@bendyline/gilde 0.1.50 → 0.1.52

package/data/craftbook-templates/im/image-batch-edit/versions/1.0.3/craftbook.json

@@ -0,0 +1,277 @@
+{
+  "id": "image-batch-edit",
+  "name": "Batch Image Transform",
+  "description": "Apply a consistent transform across a whole folder of images — resize, crop to a ratio, convert format, compress, watermark, or recolor — and verify every file came out correct. Scope the exact transform and a verification rule FIRST (target dimensions, format, quality, naming) so the batch is deterministic and reproducible, then process every file and prove the output count, dimensions, and format match the spec. Covers bulk image processing, resizing, format conversion, compression, watermarking, and folder-wide edits.\n\nA gallery craftbook generated from an archetype spec. It runs\n`phase → (per-phase gate) → … → evaluate → (loop) → finish`. Each build\nphase that produces a checkable artifact is followed by a **runtime\ngate-checkpoint** — the runtime verifies the artifact and routes with no\nmodel turn, looping back to redo the phase on a miss. The final `evaluate`\nstep holds a static deliverable gate plus a reviewer QA pass. What it adds\nover the generic `build-loop`: a specialist role per phase, a\ndomain-correct ordering, and a concrete per-phase quality bar.\n\nDeliverables marked \"artifact\" land in the project's artifacts drawer (`write_artifact` / `read_artifact`), not the shipped workspace — review output is not product source.\n\nPhases:\n\n1. Scope the transform (planner) — lock the exact transform, output spec, and verify rule → gated on artifact `{{workPath}}/scope.md` (markdown-notes)\n2. Process the folder (developer) — apply the transform to every file → gated on `out` (image-set)\n3. Verify the batch (reviewer) — confirm count, dimensions, and format per file → gated on artifact `{{workPath}}/verify.md` (markdown-notes)\n\nThe gates never advance with an unmet criterion, and loop back to the\nowning phase to fix named gaps.\n",
+  "entryStepId": "scope",
+  "triggers": [
+    "batch edit images",
+    "resize a folder of images",
+    "convert image formats",
+    "compress images",
+    "bulk image transform"
+  ],
+  "paramSchema": {
+    "type": "object",
+    "properties": {
+      "workPath": {
+        "type": "string",
+        "title": "Working folder",
+        "description": "Per-task working folder in the artifacts drawer. Defaults to this task's own folder so runs never collide; override with a stable name when you deliberately want runs to share files.",
+        "default": "{{task.dir}}"
+      }
+    }
+  },
+  "steps": [
+    {
+      "id": "scope",
+      "name": "Scope the transform",
+      "description": "lock the exact transform, output spec, and verify rule",
+      "prompt": "1) List the source folder and count/inventory the files and their current formats and sizes. 2) Lock the EXACT transform: target dimensions or max-edge, crop/fit behavior (cover vs contain vs pad), output format and quality, color/orientation handling (respect EXIF rotation), and any watermark. 3) Decide the output location and naming (in-place vs an out/ folder; suffix vs same name). 4) Write an acceptance-criteria checklist: 'every source image has exactly one output', 'all outputs are <format>', 'all outputs are <= target dimensions / exactly <ratio>', 'no output is larger in bytes than its source when compressing', 'originals are preserved if requested'. 5) Record transform + checklist via write_task_note AND to {{workPath}}/scope.md. Good looks like: a spec precise enough to script.\n\nThe deliverable `{{workPath}}/scope.md` lands in the project's artifacts drawer — write it with `write_artifact` and read it back with `read_artifact`; the shipped workspace stays untouched.",
+      "suggestedRole": "planner",
+      "advanceWhen": {
+        "file": "{{workPath}}/scope.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/scope.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/scope.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "scope",
+        "maxAttempts": 3
+      },
+      "next": "process",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "artifact",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "process",
+      "name": "Process the folder",
+      "description": "apply the transform to every file",
+      "prompt": "1) Read scope.md. 2) Prefer writing a small reproducible script (e.g. using ImageMagick/sharp/Pillow if available) over manual one-offs, so the batch is repeatable and nothing is missed. 3) Iterate every source file deterministically; apply the exact transform, honoring fit behavior and EXIF orientation. 4) Write outputs to the agreed location with the agreed names; preserve originals if requested. 5) Skip or log (never silently drop) any file that fails to decode, and report it. 6) Keep a running output count. 7) write_task_note the output folder and the in/out counts. Good looks like: one correctly-transformed output per input, produced by a rerunnable command.",
+      "suggestedRole": "developer",
+      "advanceWhen": {
+        "file": "out",
+        "minBytes": 1,
+        "sniff": "nonempty"
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "fileCount",
+            "ext": [
+              "png",
+              "jpg",
+              "jpeg",
+              "svg",
+              "webp"
+            ],
+            "min": 3,
+            "dir": "out"
+          }
+        ],
+        "onReject": "process",
+        "maxAttempts": 3
+      },
+      "next": "verify",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "artifacts",
+          "audio",
+          "browser-automation",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web"
+        ],
+        "outputMedium": "workspace",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "verify",
+      "name": "Verify the batch",
+      "description": "confirm count, dimensions, and format per file",
+      "prompt": "1) Re-read scope.md's checklist. 2) Count outputs vs inputs — they must match (or match minus any explicitly-logged failures). 3) Sample 4-5 outputs and inspect actual dimensions and format against the spec (use an identify/probe tool or open them). 4) When compressing, confirm output bytes <= source bytes. 5) Confirm originals are intact if preservation was required. 6) Write a verification note: PASS/FAIL per criterion with the measured numbers. On a loop-back the processor fixes only the failing files. write_task_note the verification result.\n\nThe deliverable `{{workPath}}/verify.md` lands in the project's artifacts drawer — write it with `write_artifact` and read it back with `read_artifact`; the shipped workspace stays untouched.",
+      "suggestedRole": "reviewer",
+      "consumes": [
+        {
+          "file": "out"
+        }
+      ],
+      "advanceWhen": {
+        "file": "{{workPath}}/verify.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "fileCount",
+            "ext": [
+              "png",
+              "jpg",
+              "jpeg",
+              "svg",
+              "webp"
+            ],
+            "min": 3,
+            "dir": "out"
+          }
+        ],
+        "onReject": "verify",
+        "maxAttempts": 4
+      },
+      "next": "evaluate",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "artifact",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "evaluate",
+      "name": "Evaluate",
+      "description": "Grade the deliverable against every acceptance criterion. All pass → finish; any fail → loop back and fix the gap.",
+      "prompt": "1) Open out/ and count files vs the source inventory in scope.md — must match (PASS/FAIL). 2) Probe 4-5 outputs: are dimensions and format exactly per spec (PASS/FAIL)? 3) For compression jobs, confirm outputs are smaller than sources. 4) Confirm originals preserved if required. 5) Read verify.md and reconcile its claims with what you actually observe. 6) Write PASS/FAIL per criterion. Any mismatch → loop back to process to fix the offending files.\n\nThen route — this is the whole point of the loop:\n\n- **Every criterion PASSES →** call `advance_task_step({ ref, stepId: \"evaluate\", next: \"finish\" })`.\n- **Any criterion FAILS →** write the specific gaps to notes, then call `advance_task_step({ ref, stepId: \"evaluate\", next: \"process\" })` to loop back. The builder fixes exactly those gaps.\n\nNever route to `finish` while any criterion is unmet. The build phase's completion gate already blocked a grossly-incomplete deliverable; your job is the judgment an automated check cannot make (does it actually work, read well, look right). After ~3 unproductive loops, stop and report DONE_WITH_CONCERNS so the user can step in.",
+      "suggestedRole": "reviewer",
+      "consumes": [
+        {
+          "file": "out"
+        }
+      ],
+      "next": "process",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "artifacts",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "task-note"
+      }
+    },
+    {
+      "id": "finish",
+      "name": "Finish",
+      "description": "All acceptance criteria met. Stamp a short summary and report DONE.",
+      "prompt": "Every acceptance criterion passed. Write a one-paragraph DONE summary to task notes via `write_task_note`: what was built, the deliverable path(s), and a one-line confirmation that each criterion is met. Then report DONE.",
