# Framework Format Reference

## Purpose

This document is the canonical contract for Promptkit prompt-definition,
profile, and schema files. The [Go package consumer guide](consumers/pkg-promptkit.md)
explains how to select these sources and invoke the engine. The
[OpenAI-compatible integration contract](integrations/openai-compatible-chat.md)
owns the resulting outbound wire behavior.

Prompt and profile sources recursively discover files ending in `.yaml` or
`.yml`. Each prompt-definition and profile file contains exactly one YAML
document; comments and trailing whitespace are allowed. YAML decoding is
strict: unknown fields are errors for the selected definition. Definitions are
selected by their YAML `id`, not their file name or directory.

## Prompt Definitions

A prompt definition describes inputs, Go-template messages, an optional
default profile, and an output contract.

```yaml
id: meeting.summary
version: "1.0.0"
default_profile: local-summary
description: Summarize a synthetic meeting note.
session_id: '{{.session}}'
inputs:
  - name: note
    required: true
    content_type: text/plain
    description: Meeting note to summarize.
messages:
  - role: system
    content: Return a concise summary.
    cache_control:
      type: ephemeral
      ttl: 1h
  - role: user
    content_file: ./summary.user.md
output:
  format: markdown
  validation_mode: basic
  repair_attempts: 0
```

| Field | Required | Meaning |
| --- | --- | --- |
| `id` | yes | Non-empty prompt identifier used by `RunRequest.PromptID`. |
| `version` | yes | Non-empty version selected by an optional `RunRequest.PromptVersion`. |
| `default_profile` | no | Non-empty profile ID used when the request omits `ProfileID`. |
| `description` | no | Human-readable description. |
| `session_id` | no | Go template rendered from request variables and input helpers. |
| `inputs` | no | Declared input metadata. |
| `messages` | yes | One or more chat-message templates. |
| `output` | yes | Output format and validation settings. |

When a request omits a version, the selected prompt ID must identify exactly
one definition. When it supplies a version, the ID and version pair must be
unique.

Exact prompt inspection uses this same configured source, strict decoding,
referenced content-file resolution, and ID/version selection. It reports the
selected definition's declared metadata without changing the prompt format or
executing the definition.

### Inputs

Each `inputs` item has these fields:

| Field | Required | Meaning |
| --- | --- | --- |
| `name` | yes | Non-empty name used by the request input map and `input` template helper. Names must be unique. |
| `required` | no | When true, preparation fails if the request omits the input. The default is false. |
| `content_type` | no | Expected media-type metadata. |
| `description` | no | Human-readable input description. |

Requests supply inputs as inline or file-backed `ArtifactRef` values. Declared
required inputs must be present. A template reference also requires the named
input to exist, whether or not it was declared. Extra request inputs are
allowed.

### Messages And Templates

Each message has a non-empty `role` and exactly one of:

- `content`, containing an inline Go template; or
- `content_file`, naming a file whose contents are the Go template.

`content_file` must be a relative path. It resolves from the directory that
contains the prompt file and must remain within the configured prompt source
root; parent components are allowed only when the resolved target remains
inside that root. Absolute paths and paths that escape the root are rejected.
Operating-system directory and single-file sources also reject symlink targets
outside the root, while injected `fs.FS` sources apply containment in that
filesystem's relative path namespace. For `WithPromptFile`, the source root is
the directory containing the selected prompt file. Promptkit uses the parsed
path text exactly after checking separately that it is not blank, so leading
and trailing whitespace can name real filesystem entries.

Request variables are the template data, so a variable named `audience` is
referenced as `{{.audience}}`. The `{{input "note"}}` helper renders the body
of a named input. Missing variables and input references are errors.

The optional `session_id` uses the same template data and input helper. Its
rendered value is trimmed, omitted when empty, and limited to 256 Unicode code
points. A nonblank direct request session ID bypasses this template completely;
a blank direct value leaves the template behavior unchanged.

### Cache Control

`cache_control` is optional and has these fields:

| Field | Required | Values |
| --- | --- | --- |
| `type` | yes | `ephemeral` |
| `ttl` | no | Empty or `1h` |

Promptkit preserves cache-control metadata on the rendered message. The
outbound integration determines its wire representation.

