Archetype 0.6¶

Archetype 0.6 is an intentional pre-1.0 boundary reset. It separates the generic ECS framework from three independently installable world libraries and removes APIs and persisted shapes that put domain behavior in the wrong owner. It also replaces the frozen parallel core-sync implementation with a separate, deliberately bounded teaching engine.

This is a clean break. There are no world-library import shims, dynamic runtime or world method aliases, deprecated world-library spellings, or automatic Research-ledger migration. Update an application to the canonical 0.6 surface before upgrading its environment.

Distribution layout¶

Install	Owns
`archetype-ecs`	Generic ECS execution, storage, migration, commands, Activities, artifacts, evaluation, runtime, API, and CLI
`archetype-smol`	Synchronous in-memory Components, DataFrame Processors, atomic steps, and snapshot queries for education
`archetype-missions`	Coding-agent missions, sandboxes, sessions, transcripts, and trajectory evidence
`archetype-physical-ai`	Physical state, local policies, and hosted physical episodes
`archetype-research`	Generic AutoResearch values, ledger state, views, and workflow

Install only the libraries an application uses, or install the complete world-library stack with archetype-ecs[all]. Each world library depends on the compatible framework release; world libraries do not depend on one another. archetype-smol is independent and is not selected by archetype-ecs[all].

Synchronous surface reset¶

ArchetypeRuntime.sync(), SyncArchetypeRuntime, SyncRuntimeWorld, and run_sync remain the supported blocking interface to the production runtime. They execute the same async engine, command gate, storage, and lifecycle contracts as the async handles.

The parallel archetype.core.sync engine and its root aliases are removed: World, Processor, System, Store, SyncWorld, SyncProcessor, SyncSystem, SyncStore, QueryManager, and UpdateManager. Applications that need production semantics use the runtime facade. Educational code may choose archetype-smol, whose smaller contract is intentionally not a drop-in replacement.

Required source changes¶

Use family-qualified imports and explicit typed adapters:

Removed pre-0.6 surface	0.6 surface
Domain values imported from `archetype`	Import from `archetype.missions`, `archetype.physical_ai`, or `archetype.research`
`runtime.missions(...)` or `RuntimeMissions(...)`	`Missions(runtime, ...)`
`world.query_trajectory(...)` or `world.grade_trajectory(...)`	`MissionWorld(world).query_trajectory(...)` or `.grade_trajectory(...)`
`world.ingest_claude_transcript(...)` or `world.transcript_rows()`	`MissionWorld(world).ingest_claude_transcript(...)` or `.transcript_rows()`
`world.run_hosted_episode(...)`	`PhysicalAI(world).run_hosted_episode(...)`
`world.autoresearch(...)`	`Research(world).autoresearch(...)`
`CandidateContext`	`ResearchCandidateContext`
`archetype.episodes` transcript and trajectory imports	`archetype.missions.trajectories`
Research runner-state loading	`archetype.missions.sessions.load_runner_sessions(...)`

World-library adapters are async in 0.6. A synchronous framework world does not grow domain methods when a library is installed.

Research ledger reset¶

Pre-0.6 Research ledgers are unsupported and are not migrated in place. Start a new 0.6 experiment ledger. If historical evidence matters, export it with the older release before upgrading and retain it as application-owned evidence.

The 0.6 ledger is deliberately generic:

RunStatus is RUNNING, SUCCEEDED, or FAILED; the runner-shaped BOOTING, STOPPING, STOPPED, and CRASHED states are gone.
Experiment no longer persists repository URL or branch fields.
Run no longer persists VM, harness, repository, branch, task, agent, workspace, or commit fields. It records the candidate world instead.
BranchHead records the incumbent candidate world rather than a Git commit.
SQLite coding-runner ingestion belongs to Missions and returns RunnerSession values.

Research can still optimize code. An application composes Missions and Research through their public adapters and stores bounded coding evidence in the opaque Research result envelope.

Episode ownership¶

There is no archetype-episodes distribution and no archetype.episodes package:

execution EpisodeConfig and EpisodeResult remain in archetype.world;
coding-session transcripts and trajectories live in archetype.missions.trajectories;
hosted physical episodes live in archetype.physical_ai; and
dataset episode identity remains in archetype.evaluation.

Shared episode identifiers permit joins without inventing a universal episode facade.

Release unit¶

The 0.6 release is one coordinated set: five wheels and five source distributions at version 0.6.0. Release evidence installs the base framework, each individual library with the framework, and the complete stack from those exact artifacts, plus Smol in isolation. A partial five-project publication is not a completed release. Before the first release, register pending Trusted Publishers for the four new project names on both PyPI and TestPyPI. Registration preconfigures OIDC; it does not reserve or claim a name, and each new name remains claimable until the first successful OIDC publication creates the project on that registry. See Repository harness for the exact, package-specific OIDC identities and the direct-workflow bootstrap required for new projects.

Release publication also requires the live, pinned Biome scenario. The release-only proof runs on the same approved, one-job ephemeral Apple Silicon Mac used for Apple Container evidence. It fail-closes on the Darwin, build-tool, and explicit-live prerequisites; builds and launches the checked-in Biome and Flecs revisions in an active WindowServer/Metal session; waits for REST readiness; proves the native mission plus durable Archetype evidence; and closes the process and port. Before the native executable can start, a runner-owned guardian acknowledges its isolated process group; parent EOF, timeout, or cancellation then triggers independent TERM/KILL and loopback-port closure proof. Since the release-workflow simplification, Biome and Apple Container run at demand cadence as local scripts on the qualifying Mac (make operational-demand-biome, then make operational-demand-apple) rather than as a release job; their receipts are demand evidence with the same installed-wheel binding, and the release evidence gate requires exactly the release-cadence receipts. See Repository harness for the current lane inventory.

For current architecture and installation contracts, continue with World libraries, Runtime, and API stability.