Files
weatherreporter/docs/internal/app-orchestration.md

229 lines
10 KiB
Markdown

# App Orchestration Internals
This document describes the workflow coordinator in `internal/app`.
## Purpose
`internal/app` coordinates the top-level use cases after CLI parsing and config
loading are complete. It resolves report definitions, collects weather data
through `internal/collect`, builds collected and derived facts, builds module
snapshots and prompt-input artifacts, invokes Scriptorium through the adapter
boundary, optionally notifies distributor through an app-owned notifier
boundary, persists managed state, runs batches, and reads existing artifacts
for inspection.
## Inputs And Outputs
Inputs:
- `GenerateRequest` for one report command
- `BatchRequest` for morning or evening batch commands
- `FetchBundleRequest` for explicit bundle collection and save workflows
- `ReportRequest` for single-report generation
- resolved report definitions from `internal/report`
- collection results from `internal/collect`
- prior snapshots loaded from `internal/state`
- optional collector, renderer, notifier, and state-store fakes for tests
Outputs:
- generated report results with JSON module snapshot, YAML data package,
preflight, report, metadata, prior snapshot, Recent Changes, Scriptorium
result details, generated-text artifact paths when applicable, and
notification result when attempted
- batch summaries with per-report status, artifact paths, error text, and
one top-level batch notification result when attempted or skipped
- saved Weather API bundle JSON for explicit bundle collection workflows
- inspection JSON values for reports, metadata, module snapshots, data
packages, prior snapshots, and source provenance
## Boundaries
`internal/app` owns workflow order and request composition. It does not parse
CLI flags, load YAML files directly, implement HTTP transport, own fact
derivation algorithms, define report periods, compare rendered Markdown, or
construct Scriptorium argv.
Report selection and report identity policy come from `internal/report`.
Collected and derived fact contracts come from `internal/facts`.
Weather API transport stays in `internal/adapters/weatherapi`, and app-facing
upstream collection stays in `internal/collect`. Scriptorium subprocess
behavior stays in `internal/adapters/scriptorium`. Distributor upload behavior
stays in `internal/adapters/distributor`. Filesystem layout and persisted
metadata stay in `internal/state`.
## Data Flow Terms
- `collect.Result` is the app-facing upstream collection result. It carries the
normalized `weatherdata.Bundle` used by report generation.
- `CollectedFacts` are normalized source facts derived from a collected Weather
API bundle and made available to derivation and module builders.
- `DerivedFacts` are deterministic calculations over collected facts, the
resolved valid period, daypart configuration, and report-specific windows.
- `module.Output` values are ordered deterministic stanzas built from collected
and derived facts for prompt input and inspection.
- `GeneratedText` is structured prose returned by Scriptorium for
generated-text-template reports and validated by `internal/generatedtext`.
- `RenderContext` is the typed template input built from report metadata,
module outputs, and validated generated text before Markdown rendering.
## Config Fields Used
- `weather_api.*` for Weather API client construction and module metadata
- `scriptorium.*` for renderer construction
- `workspace.*` for filesystem state
- `dayparts` for daily and outlook summarization
- `recent_change.*` for structured Recent Changes thresholds
- `notify.distributor.*` for optional single-report and batch notification
after report generation
Output copy flags are command request fields. They are not configuration
defaults.
## Generation Workflow
Single-report commands validate the report command, collect once through
`internal/collect`, resolve the requested report, and pass the resolved report
plus explicit collection into `GenerateReport`.
`GenerateReport` then uses this setup:
1. Create or use a filesystem store.
2. Locate any prior compatible snapshot through `internal/state`.
3. Build collected and derived facts from the supplied collection.
4. Execute configured modules and save the module snapshot.
5. Compute Recent Changes from structured prior and current module snapshots.
6. Build and save the YAML Scriptorium `data_package`.
7. Run Scriptorium render preflight.
8. Save preflight JSON when a render result is available.
9. Save metadata for inspection.
For `scriptorium_markdown` reports, generation then:
10. Runs Scriptorium report generation to the managed report path.
For `generated_text_template` reports, generation then:
10. Looks up the generated-text catalog entry for the report schema/template
IDs.
11. Runs structured Scriptorium generation to the raw generated-text JSON path.
12. Saves the structured Scriptorium run result.
13. Validates and saves normalized generated text.
14. Builds and saves a typed render context.
15. Renders Markdown from the embedded template to the managed report path.
After either mode has produced a managed Markdown report, shared finalization:
1. Copies the managed report to the requested `--out` or `--out-dir` path when
provided.
