Skip to content

learning-surface-2026-05-12.md

Latest commit

 

History

History
197 lines (158 loc) · 10.4 KB

learning-surface-2026-05-12.md

File metadata and controls

197 lines (158 loc) · 10.4 KB

Learning Surface Audit — 2026-05-12

Survey of ~/core/* (90 projects) to identify domains the hexa-forge code-LLM should learn beyond the current 7 langs (Python/Rust/TS/Go/C/ Zig/Swift) + Apple stack.

Method

  • Walked all 90 ~/core/* subdirs (excluding archive-*, anima_clm_* numbered snapshots, legacy).
  • Tagged each by: dominant file extension(s), build manifest, presence of framework files (Dockerfile, CMakeLists.txt, package.json, Cargo.toml, go.mod, pyproject.toml, Package.swift, pubspec.yaml, build.zig).
  • Read README/spec headers for the top-signal projects.
  • Discarded archives, internal-only marker streams, snapshot directories.

Findings — domains the LLM does NOT currently know

Ordered by both volume of code in the user's ecosystem AND likelihood of being asked.

Tier 1 — actively used, large volume

domain source project files (approx) why it matters
Dart / Flutter cake-wallet/ ~962 .dart The Stage-1 backend of wraith-wallet wraps cake-wallet directly. Multi-chain BTC + XMR + ETH + LTC mobile UI. User actively maintains the fork.
Zig (deep) void/ substantial .zig (build.zig + src/) void is a CMake-wrapped Zig library (libvoid-vt). Real-world Zig + Zig build system (build.zig.zon). Our current Zig coverage is Stack v1 sample only — no build.zig patterns.
PyTorch + CUDA + RunPod anima/ ~3296 .json, large Python training/ tree Production ML stack: pytorch/pytorch:2.5.1-cuda12.4 docker, requirements pinning, RunPod SSH, cloudflared tunneling, checkpoint management (checkpoints/clm_v2/).
Playwright (browser automation) browser-harness/ playwright-core node deps Cross-browser scraping / e2e test harness. Not in our SFT.
Discord bot + Anthropic SDK pixie/ node + Anthropic SDK pixie/discord-translator/ + worker/ use @anthropic-ai/sdk. Bot patterns: slash commands, interactions, embeds.
TOML schema design hexa-meta/ ~601 .toml hexa-toml-spec.json versioned 1.0.0. The user's whole stack uses TOML for project manifests (hexa.toml everywhere). Our SFT touched a few but not the patterns.
DuckDB analytics orpheus/dormant_explorer/ implicit "DuckDB main path; BigQuery driven; lost-likelihood heuristics" — chain-wide Bitcoin dormant address analysis.
BigQuery SQL orpheus/ implicit Same project, BigQuery for chain-scale aggregations.

Tier 2 — domain-specific terminology + libraries

domain source concept set
BIP-series crypto wraith-wallet, orpheus BIP39 mnemonic (24 words + wordlist checksum), HD wallet derivation, PSBT, ECDSA recovery, Schnorr/Taproot
Monero / XMR / RingCT wraith-wallet/backend_cake → backend_native RingCT, ring signatures, stealth addresses, view/spend keys
Tor stack wraith-wallet, cake-wallet built-in Tor integration for mobile wallet; .onion v3
Lightning Network wraith-wallet (relay sub-domain) HTLCs, channel state, BOLT specs
Slipstream / Eden / bloXroute orpheus, wraith-wallet private mempool / private-relay procurement
age-file key store wraith-wallet/vault/ age encryption format for keys at rest
TOTP 2FA wraith-wallet/vault/ RFC 6238 time-based OTP
Atomic swaps cake-wallet integration BTC↔XMR atomic swap protocol

Tier 3 — meta-tooling / build / ops

domain source gap
Docker (PyTorch ML) anima/Dockerfile apt-get, cloudflared, SSH-for-RunPod patterns
cloudflared tunneling anima Linux GPU rental tunneling
age encryption wraith-wallet/vault symmetric file encryption
GitHub Actions / CI hexa-forge .github/workflows (we have one workflow, but minimal SFT)
CMake → Zig bridge void/CMakeLists.txt non-trivial cross-build-system pattern

Tier 4 — narrow but distinctive

domain source notes
DSP / audio anima (portaudio19-dev, libportaudio2) audio I/O on GPU runtime
Cairo / circuit DSL (not found) not in ecosystem
Solidity / Ethereum smart contracts (mentioned in cake-wallet ETH support but no source) not in repo
QRNG / quantum entropy qmirror, qrng hardware entropy source consumption for cryptography

Specifically excluded per user direction

  • ORM — user said "ORM 은 사용안되고 있을수 있다 (may not be used)". Survey confirms: no SQLAlchemy / Prisma / Diesel / TypeORM / Drizzle / Hibernate / Active Record patterns visible. Stack uses raw SQL via DuckDB/BigQuery, or hexa.toml declarative state. SKIP ORM.

Recommended additions for v0.3.0 SFT

In priority order by user-value × LLM-leverage:

P0 (next round — high yield)

  1. Dart / Flutter — ~100 hand-crafted Q/A pairs covering:

    • StatelessWidget / StatefulWidget skeleton
    • BLoC pattern, Provider, Riverpod
    • flutter_secure_storage for keys
    • HTTP via dio / http
    • JSON via dart:convert + json_serializable
    • Flutter for desktop (cake-wallet is multi-platform)
    • Plus: Dart-side BIP39 / ECDSA helper calls (since cake-wallet uses them)

  2. PSBT + BIP39 + HD-wallet — ~50 pairs covering canonical recipes:

    • BIP39 mnemonic → seed → master xprv → derivation path
    • PSBT v0 + v2 structure (BIP-174 / BIP-370)
    • Taproot key-spend vs script-spend
    • Common scripts: P2WPKH, P2WSH, P2TR
    • Anti-pattern: hand-rolled wordlist (always use BIP39 official list)

  3. PyTorch training-loop boilerplate — ~80 pairs:

    • DataLoader + Dataset + collate_fn
    • mixed-precision (autocast + GradScaler)
    • DDP (DistributedDataParallel) skeleton
    • checkpoint save/load + strict=False for resume
    • LR schedulers (cosine + warmup, WSD)
    • PyTorch 2.5+ specific: torch.compile(), FSDP2

P1 (round after)

  1. Zig deep — ~50 pairs:

    • build.zig.zon package manifest
    • build.zig artifact / dep / pkg-config emit
    • comptime patterns
    • allocator hygiene (std.heap.GeneralPurposeAllocator)
    • error sets + errdefer

  2. Discord bot patterns — ~30 pairs (since pixie uses them):

    • discord.js Client + intents
    • slash commands via SlashCommandBuilder
    • interaction.deferReply for >3s work
    • embeds + components (buttons / select menus)
    • rate limit handling

  3. Playwright (browser-harness style) — ~30 pairs:

    • browser.newContext + storage state
    • page.waitForLoadState patterns
    • locator chaining, getByRole / getByText
    • download/upload handling
    • tracing + video for debugging

  4. TOML schema design — ~30 pairs:

    • root-table key conventions
    • inline table vs table-array
    • dotted key path semantics
    • validation via serde (Rust) / tomli (Python)

P2 (later)

  1. DuckDB + BigQuery SQL — analytical patterns (window funcs, CTEs, qualify).
  2. age encryption — file encryption recipe + key recovery.
  3. Tor / .onion integration — stem/torpy basics.

Per-project recommendation table (what to mine for SFT)

project tier mine for rough effort
cake-wallet/ P0 Dart/Flutter idioms; BIP39 in Dart 1-2 hours hand-curate
anima/ P0 PyTorch boilerplate; CUDA dockerfile patterns 1 hour
void/ P1 Zig build.zig; libvt patterns 1 hour
pixie/ P1 Discord bot + Anthropic SDK; node worker patterns 30 min
browser-harness/ P1 Playwright recipes 30 min
orpheus/ P0 BIP/crypto + DuckDB analytics 1 hour
wraith-wallet/ P0 BIP/age/TOTP/PSBT recipes 1 hour
hexa-meta/ P1 TOML schema authoring 30 min
hexa-codex/ (already covered in canon corpus) — —
hexa-lang/ (already covered in canon corpus) — —

Not worth SFT-ing

  • anima_clm_01..13 snapshot directories — historical only; no new patterns.
  • archive-* — explicitly archived.
  • state/, markers/ — large counts of .marker files are state, not code.
  • gamebox/ — release-note heavy; minimal new code patterns.
  • legacy/ — opt-out.
  • All hexa-<single-word> placeholders with mostly empty contents (hexa-arts, hexa-cosmos, etc. — 14-18 md files = roadmap stubs).

Operating note

The ~/core/* survey itself reveals an important framing fact: this ecosystem is >90% hexa + .marker state files, with the actual code-LLM training surface coming from a small set of high-density projects (cake-wallet, void, anima, pixie, browser-harness, orpheus). Future SFT iterations should mine these directly rather than treating the full ecosystem as a uniform corpus.

v0.3.0 SFT dataset proposal

Builder: tool/build_sft_dataset_v8.py (next round)

Composition (target ~2,400 rows):

  • v7 base 1,985
  • Dart/Flutter Q/A (P0) 100
  • BIP39/PSBT/HD-wallet (P0) 50
  • PyTorch training-loop (P0) 80
  • Zig deep (P1) 50
  • Discord bot (P1) 30
  • Playwright (P1) 30
  • TOML schema (P1) 30
  • Plus optional: T5 HX-codes recovery + T4 enum fix ~50

This would land the v0.2.0-r8 adapter on dancinlab with crypto-wallet ecosystem competence (cake-wallet maintainability) + ML-systems competence (anima training loops) + browser automation + Discord bot patterns.

Expected eval impact (best case):

  • hexa-eval STRICT: 60.7% → 60-65% (steady, depending on r7 retention)
  • 5-NL F1 code synth: 100% → 100% (broader code base, same baseline)
  • New ad-hoc benches possible:
    • dart-eval: Flutter widget Q/A (10-20 tasks)
    • crypto-eval: BIP/PSBT recipes (15-20 tasks)
    • pytorch-eval: training-loop Q/A (20-30 tasks)