Add Streamlit analysis stories and visual design review

.github/workflows/ci.yml

@@ -14,9 +14,10 @@ jobs:
         with:
           enable-cache: true
       - run: uv python install 3.12
-      - run: uv sync --frozen
-      - run: uv run ruff check src tests scripts
-      - run: uv run ruff format --check src tests scripts
+      - run: uv sync --frozen --extra dashboard --extra dashboard-review
+      - run: uv run playwright install --with-deps chromium
+      - run: uv run ruff check src tests scripts dashboards
+      - run: uv run ruff format --check src tests scripts dashboards
       - run: uv run vulture src --min-confidence 100
       - run: uv run pytest -q
   secrets:

.streamlit/config.toml

@@ -0,0 +1,24 @@
+[server]
+address = "127.0.0.1"
+headless = true
+fileWatcherType = "none"
+
+[browser]
+gatherUsageStats = false
+
+[client]
+toolbarMode = "minimal"
+showErrorDetails = false
+
+[theme]
+base = "light"
+primaryColor = "#176581"
+backgroundColor = "#FAF9F6"
+secondaryBackgroundColor = "#F0EEE8"
+textColor = "#203646"
+font = "sans-serif"
+headingFont = "serif"
+baseFontSize = 17
+baseRadius = "small"
+borderColor = "#D9D6CD"
+showWidgetBorder = true

AGENTS.md

@@ -33,6 +33,12 @@ infrastructure last. The agent's value is measured, not asserted.
   Owner-selected specialist default: `chatgpt-plan/gpt-6.1-sol`, medium reasoning. The `analyze`
   CLI resolves `--model`, then `BACKINTEL_ANALYSIS_MODEL`, then that built-in default.
   Scripted integration proves wiring, not live model quality; automatic delegation stays off.
+- `presentation/review.py`: optional rendered dashboard review after verified export,
+  also available through `dashboard review`. Native Anthropic reviewer, existing Sol
+  writer, at most two presentation-only revisions. Host-owned Playwright capture and
+  strict DesignPlan validation preserve numerical tables and chart bindings. Review
+  acceptance is separate from analytical completion. Install dashboard/dashboard-review
+  extras and Chromium; no browser capability is added to the offline analytical VM.
 - Contract (`contracts.py`): the analyst returns an `AnalystAnswer` with `evidence` (the exact
   SQL or Python that produced the number), `method` and `caveats`. Code re-executes the
   evidence; an answer that cannot be reproduced from its evidence is scored as unsupported.

README.md

@@ -23,6 +23,7 @@ not a product version. Code is available under the [MIT license](LICENSE).
 | Data engineering | Schema discovery, bounded profiling, read-only source queries and transformations inside a run | No production ETL orchestration, source writes or customer connector service |
 | Data science | Descriptive analysis and executable calculation checks using Python/SQL | Forecasting, causal inference and trained predictive workflows are not validated capabilities |
 | Reporting | HTML reports; bar, line, scatter, histogram and heatmap charts; PNG/SVG; offline interactive HTML; CSV and XLSX | Local file bundles with bounded tables; no hosted dashboard or automated delivery service |
+| Dashboard stories | Automatic Streamlit dashboard and business workbook from completed analysis reports | Trusted local app over report tables; hosting and arbitrary custom app generation remain planned |
 | Repeated analysis | Standing questions and historical snapshot replays that hide future observations | Replayed invocations, not a durable scheduler |
 | Improvement | Versioned cases, reproducible evidence, LangSmith traces/scores and supervised candidate comparisons | No automatic weight training, deployment or self-promotion |
 
@@ -141,6 +142,24 @@ Use a **new output directory** for every run. A valid refusal explains the missi
 it does not export a substitute report. A failed or unfinished analysis is not a successful
 delivery. The example's figures are computed from the downloaded source, not hardcoded.
 
+### Explore a dashboard story
+
+The first Streamlit presentation turns the retained taxi analysis into a business story,
+with readable measure names, focused charts, monthly exploration and a clean workbook.
+Run IDs, hashes and execution traces stay outside the business view. See
+[dashboard stories](docs/dashboard-stories.md) for setup, design guidance and current limits.
+Every successful `analyze` run now also creates a `dashboard/` directory. Open it with
+`uv run --extra dashboard backintel dashboard serve data/output/retail-report/dashboard`.
+Existing reports can be converted with `backintel dashboard generate --report /path/to/report.json
+--output /path/to/new-dashboard`. Generation makes no additional model calls.
+
+An optional **visual design reviewer** inspects the rendered dashboard at desktop and
+phone widths, checks navigation and downloads, and asks the writer for at most two
+presentation revisions. Add `--review-dashboard` to `analyze`, or review a retained
+dashboard with `backintel dashboard review`. The analyst/writer stays on Sol; the reviewer
+uses the native Anthropic account. Scores and screenshots can be traced in LangSmith.
+See [review setup, limits and commands](docs/dashboard-stories.md#design-review-and-revision).
+
 ### Configuration and credentials
 
