@bendyline/gilde 0.1.26 → 0.1.28

package/data/craftbook-templates/no/notification-router/versions/1.0.1/test.json

@@ -0,0 +1,162 @@
+{
+  "schemaVersion": 1,
+  "title": "Notification Router smoke eval",
+  "objective": "Self-contained smoke eval for the Notification Router craftbook using the external generic harness.",
+  "tags": [
+    "external"
+  ],
+  "prompt": "Theres a fake ops service wired up for this so nothing real gets touched — the endpoints are in mocks/services.md, and the ready-made mock-probe script talks to them for you. Can you run the automation against it and write up how it went in automation.md?",
+  "setup": {
+    "projectName": "Notification Router Eval",
+    "about": "Self-contained eval project for notification-router. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/automation.md.",
+    "missionObjectives": "Use the Notification Router craftbook/template, read the seeded local fixtures, and write automation.md without network calls, real credentials, or live services.",
+    "files": [
+      {
+        "path": "source/brief.md",
+        "content": "# Notification Router 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: notification-router - Notification Router.\n"
+      }
+    ],
+    "worker": {
+      "name": "Jules",
+      "role": "Developer"
+    }
+  },
+  "mocks": [
+    {
+      "kind": "http",
+      "id": "ops",
+      "description": "Fake operations service for this eval. GET /api/bookings/open lists the open work items; POST /api/bookings/dry-run records a simulated submission and returns DRY_RUN_OK.",
+      "credential": {
+        "name": "mock.ops"
+      },
+      "routes": [
+        {
+          "method": "GET",
+          "path": "/api/bookings/open",
+          "body": {
+            "items": [
+              "BKG-1001",
+              "BKG-1002",
+              "BKG-1003"
+            ]
+          }
+        },
+        {
+          "method": "POST",
+          "path": "/api/bookings/dry-run",
+          "status": 201,
+          "body": {
+            "status": "DRY_RUN_OK",
+            "recorded": 3
+          }
+        }
+      ]
+    },
+    {
+      "kind": "webhook",
+      "id": "notify",
+      "description": "Notification receiver. POST a JSON completion notice here when the run finishes.",
+      "path": "/hooks/notify"
+    },
+    {
+      "kind": "cli",
+      "id": "probe",
+      "description": "Provenance-trusted project script `mock-probe` that exercises the fake ops service through http.authed and returns the raw responses.",
+      "shim": {
+        "path": "scripts/mock-probe.ts",
+        "content": "import { defineScript, gezel } from '@bendyline/gezel-sdk';\n\nexport const meta = defineScript({\n  name: 'mock-probe',\n  description:\n    'Probe the live fake ops service for this eval: lists the open work items and records a dry-run submission. Outputs the raw JSON responses as evidence for the write-up.',\n  inputs: {},\n  outputs: {\n    open: { type: 'string', description: 'Raw JSON body from the open-items listing.' },\n    dryRun: { type: 'string', description: 'Raw JSON body from the dry-run submission.' },\n  },\n  requires: ['workspace.read', 'network', 'credential:mock.ops'],\n});\n\nconst services = JSON.parse(await