== 'both')
+        throw new Error('--client must be claude, codex or both');
+    const config = await readAuthConfig(getConfigPath(flags));
+    // Only name a profile the user chose or one that is saved: a synthetic profile
+    // would make resolveAuth reject a fresh machine before its ASSETHUB_API_KEY fallback.
+    const base = { ...flags };
+    const chosenProfile = hasFlag(flags, 'profile') ? requireFlag(flags, 'profile') : config.defaultProfile;
+    if (chosenProfile)
+        base.profile = chosenProfile;
+    else
+        delete base.profile;
+    delete base.workspace;
+    const apiKeyStdin = hasFlag(flags, 'api-key-stdin');
+    const flagKey = hasFlag(flags, 'api-key') ? requireFlag(flags, 'api-key') : undefined;
+    const envKey = env.ASSETHUB_API_KEY?.trim() || undefined;
+    const quiet = async (fn) => {
+        printSuppressed = true;
+        try {
+            return await fn();
+        }
+        finally {
+            printSuppressed = false;
+        }
+    };
+    const result = await runSetup({
+        client,
+        workspace: getFlag(flags, 'workspace'),
+        apiKeyStdin,
+        apiKey: flagKey ?? envKey,
+        apiKeySource: flagKey ? 'flag' : envKey ? 'env' : undefined,
+        noHook: hasFlag(flags, 'no-hook'),
+        noAppEnv: hasFlag(flags, 'no-app-env'),
+        saveSessions: hasFlag(flags, 'save-sessions'),
+        dryRun: hasFlag(flags, 'dry-run'),
+        yes: hasFlag(flags, 'yes'),
+        printEnv: hasFlag(flags, 'print-env'),
+    }, {
+        ...defaultFileDeps(),
+        launchAgentProgram: launchAgentProgram(process.execPath, realpathSync(argv[1] ?? fileURLToPath(import.meta.url)), ['env', 'load', ...(hasFlag(flags, 'profile') ? ['--profile', requireFlag(flags, 'profile')] : [])]),
+        interactive: Boolean(stdin.isTTY && stderr.isTTY) && !apiKeyStdin,
+        installHooks,
+        log: line => {
+            stderr.write(`${line}\n`);
+        },
+        hasWorkingKey: async () => {
+            try {
+                const auth = await resolveAuth(base);
+                if (auth.source !== 'profile')
+                    return false;
+                if (await personalAccount(auth))
+                    return true;
+                await createAssetHubClient({ apiKey: auth.apiKey, baseUrl: auth.baseUrl }).v2.listModels();
+                return true;
+            }
+            catch {
+                return false;
+            }
+        },
+        promptApiKey: () => promptHidden('AssetHub API key (input hidden): '),
+        login: ({ apiKey }) => quiet(() => {
+            const loginFlags = { ...base, profile: typeof base.profile === 'string' ? base.profile : defaultProfileName };
+            if (apiKey) {
+                delete loginFlags['api-key-stdin'];
+                loginFlags['api-key'] = apiKey;
+            }
+            return commandAuth('login', loginFlags);
+        }),
+        resolveAuth: async () => {
+            const auth = await resolveAuth(base);
+            return { apiKey: auth.apiKey, baseUrl: auth.baseUrl, profile: auth.profile, workspaceId: auth.workspaceId };
+        },
+        discoverWorkspace: async (auth) => {
+            try {
+                const capabilities = await createAssetHubClient({ apiKey: auth.apiKey, baseUrl: auth.baseUrl }).v2.getCapabilities();
+                return typeof capabilities.ownerId === 'string' && capabilities.ownerId ? capabilities.ownerId : undefined;
+            }
+            catch {
+                return undefined;
+            }
+        },
+        useWorkspace: async (id) => {
+            await quiet(() => commandWorkspace('use', ['workspace', 'use', id], base));
+            // A non-personal `workspace use` stores its key under a new default profile.
+            const saved = await readAuthConfig(getConfigPath(flags));
+            if (saved.defaultProfile)
+                base.profile = saved.defaultProfile;
+        },
+        listWorkspaces: async () => (await (await workspaceAuth(base, true)).client.list()).workspaces.map(item => ({ id: item.id, name: item.name })),
+        pickWorkspace: pickFromList,
+        confirm: confirmYes,
+        confirmOptIn: confirmNo,
+        canSaveSessions: auth => canUploadSessions(auth),
+        diagnose: () => diagnose({
+            resolveAuth: () => resolveAuth(base),
+            includeMcp: true,
+            timeoutMs: parsePositiveIntegerFlag(flags, 'timeout-ms', 15000),
+        }),
+    });
+    if (hasFlag(flags, 'json'))
+        print(result);
+    else {
+        stdout.write(formatSetupSummary(result));
+        if (result.exportLine)
+            stdout.write(`${result.exportLine}\n`);
+    }
+    if (!result.ok)
+        process.exitCode = 2;
+};
+// `env load` is what the macOS login agent installed by `setup` runs: it copies
+// the saved key into launchd's user environment so Claude and Codex, started
+// from the Dock, can expand ${ASSETHUB_API_KEY} in their MCP config.
+const commandEnv = async (subcommand, flags) => {
+    if (subcommand !== 'load')
+        throw new Error('Unknown env command. Use "env load".');
+    if (process.platform !== 'darwin')
+        throw new Error('env load is only needed on macOS.');
+    const auth = await resolveAuth(flags);
+    const loaded = await loadKeyIntoLaunchd(auth.apiKey, runProcess);
+    if (!loaded.ok)
+        throw new Error(loaded.detail);
+    stderr.write(`ASSETHUB_API_KEY loaded for apps (profile ${auth.profile}).\n`);
+};
 const commandFiles = async (subcommand, positionals, ctx) => {
     const mediaType = parseMediaType(requireFlag(ctx.flags, 'media-type'));
     if (subcommand === 'import') {
@@ -1874,11 +2068,32 @@ const executionOptions = (flags) => ({
         }
         : undefined,
 });
+/** How long `--wait` waits by default: the selected agent's run length. */
+const defaultWaitMs = (flags) => {
+    const extractor = getPartExtractorFlag(flags);
+    const option = extractor == null
+        ? undefined
+        : partExtractorByInput.get(normalizePartExtractorName(extractor));
+    return option?.apiValue === 'ah_agent_graph_harpy_assembly_v2'
+        ? harpyAssemblyV2WaitMs
+        : 900000;
+};
 const watchOptions = (flags) => ({
     intervalMs: parsePositiveIntegerFlag(flags, 'interval-ms', 5000),
-    timeoutMs: parsePositiveIntegerFlag(flags, 'timeout-ms', 900000),
-    onProgress: (run) => stderr.write(`[run] ${run.runId} ${run.status}\n`),
+    timeoutMs: parsePositiveIntegerFlag(flags, 'timeout-ms', defaultWaitMs(flags)),
+    onProgress: progressPrinter(run => `[run] ${run.runId} ${run.status}${run.progress ? ` · ${runProgressLine(run)}` : ''}`),
 });
+/** Prints a run's line only when it changed since the last poll. */
+const progressPrinter = (line) => {
+    let last = '';
+    return (run) => {
+        const next = line(run);
+        if (next === last)
+            return;
+        last = next;
+        stderr.write(`${next}\n`);
+    };
+};
 const readJsonArgument = async (value, flag) => {
     const text = value === '@-'
         ? await readStdin()
@@ -2049,14 +2264,7 @@ const commandNodeMesh = async (ctx, nodeId) => {
         throw new Error('Canvas node is unavailable; read canvas nodes again');
     const finishedOrBusy = (status) => ['complete', 'completed', 'generating'].includes(status ?? '');
     if (target.type === 'production-3d') {
-        if (!target.actions.includes('mesh.generate'))
-            throw new Error('Mesh generation is unavailable for this node');
-        const children = items.filter(item => item.sourceNodeId === nodeId &&
-            ['mesh-gen', 'preview-asset', 'part-group'].includes(item.type));
-        if (!children.length)
-            throw new Error('Production node has no saved part nodes to generate');
-        const pending = children.filter(item => item.actions.includes('mesh.generate') &&
-            !finishedOrBusy(item.meshStatus));
+        const plan = planProductionNodeMeshes(items, nodeId);
         const batchId = operationId ?? randomUUID();
         stderr.write(`[nodes] batch ${batchId}\n`);
         await printNodeMeshBatch(ctx, await executeNodeMeshBatch(session, {
@@ -2064,10 +2272,7 @@ const commandNodeMesh = async (ctx, nodeId) => {
             operationId: batchId,
             nodeId,
             body,
-            children: pending.map(item => item.nodeId),
-            skipped: children
-                .filter(item => !pending.includes(item))
-                .map(({ nodeId, meshStatus }) => ({ nodeId, meshStatus })),
+            ...plan,
         }));
         return;
     }
@@ -2163,6 +2368,49 @@ const meshRefinementModes = [
     'blender',
     'codex',
 ];
+const COMPOSER_V6_MODEL = 'part_composer_v6_auto_assemble';
+const COMPOSER_SKILL_ASSEMBLY_MODEL = 'part_composer_v5_1_skill_assembly';
+/** Same text the canvas node sends, so CLI and canvas runs read alike. */
+const COMPOSER_SKILL_ASSEMBLY_INSTRUCTION = 'Assemble these parts into one character with the Character Assembly skill.';
+const IDENTITY_TRANSFORM = [0, 0, 0, 0, 0, 0, 1, 1, 1, 1];
+const composerModelAliases = {
+    v6: COMPOSER_V6_MODEL,
+    'auto-assemble': COMPOSER_V6_MODEL,
+    'v5.1': COMPOSER_SKILL_ASSEMBLY_MODEL,
+    'skill-assembly': COMPOSER_SKILL_ASSEMBLY_MODEL,
+};
+/** Short names for the picker labels; anything else goes to the server as given. */
+const resolveComposerModel = (value) => value == null ? undefined : (composerModelAliases[value.toLowerCase()] ?? value);
+/**
+ * `--part <mesh-id>[=<part-image-id>]`: the image is the drawing the mesh was
+ * generated from. `<mesh-id>:<part-image-id>` (CLI 0.1.27+) is accepted too.
