Tooling Architecture
The v1.0 tooling architecture follows the RFC: schemas are generated from Jinja + YAML sources, writes flow through MCP tools, and indexes are extended rather than replaced.
MCP Server Shape
MCP is the stable runtime contract for agents and harnesses. Files remain human-inspectable, but canonical mutations happen through tools.
The v1.0 server surface should include:
- a processkit gateway that exposes the common read/write surface
- per-domain management tools for work, decisions, records, gates, discussions, roles, bindings, migrations, and skills
- an index-management surface for reads, search, relation traversal, and interface queries
- a schema-management surface with
regenerate_schemas - a doctor/audit surface for validation, drift, and release readiness
The gateway can aggregate tools for harness convenience, but tool ownership should remain domain-specific so validation and lifecycle rules stay close to the schema they enforce.
MCP Helper Library
Every MCP server should use shared helpers rather than reimplementing process rules. The helper layer should provide:
- ID allocation and collision checks
- generated schema loading
- draft-2020-12 JSON Schema validation
- state-machine loading and transition validation
- strict/tolerant validation-mode lookup
- frontmatter and body parsing/serialization
- atomic file writes under the canonical storage layout
- event-log emission for mutating actions
- index upsert/delete calls after successful writes
- typed error responses with actionable remediation
- Python type-signature to JSON Schema consistency checks
- golden fixture helpers for MCP contract tests
The RFC requires MCP tools to have valid Python type signatures and matching draft-2020-12 JSON Schemas before release candidate.
Schema Generation
Schema sources live under schemas/src/. Runtime tools consume a flat,
committed _generated/*.yaml tree.
Composition rules:
extends: parent.yamldeclares composition inheritance{% include %}includes T fragments into templates__merge: replace|concat|name-mergedeclares per-field merge strategy- generated schemas declare interfaces such as
RecordorVersioned - generated files are committed so agents and reviewers can diff runtime contracts directly
The required MCP endpoint is:
regenerate_schemas(kinds: list | None) -> {rebuilt, unchanged, errors}
kinds=None performs a full rebuild. A non-empty list rebuilds only the
requested generated schemas and their dependencies. aibox apply may
trigger full rebuilds by default, but schema generation must not depend
on aibox; it must be runnable in processkit tests and CI directly.
Validation Modes
Validation is phase-gated:
- migrated kinds validate strictly
- kinds still being migrated validate tolerantly and emit warnings
- validation mode is queryable through MCP per kind
- release gates fail on invalid strict entities
This lets alpha/beta users test incomplete migrations without allowing known-invalid final entities to pass unnoticed.
Indexing Database
The RFC keeps the existing SQLite/FTS5 direction and extends it with interface-level grouping. The index is an accelerator and query surface, not the source of truth. Git-backed entity files remain canonical.
The minimum index stores:
- entity identity, kind, discriminator, state, title, timestamps, and path
- declared interfaces from generated schemas
- frontmatter fields needed for common filters
- typed relation edges from Bindings and inline references
- event subjects and actors for timeline queries
- full-text rows for titles, bodies, specs, and selected metadata
- validation and generation metadata for drift checks
The required read patterns are:
get_entity(id)- search by text
- query by kind, state, owner, and container
- traverse relation edges
- backlinks or cited-by navigation
query_by_interface(interface, filters...)
query_by_interface(Record, ...) is load-bearing because it lets agents
retrieve DecisionRecords, LogEntries, measurements, approvals, and other
record-like entities without guessing every concrete kind.
Index Update Flow
Writes should follow one transaction-like path:
- MCP tool receives a typed request.
- The tool loads the generated schema and current validation mode.
- The helper validates input and transition guards.
- The entity file is written atomically.
- Required LogEntries or Events are emitted.
- The changed entity, relations, and events are upserted into SQLite.
- The response returns the entity ID, path, state, validation mode, and index update status.
If index update fails after a file write, the response must surface drift
and pk-doctor must detect it. A separate reindex tool should rebuild
SQLite from files for recovery, CI fixtures, and release checks.