feat: add date forecast type for timing questions (#5247)

## Summary

Adds a third forecast mode (`date`) alongside `binary` and `numeric`.
Date forecasts produce YYYY-MM-DD percentile estimates (p10–p90) for
"when will X happen?" questions, with prompts emphasizing delay bias and
status-quo anchoring. Output columns follow the same
`{output_field}_p{N}` naming as numeric, so `output_field` is required
and `units` is ignored.

## Changes

- **Engine** (`forecast.py`, `task_spec.py`, `agent_state.py`,
`operations.py`)
- New `DATE_FORECASTER_PROMPT` with delay bias / never-happen sentinel
(`2099-12-31`) guidance
- New `build_date_response_schema()` returning string-typed percentile
fields
- New `_combine_batched_date_results()` aggregating with **median
ordinals** (robust to the 2099 sentinel) instead of mean
- `forecast_type` literal extended to `"binary" | "numeric" | "date"`
across `DeepForecastPublicParams`, `DeepForecastFullParams`,
`ForecastBatchStateSerializable`, and `ForecastOperation`
  - Date validation requires `output_field` and ignores `units`
- **OpenAPI / SDK / MCP**
  - Regenerated OpenAPI types
- `forecast()` / `forecast_async()` SDK signatures and docstrings
updated
  - MCP `ForecastInput` and `tools.py` mode_label updated
- **everyrow-cc frontend**
- `ColumnInfo.forecastType` extended; date branch added in
`extractColumnInfo`
- New `extractDatePercentiles()` and `DatePercentileRangeBar`
(timestamp-based scaling, compact "Jun '25" labels)
- `ResearcherStreamItem`, `ResearcherDetailPanel`,
`ResearcherStreamView` rendering branches added
- **everyrow-cc agent**
- System prompt updated to mention `forecast_type="date"` for timing
questions

## Design Notes

- **Median over mean** for date aggregation: dates aren't continuous in
the same way numerics are, and median gracefully handles the 2099-12-31
"never happens" sentinel that some forecasters may emit.
- **Schema**: percentile fields use `{"type": "string"}`, mapping to
`Nullable(String)` in ClickHouse via the existing `_JSON_TO_CH` mapping.
No CH schema changes needed.
- **Backward compatible**: all changes are additive — extending
`Literal` unions and adding `elif` branches. Binary/numeric forecasts
are unaffected.

## Test plan

- [ ] `cd cohort/engine && uv run pyright src` (passes locally — 0
errors)
- [ ] `cd cohort/engine && uv run ruff check` (passes locally)
- [ ] `cd cohort/everyrow-cc/frontend && pnpm run tsc` (passes locally)
- [ ] `cd cohort/everyrow-cc/frontend && pnpm run lint` (passes locally)
- [ ] Run a date forecast end-to-end via the SDK and verify percentile
output + visualization
- [ ] Verify a numeric/binary forecast still works (regression check)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Sourced from commit a46ce2370326ec99497b3e869021b3ce4d83068d

Author: jackwildman

futuresearch-mcp/src/futuresearch_mcp/models.py

@@ -413,10 +413,12 @@ class ForecastInput(_SingleSourceInput):
         "(e.g. 'Focus on EU regulatory sources' or 'Assume resolution by end of 2027'). "
         "Leave empty when the rows are self-contained.",
     )
-    forecast_type: Literal["binary", "numeric"] = Field(
+    forecast_type: Literal["binary", "numeric", "date"] = Field(
         description="Type of forecast. 'binary': yes/no probability (0-100) for questions like "
         "'Will X happen?'. 'numeric': percentile estimates (p10-p90) for questions like "
-        "'What will the price/value/count be?'. Requires output_field when 'numeric'.",
+        "'What will the price/value/count be?'. 'date': date percentile estimates (p10-p90) "
+        "as YYYY-MM-DD strings for timing questions like 'When will X happen?'. "
+        "Requires output_field when 'numeric' or 'date'.",
     )
     output_field: str | None = Field(
         default=None,

futuresearch-mcp/src/futuresearch_mcp/tools.py

@@ -647,7 +647,7 @@ async def futuresearch_forecast(
 ) -> list[TextContent]:
     """Forecast questions about the future using deep research and multi-model ensemble.
 
-    Supports two modes:
+    Supports three modes:
 
     - **binary** (default): Forecasts probability (0-100) for YES/NO questions.
       Output columns: ``probability`` (int, 0-100) and ``rationale`` (str).
@@ -657,6 +657,11 @@ async def futuresearch_forecast(
       Output columns: ``{output_field}_p10`` through ``{output_field}_p90`` (float),
       ``units`` (str), and ``rationale`` (str).
 