+      "suggestedRole": "developer",
+      "terminal": true,
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "artifacts",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "task-note"
+      }
+    }
+  ],
+  "version": "1.0.3",
+  "releasedAt": "2026-09-01T20:00:00Z",
+  "minGezelVersion": "1.26233"
+}

package/data/craftbook-templates/im/image-batch-edit/versions/1.0.3/test.json

@@ -0,0 +1,103 @@
+{
+  "schemaVersion": 1,
+  "title": "Batch Image Transform smoke eval",
+  "objective": "Self-contained smoke eval for the Batch Image Transform craftbook using the media generic harness.",
+  "tags": [
+    "media"
+  ],
+  "prompt": "Our media files are all listd in source/media-index.csv. Can you put together the production plan for them as media-plan.json?",
+  "setup": {
+    "projectName": "Batch Image Transform Eval",
+    "about": "Self-contained eval project for image-batch-edit. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/media-plan.json.",
+    "missionObjectives": "Use the Batch Image Transform craftbook/template, read the seeded local fixtures, and write media-plan.json without network calls, real credentials, or live services.",
+    "files": [
+      {
+        "path": "source/brief.md",
+        "content": "# Batch Image Transform Eval Brief\n\nClient: Boreal Desk, a home-office accessories company.\nAudience: operations leads who need an artifact they can use this week.\n\nFixed source facts for grounding:\n- The returns desk pilot covered 18 SKUs.\n- Median first response improved from 18 hours to 6 hours.\n- Preventable refund leakage fell from 14.2% to 8.9%.\n- The top unresolved complaint is status silence after photo submission.\n- Required next actions are automated status emails, barcode-exception training, and a weekly Finance exception export.\n\nUse these facts when the task asks for prose, analysis, copy, UI content, or test data. Do not use live web services, real credentials, or current outside data.\n\nCraftbook under test: image-batch-edit - Batch Image Transform.\n"
+      },
+      {
+        "path": "source/media-index.csv",
+        "content": "filename,type,duration_or_size,notes\npilot-interview.wav,audio,00:02:15,mentions response time and status silence\nreturns-desk.png,image,1600x900,shows intake form and photo evidence step\nlaunch-clip.mp4,video,00:01:05,shows barcode exception handling\n"
+      }
+    ],
+    "worker": {
+      "name": "Jules",
+      "role": "Producer"
+    }
+  },
+  "mocks": [],
+  "success": {
+    "summary": "media-plan.json is a valid JSON media-production manifest grounded in seeded fixture metadata.",
+    "deliverables": [
+      {
+        "path": "media-plan.json",
+        "kind": "json",
+        "minBytes": 700,
+        "checks": [
+          {
+            "kind": "sniff",
+            "file": "media-plan.json",
+            "sniff": "json-valid"
+          },
+          {
+            "kind": "contains",
+            "file": "media-plan.json",
+            "pattern": "pilot-interview\\.wav|returns-desk\\.png|launch-clip\\.mp4",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "media-plan.json",
+            "pattern": "outputs|assets|steps|cues|verification",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "media-plan.json",
+            "pattern": "accessibility|alt|caption|transcript|metadata",
+            "flags": "i"
+          }
+        ]
+      }
+    ]
+  },
+  "rubric": {
+    "artifact": {
+      "path": "media-plan.json",
+      "kind": "json"
+    },
+    "axes": [
+      {
+        "name": "completeness",
+        "description": "The manifest covers sources, outputs, and processing steps end to end."
+      },
+      {
+        "name": "grounding",
+        "description": "Only seeded media fixtures are referenced, with faithful metadata."
+      },
+      {
+        "name": "accessibility",
+        "description": "Alt text, captions, or transcripts are planned wherever the class requires."
+      },
+      {
+        "name": "precision",
+        "description": "Timestamps, crops, and cues are specific enough to execute."
+      }
+    ]
+  },
+  "qualityFocus": [
+    "media fixture handling",
+    "structured manifest",
+    "verification metadata"
+  ],
+  "extensions": {
+    "legacySimulators": [
+      {
+        "id": "image-batch-edit-media-fixtures",
+        "kind": "data-source",
+        "status": "implemented",
+        "description": "Seeded local media index replacing real audio, video, image generation, and external media metadata services."
