@bendyline/gezel 1.0.3 → 1.0.4

package/dist/index.js

@@ -673,7 +673,9 @@ var GateCheckSchema = z3.discriminatedUnion("kind", [
   z3.object({
     kind: z3.literal("markdownHeadingsMatch"),
     file: z3.string().min(1),
-    outlineFile: z3.string().min(1)
+    outlineFile: z3.string().min(1),
+    /** Read outlineFile from artifacts while file uses the check's primary surface. */
+    outlineArtifact: z3.boolean().optional()
   }),
   /**
    * Named facts in `file` must come from authorized sources: each fact
@@ -997,14 +999,16 @@ var CraftbookStepSchema = z5.object({
    * Marks the parent step that triggers a declarative per-item fanout
    * (see {@link CraftbookSpawnSchema}). When this step activates on a
    * spawn-host task, the runtime reads the craftbook's `spawn.overFile`
-   * workspace JSON array and spawns one child task per item — no model
+   * JSON array on its declared surface and spawns one child task per item — no model
    * tool call. Inert unless the craftbook also declares `spawn`.
    */
   spawnFanout: z5.boolean().optional()
 });
 var CraftbookSpawnSchema = z5.object({
-  /** Workspace-relative JSON file the parent produces; its array drives the fanout. */
+  /** Surface-relative JSON file the parent produces; its array drives the fanout. */
   overFile: z5.string().min(1),
+  /** Read overFile from the artifacts drawer instead of the project workspace. */
+  overArtifact: z5.boolean().optional(),
   /** Dotted path to the array inside `overFile`. Absent → the file itself is the array. */
   itemsPath: z5.string().optional(),
   /** Entry step id of the child template. Defaults to the first `steps` entry. */
@@ -4392,7 +4396,7 @@ var ProjectSchema = z19.object({
   /** Optional gezel that acts as the project's voorman (foreman). Surfaces in
    *  the project detail pane, flows into the system prompt when a session is
    *  scoped here. For solo projects (`mode === 'solo'`) this same field
-   *  holds the project's ambachtsman — the data is unchanged, only the
+   *  holds the project's Builder — the data is unchanged, only the
    *  label flips. */
   voormanGezelId: z19.string().optional(),
   /**
@@ -4442,7 +4446,7 @@ var ProjectSchema = z19.object({
   /**
    * Project shape. `crew` (the default) is the original behavior — the
    * voorman recruits and coordinates a team of specialists. `solo` is a
-   * "job" — a single specialist (the ambachtsman, stored in
+   * "job" — a single Builder (stored in
    * `voormanGezelId`) handles the whole project themselves; team-
    * management MCP tools are stripped from their session, and the
    * Meester is instructed not to nominate other gezels. Missing → `crew`
@@ -4452,7 +4456,7 @@ var ProjectSchema = z19.object({
   mode: z19.enum(["crew", "solo"]).optional(),
   /**
    * Optional custom label for this project's lead gezel, overriding the
-   * mode-based default ("Voorman" / "Ambachtsman") everywhere the UI
+   * mode-based default ("Voorman" / "Builder") everywhere the UI
    * renders it. Set by a project type at adoption (e.g. checkers →
    * "Opponent"); absent → the mode default. The data field stays
    * `voormanGezelId` — only the label changes.
@@ -5389,6 +5393,15 @@ var TaskSchema = z23.object({
       issueRef: z23.string().regex(/^BW-[1-9]\d*$/),
       path: z23.string().min(1)
     }),
+    z23.object({
+      /**
+       * An invoke_craftbook call keyed to one persisted root chat turn.
+       * The HTTP task boundary uses this opaque digest to return the first
+       * still-active task when a provider continuation repeats the call.
+       */
+      kind: z23.literal("craftbook-invocation"),
+      key: z23.string().regex(/^craftbook-root-v1:[a-f0-9]{64}$/)
+    }),
     z23.object({
       /**
        * A host materialized from a gezel template's `suggestedCraftbooks`
@@ -5418,6 +5431,13 @@ var TaskSchema = z23.object({
     params: z23.record(z23.string(), z23.unknown()).optional(),
     at: z23.string()
   }).optional(),
+  /**
+   * Naming presentation inherited from the session that launched this
+   * workflow. Task handoffs create fresh sessions (and may be rehydrated
+   * after restart), so they cannot rely on the launcher's session record
+   * still being available when the next step starts.
+   */
+  roleBasedNameOnlyMode: z23.boolean().optional(),
   createdAt: z23.string(),
   updatedAt: z23.string(),
   createdBy: TaskAssigneeSchema
@@ -5472,6 +5492,8 @@ var CreateTaskRequestSchema = z23.object({
   }).optional(),
   fanout: NewTaskFanoutSchema.optional(),
   createdBy: TaskAssigneeSchema.optional(),
+  /** Preserve the launcher's naming presentation across task handoffs. */
+  roleBasedNameOnlyMode: z23.boolean().optional(),
   /**
    * Enqueue the entry-step handoff immediately after create — the
    * single-channel kickoff (there is no "tell a gezel about work"
@@ -5481,7 +5503,13 @@ var CreateTaskRequestSchema = z23.object({
    * on cron/fanout hosts (their children dispatch via their own
    * activation hooks — flag-dispatching the host would double-engage).
    */
-  dispatchEntry: z23.boolean().optional()
+  dispatchEntry: z23.boolean().optional(),
+  /**
+   * Internal invoke_craftbook idempotency digest. The task route converts
+   * it into service-owned Task.origin provenance; ordinary create_task
+   * callers omit it.
+   */
+  craftbookInvocationKey: z23.string().regex(/^craftbook-root-v1:[a-f0-9]{64}$/).optional()
 }).refine((v) => !!v.craftbookId !== !!(v.steps && v.steps.length > 0), {
   message: "exactly one of craftbookId or steps must be provided for the main craftbook",
   path: ["craftbookId"]
@@ -6397,7 +6425,7 @@ var ProjectTypeCompositionShape = {
   tabVisibility: ProjectTabVisibilitySchema.optional(),
   /**
    * Project shape for instances of this type. `solo` marks a single-gezel
-   * "ambachtsman" experience (games, the chat room): the one roster gezel
+   * "Builder" experience (games, the chat room): the one roster gezel
    * does everything, no separate overseer voorman is recruited, and the UI
    * collapses the crew picker to that lead. Maps onto the project's `mode`
    * at adoption. Omit for the default crew shape.
@@ -6405,9 +6433,9 @@ var ProjectTypeCompositionShape = {
   mode: z26.enum(["crew", "solo"]).optional(),
   /**
    * Custom label for this type's project lead, shown wherever the generic
-   * "Voorman" / "Ambachtsman" would appear (the chat chip, project
+   * "Voorman" / "Builder" would appear (the chat chip, project
    * settings) — e.g. checkers uses "Opponent". Aimed at solo types that
-   * want a domain word instead of "Ambachtsman". The underlying data field
+   * want a domain word instead of "Builder". The underlying data field
    * stays `voormanGezelId`; only the rendered label changes.
    */
   leadLabel: z26.string().min(1).optional(),
@@ -11250,7 +11278,7 @@ var GezelConfigSchema = z37.object({
    *    crew with granular per-step craftbooks. Right for raw-completion
    *    local models, where gezel's loop IS the agent.
    *  - `flat`: route concrete asks to a single generalist (`start_job` →
-   *    solo "ambachtsman", which already collapses the craftbook onto the
+   *    solo Builder, which already collapses the craftbook onto the
    *    one specialist). Right for self-orchestrating providers (codex-cli,
    *    anthropic-cli, copilot) that bring their own agent loop — the crew
    *    + granular steps are mostly redundant overhead for them (eval data:
@@ -12154,6 +12182,12 @@ var ListProjectsResponseSchema = z37.object({
 var CreateProjectRequestSchema = z37.object({
   name: z37.string().min(1),
   description: z37.string().optional(),
+  /**
+   * Existing local folder to use as the project workspace. Folder-backed
+   * projects are created with ambient Meester progress check-ins disabled;
+   * the user can opt back in from Project Settings.
+   */
+  workingDir: z37.string().min(1).optional(),
   // about/missionObjectives are *encouraged, not required* at the wire level.
   // The New Project dialog still enforces the 60/40 richness minimums for the
   // blank/GitHub flows (where the user is authoring context from scratch), but
@@ -12166,7 +12200,7 @@ var CreateProjectRequestSchema = z37.object({
   missionObjectives: z37.string().optional().describe('Concrete success criteria \u2014 usually a bullet list. What does "done" look like?'),
   /**
    * Project shape. `crew` (default) → traditional voorman-coordinates-
-   * specialists. `solo` → a "job": one specialist (the ambachtsman) does
+   * specialists. `solo` → a "job": one Builder does
    * everything themselves; team-management tools are filtered out for
    * sessions on this project.
