cleveragents-core/docs/reference/database_schema.md

Database Schema Reference

Overview

CleverAgents uses SQLAlchemy ORM models with SQLite as the default backend. All tables are created via Base.metadata.create_all through the init_database() helper in src/cleveragents/infrastructure/database/models.py.

Resource Registry Tables

The resource registry consists of three core tables introduced in Stage B1:

Table	Model	Description
resource_types	ResourceTypeModel	Schema-level resource type defs
resources	ResourceModel	Registered resource instances
resource_edges	ResourceEdgeModel	Parent-child DAG edges

resource_types

Primary key: name (namespaced, e.g. builtin/git-checkout). Stores kind (physical/virtual), handler references, JSON argument schemas, allowed parent/child types, auto-discovery config, and capabilities.

resources

Primary key: resource_id (26-char ULID). FK to resource_types.name. Stores optional namespaced name, location, JSON properties/metadata, and content hash for equivalence tracking.

resource_edges

Composite primary key: (parent_id, child_id). Both columns FK to resources.resource_id with CASCADE delete. Stores link_type (contains, references, derived_from) and a self-loop check constraint.

Robot Migration Smoke Suite

A Robot Framework smoke suite validates that schema creation produces the expected resource registry tables with correct columns.

Running the suite

# Via nox (recommended — runs all Robot integration tests):
nox -s integration_tests

# Directly with robot:
robot --outputdir build/reports/robot robot/resource_registry_migration.robot

Test cases

Test Case	What it checks
Schema Creation Produces Resource Registry Tables	resource_types, resources, resource_edges exist
Resource Registry Tables Have Expected Columns	Core columns present on each table
Migration Is Idempotent	Calling init_database twice is safe

The helper script (robot/helper_resource_registry_migration.py) uses init_database() with a temporary SQLite file and validates via sqlalchemy.inspect.

Behave BDD Tests

The Behave feature file features/resource_registry_tables.feature contains comprehensive scenarios covering:

• Table existence after schema creation
• CRUD operations for ResourceTypeModel, ResourceModel, ResourceEdgeModel
• Constraint enforcement (uniqueness, FK, check constraints)
• ORM relationship navigation
• Migration smoke verification

Run Behave tests via:

nox -s unit_tests

ASV Benchmarks

Performance benchmarks for resource registry operations live in benchmarks/resource_registry_migration_bench.py and measure:

• Schema creation time
• Insert throughput for types, resources, and edges
• DAG query performance

Run benchmarks via:

nox -s benchmark

Tool and Validation Registry Tables

Tables

tools

Stores registered tool and validation definitions.

Column	Type	Constraints	Description
tool_id	TEXT	PRIMARY KEY	ULID identifier
name	TEXT	NOT NULL, UNIQUE	Namespaced name (namespace/short_name)
description	TEXT		Human-readable description
tool_type	TEXT	NOT NULL, CHECK (tool/validation)	Discriminator for registry queries
source_type	TEXT	NOT NULL, CHECK (see below)	Implementation source
input_schema	TEXT		JSON Schema for tool inputs
output_schema	TEXT		JSON Schema for tool outputs
read_only	BOOLEAN	NOT NULL, DEFAULT FALSE	Tool only reads, never writes
writes	BOOLEAN	NOT NULL, DEFAULT FALSE	Tool can write to resources
checkpointable	BOOLEAN	NOT NULL, DEFAULT FALSE	Tool supports checkpoint/rollback
side_effects	BOOLEAN	NOT NULL, DEFAULT FALSE	Tool has known side effects
config_yaml	TEXT		Raw YAML config for reconstruction
created_at	TEXT	NOT NULL	ISO-8601 timestamp
updated_at	TEXT	NOT NULL	ISO-8601 timestamp

Check constraints:

• ck_tools_tool_type: tool_type IN ('tool', 'validation')
• ck_tools_source_type: source_type IN ('mcp', 'agent_skill', 'builtin', 'custom', 'wrapped')

Indexes:

• ix_tools_name on (name)
• ix_tools_tool_type on (tool_type)

---

tool_bindings

Stores resource slot bindings for tools.

Column	Type	Constraints	Description
binding_id	TEXT	PRIMARY KEY	ULID identifier
tool_id	TEXT	NOT NULL, FK → tools.tool_id CASCADE	Parent tool reference
slot_name	TEXT	NOT NULL	Named resource slot
resource_type	TEXT	NOT NULL	Expected resource type
binding_mode	TEXT	NOT NULL, CHECK (see below)	How the slot is resolved
access_level	TEXT	NOT NULL, DEFAULT read_only	Access level on bound resource
created_at	TEXT	NOT NULL	ISO-8601 timestamp

Check constraints:

• ck_tool_bindings_mode: binding_mode IN ('context', 'static', 'parameter')
• ck_tool_bindings_access: access_level IN ('read_only', 'read_write')

Unique constraints:

• uq_tool_bindings_tool_slot: UNIQUE(tool_id, slot_name)

---

validation_attachments

Links validations to resources with mode semantics.

Column	Type	Constraints	Description
attachment_id	TEXT	PRIMARY KEY	ULID identifier
resource_id	TEXT	NOT NULL, FK → resources.resource_id CASCADE	Target resource
validation_name	TEXT	NOT NULL	References tools.name (validation)
mode	TEXT	NOT NULL, DEFAULT required	required or informational
created_at	TEXT	NOT NULL	ISO-8601 timestamp

Check constraints:

• ck_validation_attachments_mode: mode IN ('required', 'informational')

Unique constraints:

• uq_validation_attachments_resource_validation: UNIQUE(resource_id, validation_name)

Indexes:

• ix_validation_attachments_resource_id on (resource_id)

---

Relationships

tools 1 ──< tool_bindings       (tool_id FK, CASCADE delete)
resources 1 ──< validation_attachments  (resource_id FK, CASCADE delete)
validation_attachments.validation_name → tools.name (logical, not enforced by FK)

---

actors

Stores actor configurations with YAML text retention, schema version tracking, and optional compiled metadata.

Column	Type	Constraints	Description
id	INTEGER	PRIMARY KEY, AUTOINCREMENT	Auto-generated row ID
name	TEXT(255)	NOT NULL, UNIQUE	Namespaced name (namespace/identifier)
provider	TEXT(255)	NOT NULL	Provider identifier (e.g. openai)
model	TEXT(255)	NOT NULL	Model identifier (e.g. gpt-4)
config_blob	JSON	NOT NULL, DEFAULT {}	Canonical actor configuration blob
config_hash	TEXT(128)	NOT NULL	SHA-256 hash of config_blob
graph_descriptor	JSON		Adapter-produced graph descriptor
yaml_text	TEXT		Original YAML source text for the actor config
schema_version	TEXT(20)	NOT NULL, DEFAULT '1.0'	Actor config schema version
compiled_metadata	JSON		Compiler-produced metadata (graph topology etc)
unsafe	BOOLEAN	NOT NULL, DEFAULT FALSE	True when actor is marked unsafe
is_built_in	BOOLEAN	NOT NULL, DEFAULT FALSE	Generated from provider registry
is_default	BOOLEAN	NOT NULL, DEFAULT FALSE	Designated default actor
created_at	DATETIME	NOT NULL	Row creation timestamp
updated_at	DATETIME	NOT NULL	Last update timestamp

Notes:

• yaml_text preserves the original YAML source so that actors can be reconstructed or exported without loss.
• schema_version tracks which version of the actor YAML schema was used to create the configuration. Defaults to 1.0.
• compiled_metadata stores JSON output from the actor compiler (graph topology, resolved tools, etc.).
• name follows the namespace/identifier convention enforced by the domain model validator (e.g. local/my-actor, openai/gpt-4).

---

Decision Tree Tables

The decision tree subsystem consists of two tables introduced in Stage M3.2:

Table	Model	Description
decisions	DecisionModel	Decision tree nodes
decision_dependencies	DecisionDependencyModel	Decision influence DAG edges

decisions

Primary key: decision_id (26-char ULID). FK to v3_plans.plan_id (CASCADE delete). Self-referential FKs for parent_decision_id, corrects_decision_id, and superseded_by (SET NULL).

Column	Type	Constraints	Description
decision_id	String(26)	PRIMARY KEY	ULID identifier
plan_id	String(26)	NOT NULL, FK -> v3_plans CASCADE	Parent plan reference
parent_decision_id	String(26)	FK -> decisions SET NULL	Parent decision in tree
sequence_number	Integer	NOT NULL	Monotonic order within plan
decision_type	String(30)	NOT NULL, CHECK (11 enum values)	Classification of the decision
question	Text	NOT NULL	What question was being answered
chosen_option	Text	NOT NULL	The option that was chosen
alternatives_considered_json	Text		JSON array of alternative options
confidence_score	Float	CHECK (0.0-1.0 or NULL)	Confidence in the decision
context_snapshot_json	Text	NOT NULL	JSON blob of ContextSnapshot
rationale	Text		Human-readable rationale
actor_reasoning	Text		Raw LLM reasoning trace
downstream_decision_ids_json	Text		JSON array of dependent decision ULIDs
downstream_plan_ids_json	Text		JSON array of spawned plan ULIDs
artifacts_produced_json	Text		JSON array of ArtifactRef objects
created_at	String(30)	NOT NULL	ISO-8601 timestamp
is_correction	Boolean	NOT NULL, DEFAULT FALSE	Whether this is a correction
corrects_decision_id	String(26)	FK -> decisions SET NULL	Decision being corrected
correction_reason	Text		Why the correction was made
superseded_by	String(26)	FK -> decisions SET NULL	Decision that supersedes this one

Check constraints:

• ck_decisions_type: decision_type IN ('prompt_definition', 'invariant_enforced', 'strategy_choice', 'implementation_choice', 'resource_selection', 'subplan_spawn', 'subplan_parallel_spawn', 'tool_invocation', 'error_recovery', 'validation_response', 'user_intervention')
• ck_decisions_confidence: confidence_score IS NULL OR (confidence_score >= 0.0 AND confidence_score <= 1.0)

Indexes:

• ix_decisions_plan on (plan_id)
• ix_decisions_parent on (parent_decision_id)
• ix_decisions_type on (decision_type)
• ix_decisions_created on (created_at)
• ix_decisions_plan_seq on (plan_id, sequence_number)

decision_dependencies

Composite primary key: (source_decision_id, target_decision_id). Both columns FK to decisions.decision_id with CASCADE delete. Stores relationship_type (default influences) and a self-loop check constraint.

Column	Type	Constraints	Description
source_decision_id	String(26)	PK, FK -> decisions CASCADE	Decision that influences
target_decision_id	String(26)	PK, FK -> decisions CASCADE	Decision that is influenced
relationship_type	String(30)	NOT NULL, DEFAULT 'influences'	Type of influence relationship
created_at	String(30)	NOT NULL	ISO-8601 timestamp

Check constraints:

• ck_decision_deps_no_self_loop: source_decision_id != target_decision_id

Indexes:

• ix_decision_deps_source on (source_decision_id)
• ix_decision_deps_target on (target_decision_id)

---

Migration

• Revision: c1_001_tool_registry
• Depends on: b0_001_projects
• File: alembic/versions/c1_001_tool_registry.py