+ */
+const composerPartFlag = (value) => {
+    const raw = assertFlagHasValue(value, '--part <mesh-asset-id>[=<part-image-id>]');
+    const separator = raw.includes('=') ? '=' : ':';
+    const index = raw.indexOf(separator);
+    const assetId = index === -1 ? raw : raw.slice(0, index);
+    const partImageAssetId = index === -1 ? undefined : raw.slice(index + 1);
+    if (!assetId || partImageAssetId === '')
+        throw new Error('--part must be <mesh-asset-id> or <mesh-asset-id>=<part-image-id>');
+    return { assetId, ...(partImageAssetId ? { partImageAssetId } : {}) };
+};
+/** V6 waits up to 81 min on ml and V5.1's sandbox up to 2 h; 15 min cannot see either finish. */
+const defaultComposerWaitTimeout = (flags, timeoutMs) => {
+    if (!hasFlag(flags, 'timeout-ms'))
+        flags['timeout-ms'] = String(timeoutMs);
+};
+/** Discovery is the only model manifest; qualify duplicate IDs by their advertised provider. */
+const selectComposerAgentModel = (models, value) => {
+    if (!value)
+        return undefined;
+    const matches = models?.filter(model => value === `${model.provider}:${model.id}` || value === model.id) ?? [];
+    if (matches.length > 1)
+        throw new Error(`Ambiguous Composer agent model ${value}; use --agent-model ${matches.map(model => `${model.provider}:${model.id}`).join(' or ')}`);
+    if (!matches[0])
+        throw new Error('Composer agent model is not available; inspect composer refine --list-modes');
+    return matches[0];
+};
 const refinementTransforms = (value, flag) => {
     if (value == null || typeof value !== 'object' || Array.isArray(value))
         throw new Error(`${flag} must be an object keyed by part asset ID`);
@@ -2196,7 +2444,8 @@ const refinementModeInfo = (capabilities, mode) => {
 const refinementContinuationFields = [
     'assemblyPolicy', 'dressingGeneration', 'effort',
     'geometryBackend', 'referenceMode', 'calibration', 'referenceViews',
-    'geometrySources', 'reviewContext', 'background',
+    'geometrySources', 'reviewContext', 'background', 'skillSelection',
+    'skillAssembly', 'skillAssemblyReasoning',
 ];
 /** Reuse the frozen sources and latest editable scene, never replay server-only receipt fields. */
 const composerInputFromRun = (previous, operation) => {
@@ -2248,6 +2497,15 @@ const composerInputFromRun = (previous, operation) => {
             if (previous.refinement?.[key] != null)
                 input[key] = previous.refinement[key];
         }
+        const generation = input.dressingGeneration;
+        if (generation && Array.isArray(input.parts)) {
+            const sources = input.geometrySources;
+            input.dressingGeneration = { ...generation, sourceImageAssetIds: Object.fromEntries(input.parts.flatMap((part) => {
+                    const image = generation.sourceImageAssetIds[part.assetId] ??
+                        generation.sourceImageAssetIds[part.sourceAssetId ?? sources?.[part.assetId] ?? ''];
+                    return typeof image === 'string' ? [[part.assetId, image]] : [];
+                })) };
+        }
         const reviewContext = previous.refinement?.report?.reviewContext;
         if (reviewContext != null)
             input.reviewContext = reviewContext;
@@ -2260,7 +2518,7 @@ const composerInputFromRun = (previous, operation) => {
     }
     return input;
 };
-const commandComposerRefine = async (ctx) => {
+const commandComposerRefine = async (ctx, options = {}) => {
     const capabilities = await ctx.client.v2.getMeshRefinementModes();
     if (hasFlag(ctx.flags, 'list-modes')) {
         print(capabilities);
@@ -2285,10 +2543,32 @@ const commandComposerRefine = async (ctx) => {
         : previous
             ? composerInputFromRun(previous, 'mesh.refine')
             : {
-                parts: parts.map(parsePartFlagValue),
+                // Refinement reads meshes and the concept only; a paired part image is not sent.
+                parts: parts.map(value => ({ assetId: composerPartFlag(value).assetId })),
                 fullBodyImageAssetId: requireFlag(ctx.flags, 'reference'),
             };
     delete input.executionContext;
+    const skillAssembly = options.skillAssembly ||
+        hasFlag(ctx.flags, 'skill-assembly') ||
+        input.skillAssembly === 'assemble-character';
+    const reasoning = parseEnumFlag(ctx.flags, 'reasoning', [
+        'low',
+        'medium',
+        'high',
+    ]);
+    if (reasoning && !skillAssembly)
+        throw new Error('--reasoning applies only to Skill assembly (V5.1)');
+    if (skillAssembly) {
+        // The bundled skill runs alone in the codex sandbox; the server refuses anything else.
+        input.skillAssembly = 'assemble-character';
+        if (reasoning)
+            input.skillAssemblyReasoning = reasoning;
+        input.mode = 'codex';
+        input.instruction ??= COMPOSER_SKILL_ASSEMBLY_INSTRUCTION;
+        if (previous?.operation === 'mesh.compose')
+            delete input.agentRuntime;
+        defaultComposerWaitTimeout(ctx.flags, 125 * 60_000);
+    }
     if (!Array.isArray(input.parts) || input.parts.length === 0)
         throw new Error('Composer refinement requires parts[]');
     const refinementParts = input.parts.map((part, index) => {
@@ -2310,6 +2590,8 @@ const commandComposerRefine = async (ctx) => {
         capabilities.defaultMode;
     if (!meshRefinementModes.includes(mode))
         throw new Error(`--mode must be one of: ${meshRefinementModes.join(', ')}`);
+    if (skillAssembly && mode !== 'codex')
+        throw new Error('Skill assembly (V5.1) runs in codex mode only');
     if (mode !== 'codex' && (input.geometryBackend || input.referenceMode === 'all_angles' || input.background))
         throw new Error('Saved Blender-backend, all-angle or background settings require codex mode; use --input-json with settings matching the new mode');
     const modeInfo = refinementModeInfo(capabilities, mode);
@@ -2341,7 +2623,9 @@ const commandComposerRefine = async (ctx) => {
     const transformsFlag = getFlag(ctx.flags, 'transforms-json');
     const transforms = transformsFlag
         ? refinementTransforms(await readJsonArgument(transformsFlag, '--transforms-json'), '--transforms-json')
-        : refinementTransforms(input.transforms, 'refinement transforms');
+        : skillAssembly && input.transforms == null
+            ? Object.fromEntries(refinementParts.map(part => [part.assetId, IDENTITY_TRANSFORM]))
+            : refinementTransforms(input.transforms, 'refinement transforms');
     const referenceTransform = input.referenceTransform == null
         ? undefined
         : refinementTransform(input.referenceTransform, 'referenceTransform');
@@ -2354,10 +2638,25 @@ const commandComposerRefine = async (ctx) => {
         canvasId,
         operation: 'mesh.refine',
     });
+    const agentModelFlag = getFlag(ctx.flags, 'agent-model');
+    if (skillAssembly &&
+        agentModelFlag?.includes(':') &&
+        !agentModelFlag.startsWith('agents-api:'))
+        throw new Error('Skill assembly runs on an Agents API model');
+    const modelOption = selectComposerAgentModel(skillAssembly
+        ? capabilities.agentModels?.filter(model => model.provider === 'agents-api')
+        : capabilities.agentModels, agentModelFlag);
     const body = {
         ...Object.fromEntries(refinementContinuationFields
             .filter(key => input[key] !== undefined)
             .map(key => [key, input[key]])),
+        ...(modelOption
+            ? { agentRuntime: { provider: modelOption.provider, model: modelOption.id } }
+            : input.agentRuntime
+                ? {
+                    agentRuntime: input.agentRuntime,
+                }
+                : {}),
         parts: refinementParts,
         fullBodyImageAssetId: input.fullBodyImageAssetId,
         transforms,
@@ -2366,15 +2665,18 @@ const commandComposerRefine = async (ctx) => {
         ...(effort ? { effort: effort } : {}),
         instruction,
         ...(maxRounds != null ? { maxRounds } : {}),
-        ...(input.agentRuntime
-            ? { agentRuntime: input.agentRuntime }
-            : {}),
     };
     await printRecorded(ctx, await executeRecorded(session, 'mesh.refine', body, {
         ...executionOptions(ctx.flags),
         ...(previous ? { parentRunId: previous.runId } : {}),
     }));
 };
+/**
+ * The Part Composer node's "Optimize output" presets: the share of geometry the full-resolution bake keeps
+ * (textures become WebP). Assembly always uses its own shrunk parts; this only sizes the delivered mesh.
+ */
+const COMPOSER_OPTIMIZE_RATIOS = { light: 0.75, medium: 0.5, heavy: 0.25 };
+const composerOptimizePresets = ['off', 'light', 'medium', 'heavy'];
 const commandComposer = async (subcommand, ctx) => {
     if (subcommand === 'models') {
         print(await ctx.client.v2.getMeshComposers());
@@ -2385,8 +2687,29 @@ const commandComposer = async (subcommand, ctx) => {
         return;
     }
     if (subcommand !== 'run')
-        throw new Error('Use composer models|run');
+        throw new Error('Use composer models|refine|run');
+    const model = resolveComposerModel(getFlag(ctx.flags, 'model'));
+    if (model === COMPOSER_SKILL_ASSEMBLY_MODEL) {
+        // V5.1 is not a compose model: the server runs it as a codex refinement.
+        if (hasFlag(ctx.flags, 'node'))
+            throw new Error('Skill assembly (V5.1) takes --part and --reference, not --node');
+        if (hasFlag(ctx.flags, 'mode'))
+            throw new Error('Skill assembly (V5.1) has no quick/quality mode');
+        if (hasFlag(ctx.flags, 'optimize'))
+            throw new Error('--optimize applies to composition, not Skill assembly (V5.1)');
+        await commandComposerRefine(ctx, { skillAssembly: true });
+        return;
+    }
+    if (hasFlag(ctx.flags, 'reasoning'))
+        throw new Error('--reasoning applies only to Skill assembly (V5.1)');
+    const optimizeFlag = getFlag(ctx.flags, 'optimize');
+    if (optimizeFlag != null &&
+        !composerOptimizePresets.includes(optimizeFlag))
+        throw new Error('--optimize must be off, light, medium or heavy');
+    const optimize = optimizeFlag;
     const nodeId = canvasNodeFlag(ctx.flags);
+    if (optimize && nodeId)
+        throw new Error("--optimize is not available with --node; set the node's Optimize output instead");
     const fromRun = getFlag(ctx.flags, 'from-run');
     const json = getFlag(ctx.flags, 'input-json');
     const parts = getFlagValues(ctx.flags, 'part');
@@ -2403,7 +2726,7 @@ const commandComposer = async (subcommand, ctx) => {
             : json
                 ? await readJsonArgument(json, '--input-json')
                 : {
-                    parts: parts.map(parsePartFlagValue),
+                    parts: parts.map(composerPartFlag),
                     fullBodyImageAssetId: requireFlag(ctx.flags, 'reference'),
                 };
     delete input.executionContext;
@@ -2417,9 +2740,34 @@ const commandComposer = async (subcommand, ctx) => {
         throw new Error('--mode must be quick or quality');
     if (mode)
         input.mode = mode;
-    const model = getFlag(ctx.flags, 'model');
     if (model)
         input.agentVersion = model;
+    if (input.agentVersion === COMPOSER_V6_MODEL) {
+        // V6 re-plans every part from its own drawing; ml fails the run without one.