+    - **date**: Forecasts date percentile estimates for timing questions.
+      Requires ``output_field`` (e.g. ``"launch_date"``).
+      Output columns: ``{output_field}_p10`` through ``{output_field}_p90``
+      (YYYY-MM-DD strings) and ``rationale`` (str).
+
     The CSV should contain at minimum a ``question`` column.  Recommended additional
     columns: ``resolution_criteria``, ``resolution_date``, ``background``.  All
     columns are passed to the research agents and forecasters.
@@ -695,9 +700,12 @@ async def futuresearch_forecast(
         task_id = str(cohort_task.task_id)
         total = len(input_data) if isinstance(input_data, pd.DataFrame) else 0
 
-    mode_label = (
-        "numeric percentile" if params.forecast_type == "numeric" else "probability"
-    )
+    if params.forecast_type == "date":
+        mode_label = "date"
+    elif params.forecast_type == "numeric":
+        mode_label = "numeric percentile"
+    else:
+        mode_label = "probability"
     return await create_tool_response(
         task_id=task_id,
         label=f"Submitted: {total} rows for {mode_label} forecasting (6 research dimensions + 3 forecasters per batch)."

src/futuresearch/generated/models/forecast_operation_forecast_type.py

@@ -4,6 +4,7 @@
 class ForecastOperationForecastType(str, Enum):
     BINARY = "binary"
     NUMERIC = "numeric"
+    DATE = "date"
 
     def __str__(self) -> str:
         return str(self.value)

src/futuresearch/ops.py

@@ -818,13 +818,13 @@ async def forecast(
     context: str | None = None,
     session: Session | None = None,
     *,
-    forecast_type: Literal["binary", "numeric"],
+    forecast_type: Literal["binary", "numeric", "date"],
     output_field: str | None = None,
     units: str | None = None,
 ) -> TableResult:
     """Forecast questions using deep research and multi-model ensemble.
 
-    Supports two modes:
+    Supports three modes:
 
     - **binary** (default): Forecasts the probability (0-100) of YES/NO questions.
       Output columns: ``probability`` (int) and ``rationale`` (str).
@@ -834,6 +834,11 @@ async def forecast(
       Output columns: ``{output_field}_p10`` through ``{output_field}_p90`` (float),
       ``units`` (str), and ``rationale`` (str).
 
+    - **date**: Forecasts percentile date estimates for timing questions.
+      Requires ``output_field`` (e.g. ``"launch_date"``).
+      Output columns: ``{output_field}_p10`` through ``{output_field}_p90``
+      (YYYY-MM-DD strings) and ``rationale`` (str).
+
     Each row is forecast using 6 parallel research agents followed by a 3-model
     forecaster ensemble, validated against FutureSearch's past-casting environment.
 
@@ -848,9 +853,9 @@ async def forecast(
             end of 2027").  Leave *None* when the rows are self-contained.
         session: Optional session. If not provided, one will be created automatically.
         forecast_type: ``"binary"`` for probability forecasts, ``"numeric"`` for
-            percentile estimates.
-        output_field: Name of the quantity being forecast (required for numeric,
-            e.g. ``"price"``, ``"count"``).
+            percentile estimates, ``"date"`` for date percentile estimates.
+        output_field: Name of the quantity being forecast (required for numeric
+            and date, e.g. ``"price"``, ``"launch_date"``).
         units: Units for numeric forecasts (e.g. ``"USD per barrel"``).
             Required when *forecast_type* is ``"numeric"``.
 
@@ -890,7 +895,7 @@ async def forecast_async(
     task: str,
     session: Session,
     input: DataFrame | UUID | TableResult,
-    forecast_type: Literal["binary", "numeric"],
+    forecast_type: Literal["binary", "numeric", "date"],
     output_field: str | None = None,
     units: str | None = None,
 ) -> EveryrowTask[BaseModel]:
@@ -900,8 +905,9 @@ async def forecast_async(
         task: Context or instructions for the forecast.
         session: Active session.
         input: Input data.
-        forecast_type: ``"binary"`` for yes/no probability, ``"numeric"`` for percentile estimates.
-        output_field: Name of the numeric quantity (required for numeric).
+        forecast_type: ``"binary"`` for yes/no probability, ``"numeric"`` for
+            percentile estimates, ``"date"`` for date percentile estimates.
+        output_field: Name of the quantity (required for numeric and date).
         units: Units for numeric forecasts (required for numeric).
 
     Returns: