Skip to content

runtime: introduce Forty-Two runtime orchestrator foundation - #1

Draft
bobbygodias wants to merge 28 commits into
mainfrom
runtime/orchestrator-v1
Draft

bobbygodias wants to merge 28 commits into
mainfrom
runtime/orchestrator-v1

Conversation

@bobbygodias

@bobbygodias bobbygodias commented Aug 18, 2026 •

Copy link
Copy Markdown
Owner

Scope

This draft introduces the first engine-agnostic Forty-Two Runtime Layer without switching the existing ModelStore execution path yet.

Included

  • capability registry separating physical hardware from engine devices
  • execution planner with automatic / CPU / balanced / strict policies
  • strict acceleration exposed only when zero-fallback execution can be proven
  • GGML adapter that intersects verified hardware with devices exposed by the concrete runtime
  • llama.rn shadow bridge preserving the current native engine behind the new abstraction
  • semantic runtime load settings
  • runtime telemetry contract
  • architecture and network/lore design notes
  • unit tests for planner invariants and GGML hardware/runtime matching

Safety

main behavior is intentionally unchanged. This PR remains draft until quality gates pass and the shadow bridge reproduces the current load path.

13/09/2026 — reconciliation checkpoint

  • reconciled runtime/orchestrator-v1 with the current main, which already contains the integrated Forty-Two chat shell;
  • preserved the current main versions of shared UI/store files instead of overwriting them with stale branch state;
  • post-reconciliation diff is clean: runtime/documentation changes only, with the branch 0 commits behind main;
  • fresh Forty-Two quality CI on the reconciled HEAD passes localization validation, font validation, TypeScript, lint and the full unit-test/coverage step;
  • the historical unit-test failure on the pre-reconciliation runtime HEAD is therefore retained only as historical evidence and must not be “fixed” with unrelated MobX/async test workarounds;
  • Android ARM64/Vulkan CI is the remaining native build gate for this checkpoint;
  • this PR intentionally stays draft after quality CI: code inspection alone does not yet prove full parity between LlamaRnBridge and every current ModelStore.initContext() load path (including resolved settings and multimodal/projection cases). That parity must be demonstrated before production cutover.

NÃO ENTRE EM PÂNICO. This is the new engine room, not a production cutover.

Bring the green chat-shell main line into runtime/orchestrator-v1 while preserving the runtime foundation and keeping the current main version of shared UI/store files. This drops the stale historical ChatHeaderTitle test state without masking it with async test workarounds.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant