Align Prompt Library¶

align_prompt_library — mvp.align_prompt_library

Cluster: Safety & Alignment | Type: component | MCP Tools: 27

Overview¶

Versioned prompt store with template rendering, full-text search, tag and category filtering, and variable validation. The direct block keeps prompts in memory unless db_path is provided; the MCP block uses SQLite persistence by default. Each store call auto-increments the version for a named prompt and populates id and timestamps; render substitutes {{variable}} placeholders from supplied values and fails when required values are missing; validate extracts placeholder names without storing. The extended AlignmentBlock sub-component adds alignment specification storage alongside prompt management and now persists specs/goal trees to SQLite when db_path is supplied. The 27-operation MCP sub-package wraps deterministic CRUD (prompts, alignment specs, goal trees, version history/rollback), adds pl_capabilities, and retains list_patterns readout of the applied deterministic reliability patterns. Every operation is fixed-rule code with no LLM in the loop, and a safety/regulated-tagged prompt is refused storage unless a human review artifact (reviewer_id, passed adversarial test, failure modes, usage examples) is supplied. That refusal is surfaced as a structured blocked-escalated envelope with warning_card and evidence rather than cosmetic success.

MVP production caveat

align_prompt_library is suitable for pilot and launch-plan use as a practical MCP-accessible prompt library, but it is not yet a fully polished production service. Before relying on it for paid or regulated customer workflows, run a smoke test through the actual MCP install/server path, verify the SQLite database location and backup policy, and confirm any existing prompt-library databases have been migrated or reseeded. The component now has persistence, capability discovery, version history, rollback, validation, tag filtering, and clearer error behavior, but the surrounding install/docs/migration experience still needs final launch hardening.

When to use:

Centralising all prompt templates used by a pipeline, with version history
Rendering parameterised prompts at runtime without string formatting scattered across the codebase
Searching and filtering prompts by tag or category before injection into an agent

Example:

from mvp.align_prompt_library import AlignPromptLibraryBlock, PromptLibraryInput

block = AlignPromptLibraryBlock(name="prompts")
block.infer(PromptLibraryInput(op="store", name="greeter", content="Hello, {{name}}!"))
result = block.infer(PromptLibraryInput(op="render", name="greeter", render_values={"name": "Alice"}))
# result.value.rendered → "Hello, Alice!"

Works well with: align_artifacts, align_specs, llm_router

Public API¶

`AlignmentBlock(AIBlock[AlignmentSpecInput, AlignmentSpecOutput, dict])`¶

In-memory alignment specification store.

Field	Type	Default
`name`	`str`	'alignment_block'
`resource_bounds`	`ResourceBounds \\| None`	None
`usage`	`ResourceUsage`	field(default_factory=ResourceUsage)
`db_path`	`str \\| None`	None

Methods:

`infer(data: AlignmentSpecInput) -> Result[AlignmentSpecOutput]`¶

`AlignmentSpecType(str, Enum)`¶

`RICEPrinciple(str, Enum)`¶

`AlignmentFailureMode(BaseModel)`¶

A documented failure mode specific to an alignment specification.

Field	Type	Default
`name`	`str`	required
`description`	`str`	required
`alignment_failure_type`	`Literal['outer', 'inner']`	'outer'
`severity`	`Literal['low', 'medium', 'high', 'critical']`	'medium'
`mitigation`	`str`	''
`detection_prompt`	`str`	''
`example`	`str`	''

`AlignmentTestCase(BaseModel)`¶

A concrete test case for an alignment specification.

Field	Type	Default
`id`	`str`	''
`input_description`	`str`	required
`expected_behavior`	`str`	required
`prohibited_behavior`	`str`	''
`enforcement_check`	`str`	''
`passes`	`bool \\| None`	None

`EnforcementMechanism(BaseModel)`¶

A mechanism that enforces an alignment specification.

Field	Type	Default
`mechanism_type`	`Literal['prompt', 'code', 'rlhf', 'rule', 'constitution']`	'prompt'
`content`	`str`	required
`description`	`str`	''
`applies_at`	`Literal['training', 'inference', 'both']`	'inference'

`AlignmentSpec(BaseModel)`¶

A formal alignment specification for a goal.