+        const missing = (Array.isArray(input.parts) ? input.parts : [])
+            .filter((part) => !part.partImageAssetId && !part.partTaskId)
+            .map((part) => part.assetId);
+        if (missing.length)
+            throw new Error(`Auto assemble (V6) needs each part's drawing; pass --part <mesh-id>=<part-image-id> for: ${missing.join(', ')}`);
+        defaultComposerWaitTimeout(ctx.flags, 95 * 60_000);
+    }
+    const agentModel = getFlag(ctx.flags, 'agent-model');
+    if (agentModel) {
+        const capabilities = await ctx.client.v2.getMeshRefinementModes();
+        const selected = selectComposerAgentModel(capabilities.agentModels, agentModel);
+        input.agentRuntime = { provider: selected.provider, model: selected.id };
+    }
+    if (optimize) {
+        // --from-run rebuilds its input without the base body, so the previous run is asked too.
+        const previousSource = previous?.resolvedInput ?? previous?.input ?? previous?.requestedInput;
+        if (optimize !== 'off' &&
+            (input.baseBodyAssetId || previousSource?.baseBodyAssetId))
+            throw new Error('--optimize cannot be combined with a base body (baseBodyAssetId)');
+        if (optimize === 'off')
+            delete input.decimateRatio;
+        else
+            input.decimateRatio = COMPOSER_OPTIMIZE_RATIOS[optimize];
+    }
     const name = getFlag(ctx.flags, 'name');
     if (name)
         input.projectName = name;
@@ -3276,38 +3624,211 @@ const commandProductionRunOneShot = async (ctx) => {
     })));
     process.exitCode = executionExitCode(finalExecution);
 };
+const productionBatchImageExtensions = new Set(['.png', '.jpg', '.jpeg', '.webp']);
+const productionBatchUnsupportedSources = [
+    'image-url', 'url', 'source-resource-id', 'resource-id', 'file-ref-json',
+    'stdin', 'stdin-base64', 'stdin-data-uri', 'stdin-json', 'source-json',
+    'data-uri', 'clipboard', 'upload-id',
+];
+/** Images in command order: --files-dir (by name), --file, --source-id, --source-url. */
+const productionBatchSources = async (flags) => {
+    const unsupported = productionBatchUnsupportedSources.find(name => hasFlag(flags, name));
+    if (unsupported)
+        throw new Error(`production batch does not take --${unsupported}; pass images with --file, --files-dir, --source-id or --source-url`);
+    const sources = [];
+    for (const dir of getFlagValues(flags, 'files-dir')) {
+        const names = (await readdir(resolve(dir), { withFileTypes: true }))
+            .filter(entry => entry.isFile() && productionBatchImageExtensions.has(extname(entry.name).toLowerCase()))
+            .map(entry => entry.name)
+            .sort();
+        if (names.length === 0)
+            throw new Error(`--files-dir ${dir} has no .png, .jpg, .jpeg or .webp images`);
+        sources.push(...names.map(name => ({ flag: 'file', value: resolve(dir, name) })));
+    }
+    for (const flag of ['file', 'source-id', 'source-url'])
+        for (const value of getFlagValues(flags, flag))
+            sources.push({ flag, value: flag === 'file' ? resolve(value) : value });
+    if (sources.length === 0)
+        throw new Error('production batch needs images: --file <path>... , --files-dir <dir>, --source-id <id>... or --source-url <url>...');
+    return sources;
+};
+/**
+ * One `production analyze` per image (× --repeat) with the same agent, on one
+ * canvas. The batch is pinned locally before the first paid request, so
+ * rerunning with its --operation-id resumes it and never starts an item twice.
+ */
+const commandProductionBatch = async (ctx) => {
+    const sources = await productionBatchSources(ctx.flags);
+    const repeat = parsePositiveIntegerFlag(ctx.flags, 'repeat', 1);
+    const entries = expandProductionBatch(sources, repeat);
+    const labels = productionBatchLabels(sources);
+    if (entries.length > 1 && !hasFlag(ctx.flags, 'yes'))
+        throw new Error(`production batch would start ${entries.length} paid runs (${sources.length} image(s) × ${repeat}). Pass --yes to start them.`);
+    const { options, projectContext } = await resolveProductionAnalyzeOptions(ctx);
+    const { agent, parentRunId } = executionOptions(ctx.flags);
+    const operationId = getFlag(ctx.flags, 'operation-id') ?? randomUUID();
+    const state = stateOptions(ctx);
+    const saved = await readProductionBatch(state.stateDir, operationId);
+    const requestedCanvasId = selectedCanvasId(ctx.flags) ?? projectContext?.canvasId;
+    if (saved && requestedCanvasId != null && requestedCanvasId !== saved.canvasId)
+        throw new Error(`Batch ${operationId} runs on canvas ${saved.canvasId}; drop --canvas or pass that canvas`);
+    const session = await productionSession(ctx, 'production.analyze', saved?.canvasId ?? requestedCanvasId);
+    stderr.write(`[batch] ${operationId} · ${entries.length} run(s) on canvas ${session.canvas.id}. Rerun with --operation-id ${operationId} to resume.\n`);
+    const wait = hasFlag(ctx.flags, 'wait');
+    const concurrency = parsePositiveIntegerFlag(ctx.flags, 'concurrency', 2);
+    const printItem = (item) => stderr.write(`[batch] ${item.key} ${item.label} ${item.status}${item.runId ? ` ${item.runId}` : ''}${item.execution?.progress ? ` · ${runProgressLine(item.execution)}` : ''}${item.message ? ` · ${item.message}` : ''}\n`);
+    const result = await runProductionBatch(session, {
+        operationId,
+        request: { sources, repeat, options, agent, parentRunId },
+        agent,
+        parentRunId,
+        buildItems: async () => {
+            const resolved = [];
+            for (const source of sources) {
+                const flags = { ...ctx.flags };
+                for (const flag of ['file', 'files-dir', 'source-id', 'source-url'])
+                    delete flags[flag];
+                resolved.push(await resolveProductionAnalyzeSource({ ...ctx, flags: { ...flags, [source.flag]: source.value } }));
+            }
+            return entries.map((entry) => {
+                const label = labels[entry.index];
+                const name = `${options.name ? `${options.name} ` : ''}${label}${repeat > 1 ? ` #${entry.repeat}` : ''}`;
+                return {
+                    key: entry.key,
+                    label,
+                    imageIndex: entry.index,
+                    repeat: entry.repeat,
+                    attempt: 0,
+                    body: { ...resolved[entry.index], ...options, name },
+                };
+            });
+        },
+        ...(wait
+            ? {
+                wait: {
+                    concurrency,
+                    intervalMs: parsePositiveIntegerFlag(ctx.flags, 'interval-ms', 5000),
+                    timeoutMs: parsePositiveIntegerFlag(ctx.flags, 'timeout-ms', defaultWaitMs(ctx.flags) * Math.ceil(entries.length / concurrency)),
+                    onProgress: printItem,
+                },
+            }
+            : {}),
+    });
+    if (!wait)
+        result.items.forEach(printItem);
+    const downloads = [];
+    if (hasFlag(ctx.flags, 'download')) {
+        const outDir = getFlag(ctx.flags, 'out-dir') ?? getFlag(ctx.flags, 'out');
+        for (const item of result.items) {
+            if (!item.execution?.outputs.length)
+                continue;
+            const downloaded = await maybeDownloadUrls({ ...ctx.flags, 'out-dir': join(outDir ?? '', item.label, `run-${item.repeat}`) }, item.execution.outputs);
+            if (downloaded)
+                downloads.push({ key: item.key, ...downloaded });
+        }
+    }
+    const count = (status) => result.items.filter(item => item.status === status).length;
+    print({
+        batchOperationId: result.batchOperationId,
+        canvasId: result.canvasId,
+        agentVersion: options.agentVersion,
+        items: result.items,
+        counts: {
+            dispatched: count('dispatched'),
+            deferred: count('deferred'),
+            pending: count('pending'),
+            failed: count('failed'),
+            finished: result.items.filter(item => item.execution && !['dispatched', 'deferred', 'failed', 'pending'].includes(item.status)).length,
+        },
+        ...(result.timedOut ? { timedOut: true } : {}),
+        ...(downloads.length ? { downloads } : {}),
+    });
+    process.exitCode = Math.max(0, ...result.items.map(item => item.status === 'failed'
+        ? 1
+        : item.status === 'deferred' || item.status === 'pending'
+            ? 3
+            : wait && item.execution && item.status !== 'dispatched'
+                ? executionExitCode(item.execution)
+                : wait
+                    ? 3
+                    : 0));
+};
+/**
+ * Everything `production analyze` sends besides the image source; shared with
+ * `production batch` so a batch item is the same request.