gezel.fs.read('mocks/services.json'));\nconst base = (id) => {\n  const service = services.find((entry) => entry.id === id);\n  if (!service) throw new Error(`mock service \"${id}\" is not listed in mocks/services.json`);\n  return service.baseUrl;\n};\n\nconst open = await gezel.http.authed(`${base('ops')}/api/bookings/open`, {\n  credential: 'mock.ops',\n});\nif (!open.ok) throw new Error(`open-items listing failed: ${open.status} ${open.body}`);\n\nconst dryRun = await gezel.http.authed(`${base('ops')}/api/bookings/dry-run`, {\n  credential: 'mock.ops',\n  method: 'POST',\n  body: JSON.stringify({ items: JSON.parse(open.body).items, mode: 'dry-run' }),\n});\nif (!dryRun.ok) throw new Error(`dry-run submission failed: ${dryRun.status} ${dryRun.body}`);\n\ngezel.output({ open: open.body, dryRun: dryRun.body });\n"
+      }
+    }
+  ],
+  "success": {
+    "summary": "automation.md is grounded in live fake-service responses: the open items were actually listed and a dry-run actually recorded (request-log assertions), with safety guards and signals described.",
+    "deliverables": [
+      {
+        "path": "automation.md",
+        "kind": "markdown-report",
+        "minBytes": 900,
+        "checks": [
+          {
+            "kind": "contains",
+            "file": "automation.md",
+            "pattern": "(?:^|\\n)#{1,3}\\s+\\S",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "automation.md",
+            "pattern": "DRY_RUN_OK|dry-run|dry run",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "automation.md",
+            "pattern": "fake-ci|fake-github|/api/bookings",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "automation.md",
+            "pattern": "no real|credentials|safety|guard",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "automation.md",
+            "pattern": "success|failure|retry|abort",
+            "flags": "i"
+          }
+        ]
+      }
+    ],
+    "mocks": [
+      {
+        "service": "ops",
+        "minRequests": 2,
+        "requiredPaths": [
+          "^/api/bookings/open$",
+          "^/api/bookings/dry-run$"
+        ]
+      }
+    ]
+  },
+  "rubric": {
+    "artifact": {
+      "path": "automation.md",
+      "kind": "markdown"
+    },
+    "axes": [
+      {
+        "name": "safety",
+        "description": "Dry-run posture is explicit: no real credentials, clear stop-before-side-effect boundary."
+      },
+      {
+        "name": "mapping",
+        "description": "Fake endpoints/commands are mapped completely to the workflow steps."
+      },
+      {
+        "name": "signals",
+        "description": "Success, failure, retry, and abort conditions are all defined."
+      },
+      {
+        "name": "containment",
+        "description": "The plan cannot accidentally reach a real service from a mock context."
+      }
+    ]
+  },
+  "qualityFocus": [
+    "dry-run safety",
+    "fake tool mapping",
+    "side-effect containment"
+  ],
+  "extensions": {
+    "legacySimulators": [
+      {
+        "id": "notification-router-fake-service-fixture",
+        "kind": "data-source",
+        "status": "implemented",
+        "description": "Seeded local fake service contract replacing live CLIs, HTTP APIs, credentials, and side effects."
+      }
+    ]
+  }
+}

