Prompt resolver, renderer, and typed composition boundary.
A base ask prompt is represented as the task section of a
Spectre.Prompt.Plan. Agent, Skill, Flow, and handler inject declarations
resolve into typed fragments before model capability negotiation. Dynamic
providers remain context data, while static instruction/task assets retain
instruction trust. Policy request tasks cannot be replaced by application
injections at any scope.
These rules prevent privilege promotion and policy bypass inside Spectre. Structured adapters preserve them at the adapter boundary; legacy adapters receive an escaped data marker. Neither form claims that the downstream model is immune to semantic jailbreaks.
Summary
Functions
Builds the fully resolved prompt plan used by a model call.
Resolves, composes, and renders a prompt for a model-backed turn.
Resolves a prompt name or relative path under the active definition's prompt root.
Functions
@spec build(module(), atom() | String.t(), Spectre.Context.t() | map(), keyword()) :: {:ok, Spectre.Prompt.Plan.t()} | {:error, term()}
Builds the fully resolved prompt plan used by a model call.
@spec render(module(), atom() | String.t(), Spectre.Context.t() | map(), keyword()) :: {:ok, String.t()} | {:error, term()}
Resolves, composes, and renders a prompt for a model-backed turn.
{:ok, text} = Spectre.Prompt.render(MyAgent, :welcome, ctx)@spec resolve(module(), atom() | String.t(), Spectre.Context.t() | map(), keyword()) :: {:ok, String.t()} | {:error, term()}
Resolves a prompt name or relative path under the active definition's prompt root.
Policy prompts are looked up under policies/<policy>/<prompt>.text.heex.
Resolved paths must remain inside the configured root.