### Output Contract

| Field | Required | Values or behavior |
| --- | --- | --- |
| `format` | yes | `text`, `markdown`, or `json`. |
| `validation_mode` | yes | `none`, `basic`, `json`, or `json_schema`. |
| `schema_path` | for `json_schema` | Path to a schema in the configured schema source. |
| `repair_attempts` | no | Integer zero or greater; omitted means zero. |

The validation modes behave as follows:

- `none` skips content validation;
- `basic` requires non-empty generated content;
- `json` requires valid JSON; and
- `json_schema` requires valid JSON that satisfies the selected schema.

`format` controls output artifact metadata. JSON Schema mode also supplies the
schema to compatible model clients as structured-output metadata. The public
engine does not install an output repairer, so its validation is single-pass
even when a positive `repair_attempts` value is present.

A request-level `OutputContract` replaces the complete prompt output contract.
It does not merge individual fields. If its format is empty, Promptkit uses
`text`.

## Profile Definitions

A profile supplies model execution settings:

```yaml
id: local-summary
backend: openrouter
model: example-model
temperature: 0.2
max_tokens: 500
top_p: 0.95
timeout_seconds: 90
service_tier: flex
reasoning_effort: medium
extra_params:
  provider_option: enabled
```

| Field | Required | Meaning |
| --- | --- | --- |
| `id` | yes | Profile identifier, trimmed before selection and publication. It must be non-empty after trimming and unique within one source after normalization. |
| `backend` | unless `endpoint` is present | Backend registry ID. It is trimmed and registry membership is checked when the profile is prepared or inspected. |
| `endpoint` | unless `backend` is present | OpenAI-compatible base URL, including an API version path when required. A nonempty value is trimmed and must be absolute HTTP or HTTPS with a host and without user information, a query, or a fragment. When both connection fields are present, this overrides the backend endpoint without changing backend identity. |
| `model` | yes | Non-empty provider model name. |
| `temperature` | no | Number from 0 through 2. |
| `max_tokens` | no | Integer zero or greater. |
| `top_p` | no | Number from 0 through 1. |
| `timeout_seconds` | no | Per-generation deadline in whole seconds; integer zero or greater. |
| `service_tier` | no | Provider-specific request tier. |
| `reasoning_effort` | no | Provider-specific reasoning setting. |
| `api_key_env` | no | Name of an environment variable containing the API key. |
| `extra_params` | no | JSON-compatible provider-specific outbound fields. |

Raw `api_key` is prohibited in profile YAML. Store only an environment
variable name in `api_key_env`.

Promptkit does not infer a backend from a model or endpoint. Endpoint-only
profiles remain supported and have no effective backend ID.
The engine always provides the built-in `openrouter` ID. Consumers can add
engine-scoped IDs with
[`WithBackend`](../backends.go); exact registration validation belongs to its
GoDoc.

`extra_params` accepts null, booleans, finite numbers, strings, arrays, and
objects with string keys. Keys must be non-empty. With the built-in client,
they also cannot collide with the standard fields listed in the
[outbound request contract](integrations/openai-compatible-chat.md#request-body).
Excessively deep or large JSON-shaped values are rejected for safety.

### Defaults And Overrides

Execution settings resolve in this order:

1. the framework timeout baseline;
2. the selected backend, when the profile names one;
3. the selected profile; and
4. request `ExecutionTargetOverride` values.

The framework baseline is:

| Setting | Default |
| --- | --- |
| `temperature` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
| `max_tokens` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
| `top_p` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
| `timeout_seconds` | `600` |

Numeric zero in a file or in-memory profile does not select a numeric value.
For `temperature`, `max_tokens`, and `top_p`, it leaves the provider control
unspecified. For `timeout_seconds`, it retains the framework deadline. Numeric
request overrides use pointers, so an explicit zero is retained and sent to
compatible providers. In particular, an explicit request `timeout_seconds` of
zero disables the per-generation deadline while leaving the caller context and
transport timeout intact.

Non-empty profile strings replace backend defaults, and non-empty request
strings replace both. Request reasoning is the exception: a nil
`ReasoningEffort` pointer inherits the profile, a pointer to a nonblank string
trims and replaces it, and a pointer to a blank string clears it. Backend
identity is retained when either layer overrides the endpoint, so the override
also retains any engine-local capacity policy configured for that backend.
Capacity configuration belongs to the Go
[`Backend` API](../backends.go), not prompt or profile YAML. A non-empty
`extra_params` map at each layer replaces the entire lower-precedence map
rather than merging keys.
The [outbound integration contract](integrations/openai-compatible-chat.md)
defines how the effective settings are serialized.

### Source And Profile Precedence

An explicit request profile ID takes precedence over the prompt's
`default_profile`. If neither is present, preparation fails. Exact profile
inspection instead takes one explicit profile ID and does not use a prompt
default.