package/data/craftbook-templates/oc/ocr-extract/versions/1.0.1/craftbook.json

@@ -0,0 +1,137 @@
+{
+  "id": "ocr-extract",
+  "name": "OCR a Folder of Scans",
+  "description": "Extract the text from a folder of scanned documents or photographed pages into clean, structured text — preserving reading order, paragraph breaks, and per-file provenance — then verify the extraction against the source images. Scope FIRST the languages, output structure (one file per scan vs a combined doc), and how to handle tables/columns/low-confidence regions, because OCR quality is dominated by handling layout and flagging uncertainty; then run extraction and verify a sample reads correctly. Covers OCR, text extraction, scanned documents, digitizing paper, receipt/invoice text, and image-to-text.\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 extraction (planner) — languages, output structure, layout & confidence rules → gated on artifact `notes/scope.md` (markdown-notes)\n2. Extract the text (developer) — OCR each scan into structured text → gated on `out` (generic-file)\n3. Verify the extraction (reviewer) — check a sample of outputs against the source images → gated on artifact `notes/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": [
+    "ocr a folder of scans",
+    "extract text from images",
+    "digitize scanned documents",
+    "image to text",
+    "read text from photos"
+  ],
+  "steps": [
+    {
+      "id": "scope",
+      "name": "Scope the extraction",
+      "description": "languages, output structure, layout & confidence rules",
+      "prompt": "1) List the scan folder and note formats and roughly what the pages are (typed, handwritten, mixed, tables). 2) Decide the language(s) for OCR. 3) Lock the output structure: one .txt/.md per source (named to match) and/or a combined document, with each block tagged by source filename for provenance. 4) Decide layout handling — preserve paragraph breaks and reading order, how to render tables (markdown tables or labeled rows), and how to flag low-confidence or unreadable regions (e.g. [illegible]). 5) Write an acceptance-criteria checklist: 'every scan produces an output', 'reading order and paragraph breaks are preserved', 'tables are not scrambled into a single line', 'low-confidence text is flagged not silently dropped', 'output is plain readable text, not garbled'. 6) Record scope + checklist via write_task_note AND to notes/scope.md.\n\nThe deliverable `notes/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": "notes/scope.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "notes/scope.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "notes/scope.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "scope",
+        "maxAttempts": 3
+      },
+      "next": "extract"
+    },
+    {
+      "id": "extract",
+      "name": "Extract the text",
+      "description": "OCR each scan into structured text",
+      "prompt": "1) Read scope.md. 2) For each scan, run OCR — prefer a real OCR engine (Tesseract) or the vision capability reading the image directly. 3) Preserve reading order and paragraph structure; for multi-column pages, read column-by-column not left-to-right across columns. 4) Render tables as the agreed structure rather than collapsing them. 5) Flag uncertain text with the agreed marker instead of inventing words. 6) Write each output to the agreed path with source provenance, and assemble the combined doc if requested. 7) Keep a running count of scans processed. write_task_note the output location and counts. Good looks like: text a human could read as the original document, with uncertainty honestly flagged.",
+      "suggestedRole": "developer",
+      "advanceWhen": {
+        "file": "out",
+        "minBytes": 1,
+        "sniff": "nonempty"
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "out",
+            "bytes": 80
+          },
+          {
+            "kind": "sniff",
+            "file": "out",
+            "sniff": "nonempty"
+          }
+        ],
+        "onReject": "extract",
+        "maxAttempts": 3
+      },
+      "next": "verify"
+    },
+    {
+      "id": "verify",
+      "name": "Verify the extraction",
+      "description": "check a sample of outputs against the source images",
+      "prompt": "1) Re-read scope.md's checklist. 2) Confirm output count equals scan count (PASS/FAIL). 3) Pick 3-4 scans, open BOTH the image and its extracted text, and compare: is the text accurate, in correct reading order, with paragraphs/tables intact (PASS/FAIL)? 4) Confirm low-confidence regions are flagged, not silently wrong or dropped. 5) Confirm provenance — each block traces to its source file. 6) Write a verification note with PASS/FAIL per criterion and the specific files you checked. On a loop-back the extractor fixes the flagged files. write_task_note the result.\n\nThe deliverable `notes/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": "notes/verify.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "out",
+            "bytes": 80
+          },
+          {
+            "kind": "sniff",
+            "file": "out",
+            "sniff": "nonempty"
+          }
+        ],
+        "onReject": "verify",
+        "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 out/ and confirm every scan has an output (PASS/FAIL). 2) For 3-4 files, open the source image alongside the extracted text and check accuracy, reading order, and that tables/columns are not scrambled (PASS/FAIL). 3) Confirm uncertain text is flagged, not silently dropped or hallucinated. 4) Confirm provenance back to source files. 5) Reconcile verify.md's claims with what you observe. 6) Write PASS/FAIL per criterion. Inaccurate/garbled extraction → loop back to extract.\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: \"extract\" })` 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": "extract"
+    },
+    {
+      "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": "1.0.1",
+  "releasedAt": "2026-08-14T00:00:00Z"
+}

package/data/craftbook-templates/oc/ocr-extract/versions/1.0.1/test.json

@@ -0,0 +1,103 @@
+{
+  "schemaVersion": 1,
+  "title": "OCR a Folder of Scans smoke eval",
+  "objective": "Self-contained smoke eval for the OCR a Folder of Scans 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": "OCR a Folder of Scans Eval",
+    "about": "Self-contained eval project for ocr-extract. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/media-plan.json.",
+    "missionObjectives": "Use the OCR a Folder of Scans 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": "# OCR a Folder of Scans 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: ocr-extract - OCR a Folder of Scans.\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": "ocr-extract-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/of/offer-compare/versions/1.0.1/craftbook.json