+ */
+const resolveProductionAnalyzeOptions = async (ctx) => {
+    const agentVersion = resolvePartExtractorForAnalyze(ctx.flags);
+    const baseBodyAssetId = resolveProductionBaseBody(ctx.flags);
+    const skillPlannerModel = resolveSkillPlannerModel(ctx.flags);
+    const autoRepair = parseBooleanFlag(ctx.flags, 'auto-repair');
+    const skillSelection = resolveProductionSkillSelection(ctx.flags, agentVersion);
+    if (skillPlannerModel !== undefined && !['auto', 'manual'].includes(skillSelection?.mode ?? ''))
+        throw new Error('--skill-planner-model requires --skill-mode auto or manual');
+    const projectContext = hasFlag(ctx.flags, 'context') || hasFlag(ctx.flags, 'context-version')
+        ? {
+            canvasId: Number(requireFlag(ctx.flags, 'context')),
+            version: Number(requireFlag(ctx.flags, 'context-version')),
+        }
+        : undefined;
+    if (projectContext) {
+        if (Object.values(projectContext).some(value => !Number.isSafeInteger(value) || value <= 0))
+            throw new Error('--context and --context-version must be positive integers');
+        const canvasId = selectedCanvasId(ctx.flags);
+        if (canvasId != null && canvasId !== projectContext.canvasId)
+            throw new Error('--context must match --canvas');
+    }
+    const pipelineDepth = getFlag(ctx.flags, 'pipeline-depth');
+    if (pipelineDepth !== undefined &&
+        pipelineDepth !== 'parts' &&
+        pipelineDepth !== 'mesh' &&
+        pipelineDepth !== 'composition')
+        throw new Error('--pipeline-depth must be parts, mesh or composition');
+    const assemblyPolicy = getFlag(ctx.flags, 'assembly-policy');
+    const meshGenerationInput = getFlag(ctx.flags, 'mesh-generation-json');
+    const meshGeneration = meshGenerationInput === undefined
+        ? undefined
+        : await readJsonArgument(meshGenerationInput, '--mesh-generation-json');
+    if (assemblyPolicy !== undefined &&
+        assemblyPolicy !== 'concept-to-character-v1')
+        throw new Error('--assembly-policy must be concept-to-character-v1');
+    if (assemblyPolicy && pipelineDepth !== 'composition')
+        throw new Error('--assembly-policy requires --pipeline-depth composition');
+    return {
+        projectContext,
+        options: {
+            agentVersion,
+            name: getFlag(ctx.flags, 'name'),
+            ...(baseBodyAssetId === undefined ? {} : { baseBodyAssetId }),
+            ...(skillPlannerModel === undefined ? {} : { skillPlannerModel }),
+            ...(autoRepair === undefined ? {} : { autoRepair }),
+            ...(skillSelection == null ? {} : { skillSelection }),
+            ...(projectContext == null ? {} : { projectContext }),
+            ...(pipelineDepth === undefined ? {} : { pipelineDepth }),
+            ...(assemblyPolicy === undefined ? {} : { assemblyPolicy }),
+            ...(meshGeneration === undefined ? {} : { meshGeneration }),
+        },
+    };
+};
 const commandProduction = async (subcommand, positionals, ctx) => {
     if (subcommand === 'agents') {
         print(sanitizeProductionOutput(await ctx.client.v1.getProductionAgents()));
         return;
     }
     if (subcommand === 'analyze') {
-        const agentVersion = resolvePartExtractorForAnalyze(ctx.flags);
-        const skillSelection = resolveProductionSkillSelection(ctx.flags, agentVersion);
-        const projectContext = hasFlag(ctx.flags, 'context') || hasFlag(ctx.flags, 'context-version')
-            ? {
-                canvasId: Number(requireFlag(ctx.flags, 'context')),
-                version: Number(requireFlag(ctx.flags, 'context-version')),
-            }
-            : undefined;
-        if (projectContext) {
-            if (Object.values(projectContext).some(value => !Number.isSafeInteger(value) || value <= 0))
-                throw new Error('--context and --context-version must be positive integers');
-            const canvasId = selectedCanvasId(ctx.flags);
-            if (canvasId != null && canvasId !== projectContext.canvasId)
-                throw new Error('--context must match --canvas');
-        }
+        const { options, projectContext } = await resolveProductionAnalyzeOptions(ctx);
         const session = await productionSession(ctx, 'production.analyze', projectContext?.canvasId);
         const analyzed = await executeRecorded(session, 'production.analyze', {
             ...(await resolveProductionAnalyzeSource(ctx)),
-            agentVersion,
-            name: getFlag(ctx.flags, 'name'),
-            ...(skillSelection == null ? {} : { skillSelection }),
-            ...(projectContext == null ? {} : { projectContext }),
+            ...options,
         }, executionOptions(ctx.flags));
         await printRecorded(ctx, analyzed);
         return;
     }
+    if (subcommand === 'batch') {
+        await commandProductionBatch(ctx);
+        return;
+    }
     if (subcommand === 'automation') {
         const session = await productionSession(ctx, 'production.automation');
         const input = getFlag(ctx.flags, 'input-json');
@@ -3430,6 +3951,39 @@ const commandProduction = async (subcommand, positionals, ctx) => {
         print(await withOptionalDownloads(ctx.flags, sanitizeProductionOutput({ started, status })));
         return;
     }
+    if (subcommand === 'resume') {
+        const id = requirePositional(positionals, 2, 'run-id|order-id');
+        const expected = getFlag(ctx.flags, 'expected-resume-count');
+        const expectedResumeCount = expected === undefined ? undefined : Number(expected);
+        if (expectedResumeCount !== undefined &&
+            (!Number.isSafeInteger(expectedResumeCount) || expectedResumeCount < 0))
+            throw new Error('--expected-resume-count must be a whole number');
+        // A run ID resumes that run's graph and can be followed with --wait; an
+        // order ID resumes the graph only. Without --wait a failed run lookup
+        // (not a run, or no access to runs) falls back to the order ID as before.
+        const wait = hasFlag(ctx.flags, 'wait');
+        const run = await ctx.client.v2.getRun(id).catch((error) => {
+            if (!wait ||
+                (error instanceof AssetHubApiError && error.status === 404))
+                return undefined;
+            throw error;
+        });
+        const resumed = await ctx.client.v2.resumeProduction(run?.orderIds[0] ?? id, expectedResumeCount === undefined ? {} : { expectedResumeCount });
+        if (!wait) {
+            print(run ? { ...resumed, runId: run.runId } : resumed);
+            return;
+        }
+        if (!run)
+            throw new Error('--wait follows a run: pass the run ID (runs get shows it) instead of the order ID.');
+        const execution = await waitForExecution(ctx.client, run.runId, {
+            ...watchOptions(ctx.flags),
+            timeoutMs: parsePositiveIntegerFlag(ctx.flags, 'timeout-ms', harpyAssemblyV2WaitMs),
+        });
+        const downloads = await maybeDownloadUrls(ctx.flags, execution.outputs);
+        print({ resume: resumed, execution, ...(downloads ? { downloads } : {}) });
+        process.exitCode = executionExitCode(execution);
+        return;
+    }
     if (subcommand === 'interventions') {
         const orderId = requirePositional(positionals, 2, 'order-id');
         print(await ctx.client.v2.listInterventions(orderId));
@@ -3465,7 +4019,7 @@ const commandProduction = async (subcommand, positionals, ctx) => {
         print(await withOptionalDownloads(ctx.flags, sanitizeProductionOutput({ status })));
         return;
     }
-    throw new Error('Unknown production command. Use "production agents", "production analyze", "production automation", "production run", "production watch", "production status", "production execute", "production intervene", or "production interventions".');
+    throw new Error('Unknown production command. Use "production agents", "production analyze", "production batch", "production automation", "production run", "production watch", "production status", "production execute", "production intervene", "production interventions", or "production resume".');
 };
 // Multi-phase commands may contain several receipts; only their durable outputs
 // are downloadable, never canvas links, source URLs or evaluation evidence URLs.
@@ -3491,8 +4045,17 @@ const isArtifactGraphPartSplit = (value, execution) => (value != null &&
     artifactGraphPartExtractorApiValues.has(String(execution.resolvedInput?.agentVersion));
 const commandParts = async (subcommand, ctx) => {
     const runPartSplit = async (source, operationId = getFlag(ctx.flags, 'operation-id') ?? randomUUID()) => {
+        const autoRepair = parseBooleanFlag(ctx.flags, 'auto-repair');
+        const baseBodyAssetId = resolveProductionBaseBody(ctx.flags);
+        const skillPlannerModel = resolveSkillPlannerModel(ctx.flags);
         let orderId = source == null ? getFlag(ctx.flags, 'order-id') : undefined;
         if (orderId != null) {
+            if (autoRepair !== undefined)
+                throw new Error('--auto-repair can only be set when starting a new production run');
+            if (baseBodyAssetId !== undefined)
+                throw new Error('--base-body can only be set when starting a new production run');
+            if (skillPlannerModel !== undefined)
+                throw new Error('--skill-planner-model can only be set when starting a new production run');
             orderId = assertFlagHasValue(orderId, '--order-id <id>');
             if (hasFlag(ctx.flags, 'skill-mode') || hasFlag(ctx.flags, 'skill-id'))
                 throw new Error('--skill-mode and --skill-id cannot be changed with --order-id; use runs resume to preserve the saved selection');
@@ -3501,9 +4064,11 @@ const commandParts = async (subcommand, ctx) => {
         const requestedAgentVersion = orderId == null || partExtractorFlag != null
             ? resolvePartExtractorApiValue(partExtractorFlag ?? defaultPartExtractorName)
             : undefined;
-        const skillSelection = requestedAgentVersion == null
-            ? undefined
-            : resolveProductionSkillSelection(ctx.flags, requestedAgentVersion);
+        const skillSelection = orderId == null && requestedAgentVersion != null
+            ? resolveProductionSkillSelection(ctx.flags, requestedAgentVersion)
+            : undefined;
+        if (skillPlannerModel !== undefined && !['auto', 'manual'].includes(skillSelection?.mode ?? ''))
+            throw new Error('--skill-planner-model requires --skill-mode auto or manual');
         if (requestedAgentVersion != null &&
             artifactGraphPartExtractorApiValues.has(requestedAgentVersion)) {
             const publicName = publicPartExtractorName(requestedAgentVersion);
@@ -3523,6 +4088,9 @@ const commandParts = async (subcommand, ctx) => {
                 ...(source ?? (await resolveProductionAnalyzeSource(ctx))),
                 agentVersion: resolvePartExtractorForAnalyze(ctx.flags),
                 name: getFlag(ctx.flags, 'name'),
+                ...(baseBodyAssetId === undefined ? {} : { baseBodyAssetId }),
+                ...(skillPlannerModel === undefined ? {} : { skillPlannerModel }),
+                ...(autoRepair === undefined ? {} : { autoRepair }),
                 ...(skillSelection == null ? {} : { skillSelection }),
             }, { ...executionOptions(ctx.flags), operationId });
             analyzed = result;
@@ -3538,7 +4106,7 @@ const commandParts = async (subcommand, ctx) => {
             if (!orderId)
                 throw new CliExecutionError('Analysis receipt omitted its order', 3, operationId, execution);
             stderr.write(`[run] ${execution.runId} order=${orderId}\n`);
-            const executeOperationId = childOperationId(operationId, 'execute');
+            const executeOperationId = await childOperationId(operationId, 'execute');
             if ((hasFlag(ctx.flags, 'all-ready') ||
                 getFlagValues(ctx.flags, 'task-id').length > 0) &&
                 (await hasRecordedOperation(session.stateDir, executeOperationId))) {
@@ -3613,7 +4181,7 @@ const commandParts = async (subcommand, ctx) => {
             }, {
                 ...executionOptions(ctx.flags),
                 operationId: execution
-                    ? childOperationId(operationId, 'execute')
+                    ? await childOperationId(operationId, 'execute')
                     : operationId,
                 parentRunId: execution?.runId ?? executionOptions(ctx.flags).parentRunId,
             });
@@ -3682,7 +4250,7 @@ const commandParts = async (subcommand, ctx) => {
         const direct = await runPartSplit(durableInput, compareOperationId);
         let preprocessed = null;
         if (preprocessPrompt != null && preprocessPrompt !== 'true') {
-            const preprocessOperationId = childOperationId(compareOperationId, 'preprocess');
+            const preprocessOperationId = await childOperationId(compareOperationId, 'preprocess');
             try {
                 const queued = await executeRecorded(compareSession, 'image.generate', {
                     source: sourceFromProductionAnalyzeSource(durableInput),
@@ -3719,7 +4287,7 @@ const commandParts = async (subcommand, ctx) => {
                         job,
                         source: generatedSource,
                     },
-                    split: await runPartSplit(generatedSource, childOperationId(compareOperationId, 'preprocessed-split')),
+                    split: await runPartSplit(generatedSource, await childOperationId(compareOperationId, 'preprocessed-split')),
                 };
             }
             catch (error) {
@@ -3757,6 +4325,11 @@ const canvasForRead = (ctx) => openExecutionSession({
     create: false,
 });
 const commandCanvas = async (subcommand, positionals, ctx) => {
+    if (subcommand === 'graph-id') {
+        const canvasId = Number(requirePositional(positionals, 2, 'canvas-id'));
+        print(await ctx.client.v2.getCanvasGraphIds(canvasId));
+        return;
+    }
     if (subcommand === 'nodes') {
         const { canvas } = await canvasForRead(ctx);
         print(await ctx.client.v2.listCanvasNodes(canvas.id));
@@ -3968,7 +4541,44 @@ const commandMoodboard = async (subcommand, positionals, ctx) => {
     if (revision.status === 'failed')
         process.exitCode = 1;
 };
-const commandGraph = async (subcommand, ctx) => {
+export const commandGraph = async (subcommand, positionals, ctx) => {
+    if (subcommand === 'export-canvas') {
+        await runInternalCommand(() => runExportCanvas(ctx));
+        return;
+    }
+    // Memorized-graph reads: the same store the Skill Builder agent's own tools
+    // read, distinct from `graph show|lineage|export` below (the
+    // authority/generated store via getGraph/getCanvasGraph).