    */
@@ -23119,6 +23153,55 @@ function firstActionForKind(kind, path) {
 var scorecard_default = {
   schemaVersion: 1,
   runs: [
+    {
+      id: "2026-08-13-m4max-llamacpp",
+      provenance: {
+        startedAt: "2026-08-13T07:05:36.620Z",
+        device: {
+          label: "Mac \xB7 Apple M4 Max",
+          platform: "darwin",
+          arch: "arm64",
+          memoryGb: 64,
+          osRelease: "darwin 25.5.0",
+          cpuModel: "Apple M4 Max"
+        },
+        harnessCommit: "c5904085",
+        gildeVersion: "0.1.23",
+        count: 3,
+        judgeModelId: null
+      },
+      suites: ["core", "productivity"],
+      scenariosBySuite: {
+        core: [
+          "tictactoe",
+          "petshop",
+          "tankcombat",
+          "schema-migration",
+          "failing-tests-spec",
+          "symptom-debug",
+          "data-wrangle",
+          "incident-postmortem",
+          "ops-runbook-anomaly",
+          "plan-and-estimate",
+          "conflict-synthesis"
+        ],
+        productivity: [
+          "constrained-comms",
+          "craftbook-week-plan",
+          "craftbook-ab-test-readout",
+          "craftbook-annotated-bibliography",
+          "records-intake",
+          "plan-and-estimate",
+          "meeting-followup",
+          "craftbook-spreadsheet-model",
+          "conflict-synthesis",
+          "docblocks-theme-roundtrip",
+          "craftbook-research-to-document",
+          "craftbook-powerpoint-deck",
+          "wikipedia-research-brief"
+        ]
+      }
+    },
     {
       id: "2026-08-11-m4max-llamacpp",
       provenance: {
@@ -23220,6 +23303,878 @@ var scorecard_default = {
     }
   ],
   results: [
+    {
+      modelId: "gemma4-26b-q4",
+      label: "gemma4-26b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "25.2B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "core",
+      performance: {
+        prefillTokensPerSec: 1264,
+        decodeTokensPerSec: 107.5,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 120832,
+        peakMemoryMb: 42690
+      },
+      judge: {
+        meanScore: 5.7,
+        artifacts: 9,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "tictactoe",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 127411
+        },
+        {
+          scenarioId: "petshop",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 482159
+        },
+        {
+          scenarioId: "tankcombat",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 290618
+        },
+        {
+          scenarioId: "schema-migration",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 69039
+        },
+        {
+          scenarioId: "failing-tests-spec",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 28770
+        },
+        {
+          scenarioId: "symptom-debug",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 83940
+        },
+        {
+          scenarioId: "data-wrangle",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 48055
+        },
+        {
+          scenarioId: "incident-postmortem",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 189005
+        },
+        {
+          scenarioId: "ops-runbook-anomaly",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 233624
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 17679
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 52710
+        }
+      ]
+    },
+    {
+      modelId: "muse-glimmer-30b-q4",
+      label: "muse-glimmer-30b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "30B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "core",
+      performance: {
+        prefillTokensPerSec: 226,
+        decodeTokensPerSec: 26,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 131072,
+        peakMemoryMb: 19638
+      },
+      judge: {
+        meanScore: 6.5,
+        artifacts: 9,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "tictactoe",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 210652
+        },
+        {
+          scenarioId: "petshop",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 356601
+        },
+        {
+          scenarioId: "tankcombat",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 272117
+        },
+        {
+          scenarioId: "schema-migration",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 269225
+        },
+        {
+          scenarioId: "failing-tests-spec",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 279796
+        },
+        {
+          scenarioId: "symptom-debug",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 239896
+        },
+        {
+          scenarioId: "data-wrangle",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 563414
+        },
+        {
+          scenarioId: "incident-postmortem",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 4467148
+        },
+        {
+          scenarioId: "ops-runbook-anomaly",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 474336
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 150465
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 404097
+        }
+      ]
+    },
+    {
+      modelId: "nemotron3.5-lightning-30b-q4",
+      label: "nemotron3.5-lightning-30b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "30B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "core",
+      performance: {
+        prefillTokensPerSec: 1030,
+        decodeTokensPerSec: 93.2,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 1048576,
+        peakMemoryMb: 30085
+      },
+      judge: {
+        meanScore: 5.6,
+        artifacts: 9,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "tictactoe",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 90179
+        },
+        {
+          scenarioId: "petshop",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 899922
+        },
+        {