2. Saves final metadata with the managed report path and any generated-text
artifact paths already produced.
3. If distributor notification is enabled, notifies using the managed report
path as the source file.
4. Saves a distributor notification debug artifact and updates metadata with
its path.
If render preflight returns both a result and an error, preflight JSON and
metadata are persisted before the error is returned. If Scriptorium report
generation returns an error after writing output, the managed report and
metadata remain inspectable. Notification is not attempted after collection,
module snapshot, prompt input, render, Scriptorium run, or metadata-save
failures.
Generated-text report failures are returned with report ID, RunID, and the
failed operation. When available, the app preserves the latest generated-text
artifacts already reached by the workflow: preflight output, structured run
result, raw generated text, validated generated text, and render context.
When notification is attempted, the debug artifact records request identity,
including rendered pipeline ID, bundle paths, accepted upload fields,
distributor status fields, raw status report JSON when available, and redacted
failure context.
`--out` copies are never used as notification source files.
## Batch Workflow
`run morning` collects once, plans Today Report, Tomorrow Report, and eligible
future Daily Reports from the collected hourly forecast, then passes the same
collection into each report generation. `run evening` uses the same collection
and planning rules, but starts with Tomorrow Report. Future Daily reports start
with the day after tomorrow and require complete hourly forecast coverage for
the target local civil day. Dynamic Daily `--out-dir` copies use
`daily-YYYY-MM-DD.md`; other batch copies use report definition output names.
A collection failure stops the batch before planning or report generation.
After planning succeeds, batch generation continues independent reports after a
failure, records each result, writes compact status lines to stderr, emits a
JSON summary to stdout, and returns an aggregate error when any report failed.
Batch report generation suppresses per-report distributor notification. After
all planned reports finish, app orchestration evaluates batch notification:
1. If distributor notification is disabled, the batch notification result is
omitted.
2. If batch notification is disabled, the batch notification result is omitted
and there is no per-report fallback upload.
3. If any planned report failed, the batch notification result is `skipped`
with reason `one or more reports failed`, and distributor is not called.
4. If every report succeeded, app orchestration renders batch pipeline, bundle
ID, and idempotency key templates, renders `report_path_templates` for each
included report, validates every managed source path and bundle path, checks
duplicate bundle paths across the batch, calls the notifier once with a
multi-file request, and saves a batch notification debug artifact.
Batch notification failure records a top-level failed notification, increments
the aggregate batch failure count, and returns an aggregate batch error without
marking individual report items failed. `--out-dir` copies are never used as
notification source files.
## Inspection Workflow
Inspection workflows load existing filesystem state only. They do not fetch
weather data or invoke Scriptorium. Run-specific inspect commands share the same
store and metadata lookup path, then load the requested artifact or derived
inspection view.
## Failure Behavior
- Resolve errors stop the requested workflow before collection.
- Collection and module execution errors stop that report before Scriptorium
runs.
- Prompt input validation fails before render preflight.
- Render and run errors preserve Scriptorium stderr and exit-code context.
- Generated-text report errors preserve available intermediate artifacts and do
not create extra output copies.
- Single-report notification errors are wrapped with report ID, RunID, and
managed report path context.
- Batch notification errors are recorded on the top-level batch notification
result and do not change individual report item status.
- Metadata and artifact path errors include filesystem context.
- Batch failures are recorded per report and surfaced through an aggregate
batch error.
## Tests
Inspect:
- `internal/app/app_test.go`
- `internal/app/batch_plan_test.go`
- `internal/collect/collect_test.go`
- `internal/cli/root_test.go`
- `internal/state/filesystem_test.go`
## Invariants
- Report behavior is resolved through `internal/report`.
- Generate and run commands collect once before report generation.
- Batch planning is app-owned because future Daily membership depends on
collected hourly forecast coverage.
- Generated reports use the same app request and result types regardless of
report ID.
- Render preflight precedes Scriptorium report generation.
- Generated-text reports render Markdown from a curated render context, not from
a raw data package.
- Recent Changes are computed from structured module snapshots.
- Metadata links artifacts produced for a run.
- Single-report distributor notification maps the managed Markdown report path
to configured bundle paths.
- Batch distributor notification maps each included managed Markdown report
path to bundle paths rendered for that report and uploads once for the
batch.
- Extra output copies are not upload sources.