+    if (subcommand === 'snapshot') {
+        const id = requirePositional(positionals, 2, 'graph-id');
+        print(await ctx.client.v2.getGraphSnapshot(id, {
+            offset: parseNonNegativeIntegerFlag(ctx.flags, 'offset'),
+            limit: parsePositiveIntegerFlag(ctx.flags, 'limit', 80),
+        }));
+        return;
+    }
+    if (subcommand === 'node') {
+        const id = requirePositional(positionals, 2, 'graph-id');
+        const nodeId = requirePositional(positionals, 3, 'node-id');
+        print(await ctx.client.v2.getGraphNode(id, nodeId));
+        return;
+    }
+    if (subcommand === 'image') {
+        const id = requirePositional(positionals, 2, 'graph-id');
+        const nodeId = requirePositional(positionals, 3, 'node-id');
+        const image = await ctx.client.v2.getGraphNodeImage(id, nodeId);
+        const out = getFlag(ctx.flags, 'out');
+        if (out) {
+            await mkdir(dirname(resolve(out)), { recursive: true });
+            await writeFile(resolve(out), Buffer.from(image.data, 'base64'));
+            print({ mediaType: image.mediaType, bytes: image.bytes, path: resolve(out) });
+            return;
+        }
+        // Without --out, print the metadata WITHOUT the base64 blob: a megabyte
+        // of base64 in a terminal is not a result.
+        print({ mediaType: image.mediaType, bytes: image.bytes });
+        return;
+    }
     const graphId = getFlag(ctx.flags, 'graph');
     if (getFlag(ctx.flags, 'source') && !graphId)
         throw new Error('--source requires --graph');
@@ -3988,7 +4598,10 @@ const commandGraph = async (subcommand, ctx) => {
         : getFlag(ctx.flags, 'artifact');
     const graph = graphId
         ? await ctx.client.v2.getGraph(graphId, {
-            source: parseEnumFlag(ctx.flags, 'source', ['generated', 'upload']),
+            source: parseEnumFlag(ctx.flags, 'source', [
+                'generated',
+                'upload',
+            ]),
             artifactId,
             direction: direction,
             depth: parsePositiveIntegerFlag(ctx.flags, 'depth', 5),
@@ -4097,7 +4710,8 @@ const commandRuns = async (subcommand, positionals, ctx) => {
             stdout.write(`${formatExecutionSummaryLines(execution).join('\n')}\n`);
         }
         else {
-            print({ execution });
+            const downloads = await maybeDownloadUrls(ctx.flags, execution.outputs);
+            print({ execution, ...(downloads ? { downloads } : {}) });
         }
         process.exitCode = executionExitCode(execution);
         return;
@@ -4217,6 +4831,734 @@ const parseNonNegativeIntegerFlag = (flags, name) => {
     }
     return parsed;
 };
+/**
+ * `assethub memory memorize` and `assethub memory replay`.
+ *
+ * Both drive the trajectory-memory feature: memorize captures a canvas as a
+ * reusable recipe, replay re-runs a memory's recipe onto a new image. Neither
+ * route is documented at /docs/api — the four replay addresses are listed in
+ * `INTERNAL_ONLY_EXPOSED_API_PATHS` and memorize does not exist yet — so
+ * neither belongs in `CLI_OPERATIONS`. See the docblock on that file for the
+ * full reasoning.
+ *
+ * `@assethub/api-client` has no memory methods, so this module talks to the
+ * API directly with `fetch`, reusing the client's own envelope types and
+ * error class rather than inventing new ones.
+ */
+/** `POST /memory/memorize` rejects a `--nodes` selection larger than this. */
+const MAX_NODE_SELECTION = 256;
+/**
+ * How many times `/memory/start` may answer `preparing` in a row before this
+ * command gives up. Each round is a fresh pick-scoped search dispatched by the
+ * server, not a client retry, so a bound here is what stops a
+ * `preparing` -> `preparing` cycle from spinning forever.
+ */
+const MAX_PREPARE_ROUNDS = 5;
+/** Default `--timeout`, in seconds. Matches `pollJob`'s 900_000ms default. */
+const DEFAULT_TIMEOUT_SECONDS = 900;
+/** Spacing between polls of a search, a pick-scoped search, or a run. */
+const MEMORY_POLL_INTERVAL_MS = 3000;
+/**
+ * Every status this module treats as an input, auth, or capability problem
+ * rather than a platform fault — see the "Exit codes" section of the brief.
+ * A status outside this set (5xx, network) is left to propagate to the
+ * top-level catch, which is the existing exit(1) path every other command
+ * already uses for an infrastructure failure.
+ */
+const CALLER_ERROR_STATUSES = new Set([400, 401, 403, 404, 409, 410, 422]);
+/**
+ * Why 404 and 403 are not told apart here, and must not be.
+ *
+ * A canvas belonging to another organization answers 404, not 403 — the same
+ * answer a canvas that does not exist gets, and the same choice `/memory/start`
+ * makes. Canvas ids are sequential enough that a distinguishable 403 would let
+ * a caller walk the id space and learn which ones exist, so the server refuses
+ * to be that oracle.
+ *
+ * The CLI's job is to not undo that. Both statuses land in the set above, both
+ * print the server's own message, and both exit 2. Do not add a branch that
+ * words them differently, and do not "improve" the message by asserting the
+ * canvas is missing — from here the two cases are genuinely indistinguishable,
+ * which is the point.
+ */
+const internalV2Request = async (ctx, method, path, body) => {
+    const baseUrl = ctx.auth.baseUrl.replace(/\/+$/, '');
+    const response = await fetch(`${baseUrl}/api/v2${path}`, {
+        method,
+        headers: {
+            Authorization: `Bearer ${ctx.auth.apiKey}`,
+            ...(ctx.auth.workspaceId ? { 'X-AssetHub-Workspace': ctx.auth.workspaceId } : {}),
+            'Content-Type': 'application/json',
+        },
+        body: body === undefined ? undefined : JSON.stringify(body),
+    });
+    const payload = (await response.json().catch(() => ({})));
+    if (!response.ok || payload.success !== true) {
+        const errorPayload = payload;
+        throw new AssetHubApiError({
+            status: response.status,
+            code: errorPayload.error?.code ?? 'UNKNOWN_ERROR',
+            message: errorPayload.error?.message ??
+                `AssetHub API request failed with status ${response.status}`,
+            payload,
+            requestId: errorPayload.error?.requestId,
+        });
+    }
+    return payload.data;
+};
+const parseNodeIds = (raw) => {
+    const nodeIds = raw
+        .split(',')
+        .map(id => id.trim())
+        .filter(id => id.length > 0);
+    if (nodeIds.length > MAX_NODE_SELECTION) {
+        throw new Error(`--nodes accepts at most ${MAX_NODE_SELECTION} node ids, got ${nodeIds.length}.`);
+    }
+    return nodeIds;
+};
+const runMemorize = async (ctx) => {
+    const canvasId = Number(requireFlag(ctx.flags, 'canvas'));
+    if (!Number.isSafeInteger(canvasId) || canvasId <= 0)
+        throw new Error('--canvas must be a positive integer');
+    const nodesFlag = getFlag(ctx.flags, 'nodes');
+    const selection = nodesFlag == null
+        ? { kind: 'whole' }
+        : { kind: 'nodes', nodeIds: parseNodeIds(nodesFlag) };
+    const started = await internalV2Request(ctx, 'POST', '/memory/memorize', {
+        canvasId,
+        selection,
+    });
+    if (!hasFlag(ctx.flags, 'wait')) {
+        print(started);
+        return;
+    }
+    const deadline = Date.now() +
+        parsePositiveIntegerFlag(ctx.flags, 'timeout', DEFAULT_TIMEOUT_SECONDS) *
+            1000;
+    let job = await internalV2Request(ctx, 'GET', `/memory/memorize/${started.memorizeJobId}`);
+    while (job.status !== 'completed' && job.status !== 'failed') {
+        if (Date.now() >= deadline) {
+            print({ ...started, job, timedOut: true });
+            process.exitCode = 3;
+            return;
+        }
+        stderr.write(`[poll] memorizeJobId=${started.memorizeJobId} status=${job.status}\n`);
+        await delay(MEMORY_POLL_INTERVAL_MS);
+        job = await internalV2Request(ctx, 'GET', `/memory/memorize/${started.memorizeJobId}`);
+    }
+    print({ ...started, job });
+    if (job.status === 'failed') {
+        process.exitCode = 1;
+    }
+};
+/**
+ * Resolves a `searching` match to a terminal one (`ready` or `failed`),
+ * polling `GET /memory/match/{matchId}` on `MEMORY_POLL_INTERVAL_MS`. Returns
+ * `undefined` once `deadline` passes without a terminal answer — the caller
+ * has already been told (`print` + `process.exitCode = 3`) and should return.
