Session 2 — 12 Sep 2026

What happened

  • Walked the Open Questions page one at a time. Q1–Q10 answered, all in: Q1 comments #, Q2 Ruby strings, Q3 units as dot-call functions, Q4 sized ints and crash-on-overflow in every build, Q5 Option with or as the only shorthand, Q6 toolchain-computed verified: line, Q7 typed handles and declared supervisors, Q8 three tiers, Q9 structured diagnostics, Q10 ID-addressed editing.
  • Robert asked for Q10 unpacked (“fascinating”). The unpack became id-addressed-editing and direction 29.
  • Robert added a standing principle after Q4: nothing is final until measured → direction 28. Every performance claim is a hypothesis with a named way to check it.
  • Robert raised package management and supply-chain security as very important → Q17 and direction 30.
  • Q11–Q17 given answer lines; Robert was reading them on his phone. Still pending at session end.
  • Notion was organized under a hub page, then Robert flagged that this was his work Notion and the project must be under his own control. Options weighed: personal Notion / Obsidian vault in the repo / Karpathy’s LLM-wiki pattern on top of that vault. Robert: in on the vault + LLM-wiki conventions, reading on the Obsidian mobile app, private GitHub for sync.
  • Built this vault: both Notion pages exported verbatim to raw/notion/, 80 wiki pages generated from them, SCHEMA.md, index.md, log.md, tools/lint.py. Installed qmd (Robert’s suggestion) with the repo as a collection.
  • Robert accidentally trashed .claude/ (the llm-wiki skill); restored with trash-restore.

Later in the session (after the vault was built)

  • Q11–Q17 answered: Q11 in (after unpacking “platform”), Q12 counter: 70 lines, not 40, Q13 in, Q14 in plus a program-menu, Q15 in (keep Mo, no story), Q16 in, Q17 ordering: research first.
  • Robert: research before the design doc, not after; a broader survey of languages old, obscure, and brand-new; papers included; he runs deep-research prompts himself (Perplexity) and I write them → two-lane research, research/prompts/ and raw/research-runs/.
  • Landscape survey → language-landscape; agentlanguages.dev catalogue (42 agent-native languages) ingested; Exa wired in (tools/exa.py).
  • Robert: use a cheaper model for mechanical work. An Opus worker session run via Herdr wrote the 13 comparison pages from comparison-pass, the comparison-synthesis-draft (7 ⚠️ tensions, 10 steals, ~20 questions), then supply-chain-defenses and landscape-second-lane from Robert’s six Perplexity runs, then Motoko, case-against-new-languages, capability-module-lineage, and refreshes.
  • Robert’s lane caught two stale facts in mine (Gleam version; MoonBit’s Rust benchmark, refuted) and produced the “three kinds of cost” framing: annotation cost collapses when agents write code; proof, runtime, and cultural costs don’t.

Decided about how we work

  • Vault in the repo is the record. Notion is retired for this project once Robert deletes the work-Notion pages.
  • Checkpoint = edit pages, bump updated:, index, one log entry, commit. See SCHEMA.md for the flow.

Next

  1. Fable walks Robert through the 7 ⚠️ tensions in comparison-synthesis-draft, one per message, with a recommendation each; then the open questions it lists.
  2. Fable finishes the synthesis and the Q17 recommendation from supply-chain-defenses.
  3. Then roadmap step 1: spec/design-v0.md (now the folder spec/design-v0/), which must also answer case-against-new-languages.
  4. Notion pages: still to be deleted by Robert (left alone on his request).