 BackIntel reads environment variables; it does not load `.env` files. Supply credentials

dashboards/taxi/presentation.json

@@ -0,0 +1,82 @@
+{
+  "title": "Planning taxi shifts with trip data",
+  "introduction": "The first proposed test is evening shift coverage. Compare airport and city pickup areas before moving vehicles.",
+  "context": {
+    "audience": "A historical case study for a taxi fleet manager.",
+    "decision": "The decision: when and where should we test different shift coverage?",
+    "coverage": "NYC Taxi and Limousine Commission yellow-taxi records, Jan–Mar 2025: 10.46 million trips after payment, fare and duration checks. This is citywide activity, not one company's fleet.",
+    "limitation": "Recorded trips do not reveal unmet demand, idle vehicles or operating costs. These records cannot establish profit."
+  },
+  "sections": [
+    {
+      "title": "1. The starting point",
+      "takeaway": "Daily trips increased across the quarter. Daily averages make months of different lengths comparable.",
+      "columns": 1,
+      "charts": [{
+        "chart": 0,
+        "title": "Average recorded trips per day",
+        "reading_guide": "One column per month; height shows its trip count divided by calendar days.",
+        "caption": "About 105,600 trips per day in January versus 124,100 in March. This establishes a busier historical baseline; it does not explain the increase or forecast the next quarter. Next, look at when those trips occurred.",
+        "encoding": "column",
+        "x_label": "Month",
+        "y_label": "Average recorded trips"
+      }]
+    },
+    {
+      "title": "2. When trips happen",
+      "takeaway": "Evenings are the first timing pattern to investigate. A small shift overlap is a more testable proposal than adding vehicles based on trip counts alone.",
+      "columns": 1,
+      "charts": [{
+        "chart": 1,
+        "title": "Which hours were busiest?",
+        "reading_guide": "Read down for the hour and across for the weekday. Darker cells mean more recorded pickups in an average occurrence of that hour. Times are New York local time.",
+        "caption": "Thursday 6–7 pm averaged 10,289 pickups across the city; Friday 6–7 pm averaged 9,761. These are averages across the quarter, not counts for one date. A suggested test is a limited evening overlap, measuring pickup waits and net receipts per available vehicle-hour against unchanged shifts.",
+        "encoding": "original",
+        "x_label": "Hour",
+        "y_label": "Day"
+      }]
+    },
+    {
+      "title": "3. Where to compare",
+      "takeaway": "Airports generate large fares, but a long airport queue could offset that advantage. Compare an airport waiting location with a city pickup area before moving vehicles.",
+      "columns": 1,
+      "charts": [{
+        "chart": 2,
+        "title": "Ten pickup areas with the most metered fares",
+        "reading_guide": "Longer bars mean a larger sum of meter fares from trips starting in that area. This measures fare volume, not the number of passengers or driver profit.",
+        "caption": "JFK contributed $24.33m and LaGuardia $11.49m. Yet Midtown Center had 485,609 trip records versus JFK's 383,928. An airport-versus-city test must include waiting, driving and repositioning time. The chart shows only the ten leading areas.",
+        "encoding": "original",
+        "x_label": "Pickup area",
+        "y_label": "Metered fares"
+      }]
+    },
+    {
+      "title": "4. The everyday trips",
+      "takeaway": "Do not focus only on high fares. Trips with meter fares below $20 make up most recorded trips, so a city route remains a useful comparison for any airport test.",
+      "columns": 1,
+      "charts": [{
+        "chart": 3,
+        "title": "How much did a typical trip cost on the meter?",
+        "reading_guide": "Each bar is the percentage of all eligible trip records in a meter-fare band. $10–<20 means at least $10 and less than $20. A fare band is not a distance band.",
+        "caption": "44.71% of trips had meter fares from $10 to under $20; another 28.45% were below $10. High ticket size alone does not establish productivity. Meter fares exclude separate taxes, tolls, tips and surcharges.",
+        "encoding": "original",
+        "x_label": "Meter fare",
+        "y_label": "Share of trip records"
+      }]
+    },
+    {
+      "title": "5. Check the monthly mix",
+      "takeaway": "Airport contribution changes by month. Recheck the mix when evaluating a trial instead of assuming the same airport share all year.",
+      "columns": 1,
+      "charts": [{
+        "chart": 4,
+        "title": "Where each month's meter fares came from",
+        "reading_guide": "Each cell is a percentage of that month's total meter fares. Read across a row to compare months. Use the printed percentages for the two airport rows; other pickup areas account for most fares.",
+        "caption": "JFK's share was 13.93% in January, 11.59% in February and 12.28% in March. These are shares of fares, not trips or profit. The next operational step is a small timing or location test with fresh data and measured costs—not a fleet-wide rollout from this historical quarter.",
+        "encoding": "original",
+        "x_label": "Month",
+        "y_label": "Pickup group"
+      }]
+    }
+  ]
+}