+ */
+const resolveMatch = async (ctx, match, deadline) => {
+    let current = match;
+    while (current.status === 'searching') {
+        if (Date.now() >= deadline) {
+            print({ matchId: current.matchId, status: 'searching', timedOut: true });
+            process.exitCode = 3;
+            return undefined;
+        }
+        stderr.write(`[poll] matchId=${current.matchId} status=searching\n`);
+        await delay(MEMORY_POLL_INTERVAL_MS);
+        current = await internalV2Request(ctx, 'GET', `/memory/match/${current.matchId}`);
+    }
+    return current;
+};
+const runReplay = async (ctx, positionals) => {
+    const memoryId = requirePositional(positionals, 2, 'memory-id');
+    const source = requireFlag(ctx.flags, 'source');
+    const wait = hasFlag(ctx.flags, 'wait');
+    const deadline = Date.now() +
+        parsePositiveIntegerFlag(ctx.flags, 'timeout', DEFAULT_TIMEOUT_SECONDS) *
+            1000;
+    const initialMatch = await internalV2Request(ctx, 'POST', '/memory/match', {
+        source: { resourceId: source },
+    });
+    const match = await resolveMatch(ctx, initialMatch, deadline);
+    if (match == null) {
+        return;
+    }
+    if (match.status === 'failed') {
+        print({ matchId: match.matchId, status: 'failed', reason: match.reason });
+        process.exitCode = 1;
+        return;
+    }
+    if (match.preparedMemoryId == null) {
+        print({
+            matchId: match.matchId,
+            status: 'ready',
+            matched: false,
+            message: 'The organization has curated no memories to replay.',
+            candidates: match.candidates ?? [],
+        });
+        return;
+    }
+    let matchId = match.matchId;
+    let dispatched;
+    for (let round = 0; round < MAX_PREPARE_ROUNDS; round += 1) {
+        const started = await internalV2Request(ctx, 'POST', '/memory/start', {
+            matchId,
+            memoryId,
+        });
+        if (started.status === 'queued') {
+            dispatched = started;
+            break;
+        }
+        // `preparing`: the server dispatched a fresh, pick-scoped search rather
+        // than a walk. Poll it to `ready`, then loop to call /start again.
+        stderr.write(`[poll] matchId=${started.matchId} status=preparing (pick-scoped search for ${memoryId})\n`);
+        const preparedMatch = await resolveMatch(ctx, await internalV2Request(ctx, 'GET', `/memory/match/${started.matchId}`), deadline);
+        if (preparedMatch == null) {
+            return;
+        }
+        if (preparedMatch.status === 'failed') {
+            print({
+                matchId: preparedMatch.matchId,
+                status: 'failed',
+                reason: preparedMatch.reason,
+            });
+            process.exitCode = 1;
+            return;
+        }
+        matchId = preparedMatch.matchId;
+    }
+    if (dispatched == null) {
+        print({
+            status: 'timeout',
+            message: `Could not confirm memoryId ${memoryId} after ${MAX_PREPARE_ROUNDS} pick-scoped search rounds.`,
+            matchId,
+        });
+        process.exitCode = 3;
+        return;
+    }
+    if (!wait) {
+        print(dispatched);
+        return;
+    }
+    let run = await internalV2Request(ctx, 'GET', `/memory/runs/${dispatched.runId}`);
+    while (run.status !== 'completed' && run.status !== 'failed') {
+        if (Date.now() >= deadline) {
+            print({ ...run, timedOut: true });
+            process.exitCode = 3;
+            return;
+        }
+        stderr.write(`[poll] runId=${run.runId} status=${run.status}\n`);
+        await delay(MEMORY_POLL_INTERVAL_MS);
+        run = await internalV2Request(ctx, 'GET', `/memory/runs/${dispatched.runId}`);
+    }
+    print(run);
+    if (run.status === 'failed') {
+        process.exitCode = 1;
+    }
+};
+export const commandMemory = async (subcommand, positionals, ctx) => {
+    try {
+        if (subcommand === 'memorize') {
+            await runMemorize(ctx);
+            return;
+        }
+        if (subcommand === 'replay') {
+            await runReplay(ctx, positionals);
+            return;
+        }
+        if (subcommand === 'match') {
+            // Finding a memory id is the step BEFORE `skills build --memory`:
+            // without it that source is only usable by someone who already knows
+            // the ids, which nobody does.
+            print(await ctx.client.v2.matchMemory({
+                goal: getFlag(ctx.flags, 'goal'),
+                imageAssetId: getFlag(ctx.flags, 'image-asset'),
+                imageUrl: getFlag(ctx.flags, 'image-url'),
+            }));
+            return;
+        }
+    }
+    catch (error) {
+        if (error instanceof AssetHubApiError &&
+            CALLER_ERROR_STATUSES.has(error.status)) {
+            print({
+                error: { code: error.code, message: error.message, status: error.status },
+            });
+            process.exitCode = 2;
+            return;
+        }
+        throw error;
+    }
+    throw new Error('Unknown memory command. Use "memory memorize" or "memory replay".');
+};
+/**
+ * Runs one internal-API command with the same caller-error boundary
+ * `commandMemory` uses: a 4xx in `CALLER_ERROR_STATUSES` prints the server's
+ * own error and exits 2 (the 404 a non-internal key gets included — see the
+ * note above `internalV2Request` on why it is not reworded); anything else
+ * propagates to the top-level catch.
+ */
+const runInternalCommand = async (operation) => {
+    try {
+        await operation();
+    }
+    catch (error) {
+        if (error instanceof AssetHubApiError &&
+            CALLER_ERROR_STATUSES.has(error.status)) {
+            print({
+                error: { code: error.code, message: error.message, status: error.status },
+            });
+            process.exitCode = 2;
+            return;
+        }
+        throw error;
+    }
+};
+/* ─── Canvas → artifact-graph export ─────────────────────────────────── */
+/** Mirrors `MAX_EXPORTS_PER_CALL` on the server: a longer POST is a 400, so
+ *  `export-canvas` chunks instead of letting the API refuse the batch. */
+const CANVAS_EXPORT_CHUNK_SIZE = 50;
+/** An export is a Trigger task that reads a room, builds a graph and zips its
+ *  blobs — minutes, not seconds — so poll slower than the memory family. */
+const CANVAS_EXPORT_POLL_INTERVAL_MS = 5000;
+/** Default `--timeout` for `export-canvas --wait`, in seconds. A 50-canvas
+ *  batch of large rooms can take a while; twice the memory default. */
+const CANVAS_EXPORT_DEFAULT_TIMEOUT_SECONDS = 1800;
+/** Per-file download deadline. The signal covers the WHOLE body, and a
+ *  canvas zip carries up to 400 MiB of blobs — the `canvas download` deadline
+ *  (120 s) would need >3 MB/s to finish one, so give a file the same budget
+ *  as the batch. */
+const CANVAS_EXPORT_DOWNLOAD_TIMEOUT_MS = 1_800_000;
+const CANVAS_EXPORT_TERMINAL_STATUSES = new Set(['ready', 'failed']);
+/** Global `setTimeout` rather than `timers/promises`' `delay`, so a test's
+ *  fake clock can drive the poll loop instead of waiting the real interval. */
+const sleep = (ms) => new Promise(resolvePromise => setTimeout(resolvePromise, ms));
+/** What `export-canvas` prints for a job: the status row WITHOUT the signed
+ *  archive URL. The zip is already on disk (or recorded under `failed`), and a
+ *  signed link in stdout is a credential in a log — same rule as
+ *  `canvas download`, which never writes signed URLs into its bundle. */
+const printableExportJob = (job) => {
+    if (!('download' in job) || job.download == null)
+        return job;
+    const { url: _url, ...download } = job.download;
+    return { ...job, download };
+};
+/** `--canvas 1,2,3` → `[1, 2, 3]`: positive integers, deduplicated, in order. */
+const parseCanvasIdList = (raw) => {
+    const ids = raw
+        .split(',')
+        .map(part => part.trim())
+        .filter(part => part.length > 0)
+        .map(part => {
+        const id = Number(part);
+        if (!Number.isInteger(id) || id <= 0)
+            throw new Error(`--canvas ids must be positive integers, got "${part}"`);
+        return id;
+    });
+    const unique = [...new Set(ids)];
+    if (unique.length === 0)
+        throw new Error('--canvas needs at least one canvas id');
+    return unique;
+};
+const chunk = (items, size) => {
+    const chunks = [];
+    for (let index = 0; index < items.length; index += size)
+        chunks.push(items.slice(index, index + size));
+    return chunks;
+};
+/**
+ * `assethub graph export-canvas --org <orgId> --canvas <id>[,<id>...]`.
+ *
+ * Without `--wait` it prints the queued jobs. With it, it polls every job to
+ * a terminal status, downloads each ready zip to `<out-dir>/<projectId>-rev<rev>.zip`
+ * and exits 1 if any job failed or any download failed, 3 on the deadline
+ * (with the partial job list printed), 0 otherwise.
+ */
+const runExportCanvas = async (ctx) => {
+    const orgId = requireFlag(ctx.flags, 'org');
+    const canvasIds = parseCanvasIdList(requireFlag(ctx.flags, 'canvas'));
+    const jobs = [];
+    for (const ids of chunk(canvasIds, CANVAS_EXPORT_CHUNK_SIZE)) {
+        const created = await internalV2Request(ctx, 'POST', '/artifact-graph/canvas-exports', { orgId, projectIds: ids });
+        jobs.push(...created.jobs);
+    }
+    // The API silently drops an id that is not a workflow canvas owned by
+    // `--org` (missing, foreign, wrong type). For a research batch that is an
+    // incomplete export, not a success: name the ids and exit 1.
+    const queuedIds = new Set(jobs.map(job => job.projectId));
+    const omitted = canvasIds.filter(id => !queuedIds.has(id));
+    if (omitted.length > 0)
+        process.exitCode = 1;
+    if (!hasFlag(ctx.flags, 'wait')) {
+        print({ jobs, omitted });
+        return;
+    }
+    const deadline = Date.now() +
+        parsePositiveIntegerFlag(ctx.flags, 'timeout', CANVAS_EXPORT_DEFAULT_TIMEOUT_SECONDS) *
+            1000;
+    const statuses = new Map();
+    const pending = () => jobs.filter(job => {
+        const status = statuses.get(job.jobId)?.status;
+        return status == null || !CANVAS_EXPORT_TERMINAL_STATUSES.has(status);
+    });
+    let remaining = pending();
+    while (remaining.length > 0) {
+        for (const job of remaining) {
+            statuses.set(job.jobId, await internalV2Request(ctx, 'GET', `/artifact-graph/canvas-exports/${job.jobId}`));
+        }
+        remaining = pending();
+        if (remaining.length === 0)
+            break;
+        if (Date.now() >= deadline) {
+            print({
+                jobs: jobs.map(job => printableExportJob(statuses.get(job.jobId) ?? job)),
+                omitted,
+                timedOut: true,
+            });
+            process.exitCode = 3;
+            return;
+        }
+        stderr.write(`[poll] ${remaining.length}/${jobs.length} canvas exports still running\n`);
+        await sleep(CANVAS_EXPORT_POLL_INTERVAL_MS);
+    }
+    const outDir = resolve(getFlag(ctx.flags, 'out-dir') ?? '.');
+    const downloaded = [];
+    const failed = [];
+    const finalJobs = jobs.map(job => statuses.get(job.jobId) ?? job);
+    for (const job of finalJobs) {
+        if (!('status' in job) || job.status !== 'ready')
+            continue;
+        const fileName = `${job.projectId ?? job.jobId}-rev${job.rev ?? 0}.zip`;
+        try {
+            // The poll stops re-reading a job once it is `ready`, so its signed URL
+            // dates from that moment; by the time a long batch reaches this file
+            // (every zip is downloaded after the LAST job settles) the hour may be
+            // over. Re-read the row for a URL signed now.