+      }
+    ]
+  }
+}

package/data/craftbook-templates/im/image-dedup-cluster/versions/1.0.2/test.json

@@ -23,9 +23,6 @@
     "worker": {
       "name": "Jules",
       "role": "Producer"
-    },
-    "craftbookParams": {
-      "workPath": "tasks/eval"
     }
   },
   "mocks": [],

package/data/craftbook-templates/im/image-dedup-cluster/versions/1.0.3/craftbook.json

@@ -0,0 +1,276 @@
+{
+  "id": "image-dedup-cluster",
+  "name": "Image Dedup & Cluster",
+  "description": "Find and group near-duplicate images across a folder — exact dupes, resized/re-encoded copies, and visually similar shots from the same burst — using perceptual hashing or visual similarity, then report clusters with a recommended keeper per group. Lock the similarity definition and thresholds FIRST (what counts as a duplicate vs merely similar) because the whole result hinges on that line; then analyze pairwise, cluster, and produce a human-reviewable report. Covers deduplication, near-duplicate detection, perceptual hashing, image clustering, burst grouping, and library cleanup.\n\nA gallery craftbook generated from an archetype spec. It runs\n`phase → (per-phase gate) → … → evaluate → (loop) → finish`. Each build\nphase that produces a checkable artifact is followed by a **runtime\ngate-checkpoint** — the runtime verifies the artifact and routes with no\nmodel turn, looping back to redo the phase on a miss. The final `evaluate`\nstep holds a static deliverable gate plus a reviewer QA pass. What it adds\nover the generic `build-loop`: a specialist role per phase, a\ndomain-correct ordering, and a concrete per-phase quality bar.\n\nDeliverables marked \"artifact\" land in the project's artifacts drawer (`write_artifact` / `read_artifact`), not the shipped workspace — review output is not product source.\n\nPhases:\n\n1. Scope the dedup (planner) — define similarity, thresholds, and the keeper rule → gated on artifact `{{workPath}}/scope.md` (markdown-notes)\n2. Analyze similarity (developer) — hash, compare, and cluster the images → gated on artifact `{{workPath}}/analyze.md` (markdown-notes)\n3. Write the report (developer) — produce the reviewable dedup report → gated on artifact `{{workPath}}/report.md` (markdown-report)\n\nThe gates never advance with an unmet criterion, and loop back to the\nowning phase to fix named gaps.\n",
+  "entryStepId": "scope",
+  "triggers": [
+    "find duplicate images",
+    "cluster similar photos",
+    "dedupe my image folder",
+    "near-duplicate detection",
+    "group similar images"
+  ],
+  "paramSchema": {
+    "type": "object",
+    "properties": {
+      "workPath": {
+        "type": "string",
+        "title": "Working folder",
+        "description": "Per-task working folder in the artifacts drawer. Defaults to this task's own folder so runs never collide; override with a stable name when you deliberately want runs to share files.",
+        "default": "{{task.dir}}"
+      }
+    }
+  },
+  "steps": [
+    {
+      "id": "scope",
+      "name": "Scope the dedup",
+      "description": "define similarity, thresholds, and the keeper rule",
+      "prompt": "1) List the folder and count the images. 2) Define the similarity method: exact-bytes hash for true dupes PLUS a perceptual hash (aHash/pHash/dHash) or embedding distance for near-dupes, and set the THRESHOLD (e.g. Hamming distance <= N) that separates 'duplicate' from 'merely similar'. 3) Define the KEEPER rule for a cluster (e.g. highest resolution, largest file, earliest timestamp). 4) Decide the report shape: clusters with member filenames, the chosen keeper, and the would-free byte savings. 5) Write an acceptance-criteria checklist: 'every image appears in exactly one cluster (singletons allowed)', 'the threshold is stated and applied consistently', 'each multi-member cluster names a keeper', 'reported savings are computed from real file sizes', 'no false-merge of clearly different images'. 6) Record method + thresholds + checklist via write_task_note AND to {{workPath}}/scope.md.\n\nThe deliverable `{{workPath}}/scope.md` lands in the project's artifacts drawer — write it with `write_artifact` and read it back with `read_artifact`; the shipped workspace stays untouched.",
+      "suggestedRole": "planner",
+      "advanceWhen": {
+        "file": "{{workPath}}/scope.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/scope.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/scope.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "scope",
+        "maxAttempts": 3
+      },
+      "next": "analyze",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "artifact",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "analyze",
+      "name": "Analyze similarity",
+      "description": "hash, compare, and cluster the images",
+      "prompt": "1) Read scope.md. 2) Compute the exact and perceptual hashes for every image (prefer a small rerunnable script; fall back to visual comparison via the vision capability for small sets). 3) Compare pairwise and group images whose distance is within the threshold into clusters (transitively — if A~B and B~C, they share a cluster). 4) Apply the keeper rule to each multi-member cluster. 5) Spot-verify a couple of clusters by actually looking at the images to catch false merges. 6) Stage the cluster assignments and per-image metadata (hash, size, dimensions) to a working note. write_task_note the cluster count, duplicate count, and any uncertain merges. Good looks like: tight clusters where members truly are the same shot.\n\nThe deliverable `{{workPath}}/analyze.md` lands in the project's artifacts drawer — write it with `write_artifact` and read it back with `read_artifact`; the shipped workspace stays untouched.",
