Complete Program¶

This compact application combines typed extraction, deterministic Python, procedure tools, profiles, command-line input, tests, and evals without hiding the boundaries between them.

incident_report.kedi¶

```
INCIDENTS = {
    42: "Checkout requests time out. Payments owns mitigation.",
    77: "Search results are stale. Search owns reindexing.",
}
```

~Incident(
  id: int,
  title: str,
  owner: str,
  severity: Literal["low", "medium", "high", "critical"]
)

@lookup_incident(incident_id: int) -> str:
  ###
  Return the source report for one incident identifier.
  ###
  = `INCIDENTS[incident_id]`

@format_incident(incident: Incident) -> str:
  = `f"#{incident.id} [{incident.severity}] {incident.title} - {incident.owner}"`

> profile: analyst:
    > adapter: pydantic
    > model: openai:gpt-5.6-luna
    > system: Look up the requested identifier. Infer severity from the report, not invented facts.
    > use: lookup_incident
    > approval: allow

@extract_incident(incident_id: int) -> Incident:
  > use: analyst
  >> According to lookup_incident, incident <incident_id> is [incident: Incident].
  = `incident`

@test: format_incident:
  > case: formats_all_fields:
    ```
    incident = Incident(id=1, title="Latency", owner="Platform", severity="high")
    assert format_incident(incident) == "#1 [high] Latency - Platform"
    ```

[incident_id: int] = `int(args.incident_id)`
[incident: Incident] = `extract_incident(incident_id)`

= `format_incident(incident)`

@eval: extract_incident:
  > data: reports:
    = ```
    return [
      (42, {"owner": "Payments"}),
      (77, {"owner": "Search"}),
    ]
    ```
  > metric: owner_accuracy(reports):
    = `extract_incident(reports).owner == expected["owner"]`

Run incident 42:

kedi incident_report.kedi --incident-id 42

Unknown CLI options become fields on the reserved args binding; hyphens are normalized to underscores. The first repeated option wins.

Why Each Boundary Is Explicit¶

• lookup_incident is deterministic and should not be a model call.
• extract_incident uses a typed capture because understanding prose requires a model.
• = `incident` returns the native custom type.
• format_incident renders only at the final presentation boundary.
• The profile captures a stable adapter, model, instructions, tool, and policy.
• The test checks deterministic formatting without spending model tokens.
• The eval scores semantic extraction with explicit expected data.

The top-level initialization passes native values through Python expressions. Writing <incident> inside a prompt would intentionally serialize it for the model instead.

The analyst is used by extract_incident, on both the CLI and eval paths. Its allow policy is limited to this trusted, in-memory procedure tool; procedure tools otherwise default to mutating risk. It is not a recommended policy for arbitrary filesystem tools. Missing identifiers raise a lookup error, and severity remains model judgment rather than a verified incident fact.

Configure the selected provider's credentials before normal execution or eval. The documentation tests replace only the model with a local FunctionModel, then assert that lookup really runs before a typed result reaches the formatter. That verifies orchestration, not extraction accuracy with Gemini.

Verification Commands¶

kedi incident_report.kedi --parse
kedi incident_report.kedi --test
kedi incident_report.kedi --eval

--test and --eval are validation modes; they do not continue into ordinary top-level execution. Keep deterministic tests separate from model-backed evals so failures identify the correct layer.