The runtime
A map of content for layer 1: the process model, the failure model, the two runtimes and the C backend, and every step that changed them under a program’s pressure. The runtime carries the thesis’s reliability claim, so this is where most of the measuring happened.
The model
- 03-semantics — chapter 3: values, functions, effects, processes, the failure model (what a crash discards, what a timeout leaves, what a restart loses), the runtime surface
- 07-toolchain — chapter 7: the interpreter, the compiler, the agent interface, the bets that carry numbers
- state-model, errors-and-failure, effects-and-capabilities — the deep dives behind chapter 3
- d14-processes-are-the-only-identity, d12-concurrency-at-the-edges, d16-direct-style-io, d33-bounded-mailboxes, d36-vm-first-runtime, d37-runtime-mcp-surface, d40-structured-runtime-events, d38-time-travel-debugging, d39-hot-code-reload
- vm-first-vs-c-first and agent-native-runtime-features — the two deep dives that filed directions 36 to 40
- q07-process-api, q18-main-and-the-platform, q11-platform-and-stdlib
The steps, in the order they landed
- interpreter-step-4 — the milestone:
Mo.Sim,updateas a transaction, invariants, bounded mailboxes, supervisors, the crash report - interpreter-step-9 — seeds and faults
- interpreter-step-11 —
Net, processes undermo run - interpreter-step-13 and interpreter-step-15 — the C backend, then processes and
Netin it - interpreter-step-16 —
Http - interpreter-step-18 — the outside review’s fixes:
invariantafter every update, handles as capabilities, the depth bound - interpreter-step-19 and interpreter-step-20 — a process that nothing holds ends; the runtime owns the loop
- interpreter-step-21 — memory and green threads: a process is a fiber, 65,530 idle connections
- interpreter-step-22 — the derived deadline,
reply_by - interpreter-step-23 — the runtime surface as a capability, structured events
- interpreter-step-24 — a delayed send, handles in state, the read-only
Fshole closed - interpreter-step-28 — writes in place, regions giving pages back
- interpreter-step-29 and interpreter-step-29b —
restart: :neverhonoured; replay memory bounded on a real log - interpreter-step-30 — a scheduler per core,
fsyncoff the scheduler; the Mac table - interpreter-step-31 — the deferred reply
- mac-scaling-run — nothing scales on the M3 Max; the row that led to step 34
- interpreter-step-32 — crash reports kept apart from the ring; the reopening restart in the corpus
- interpreter-step-33 — the crash report freed after printing (46 MiB a restart to under 0.3); the interpreter’s abort on a large log
- interpreter-step-34 — placement with the starter, the crossing made cheap (100,000 cross-scheduler asks 4 s to 0.14), measured at 1, 4, and 14 cores
Current harness-driven work
- interpreter-step-42 — in progress: stress compaction, poisoned memory, owned parcels and process-group cleanup
- interpreter-step-44 — in progress: bounded binary connection chunks to unblock mo-workspace-server-4a without changing clients
The runtime’s own row
A crashed process with the service still answering, P6: control-run-8 (Mo’s outage, a wait hidden as a message pattern), control-run-10 (the BEAM restores service in under 600 ms, and exits on the fourth kill or on a full disk), erosion-round (generation two answers through a full disk in Mo, Go, and Python; Elixir does not), 10-language-after-the-rounds §1 and §2 (what the language should say about it).