Skip to content

RELEASE-v0.4.0.md

Latest commit

 

History

History
96 lines (77 loc) · 3.83 KB

RELEASE-v0.4.0.md

File metadata and controls

96 lines (77 loc) · 3.83 KB
Raw

v0.4.0 — Agent Substrate

The smallest possible surface for treating Quilt as a substrate for agents, planners, and runtimes. Five primitives, one CLI command, one manifest schema.

What ships

schemas/manifest.schema.json (new)

  • Backward-compatible extension of the Quilt sheet format
  • Optional fields: inputs, outputs, preconditions, postconditions, rollback, resource_hints, model_requirements, provenance_tags, irreversible
  • All fields are optional — existing sheets continue to work unchanged
  • Full JSON Schema with examples, $id, and description on every field

@quilt/sdk (new package, v0.1.0)

  • 5 primitives for the agent substrate:
    1. resolveTemplate(uri, ctx) — {{var}} substitution (snake/camel-insensitive)
    2. resolveArtifact(uri, ctx, store) — pin :latest, get provenance
    3. validateManifest(manifest, opts) — JSON Schema + optional precondition checks
    4. publishArtifact(src, meta, store) — content-addressed upload
    5. publishRunTrace(trace, store) — immutable execution trace
  • InMemoryArtifactStore — test/demo store, swappable for R2/S3/etc.
  • 28 tests passing (all primitives + end-to-end planner flow)
  • Heavy header comments on every export
  • TypeScript with full type definitions
  • Apache 2.0

@quilt/cli (updated to v0.4.0)

  • New quilt validate <file> command
    • JSON Schema validation
    • --check-exists for precondition verification
    • --run-id <id> for templating {{run_id}} in preconditions
  • New quilt resolve <uri> command
    • Template substitution
    • Version pinning display
  • Existing commands unchanged

examples/train-classifier.manifest.yaml (new)

  • A fully-decorated example manifest demonstrating all schema fields
  • Works with quilt validate out of the box

landing/agent-substrate.html (new, 56 KB)

  • 8-section landing page documenting the agent substrate
  • Interactive playground: type a URI, paste a manifest, click validate
  • Maps the spec to Quilt's existing primitives
  • Visual: the 5 primitives, the QuiltAgent class, reactive vs imperative

Why this matters

Before v0.4.0, Quilt was a reactive engine. Now Quilt is also a substrate that any agent, planner, or runtime can compose against. The five primitives are the minimum surface that lets:

  • Planners compile goals into manifest DAGs
  • Runtimes execute them transactionally with rollback
  • Agents learn from the past via run traces
  • UIs show users what is happening with full provenance

The same shape that works for "train a classifier" works for "answer a question with RAG", "coordinate three agents", "simulate a universe". The schema is the contract.

Mapping to existing Quilt

New concept Already in Quilt
Manifest (declarative primitive) A Quilt sheet (YAML)
Container entrypoint A cell's evaluate()
Inputs/outputs derived_from + reactive propagation
Pre/postconditions kind: 'sensor' + validate listeners
Rollback @quilt/evolve revert step
model_requirements kind: 'ai' + ai_kind + provider
Run trace engine.serialize() + reactive history
Agent substrate @quilt/mcp exposes Quilt to any agent

Test totals

  • @quilt/sdk — 28 tests (all passing)
  • @quilt/core — 15 tests (all passing, unchanged)
  • quilt validate CLI — verified against 2 example manifests
  • quilt resolve CLI — verified with 3 URI patterns

Migration

Zero. All new fields are optional. All existing sheets work unchanged. @quilt/sdk is a new package, not a breaking change.

Next

  • quilt-runner (new repo) — reference runtime that executes Quilt sheets as transactional plans
  • quilt-rag (new repo) — production RAG with Vectorize
  • quilt-agents (new repo) — multi-agent universes with @quilt/evolve
  • quilt-demo-classifier-dag — end-to-end example