Profile sources resolve matching IDs in this order:

1. in-memory profiles supplied with `WithProfiles`;
2. the ordinary configured source selected by a profile file, `fs.FS`, or
   configured profile directory;
3. application fallback profiles supplied with `WithFallbackProfileFS`; and
4. embedded built-in profiles.

A profile source supplies a complete definition; definitions and their fields
are not merged across sources. A higher-precedence source falls back only when
the requested profile ID is absent. An invalid matching profile is an error and
does not fall back. In-memory `Profile` values follow the same ranges as YAML
profiles. They use `APIKeyRequired` for request-scoped credentials instead of
`api_key_env`. Preparation and exact profile inspection use this same source
precedence.

## Built-In Profile Catalog

Every built-in selects the `openrouter` backend. The engine's built-in backend
registry supplies `https://openrouter.ai/api/v1` and the environment-variable
name `OPENROUTER_API_KEY`, so individual profiles contain only model and
generation settings. Built-in profile files do not repeat those connection
values. A configured, application fallback, or in-memory profile with the same
profile ID takes precedence.

| Provider | ID | Model |
| --- | --- | --- |
| aion-labs | `aion-2` | `aion-labs/aion-2.0` |
| anthropic | `claude-fable-latest` | `~anthropic/claude-fable-latest` |
| anthropic | `claude-haiku-latest` | `~anthropic/claude-haiku-latest` |
| anthropic | `claude-opus-latest` | `~anthropic/claude-opus-latest` |
| anthropic | `claude-sonnet-latest` | `~anthropic/claude-sonnet-latest` |
| deepseek | `deepseek-3-2` | `deepseek/deepseek-v3.2` |
| deepseek | `deepseek-4-flash` | `deepseek/deepseek-v4-flash` |
| deepseek | `deepseek-4-pro` | `deepseek/deepseek-v4-pro` |
| google | `gemini-2-flash` | `google/gemini-2.5-flash` |
| google | `gemini-2-flash-lite` | `google/gemini-2.5-flash-lite` |
| google | `gemini-2-pro` | `google/gemini-2.5-pro` |
| google | `gemini-3-flash-lite` | `google/gemini-3.1-flash-lite` |
| google | `gemini-flash-latest` | `~google/gemini-flash-latest` |
| google | `gemini-pro-latest` | `~google/gemini-pro-latest` |
| google | `gemma-4-31b` | `google/gemma-4-31b-it:exacto` |
| minimax | `minimax-m2` | `minimax/minimax-m2.5` |
| minimax | `minimax-m3` | `minimax/minimax-m3` |
| mistral | `mistral-large-2512` | `mistralai/mistral-large-2512` |
| mistral | `mistral-medium-3-5` | `mistralai/mistral-medium-3-5` |
| mistral | `mistral-small-3` | `mistralai/mistral-small-3.2-24b-instruct` |
| mistral | `mistral-small-4` | `mistralai/mistral-small-2603` |
| nvidia | `nemotron-3-ultra` | `nvidia/nemotron-3-ultra-550b-a55b` |
| openai | `gpt-5-mini` | `openai/gpt-5.4-mini` |
| openai | `gpt-5-nano` | `openai/gpt-5.4-nano` |

## Schemas

Schemas are JSON documents selected by a prompt or request
`schema_path`. For a directory or `fs.FS` source, paths resolve within the
configured source root. Referenced nested schemas resolve relative to the
owning schema document. `WithSchemaFile` exposes one schema, addressed by its
base name.

An unreadable, invalid, or unresolvable schema produces an operational
validation error. Generated content that is valid JSON but does not satisfy the
schema produces a failed validation result.

## Credentials

Credential values belong at the request or environment boundary, never in
prompt, profile, schema, or example files:

- a file profile names an environment variable with `api_key_env`;
- an in-memory profile may set `APIKeyRequired`;
- a request can provide a direct `APIKey` or override `APIKeyEnv`; and
- a direct request key takes precedence over environment lookup.

After a direct request key, the credential-source precedence is request
`APIKeyEnv`, profile `api_key_env`, then the backend default. An in-memory
profile with `APIKeyRequired` clears an inherited backend environment name and
requires a direct key unless the request explicitly supplies `APIKeyEnv`.
Promptkit validates required credential availability during preparation.
Direct keys are excluded from JSON results and redacted by public string
formatters. Environment-variable names may appear in prepared metadata, but
their values do not.