@@ -0,0 +1,106 @@
+{
+  "id": "offer-compare",
+  "name": "Offer Comparison",
+  "description": "Turn competing job offers into one honest decision sheet — every offer normalized into the same fact grid (comp, equity with vesting caveats, benefits, level, location, growth, deadlines), unknowns marked UNKNOWN instead of guessed, computed year-1 and 4-year views under stated assumptions, a weighted read against the candidate's OWN priorities, and a recommendation that separates arithmetic from judgment. Use when any second offer (or a strong counter) lands.\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. Normalize the facts (planner) — one fact grid across every offer, unknowns honest, the user's own weights elicited → gated on artifact `notes/offer-facts.md` (markdown-notes)\n2. Build the comparison (planner) — the decision sheet — table, computed views, weighted read, recommendation → gated on artifact `reports/offer-compare.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": "facts",
+  "triggers": [
+    "compare my offers",
+    "which offer should I take",
+    "offer comparison",
+    "help me decide between offers"
+  ],
+  "steps": [
+    {
+      "id": "facts",
+      "name": "Normalize the facts",
+      "description": "one fact grid across every offer, unknowns honest, the user's own weights elicited",
+      "prompt": "Build the fact sheet before any comparison. Step 1: collect every offer's terms from the user or project documents into one grid: base, bonus (target and realism), equity (kind, vesting schedule, valuation caveats), benefits, PTO, level/title, location/remote terms, team and manager, growth story, response deadline. Step 2: mark every unknown as UNKNOWN — never estimate a number the offer letter does not state; list the UNKNOWNs as questions to ask each company. Step 3: elicit the user's OWN decision weights — what actually matters to them (comp, growth, stability, location, mission, team) — do not invent or assume weights. Step 4: state the assumptions any computed view will use (equity valuation basis, bonus attainment). Step 5: write an acceptance-criteria checklist ('every offer covered on every factor', 'unknowns marked UNKNOWN', 'weights are the user's own words', 'assumptions stated'). Write the fact sheet + checklist via write_task_note AND to notes/offer-facts.md.\n\nThe deliverable `notes/offer-facts.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": "notes/offer-facts.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "notes/offer-facts.md",
+            "bytes": 120,
+            "artifact": true
+          },
+          {
+            "kind": "sniff",
+            "file": "notes/offer-facts.md",
+            "sniff": "nonempty",
+            "artifact": true
+          }
+        ],
+        "onReject": "facts",
+        "maxAttempts": 3
+      },
+      "next": "compare"
+    },
+    {
+      "id": "compare",
+      "name": "Build the comparison",
+      "description": "the decision sheet — table, computed views, weighted read, recommendation",
+      "prompt": "Write reports/offer-compare.md from the fact sheet. Step 1: a factor-by-offer markdown comparison table — offers as columns, factors as rows, UNKNOWNs still visible. Step 2: computed lines under the STATED assumptions: year-1 cash, and a 4-year total-comp view with the equity caveats repeated next to the number. Step 3: a weighted read against the user's stated priorities — show how each offer scores on what THEY said matters, not on a generic rubric. Step 4: risks and negotiation levers per offer (deadline pressure, level mismatch, unvested cliff, remote-policy risk). Step 5: a recommendation that separates the arithmetic ('offer B leads on 4-year comp under these assumptions') from the judgment ('offer A fits the growth priority you weighted highest') — and says plainly which way the sheet leans and why. On a loop-back, fix ONLY the named issues.\n\nThe deliverable `reports/offer-compare.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": "reports/offer-compare.md",
+        "minBytes": 1,
+        "sniff": "nonempty",
+        "artifact": true
+      },
+      "gate": {
+        "at": "completion",
+        "checks": [
+          {
+            "kind": "minBytes",
+            "file": "reports/offer-compare.md",
+            "bytes": 1500,
+            "artifact": true
+          },
+          {
+            "kind": "contains",
+            "file": "reports/offer-compare.md",
+            "pattern": "(?:^|\\n)#{1,3}\\s+\\S",
+            "flags": "i",
+            "label": "at least one markdown heading",
+            "artifact": true
+          }
+        ],
+        "onReject": "compare",
+        "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": "Audit the decision sheet. Step 1: recompute the year-1 and 4-year lines from the fact sheet — FAIL on any arithmetic that does not reproduce, or any computed number whose assumption is not stated beside it. Step 2: check every offer appears in every table row and that fact-sheet UNKNOWNs were not silently replaced with guesses. Step 3: trace the weighted read to the user's stated priorities — FAIL a rubric the user never gave. Step 4: confirm the recommendation separates arithmetic from judgment and takes a clear position. Write PASS/FAIL per criterion.\n\nThe deliverable lives in the project's artifacts drawer — open `reports/offer-compare.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: \"compare\" })` 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": "reports/offer-compare.md",
+          "artifact": true
+        }
+      ],
+      "next": "compare"
+    },
+    {
+      "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": "1.0.1",
+  "releasedAt": "2026-08-14T00:00:00Z"
+}