+          scenarioId: "tankcombat",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 65485
+        },
+        {
+          scenarioId: "schema-migration",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 133904
+        },
+        {
+          scenarioId: "failing-tests-spec",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 230269
+        },
+        {
+          scenarioId: "symptom-debug",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 94023
+        },
+        {
+          scenarioId: "data-wrangle",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 319197
+        },
+        {
+          scenarioId: "incident-postmortem",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 223455
+        },
+        {
+          scenarioId: "ops-runbook-anomaly",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 93099
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 68902
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 254006
+        }
+      ]
+    },
+    {
+      modelId: "qwen3.6-35b-a3b-q4",
+      label: "qwen3.6-35b-a3b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "35B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "core",
+      performance: {
+        prefillTokensPerSec: 1104,
+        decodeTokensPerSec: 71.3,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 262144,
+        peakMemoryMb: 26502
+      },
+      judge: {
+        meanScore: 8.1,
+        artifacts: 9,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "tictactoe",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 92213
+        },
+        {
+          scenarioId: "petshop",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 162363
+        },
+        {
+          scenarioId: "tankcombat",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 247709
+        },
+        {
+          scenarioId: "schema-migration",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 186639
+        },
+        {
+          scenarioId: "failing-tests-spec",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 28425
+        },
+        {
+          scenarioId: "symptom-debug",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 37627
+        },
+        {
+          scenarioId: "data-wrangle",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 47569
+        },
+        {
+          scenarioId: "incident-postmortem",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 73048
+        },
+        {
+          scenarioId: "ops-runbook-anomaly",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 98132
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 27443
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 82923
+        }
+      ]
+    },
+    {
+      modelId: "gemma4-26b-q4",
+      label: "gemma4-26b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "25.2B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "productivity",
+      performance: {
+        prefillTokensPerSec: 1264,
+        decodeTokensPerSec: 107.5,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 120832,
+        peakMemoryMb: 41762
+      },
+      judge: {
+        meanScore: 6,
+        artifacts: 19,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "constrained-comms",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 22693
+        },
+        {
+          scenarioId: "craftbook-week-plan",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 58001
+        },
+        {
+          scenarioId: "craftbook-ab-test-readout",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 196546
+        },
+        {
+          scenarioId: "craftbook-annotated-bibliography",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 33179
+        },
+        {
+          scenarioId: "records-intake",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 58377
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 17469
+        },
+        {
+          scenarioId: "meeting-followup",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 48022
+        },
+        {
+          scenarioId: "craftbook-spreadsheet-model",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 191950
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 48088
+        },
+        {
+          scenarioId: "docblocks-theme-roundtrip",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 27505
+        },
+        {
+          scenarioId: "craftbook-research-to-document",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 32944
+        },
+        {
+          scenarioId: "craftbook-powerpoint-deck",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 937028
+        },
+        {
+          scenarioId: "wikipedia-research-brief",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 81272
+        }
+      ]
+    },
+    {
+      modelId: "muse-glimmer-30b-q4",
+      label: "muse-glimmer-30b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "30B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "productivity",
+      performance: {
+        prefillTokensPerSec: 228,
+        decodeTokensPerSec: 26,
+        samples: 2
+      },
+      runtime: {
+        contextTokens: 131072,
+        peakMemoryMb: 19648
+      },
+      judge: {
+        meanScore: 5.5,
+        artifacts: 25,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "constrained-comms",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 168777
+        },
+        {
+          scenarioId: "craftbook-week-plan",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 635096
+        },
+        {
+          scenarioId: "craftbook-ab-test-readout",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 505316
+        },
+        {
+          scenarioId: "craftbook-annotated-bibliography",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 153322
+        },
+        {
+          scenarioId: "records-intake",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 2593201
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 185158
