@bendyline/gilde 0.1.38 → 0.1.39

package/data/craftbook-templates/pe/pest-diagnosis/versions/1.0.2/test.json

@@ -0,0 +1,125 @@
+{
+  "schemaVersion": 1,
+  "title": "Pest & Problem Diagnosis smoke eval",
+  "objective": "Task-specific eval for the Pest & Problem Diagnosis craftbook: two pre-extracted symptom photo notes become a ranked differential with a gentle-first treatment ladder, a monitoring plan, and a named escalation line.",
+  "tags": [
+    "corpus"
+  ],
+  "prompt": "somethings eating the kale and the tomato leaves have yellow spots — photo notes are in photos/. what do you thinck it is and what do we do?",
+  "setup": {
+    "projectName": "Garden Almanac — Pest Diagnosis Eval",
+    "about": "Self-contained eval project for pest-diagnosis. The symptom photos have already been pre-extracted into plain-text photo notes under photos/. The garden's plantings live at garden.json in the workspace; the treatment and monitoring plan goes to tasks/eval/report.md.",
+    "missionObjectives": "Use the Pest & Problem Diagnosis craftbook/template, reason only from the seeded photo notes and plantings, and write tasks/eval/report.md — no network calls, real credentials, or live services.",
+    "files": [
+      {
+        "path": "photos/kale-damage.jpg.txt",
+        "content": "PHOTO NOTE (pre-extracted from kale-damage.jpg)\n\nRagged irregular holes across the outer kale leaves, edges chewed inward.\nSmall velvety green caterpillars, about 2 cm, found on the leaf undersides.\nDark green droppings collected in the leaf folds.\nSmall white butterflies seen fluttering around the bed in the afternoons.\n"
+      },
+      {
+        "path": "photos/tomato-leaf.jpg.txt",
+        "content": "PHOTO NOTE (pre-extracted from tomato-leaf.jpg)\n\nLower tomato leaves show yellow blotches with brown centers.\nSeveral spots have faint concentric rings, target-like.\nAffected leaves are the lowest on the plant, spreading slowly upward.\nStems clean so far; fruit unaffected.\nOwner notes the bed gets overhead watering in the evenings.\n"
+      },
+      {
+        "path": "garden.json",
+        "content": "{\n  \"zone\": \"8b\",\n  \"lastFrost\": \"2026-04-05\",\n  \"firstFrost\": \"2026-11-05\",\n  \"beds\": [\n    {\n      \"name\": \"North Bed\",\n      \"plantings\": [\n        { \"plant\": \"tomatoes\", \"variety\": \"San Marzano\", \"planted\": \"2026-05-02\" },\n        { \"plant\": \"basil\", \"planted\": \"2026-05-16\" }\n      ]\n    },\n    {\n      \"name\": \"Greens Bed\",\n      \"plantings\": [\n        { \"plant\": \"kale\", \"variety\": \"Lacinato\", \"planted\": \"2026-06-01\" },\n        { \"plant\": \"chard\", \"planted\": \"2026-06-01\" }\n      ]\n    }\n  ]\n}\n"
+      }
+    ],
+    "worker": {
+      "name": "Wim",
+      "role": "Gardener"
+    },
+    "craftbookParams": {
+      "workPath": "tasks/eval"
+    }
+  },
+  "mocks": [],
+  "success": {
+    "summary": "tasks/eval/report.md ranks candidates for both problems with the photo-note evidence, offers the gentlest treatment first, sets a monitoring plan, names the escalation line, and never claims certainty.",
+    "deliverables": [
+      {
+        "path": "tasks/eval/report.md",
+        "kind": "markdown-report",
+        "minBytes": 800,
+        "checks": [
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "(?:^|\\n)#{1,3}\\s+\\S",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "likely|probable|possible|suspect",
+            "flags": "i",
+            "label": "candidates stated as likelihoods, not verdicts"
+          },
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "caterpillar",
+            "flags": "i",
+            "label": "engages the caterpillar evidence from the kale photo note"
+          },
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "treat",
+            "flags": "i",
+            "label": "treatment ladder present"
+          },
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "monitor|watch|check",
+            "flags": "i",
+            "label": "monitoring plan present"
+          },
+          {
+            "kind": "notContains",
+            "file": "tasks/eval/report.md",
+            "pattern": "(definitely|guaranteed|certainly) (is|will cure)",
+            "flags": "i",
+            "label": "diagnosis stays hedged, never absolute"
+          },
+          {
+            "kind": "contains",
+            "file": "tasks/eval/report.md",
+            "pattern": "extension|professional|nursery|expert",
+            "flags": "i",
+            "label": "escalation path named"
+          }
+        ]
+      }
+    ]
+  },
+  "rubric": {
+    "artifact": {
+      "path": "tasks/eval/report.md",
+      "kind": "markdown"
+    },
+    "axes": [
+      {
+        "name": "diagnostic-reasoning",
+        "description": "Each problem carries at least two ranked candidates whose evidence traces to the photo notes, including what does not fit."
+      },
+      {
+        "name": "hedging",
+        "description": "Confidence stays honest — a photo-note diagnosis is presented as a strong hunch, never a lab result."
+      },
+      {
+        "name": "actionability",
+        "description": "The treatment ladder starts gentle with concrete how-and-when, and the monitoring plan has dates."
+      },
+      {
+        "name": "safety",
+        "description": "Lowest-impact options come first, and the escalation threshold to a professional is specific."
+      }
+    ]
+  },
+  "qualityFocus": [
+    "ranked differential",
+    "gentle-first treatment",
+    "honest confidence"
+  ]
+}