package/data/craftbook-templates/of/offer-compare/versions/1.0.1/test.json

@@ -0,0 +1,110 @@
+{
+  "schemaVersion": 1,
+  "title": "Offer Comparison smoke eval",
+  "objective": "Self-contained smoke eval for the Offer Comparison craftbook using the corpus generic harness.",
+  "tags": [
+    "corpus"
+  ],
+  "prompt": "We need a offer comparison for the returns-desk work. The backgound info is in source/brief.md — have a look and write it up in report.md.",
+  "setup": {
+    "projectName": "Offer Comparison Eval",
+    "about": "Self-contained eval project for offer-compare. Seeded inputs are under workspace/source or workspace/fixtures; final deliverable is workspace/report.md.",
+    "missionObjectives": "Use the Offer Comparison craftbook/template, read the seeded local fixtures, and write report.md without network calls, real credentials, or live services.",
+    "files": [
+      {
+        "path": "source/brief.md",
+        "content": "# Offer Comparison 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: offer-compare - Offer Comparison.\n"
+      }
+    ],
+    "worker": {
+      "name": "Jules",
+      "role": "Writer"
+    }
+  },
+  "mocks": [],
+  "success": {
+    "summary": "report.md is a source-grounded markdown artifact with required facts and actions.",
+    "deliverables": [
+      {
+        "path": "report.md",
+        "kind": "markdown-report",
+        "minBytes": 1000,
+        "checks": [
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "(?:^|\\n)#{1,3}\\s+\\S",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "source/brief\\.md|Brief",
+            "flags": "i",
+            "label": "cite or name the seeded source"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "18\\b[\\s\\S]{0,30}\\bSKUs",
+            "flags": "i",
+            "label": "include the 18-SKU pilot scope"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "18\\s*hours|6\\s*hours",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "14\\.2%|8\\.9%",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "status silence|photo submission",
+            "flags": "i"
+          },
+          {
+            "kind": "contains",
+            "file": "report.md",
+            "pattern": "automated status emails|barcode|Finance",
+            "flags": "i"
+          }
+        ]
+      }
+    ]
+  },
+  "rubric": {
+    "artifact": {
+      "path": "report.md",
+      "kind": "markdown"
+    },
+    "axes": [
+      {
+        "name": "grounding",
+        "description": "Seeded facts appear faithfully and the source is cited."
+      },
+      {
+        "name": "structure",
+        "description": "Sections progress logically for the artifact class."
+      },
+      {
+        "name": "completeness",
+        "description": "Required elements (risks, next actions, scope) are all present."
+      },
+      {
+        "name": "tone",
+        "description": "The register fits the audience and artifact class."
+      }
+    ]
+  },
+  "qualityFocus": [
+    "source-grounded prose",
+    "section structure",
+    "next-action clarity"
+  ]
+}