Field	Type	Default
`id`	`str`	''
`name`	`str`	required
`spec_type`	`AlignmentSpecType`	AlignmentSpecType.SAFETY_CONSTRAINT
`version`	`int`	1
`description`	`str`	''
`natural_language_spec`	`str`	''
`formal_spec`	`str`	''
`outer_alignment_goal`	`str`	''
`inner_alignment_notes`	`str`	''
`rice_principles`	`list[RICEPrinciple]`	Field(default_factory=list)
`enforcement_mechanisms`	`list[EnforcementMechanism]`	Field(default_factory=list)
`test_cases`	`list[AlignmentTestCase]`	Field(default_factory=list)
`failure_modes`	`list[AlignmentFailureMode]`	Field(default_factory=list)
`tags`	`list[str]`	Field(default_factory=list)
`category`	`str`	'general'
`domain`	`str`	''
`recommended_backends`	`list[str]`	Field(default_factory=list)
`goal_key`	`str`	''
`created_at`	`str`	''
`updated_at`	`str`	''
`author`	`str \\| None`	None
`metadata`	`dict[str, str]`	Field(default_factory=dict)

`AlignedGoalNode(BaseModel)`¶

A single node in an AlignedGoalTree.

Field	Type	Default
`goal`	`str`	required
`spec_name`	`str`	''
`depth`	`int`	0
`parent_goal`	`str \\| None`	None
`child_goals`	`list[str]`	Field(default_factory=list)

`AlignedGoalTree(BaseModel)`¶

A goal tree where each node maps a goal (str key) to an AlignmentSpec (value).

Field	Type	Default
`id`	`str`	''
`name`	`str`	required
`root_goal`	`str`	required
`nodes`	`dict[str, AlignedGoalNode]`	Field(default_factory=dict)
`specs`	`dict[str, AlignmentSpec]`	Field(default_factory=dict)
`metadata`	`dict[str, str]`	Field(default_factory=dict)
`created_at`	`str`	''
`updated_at`	`str`	''

`AlignmentSpecInput(BaseModel)`¶

Input to AlignmentBlock.

Field	Type	Default
`op`	`Literal['store', 'retrieve', 'list', 'delete', 'search', 'validate', 'create_tree', 'get_tree', 'capabilities']`	required
`name`	`str`	''
`spec_type`	`AlignmentSpecType`	AlignmentSpecType.SAFETY_CONSTRAINT
`description`	`str`	''
`natural_language_spec`	`str`	''
`formal_spec`	`str`	''
`outer_alignment_goal`	`str`	''
`inner_alignment_notes`	`str`	''
`rice_principles`	`list[RICEPrinciple]`	Field(default_factory=list)
`enforcement_mechanisms`	`list[dict]`	Field(default_factory=list)
`test_cases`	`list[dict]`	Field(default_factory=list)
`failure_modes`	`list[dict]`	Field(default_factory=list)
`tags`	`list[str]`	Field(default_factory=list)
`category`	`str`	'general'
`domain`	`str`	''
`recommended_backends`	`list[str]`	Field(default_factory=list)
`goal_key`	`str`	''
`author`	`str \\| None`	None
`metadata`	`dict[str, str]`	Field(default_factory=dict)
`query`	`str`	''
`tree_name`	`str`	''
`root_goal`	`str`	''
`goal_spec_pairs`	`list[dict]`	Field(default_factory=list)

`AlignmentSpecOutput(BaseModel)`¶

Output from AlignmentBlock.

Field	Type	Default
`op`	`str`	required
`specs`	`list[AlignmentSpec]`	Field(default_factory=list)
`retrieved`	`AlignmentSpec \\| None`	None
`tree`	`AlignedGoalTree \\| None`	None
`message`	`str`	''
`degraded`	`bool`	False
`degradation_reason`	`str`	''
`completion_state`	`Literal['verified', 'qualified-draft', 'blocked-escalated', 'failed']`	'qualified-draft'
`warning_card`	`str \\| None`	None
`evidence`	`dict[str, Any]`	Field(default_factory=dict)
`request_id`	`str \\| None`	None
`run_id`	`str \\| None`	None

`MissingVariableError(ValueError)`¶

Raised when a required template variable is absent in strict rendering mode.

Constructor:

Parameter	Type	Default
`missing`	`list[str]`	required
`template`	`str`	required