package/data/craftbook-templates/ph/photo-cull/versions/1.0.2/craftbook.json

@@ -0,0 +1,152 @@
+{
+  "id": "photo-cull",
+  "name": "Photo Cull & Rank",
+  "description": "Rank and cull a set of photos by objective quality — sharpness/focus, exposure, composition, eyes-open/expression for people, and duplicate-of-a-better-shot — to surface the keepers and justify the rejects. Lock the scoring CRITERIA and weights FIRST (what 'good' means for this set: a portrait shoot weights expression, a landscape weights sharpness and horizon) because an unstated rubric makes the cull arbitrary; then score every photo and produce a ranked report with keep/reject and reasons. Covers photo culling, image quality ranking, best-shot selection, photo curation, and burst/best-of selection.\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. Scoring criteria (planner) — lock the quality rubric, weights, and keep threshold → gated on artifact `{{workPath}}/criteria.md` (markdown-notes)\n2. Score the photos (developer) — evaluate every photo against the rubric → gated on artifact `{{workPath}}/score.md` (markdown-notes)\n3. Write the cull report (developer) — produce the ranked keep/reject 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": "criteria",
+  "triggers": [
+    "cull my photos",
+    "rank photos by quality",
+    "pick the best shots",
+    "curate a photo set",
+    "find the keepers"
+  ],
+  "steps": [
+    {
+      "id": "criteria",
+      "name": "Scoring criteria",
+      "description": "lock the quality rubric, weights, and keep threshold",
+      "prompt": "1) Understand the set (portraits, landscapes, product, event) and list the photos. 2) Lock the scoring RUBRIC: the dimensions (sharpness/focus, exposure, composition/framing, subject expression & eyes-open for people, noise) and a WEIGHT per dimension appropriate to the set. 3) Define the score scale (e.g. 1-5 per dimension) and the KEEP threshold (and how many keepers are wanted, if capped). 4) Decide near-duplicate handling: among similar shots, keep only the best-scoring one. 5) Write an acceptance-criteria checklist: 'every photo gets a score on every dimension', 'weights are stated and applied', 'keep/reject follows the threshold consistently', 'each reject has a concrete reason', 'among near-dupes only the best is kept'. 6) Record rubric + weights + checklist via write_task_note AND to {{workPath}}/criteria.md.\n\nThe deliverable `{{workPath}}/criteria.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}}/criteria.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/criteria.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/criteria.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "criteria",
+        "maxAttempts": 3
+      },
+      "next": "score"
+    },
+    {
+      "id": "score",
+      "name": "Score the photos",
+      "description": "evaluate every photo against the rubric",
+      "prompt": "1) Read criteria.md. 2) For each photo, actually LOOK at it via the vision capability (and use a sharpness/blur metric if a tool is available) and assign a 1-5 on each rubric dimension with a one-line justification grounded in what's visible (e.g. 'soft focus on eyes', 'blown highlights sky'). 3) Compute the weighted total per photo. 4) Group near-duplicates and mark the best of each group. 5) Apply the keep threshold (respecting any keeper cap) to set keep/reject. 6) Stage a table of filename → per-dimension scores → total → keep/reject → reason. write_task_note the keep/reject counts. Good looks like: defensible scores a photographer would mostly agree with.\n\nThe deliverable `{{workPath}}/score.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}}/score.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/score.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/score.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "score",
+        "maxAttempts": 3
+      },
+      "next": "report"
+    },
+    {
+      "id": "report",
+      "name": "Write the cull report",
+      "description": "produce the ranked keep/reject report",
+      "prompt": "1) Read the scored table. 2) Write {{workPath}}/report.md: a summary (N photos, K keepers, the rubric + weights used), a ranked KEEPERS section (best first, with their scores and a one-line strength), and a REJECTS section grouped by reason (out of focus, underexposed, weaker duplicate, etc.) each with the photo and the concrete reason. 3) Call out near-duplicate groups, showing the kept shot vs the rejected siblings. 4) Make the cull reversible — list filenames so a human can act on it. 5) On a loop-back, fix only the named issues (a misjudged score, an inconsistent threshold). 6) 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"
+    },
+    {
+      "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 rubric + weights and a keeper/reject summary (PASS/FAIL). 2) Verify every photo is scored on every dimension and keep/reject follows the threshold consistently (PASS/FAIL). 3) Open 3-4 photos and sanity-check their scores against what you see — are obviously soft/blown shots rejected and crisp ones kept (PASS/FAIL)? 4) Confirm each reject has a concrete reason and near-dupe groups keep only the best. 5) Write PASS/FAIL per criterion. Misjudged or inconsistent culls → loop back to report (or score if the scoring 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"
+    },
+    {
+      "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
+    }
+  ],
+  "version": "1.0.2",
+  "releasedAt": "2026-08-21T00:00:00Z",
+  "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}}"
+      }
+    }
+  },
+  "minGezelVersion": "1.26233"
+}

