weatherreporter/docs/integrations/promptkit.md

# Promptkit Integration

Weatherreporter uses Promptkit for all generated-text reports. The four logical prompts are `weather.daily_generated_text`, `weather.today_generated_text`, `weather.tomorrow_generated_text`, and `weather.hourly_generated_text`, each at version `2.0.0`. Their prompt assets, generated-text JSON Schemas, and Weatherreporter profile catalog are embedded by `internal/promptassets`.

## Logical Profile Catalog

Prompt definitions select a stable Weatherreporter profile ID. The embedded definitions currently use Promptkit's `openrouter` backend:

| Profile ID | Model | Reasoning effort | Timeout | Service tier | Default reports |
| --- | --- | --- | --- | --- | --- |
| `weather-light` | `deepseek/deepseek-v4-flash` | Provider default | 180 seconds | `flex` | Hourly |
| `weather-balanced` | `~google/gemini-flash-latest` | `high` | 240 seconds | `flex` | Daily, Today, Tomorrow |
| `weather-deep` | `~anthropic/claude-sonnet-latest` | `high` | 240 seconds | `flex` | None |

The `~` prefix is part of each OpenRouter rolling-alias model ID. The embedded profiles intentionally omit endpoints, credentials, temperature, `top_p`, and output-token limits.

## Selection And Active Execution

Before weather collection, Weatherreporter validates the report's exact generated-text report/schema/template catalog binding, prompt version and hash, output contract, and selected profile. Active profiles must resolve a nonblank backend and model identity. A nonblank `promptkit.profile` selects one profile ID for every report in the command; otherwise the prompt's declared default selects it. Promptkit resolves the selected definition in this order:

1. explicit in-memory profiles used by an embedding consumer or test;
2. the configured `profile_file` or `profile_dir`;
3. Weatherreporter's embedded fallback profiles; and
4. Promptkit's built-in catalog.

A source falls through only when the selected ID is absent. Each source supplies a complete definition, so profile fields are not merged. A malformed matching operator definition is an error and does not fall back.

Profiles that require a direct API key are unsupported; a profile that reports `APIKeyEnv` requires a nonblank value in that environment variable. Active results retain the selected logical profile ID and resolved backend and model. Ordinary errors, summaries, logs, and outputs exclude endpoints, credentials, rendered messages, schemas, request bodies, response bodies, and complete parameter maps.

Promptkit receives the YAML data package as an inline input and returns structured JSON that Weatherreporter validates before rendering its own Markdown template. Before accepting that JSON, Weatherreporter requires exactly one preparation callback and reconciles its prompt/profile/backend/model and rendered/input hashes with the inspected identity and completed result. The callback output contract and completed validation must use the report's expected JSON Schema mode and path. The package contains only reviewed prompt-facing warning summaries, never source transport or provenance details. Safe active provenance remains in memory. Content-rich diagnostics are opt-in through `--llm-debug-dir`; see [operations](../operations.md) for retention and permissions.

When capture is enabled, its preparation artifact projects a provider endpoint
to its scheme and host and retains only reviewed execution settings. Provider
extras and URL user information, paths, queries, and fragments are omitted.
Capture storage remains confined to the operator-selected debug root; an unsafe
filesystem path causes the requested execution to fail. Host availability and
operator handling are documented in the
[operations guide](../operations.md#optional-prompt-debug-capture).

## Comparison Execution

For `compare`, Weatherreporter validates the report's generated-text catalog
binding, one exact prompt, and every explicitly selected profile before weather
collection. It prepares one deterministic YAML
data package, retains immutable copies of the report inputs, and executes every
profile against the same exact data-package bytes. Each profile remains an
independent Promptkit execution: one provider, provenance, or validation failure does not
stop its peers, while caller cancellation applies to every in-flight execution.

Weatherreporter starts selected profile executions concurrently and does not
add an application-level concurrency limit. Promptkit owns backend capacity and
any profile or backend concurrency policy. A shared Weatherreporter executor
must safely accept those concurrent `Execute` calls. The durable comparison output and
its compatibility rules are defined by the
[comparison bundle contract](comparison-bundle.md); the user-facing command
contract is in the [CLI reference](../cli.md).

The generated-text schemas require `summary`, `forecast_discussion`, and `precipitation_timing`, and reject additional properties. Promptkit results are accepted only when their raw JSON is at most 64 KiB; the adapter drops larger results before copying them into Weatherreporter's execution values or debug artifacts. The validator also limits total generated prose to 20,000 characters, with 4,000-character summary and timing fields, a 12,000-character Hourly discussion, and at most 12 day-style paragraphs of 4,000 characters each. Prompts return an empty string for `precipitation_timing` when the deterministic package contains no precipitation windows.

Prompt/profile configuration and the maintained local override example are owned by the [configuration reference](../config.md). Adapter construction and mapping are documented in the [Promptkit adapter internals](../internal/promptkit-adapter.md).