+            const fresh = await internalV2Request(ctx, 'GET', `/artifact-graph/canvas-exports/${job.jobId}`);
+            if (fresh.download == null)
+                throw new Error('No download URL is available');
+            const file = await downloadCanvasGraphZip({
+                url: fresh.download.url,
+                destinationPath: join(outDir, fileName),
+                timeoutMs: CANVAS_EXPORT_DOWNLOAD_TIMEOUT_MS,
+            });
+            downloaded.push({
+                jobId: job.jobId,
+                projectId: job.projectId,
+                rev: job.rev,
+                ...file,
+            });
+        }
+        catch (error) {
+            failed.push({
+                jobId: job.jobId,
+                projectId: job.projectId,
+                error: error instanceof Error ? error.message : 'File download failed',
+            });
+        }
+    }
+    print({ jobs: finalJobs.map(printableExportJob), omitted, downloaded, failed });
+    const anyJobFailed = finalJobs.some(job => 'status' in job && job.status === 'failed');
+    if (anyJobFailed || failed.length > 0 || omitted.length > 0)
+        process.exitCode = 1;
+};
+/**
+ * Poll a build until it stops changing.
+ *
+ * `ready` counts as done: the agent has finished and a person now has
+ * something to review. Waiting past it would block forever on a human.
+ */
+const waitForSkillBuild = async (ctx, buildId) => {
+    const deadline = Date.now() + 20 * 60_000;
+    for (;;) {
+        const build = await ctx.client.v2.getWorkspaceSkillBuild(buildId);
+        if (build.status === 'ready' ||
+            TERMINAL_WORKSPACE_SKILL_BUILD_STATUSES.includes(build.status))
+            return build;
+        if (Date.now() > deadline)
+            throw new Error(`Timed out waiting for build ${buildId}; it is still ${build.status}. Re-run: skills build-status ${buildId}`);
+        await new Promise(resolve => setTimeout(resolve, 5_000));
+    }
+};
+export const commandSkills = async (subcommand, positionals, ctx) => {
+    if (subcommand === 'list') {
+        print(await ctx.client.v2.listWorkspaceSkills({
+            cursor: getFlag(ctx.flags, 'cursor'),
+        }));
+        return;
+    }
+    if (subcommand === 'get') {
+        print(await ctx.client.v2.getWorkspaceSkill(requirePositional(positionals, 2, 'skill-id'), {
+            revision: hasFlag(ctx.flags, 'revision')
+                ? parsePositiveIntegerFlag(ctx.flags, 'revision', 1)
+                : undefined,
+        }));
+        return;
+    }
+    if (subcommand === 'memory') {
+        // `skills get` hands back workflowRef and evidence[] as bare ids. This
+        // resolves each one into the memory it points at, and gives every cited
+        // memory a memoryId that `memory replay` takes directly.
+        const skillId = requirePositional(positionals, 2, 'skill-id');
+        const links = await ctx.client.v2.getSkillMemoryLinks(skillId, {
+            revision: hasFlag(ctx.flags, 'revision')
+                ? parsePositiveIntegerFlag(ctx.flags, 'revision', 1)
+                : undefined,
+        });
+        const imagesDir = getFlag(ctx.flags, 'images');
+        if (!imagesDir) {
+            print(links);
+            return;
+        }
+        // Images are fetched one node at a time through the EXISTING image
+        // operation rather than inlined into the resolve response: a skill citing
+        // eight memories would otherwise be a multi-megabyte JSON body on every
+        // read. A node whose image is over that endpoint's cap answers 413 and is
+        // recorded as a failure here, not silently skipped.
+        const downloaded = [];
+        const failed = [];
+        // Every citing shape, flattened to the (graph, node) the image operation
+        // takes: workflowRef, evidence[], and a built skill's references[].nodes,
+        // whose graph id lives on the parent reference rather than the node.
+        const candidates = [
+            ...(links.workflowRef
+                ? [
+                    {
+                        ref: links.workflowRef,
+                        graphId: links.workflowRef.graphId,
+                        nodeId: links.workflowRef.nodeId,
+                    },
+                ]
+                : []),
+            ...links.evidence.map(ref => ({
+                ref,
+                graphId: ref.graphId ?? '',
+                nodeId: ref.artifactId ?? '',
+            })),
+            ...links.references.flatMap(reference => reference.nodes.map(node => ({
+                ref: node,
+                graphId: reference.graphId ?? '',
+                nodeId: node.nodeId,
+            }))),
+        ]
+            .filter(entry => entry.ref.status === 'resolved' &&
+            entry.ref.node?.hasImage === true &&
+            entry.graphId !== '' &&
+            entry.nodeId !== '')
+            .map(entry => ({
+            memoryId: entry.ref.memoryId ?? '',
+            graphId: entry.graphId,
+            nodeId: entry.nodeId,
+        }));
+        await mkdir(resolve(imagesDir), { recursive: true });
+        for (const ref of candidates) {
+            const memoryId = ref.memoryId;
+            try {
+                const image = await ctx.client.v2.getGraphNodeImage(ref.graphId, ref.nodeId);
+                // The memory id carries colons and, for a node id, slashes are
+                // possible — neither is safe in a file name.
+                const name = `${memoryId.replaceAll(/[^A-Za-z0-9._-]/gu, '_')}.${image.mediaType.split('/')[1] ?? 'bin'}`;
+                const path = resolve(imagesDir, name);
+                await writeFile(path, Buffer.from(image.data, 'base64'));
+                downloaded.push({ memoryId, path, bytes: image.bytes });
+            }
+            catch (error) {
+                failed.push({
+                    memoryId,
+                    error: error instanceof Error ? error.message : 'download failed',
+                });
+            }
+        }
+        print({ ...links, images: { downloaded, failed } });
+        if (failed.length > 0)
+            process.exitCode = 1;
+        return;
+    }
+    if (subcommand === 'controls') {
+        const mode = parseEnumFlag(ctx.flags, 'mode', [
+            'automatic',
+            'manual',
+            'off',
+        ]);
+        if (!mode)
+            throw new Error('Missing required flag: --mode');
+        requireFlag(ctx.flags, 'revision');
+        requireFlag(ctx.flags, 'grant-revision');
+        print(await ctx.client.v2.setWorkspaceSkillControls(requirePositional(positionals, 2, 'skill-id'), {
+            expectedRevision: parsePositiveIntegerFlag(ctx.flags, 'revision', 1),
+            expectedGrantRevision: parsePositiveIntegerFlag(ctx.flags, 'grant-revision', 1),
+            mode,
+            ...(hasFlag(ctx.flags, 'share') ? { share: true } : {}),
+        }));
+        return;
+    }
+    if (subcommand === 'proposal') {
+        print(await ctx.client.v2.getWorkspaceSkillProposal(requirePositional(positionals, 2, 'proposal-id')));
+        return;
+    }
+    if (['learn', 'prepare'].includes(subcommand ?? '')) {
+        const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
+        const operationId = subcommand === 'learn'
+            ? requireFlag(ctx.flags, 'operation-id')
+            : undefined;
+        if (operationId != null &&
+            !/^[0-9a-f]{8}-[0-9a-f]{4}-[1-8][0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$/i.test(operationId))
+            throw new Error('--operation-id must be a UUID; reuse the same ID and input after an uncertain response');
+        if (operationId != null)
+            activeExecution.operationId = operationId;
+        print(subcommand === 'learn'
+            ? await ctx.client.v2.learnWorkspaceSkill(input, { idempotencyKey: operationId })
+            : await ctx.client.v2.prepareWorkspaceSkillProposal(input));
+        return;
+    }
+    if (['update', 'accept'].includes(subcommand ?? '')) {
+        const id = requirePositional(positionals, 2, subcommand === 'update' ? 'skill-id' : 'proposal-id');
+        const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
+        print(subcommand === 'update'
+            ? await ctx.client.v2.updateWorkspaceSkill(id, input)
+            : await ctx.client.v2.acceptWorkspaceSkillProposal(id, input));
+        return;
+    }
+    if (subcommand === 'official') {
+        const action = requirePositional(positionals, 2, 'list|install|prepare|customize');
+        if (action === 'list') {
+            print(await ctx.client.v2.listOfficialWorkspaceSkills());
+            return;
+        }
+        const skillId = requirePositional(positionals, 3, 'skill-id');
+        if (action === 'install') {
+            requireFlag(ctx.flags, 'revision');
+            print(await ctx.client.v2.installOfficialWorkspaceSkill(skillId, {
+                revision: parsePositiveIntegerFlag(ctx.flags, 'revision', 1),
+                expectedRevision: hasFlag(ctx.flags, 'expected-revision')
+                    ? parsePositiveIntegerFlag(ctx.flags, 'expected-revision', 1)
+                    : null,
+                expectedGrantRevision: hasFlag(ctx.flags, 'expected-grant-revision')
+                    ? parsePositiveIntegerFlag(ctx.flags, 'expected-grant-revision', 1)
+                    : null,
+                confirmedContentSha256: requireFlag(ctx.flags, 'content-sha256'),
+            }));
+            return;
+        }
+        if (action === 'prepare') {
+            const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
+            print(await ctx.client.v2.prepareOfficialWorkspaceSkillCustomization(skillId, input));
+            return;
+        }
+        // `customize` submits the exact body `prepare` returned, unchanged: the
+        // server checks the confirmed hash against a fresh re-read of the
+        // installed source, so a hand-edited body fails the hash match.
+        if (action === 'customize') {
+            const input = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
+            print(await ctx.client.v2.customizeOfficialWorkspaceSkill(skillId, input));
+            return;
+        }
+        throw new Error('Use skills official list|install|prepare|customize');
+    }
+    if (subcommand === 'build') {
+        const canvasIds = getFlagValues(ctx.flags, 'canvas').map(value => {
+            const parsed = Number(value);
+            if (!Number.isInteger(parsed) || parsed <= 0)
+                throw new Error('--canvas must be a positive canvas id');
+            return parsed;
+        });
+        const memoryIds = getFlagValues(ctx.flags, 'memory');
+        // Exactly one source. Accepting both would leave the server to pick, and
+        // the caller would not know which half of their request was ignored.
+        if (canvasIds.length > 0 === memoryIds.length > 0)
+            throw new Error('Pass either --canvas (1-5) or --memory (1-8), not both.');
+        const request = {
+            goal: requireFlag(ctx.flags, 'goal'),
+            instructions: getFlag(ctx.flags, 'instructions'),
+            taskKind: 'part_separation',
+            source: canvasIds.length > 0
+                ? { kind: 'canvas_graph', canvasIds }
+                : { kind: 'memory_path', memoryIds },
+        };
+        // FREE preview: no build row, no reservation, no dispatch. Skips
+        // --wait/--operation-id entirely -- there is nothing to poll or replay.
+        if (hasFlag(ctx.flags, 'dry-run')) {
+            print(await ctx.client.v2.dryRunWorkspaceSkillBuild(request));
+            return;
+        }
+        const started = await ctx.client.v2.startWorkspaceSkillBuild(request, 
+        // Reusing --operation-id after a lost response returns the SAME build
+        // instead of starting a second paid run.
+        { idempotencyKey: getFlag(ctx.flags, 'operation-id') ?? randomUUID() });
+        print(hasFlag(ctx.flags, 'wait')
+            ? await waitForSkillBuild(ctx, started.buildId)
+            : started);
+        return;
+    }
+    if (subcommand === 'build-status') {
+        const buildId = requirePositional(positionals, 2, 'build-id');
+        print(hasFlag(ctx.flags, 'wait')
+            ? await waitForSkillBuild(ctx, buildId)
+            : await ctx.client.v2.getWorkspaceSkillBuild(buildId));
+        return;
+    }
+    if (subcommand === 'build-accept') {
+        const buildId = requirePositional(positionals, 2, 'build-id');
+        print(await ctx.client.v2.acceptWorkspaceSkillBuild(buildId, { expectedDraftSha256: requireFlag(ctx.flags, 'draft-sha256') }, { idempotencyKey: getFlag(ctx.flags, 'operation-id') ?? randomUUID() }));
+        return;
+    }
+    if (subcommand === 'build-enhance') {
+        // Prints a SUGGESTION. Nothing is written — apply it yourself and then
+        // accept, so a rewrite can never publish something nobody read.
+        print(await ctx.client.v2.enhanceWorkspaceSkillBuildSection(requirePositional(positionals, 2, 'build-id'), {
+            section: parseEnumFlag(ctx.flags, 'section', [
+                ...WORKSPACE_SKILL_ENHANCE_SECTIONS,
+            ]),
+            current: JSON.parse(requireFlag(ctx.flags, 'current')),
+            note: getFlag(ctx.flags, 'note'),
+        }));
+        return;
+    }
+    if (subcommand === 'build-discard') {
+        print(await ctx.client.v2.discardWorkspaceSkillBuild(requirePositional(positionals, 2, 'build-id')));
+        return;
+    }
+    // FREE pre-flight checks: neither reserves credits, inserts a build row,
+    // or dispatches anything.
+    if (subcommand === 'validate') {
+        const skill = await readJsonArgument(requireFlag(ctx.flags, 'input-json'), '--input-json');
+        const buildId = getFlag(ctx.flags, 'build');
+        print(await ctx.client.v2.validateWorkspaceSkillDraft({
+            skill: skill,
+            ...(buildId ? { buildId } : {}),
+        }));
+        return;
+    }
+    if (subcommand === 'schema') {
+        print(await ctx.client.v2.getWorkspaceSkillDraftSchema());
+        return;
+    }
+    throw new Error('Use skills list|get|memory|learn|update|controls|prepare|proposal|accept|validate|schema|build|build-status|build-enhance|build-accept|build-discard|official');
+};
+/**
+ * `assethub org search <query>` and `assethub org canvases --org <orgId>`:
+ * the two reads an operator needs before `graph export-canvas`. Internal-only
+ * on the server — a non-internal key gets 404, printed as-is with exit 2.
+ */
+export const commandOrg = async (subcommand, positionals, ctx) => {
+    if (subcommand === 'search') {
+        const query = requirePositional(positionals, 2, 'query');
+        await runInternalCommand(async () => {
+            print(await internalV2Request(ctx, 'GET', `/orgs/search?q=${encodeURIComponent(query)}`));
+        });
+        return;
+    }
+    if (subcommand === 'canvases') {
+        const orgId = requireFlag(ctx.flags, 'org');
+        await runInternalCommand(async () => {
+            print(await internalV2Request(ctx, 'GET', `/orgs/${encodeURIComponent(orgId)}/workflow-canvases`));
+        });
+        return;
+    }
+    throw new Error('Unknown org command. Use "org search" or "org canvases".');
+};
 /** Base URL and workspace an MCP entry should target: explicit flags, then the selected profile, then defaults. */
 const resolveMcpTarget = async (flags) => {
     const config = await readAuthConfig(getConfigPath(flags));
@@ -4263,7 +5605,13 @@ const run = async () => {
                 if (!hasFlag(parsed.flags, 'account'))
                     return resolveAuth(parsed.flags);
                 const auth = await workspaceAuth(parsed.flags);
-                return { apiKey: auth.accessToken, workspaceMfaToken: auth.workspaceMfaToken, baseUrl: auth.baseUrl, profile: auth.profile, source: 'user' };
+                return {
+                    apiKey: auth.accessToken,
+                    workspaceMfaToken: auth.workspaceMfaToken,
+                    baseUrl: auth.baseUrl,
+                    profile: auth.profile,
+                    source: 'user',
+                };
             },
             account: hasFlag(parsed.flags, 'account'),
             includeMcp: true,
@@ -4276,7 +5624,7 @@ const run = async () => {
             return;
         }
         const check = report.checks.find(check => check.name === 'mcp');
-        const tools = check && 'tools' in check ? check.tools ?? [] : [];
+        const tools = check && 'tools' in check ? (check.tools ?? []) : [];
         const name = parsed.positionals[2];
         if (name) {
             const tool = tools.find(tool => tool.name === name);
@@ -4285,17 +5633,43 @@ const run = async () => {
             print({ tool });
         }
         else {
-            print({ total: tools.length, tools: tools.map(({ name, title, annotations }) => ({ name, title, annotations })) });
+            print({
+                total: tools.length,
+                tools: tools.map(({ name, title, annotations }) => ({
+                    name,
+                    title,
+                    annotations,
+                })),
+            });
         }
         return;
     }
     if (command === 'mcp') {
         if (subcommand !== 'config')
             throw new Error('Use mcp tools [tool-name] or mcp config --client cursor|codex');
-        const { baseUrl, workspaceId } = await resolveMcpTarget(parsed.flags);
+        const config = await readAuthConfig(getConfigPath(parsed.flags));
+        const explicitProfile = getFlag(parsed.flags, 'profile');
+        const stored = config.profiles?.[explicitProfile ?? config.defaultProfile ?? defaultProfileName];
+        if (explicitProfile && !stored)
+            throw new Error('Selected profile does not exist');
+        const selected = explicitProfile || stored?.workspaceId || isPersonalProfile(stored);
+        const baseUrl = getFlag(parsed.flags, 'base-url') ??
+            (selected ? stored?.baseUrl : env.ASSETHUB_API_BASE_URL ?? stored?.baseUrl) ?? defaultBaseUrl;
+        const workspaceId = getFlag(parsed.flags, 'workspace') ??
+            (stored && new URL(baseUrl).origin === new URL(stored.baseUrl).origin ? stored.workspaceId : undefined);
         stdout.write(mcpConfig(requireFlag(parsed.flags, 'client'), baseUrl, hasFlag(parsed.flags, 'account'), workspaceId));
         return;
     }
+    if (command === 'update') {
+        // The verifying doctor checks the same login this command was given.
+        const forwardFlags = ['profile', 'config', 'base-url', 'workspace'].flatMap(name => {
+            const value = getFlag(parsed.flags, name);
+            return value ? [`--${name}`, value] : [];
+        });
+        const apiKey = getFlag(parsed.flags, 'api-key');
+        print(await runUpdate({ check: hasFlag(parsed.flags, 'check'), dryRun: hasFlag(parsed.flags, 'dry-run'), yes: hasFlag(parsed.flags, 'yes') }, createNodeUpdateDeps(forwardFlags, apiKey ? { ASSETHUB_API_KEY: apiKey } : {})));
+        return;
+    }
     if (command === 'doctor') {
         stderr.write('Checking AssetHub connection…\n');
         const report = await diagnose({
@@ -4325,7 +5699,16 @@ const run = async () => {
         return;
     }
     if (command === 'workspace') {
-        if (['list', 'get', 'create', 'use', 'members', 'invite', 'set-role', 'remove-member'].includes(subcommand ?? '')) {
+        if ([
+            'list',
+            'get',
+            'create',
+            'use',
+            'members',
+            'invite',
+            'set-role',
+            'remove-member',
+        ].includes(subcommand ?? '')) {
             await commandWorkspace(subcommand, parsed.positionals, parsed.flags);
             return;
         }
@@ -4358,6 +5741,21 @@ const run = async () => {
         });
         return;
     }
+    if (command === 'setup') {
+        await commandSetup(parsed.flags, installSessionHooks);
+        return;
+    }
+    if (command === 'env') {
+        await commandEnv(subcommand, parsed.flags);
+        return;
+    }
+    if (command === 'hooks') {
+        // hooks parses its own flags: `save` reads the Claude Code hook JSON on stdin.
+        const code = await runHooksCommand(argv.slice(3));
+        if (code !== 0)
+            process.exitCode = code;
+        return;
+    }
     if (command === 'auth') {
         await commandAuth(subcommand, parsed.flags);
         return;
@@ -4370,9 +5768,6 @@ const run = async () => {
         case 'api':
             await commandApi(subcommand, parsed.positionals, ctx);
             return;
-        case 'skills':
-            await commandSkills(subcommand, parsed.positionals, ctx);
-            return;
         case 'composer':
             await commandComposer(subcommand, ctx);
             return;
@@ -4386,7 +5781,7 @@ const run = async () => {
             await commandCanvas(subcommand, parsed.positionals, ctx);
             return;
         case 'graph':
-            await commandGraph(subcommand, ctx);
+            await commandGraph(subcommand, parsed.positionals, ctx);
             return;
         case 'evaluations':
             await commandEvaluations(subcommand, parsed.positionals, ctx);
@@ -4429,6 +5824,15 @@ const run = async () => {
         case 'runs':
             await commandRuns(subcommand, parsed.positionals, ctx);
             return;
+        case 'memory':
+            await commandMemory(subcommand, parsed.positionals, ctx);
+            return;
+        case 'skills':
+            await commandSkills(subcommand, parsed.positionals, ctx);
+            return;
+        case 'org':
+            await commandOrg(subcommand, parsed.positionals, ctx);
+            return;
         case 'parts':
             await commandParts(subcommand, ctx);
             return;