package/data/craftbook-templates/ph/photo-cull/versions/1.0.2/test.json

@@ -0,0 +1,106 @@
+{
+  "schemaVersion": 1,
+  "title": "Photo Cull & Rank smoke eval",
+  "objective": "Self-contained smoke eval for the Photo Cull & Rank 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": "Photo Cull & Rank Eval",
+    "about": "Self-contained eval project for photo-cull. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/media-plan.json.",
+    "missionObjectives": "Use the Photo Cull & Rank 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": "# Photo Cull & Rank 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: photo-cull - Photo Cull & Rank.\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"
+    },
+    "craftbookParams": {
+      "workPath": "tasks/eval"
+    }
+  },
+  "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": "photo-cull-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/ph/physics-toy/versions/1.0.2/craftbook.json

@@ -0,0 +1,122 @@
+{
+  "id": "physics-toy",
+  "name": "Interactive Physics Toy",
+  "description": "Build a single-file interactive canvas physics or particle toy such as a bouncing-ball sandbox, gravity wells, springs, cloth, fluid, or a boids flock that the user can poke with the mouse. Defines the simulation model — forces, integration, collisions, and the interaction verbs — first, then implements a stable per-frame integrator. Physics toys feel magical only when the numerics are stable and the mouse interaction is responsive, so the concept phase pins the forces and integration scheme before any drawing, the ordering that keeps a small model from shipping a jittery or exploding sim.\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. Concept (designer) — forces, integration, collisions, interaction verbs → gated on artifact `{{workPath}}/concept.md` (markdown-notes)\n2. Build (developer) — implement the integrator + interaction in index.html → gated on `index.html` (html-page)\n\nThe gates never advance with an unmet criterion, and loop back to the\nowning phase to fix named gaps.\n",
+  "entryStepId": "concept",
+  "triggers": [
+    "physics simulation",
+    "particle toy",
+    "bouncing balls",
+    "gravity sandbox",
+    "interactive canvas",
+    "boids"
+  ],
+  "steps": [
+    {
+      "id": "concept",
+      "name": "Concept",
+      "description": "forces, integration, collisions, interaction verbs",
+      "prompt": "Define the SIMULATION and its feel. (1) Name the system (bouncing balls, gravity wells, spring lattice, boids) and the entities with their state (position, velocity, mass/radius). (2) Enumerate the forces and the integration step (explicit Euler or Verlet with a fixed dt) and how you keep it stable (clamp velocities, damping). (3) Define collisions/boundaries (walls bounce with restitution, particle-particle if any). (4) Specify the interaction verbs: what the mouse does (attract, repel, spawn, drag) and any controls (gravity slider, reset). (5) Write an acceptance-criteria checklist of 5-8 items, e.g. 'particles move continuously without exploding off-screen', 'particles bounce off the canvas edges', 'the mouse visibly affects nearby particles', 'reset clears and re-seeds', 'runs smoothly with N particles', 'no console errors'. Write to {{workPath}}/concept.md and via write_task_note. No code yet.\n\nThe deliverable `{{workPath}}/concept.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": "designer",
+      "advanceWhen": {
+        "file": "{{workPath}}/concept.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "{{workPath}}/concept.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "{{workPath}}/concept.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "concept",
+        "maxAttempts": 3
+      },
+      "next": "build"
+    },
+    {
+      "id": "build",
+      "name": "Build",
+      "description": "implement the integrator + interaction in index.html",
+      "prompt": "Implement the physics toy in a single index.html (inline CSS + JS) satisfying every criterion. (1) Set up a full-viewport <canvas> 2d context and seed the entities. (2) Run a requestAnimationFrame loop that applies forces, integrates with the chosen scheme and a fixed dt, resolves boundary/collision, then redraws. (3) Wire mouse/touch to the interaction verb (compute distance-based attraction/repulsion or spawn/drag). (4) Add the specified controls (sliders/reset). (5) Keep it numerically stable — clamp or damp so nothing explodes; choose a palette and trails or solid fills. Vanilla JS, no CDN. On a loop-back, fix only the named gaps without regressing passing criteria. write_task_note the path and passing criteria.",
+      "suggestedRole": "developer",
+      "advanceWhen": {
+        "file": "index.html",
+        "minBytes": 1,
+        "sniff": "html-complete"
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "index.html",
+            "bytes": 800
+          },
+          {
+            "kind": "htmlLint",
+            "file": "index.html"
+          }
+        ],
+        "scripts": [
+          {
+            "name": "checkHtmlComplete",
+            "scope": "standard",
+            "inputs": {
+              "file": "index.html"
+            }
+          }
+        ],
+        "onReject": "build",
+        "maxAttempts": 4
+      },
+      "next": "evaluate"
+    },
+    {
+      "id": "evaluate",
+      "name": "Evaluate",
+      "description": "Grade the deliverable against every acceptance criterion. All pass → finish; any fail → loop back and fix the gap.",
+      "prompt": "Interact with the toy. Open index.html if possible, else trace the integrator and mouse handler. Check EACH criterion from {{workPath}}/concept.md and write PASS/FAIL with a reason: do particles move without flying off or freezing, do they bounce off edges, does the mouse visibly attract/repel/spawn, does reset re-seed, is it smooth, zero console errors? Let it run 30+ seconds to confirm it stays stable (no NaN explosion). An exploding or frozen sim is a FAIL — loop back.\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: \"build\" })` 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": "index.html"
+        }
+      ],
+      "next": "build"
+    },
+    {
+      "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
+    }
+  ],
+  "version": "1.0.2",
+  "releasedAt": "2026-08-21T00:00:00Z",
+  "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}}"
+      }
+    }
+  },
+  "minGezelVersion": "1.26233"
+}

