Every non-trivial design decision in nuthatch is written down first, as a numbered RFC in
docs/rfcs/. They’re numbered
in build order, each states its dependencies and what it blocks, and the status lifecycle is
Draft → Accepted → Implemented → (Superseded / Parked). Statuses are reconciled against the
progress log; measured numbers are cited, and targets are labeled as targets, never as results.
The series
- 0001 Generalized decode & nests (Implemented) - the foundation: multi-contract nests, the decode registry.
- 0002 The Horizon nest (Implemented) - the first real-world nest.
- 0003 reth ExEx tip mode (Accepted; deferred) - colocated-node ingestion.
- 0004 Backfill throughput (Implemented) - measure first, optimise second; seal-direct.
- 0005 Release engineering (Implemented) - the v0.1.0 bar and beyond.
- 0006 Grant funding / 0007 Launch & validation (Accepted; process) - the non-engineering record.
- 0008 The compliance pack (Implemented) - labels, lists, screening, flags, exposure, the signed audit pack.
- 0009 Factories (Implemented) - dynamic child-contract discovery.
- 0010 Admin UI & webhooks (Implemented) - ease-of-use parity.
- 0011 The graph-network nest (Parked after pilot) - the wedge proven in prod.
- 0012 Multi-nest runtime & packaging (Implemented) - runtimes and content-addressed bundles.
- 0013 Storage & query-engine direction (Accepted; the gate was run) - the DuckDB union shipped. DataFusion convergence was benchmark-gated, and DataFusion did not meet the gate: 1.6-2.7× DuckDB’s latency on the fold that matters, widening as segments grow, at exact result parity. DuckDB stays in both modes; the destination is unmet, not repudiated.
- 0014 Firehose-class extraction (Draft; deferred) - traces and state diffs via ExEx.
- 0015 The delightful core (Implemented) - the REPL, magical init, live feedback,
add, the MCP one-liner. - 0016 The semantic layer & agent-grade MCP (Implemented) -
semantic.toml, errors-as- prompts,explain, result shaping, resources & prompts, the eval harness. - 0017 The builder skill (Implemented) - the generated, drift-gated CLI reference.
- 0018 What a nest is (§1 implemented; §2 retired; §3 deferred) - authored SQL views; the Starlark front-end, retired.
- 0019 The nest registry (Implemented) - publish and pull by
name@version, and workers pull the nests they are assigned - by content address when the fleet pins abundle_hash, so re-tagging a version in a registry cannot change what a fleet runs. - 0020 Nest lifecycle & the N-1 upgrade (Implemented) -
diff, hot-swap, deprecation, segment reuse. The resync tax, killed. - 0021 The multichain runtime (Accepted; slice 1 shipped, live two-chain run done) - one runtime, one isolated cursor per chain.
- 0022 Distributed scaled mode (Implemented - control plane and ingestion both) - read/write planes for operators, proven across real machines including 377 blocks indexed through a 90-second control-plane outage. Until v0.9.3 the writer pool took leases and ran no indexing at all (#250); ten level-5 checks passed throughout because every one tested the control plane and none asserted a row appears.
- 0023 Contract state, derive-first (Accepted; tiers 1-2 shipped) - the
eth_callyou don’t need: derived-view recipes and the immutable-metadata cache. Tier 3 is a foundation rather than a feature -[[calls]]parses and validates, and nothing executes it yet (#262). - 0024 The eth_call execution engine (Draft) - a demand-driven state cache, if the residue demands it.
- 0025 Adaptive MCP tool advertisement (Implemented) - advertise only the tools a nest can
answer, so an agent is never handed an inert tool that returns
{"count":0}. - 0026 Fault quarantine & partial health (Implemented) - a runtime survives its sick nests: a nest’s error no longer kills its cursor, and a cursor’s death no longer kills the runtime.
- 0027 The live runtime (Implemented; all 7 slices) - mounting and unmounting nests without a restart, so onboarding one tenant doesn’t restart every co-tenant.
- 0028 Adaptive log-range control (Implemented) - a fix pack for
eth_getLogsrange control: classify RPC failures properly rather than retrying an auth rejection forever. - 0029 The fastest indexer (Implemented; all 5 slices) - found by running someone else’s benchmark, Sentio’s OBIB. Case 1 did not merely run slowly, it never finished: Alchemy returns its oversized-range refusal as HTTP 400, which the classifier did not enumerate, so a window that needed splitting was retried unchanged forever. It now completes in 74.8 s for 294,278 events in 321 RPC requests - the record count matching Sentio’s own README.
Conventions
Every RFC honours the non-negotiables (single static binary, the ≤2 GB budget, no phone-home,
determinism in the core, MIT OR Apache-2.0) and carries the standard structure: Abstract, Motivation,
Goals/Non-goals, Design, Implementation, Testing, Risks, Alternatives, Open questions. Companions
in docs/: backlog.md (everything deferred across the series), prod-readiness.md (the bar
a release clears before it’s pointed at a real workload unattended), and the progress log (the
running narrative the statuses are reconciled against).
Proposing a change? Open a discussion first; if it survives contact, it becomes the next number.