+      "suggestedRole": "developer",
+      "advanceWhen": {
+        "file": "{{workPath}}/analyze.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/analyze.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/analyze.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "analyze",
+        "maxAttempts": 3
+      },
+      "next": "report",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "artifact",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "report",
+      "name": "Write the report",
+      "description": "produce the reviewable dedup report",
+      "prompt": "1) Read the cluster assignments. 2) Write {{workPath}}/report.md: a summary (total images, N clusters, M duplicates found, estimated bytes freed), then one section per multi-member cluster listing its members (filename, dimensions, size), the recommended KEEPER (and why), and the ones to remove. 3) Include a short methodology line (method + threshold used) so the result is auditable. 4) List singletons compactly at the end. 5) Be explicit about any low-confidence clusters the human should eyeball. 6) On a loop-back, fix only the named issues (a bad merge, a wrong keeper). 7) write_task_note the report path. The gated deliverable is {{workPath}}/report.md.\n\nThe deliverable `{{workPath}}/report.md` lands in the project's artifacts drawer — write it with `write_artifact` and read it back with `read_artifact`; the shipped workspace stays untouched.",
+      "suggestedRole": "developer",
+      "advanceWhen": {
+        "file": "{{workPath}}/report.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/report.md",
+            "bytes": 1500,
+            "artifact": true
+          },
+          {
+            "kind": "contains",
+            "file": "{{workPath}}/report.md",
+            "pattern": "(?:^|\\n)#{1,3}\\s+\\S",
+            "flags": "i",
+            "label": "at least one markdown heading",
+            "artifact": true
+          }
+        ],
+        "onReject": "report",
+        "maxAttempts": 4
+      },
+      "next": "evaluate",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "artifact",
+        "additionalOutputMedia": [
+          "task-note"
+        ]
+      }
+    },
+    {
+      "id": "evaluate",
+      "name": "Evaluate",
+      "description": "Grade the deliverable against every acceptance criterion. All pass → finish; any fail → loop back and fix the gap.",
+      "prompt": "1) Open {{workPath}}/report.md and confirm it states the method + threshold and a summary with cluster/duplicate counts (PASS/FAIL). 2) Verify every image is accounted for in exactly one cluster or as a singleton (PASS/FAIL). 3) Open 2-3 reported clusters' member images — are they genuinely the same/near-same shot (no false merges) and is the keeper choice sensible (PASS/FAIL)? 4) Confirm byte-savings are computed from real sizes. 5) Write PASS/FAIL per criterion. False merges or unaccounted images → loop back to report (or analyze if clustering is wrong).\n\nThe deliverable lives in the project's artifacts drawer — open `{{workPath}}/report.md` with `read_artifact`, not `read_file`.\n\nThen route — this is the whole point of the loop:\n\n- **Every criterion PASSES →** call `advance_task_step({ ref, stepId: \"evaluate\", next: \"finish\" })`.\n- **Any criterion FAILS →** write the specific gaps to notes, then call `advance_task_step({ ref, stepId: \"evaluate\", next: \"report\" })` to loop back. The builder fixes exactly those gaps.\n\nNever route to `finish` while any criterion is unmet. The build phase's completion gate already blocked a grossly-incomplete deliverable; your job is the judgment an automated check cannot make (does it actually work, read well, look right). After ~3 unproductive loops, stop and report DONE_WITH_CONCERNS so the user can step in.",
+      "suggestedRole": "reviewer",
+      "consumes": [
+        {
+          "file": "{{workPath}}/report.md",
+          "artifact": true
+        }
+      ],
+      "next": "report",
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "task-note"
+      }
+    },
+    {
+      "id": "finish",
+      "name": "Finish",
+      "description": "All acceptance criteria met. Stamp a short summary and report DONE.",
+      "prompt": "Every acceptance criterion passed. Write a one-paragraph DONE summary to task notes via `write_task_note`: what was built, the deliverable path(s), and a one-line confirmation that each criterion is met. Then report DONE.",
+      "suggestedRole": "developer",
+      "terminal": true,
+      "toolPolicy": {
+        "disallowBuiltinToolsets": [
+          "ai-apps",
+          "archives",
+          "artifacts",
+          "audio",
+          "browser-automation",
+          "code-execution",
+          "craftbooks",
+          "data-tables",
+          "entity-intel",
+          "git",
+          "image-intel",
+          "images",
+          "role-delegation",
+          "role-delegation-escalation",
+          "security-intel",
+          "team-management",
+          "videos",
+          "web",
+          "workspace-fs-write"
+        ],
+        "outputMedium": "task-note"
+      }
+    }
+  ],
+  "version": "1.0.3",
+  "releasedAt": "2026-09-01T20:00:00Z",
+  "minGezelVersion": "1.26233"
+}