package/data/craftbook-templates/ph/physics-toy/versions/1.0.2/test.json

@@ -0,0 +1,102 @@
+{
+  "schemaVersion": 1,
+  "title": "Interactive Physics Toy smoke eval",
+  "objective": "Self-contained smoke eval for the Interactive Physics Toy craftbook using the html-game generic harness.",
+  "tags": [
+    "html-game"
+  ],
+  "prompt": "Make us a fun litle browser game around the returns-desk theme — just one file, index.html, that we can open and play.",
+  "setup": {
+    "projectName": "Interactive Physics Toy Eval",
+    "about": "Self-contained eval project for physics-toy. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/index.html.",
+    "missionObjectives": "Use the Interactive Physics Toy craftbook/template, read the seeded local fixtures, and write index.html without network calls, real credentials, or live services.",
+    "files": [
+      {
+        "path": "source/brief.md",
+        "content": "# Interactive Physics Toy 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: physics-toy - Interactive Physics Toy.\n"
+      }
+    ],
+    "worker": {
+      "name": "Jules",
+      "role": "Developer"
+    },
+    "craftbookParams": {
+      "workPath": "tasks/eval"
+    }
+  },
+  "mocks": [],
+  "success": {
+    "summary": "index.html is a complete self-contained browser game smoke artifact.",
+    "deliverables": [
+      {
+        "path": "index.html",
+        "kind": "html-game",
+        "minBytes": 3200,
+        "checks": [
+          {
+            "kind": "cssMinBytes",
+            "bytes": 600,
+            "file": "index.html"
+          },
+          {
+            "kind": "jsParses",
+            "file": "index.html"
+          },
+          {
+            "kind": "contains",
+            "file": "index.html",
+            "pattern": "<canvas|<svg|requestAnimationFrame",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "index.html",
+            "pattern": "score|state|level|turn",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "index.html",
+            "pattern": "restart|play again|new game|game over|victory|you win|you lose",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "index.html",
+            "pattern": "Boreal|returns|desk",
+            "flags": "i"
+          }
+        ]
+      }
+    ]
+  },
+  "rubric": {
+    "artifact": {
+      "path": "index.html",
+      "kind": "html"
+    },
+    "axes": [
+      {
+        "name": "playability",
+        "description": "The game is actually playable: input works, state advances, and a session can be completed."
+      },
+      {
+        "name": "completeness",
+        "description": "Core loop, scoring/win-lose, and restart are all present rather than stubbed."
+      },
+      {
+        "name": "polish",
+        "description": "Visual design and feedback (motion, states, messages) feel deliberate, not default."
+      },
+      {
+        "name": "robustness",
+        "description": "No console errors, dead controls, or states the player can get stuck in."
+      }
+    ]
+  },
+  "qualityFocus": [
+    "runtime surface",
+    "interaction loop",
+    "self-contained game"
+  ]
+}