+        },
+        {
+          scenarioId: "meeting-followup",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 680419
+        },
+        {
+          scenarioId: "craftbook-spreadsheet-model",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 541672
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 2,
+          nonModelFailures: 0,
+          medianDurationMs: 344063
+        },
+        {
+          scenarioId: "docblocks-theme-roundtrip",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 248997
+        },
+        {
+          scenarioId: "craftbook-research-to-document",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 274010
+        },
+        {
+          scenarioId: "craftbook-powerpoint-deck",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 3009529
+        },
+        {
+          scenarioId: "wikipedia-research-brief",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 353289
+        }
+      ]
+    },
+    {
+      modelId: "nemotron3.5-lightning-30b-q4",
+      label: "nemotron3.5-lightning-30b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "30B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "productivity",
+      performance: {
+        prefillTokensPerSec: 1030,
+        decodeTokensPerSec: 93.2,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 1048576,
+        peakMemoryMb: 29452
+      },
+      judge: {
+        meanScore: 5.6,
+        artifacts: 27,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "constrained-comms",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 91133
+        },
+        {
+          scenarioId: "craftbook-week-plan",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 113088
+        },
+        {
+          scenarioId: "craftbook-ab-test-readout",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 283896
+        },
+        {
+          scenarioId: "craftbook-annotated-bibliography",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 73049
+        },
+        {
+          scenarioId: "records-intake",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 395444
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 105986
+        },
+        {
+          scenarioId: "meeting-followup",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 113144
+        },
+        {
+          scenarioId: "craftbook-spreadsheet-model",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 233654
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 469600
+        },
+        {
+          scenarioId: "docblocks-theme-roundtrip",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 109277
+        },
+        {
+          scenarioId: "craftbook-research-to-document",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 130992
+        },
+        {
+          scenarioId: "craftbook-powerpoint-deck",
+          trials: 3,
+          successes: 0,
+          nonModelFailures: 0,
+          medianDurationMs: 791082
+        },
+        {
+          scenarioId: "wikipedia-research-brief",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 136065
+        }
+      ]
+    },
+    {
+      modelId: "qwen3.6-35b-a3b-q4",
+      label: "qwen3.6-35b-a3b-q4",
+      engine: "llama-cpp",
+      tier: "medium",
+      parameterSize: "35B",
+      runId: "2026-08-13-m4max-llamacpp",
+      suiteId: "productivity",
+      performance: {
+        prefillTokensPerSec: 1104,
+        decodeTokensPerSec: 71.3,
+        samples: 1
+      },
+      runtime: {
+        contextTokens: 262144,
+        peakMemoryMb: 25657
+      },
+      judge: {
+        meanScore: 6.6,
+        artifacts: 27,
+        judgeModel: "claude-sonnet-4-6"
+      },
+      cells: [
+        {
+          scenarioId: "constrained-comms",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 22314
+        },
+        {
+          scenarioId: "craftbook-week-plan",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 88382
+        },
+        {
+          scenarioId: "craftbook-ab-test-readout",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 303967
+        },
+        {
+          scenarioId: "craftbook-annotated-bibliography",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 47530
+        },
+        {
+          scenarioId: "records-intake",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 22365
+        },
+        {
+          scenarioId: "plan-and-estimate",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 32491
+        },
+        {
+          scenarioId: "meeting-followup",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 89751
+        },
+        {
+          scenarioId: "craftbook-spreadsheet-model",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 228470
+        },
+        {
+          scenarioId: "conflict-synthesis",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 211329
+        },
+        {
+          scenarioId: "docblocks-theme-roundtrip",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 17446
+        },
+        {
+          scenarioId: "craftbook-research-to-document",
+          trials: 3,
+          successes: 3,
+          nonModelFailures: 0,
+          medianDurationMs: 22407
+        },
+        {
+          scenarioId: "craftbook-powerpoint-deck",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 1493404
+        },
+        {
+          scenarioId: "wikipedia-research-brief",
+          trials: 3,
+          successes: 1,
+          nonModelFailures: 0,
+          medianDurationMs: 125222
+        }
+      ]
+    },
     {
       modelId: "gemma4-26b-q4",
       label: "gemma4-26b-q4",
@@ -27094,7 +28049,7 @@ function visibleCatalogItems(items, showWorkInProgressFeatures, connectorIds = c
 }
 
 // src/index.ts
-var GEZEL_VERSION = "1.0.3";
+var GEZEL_VERSION = "1.0.4";
 var GEZEL_CONTENT_COMPAT = "1.26226";
 function nowIso() {
   return (/* @__PURE__ */ new Date()).toISOString();