# Handoff — 2026-09-19 (written before a full /clear)

Read this, then delete it. Style: user wants extreme concision, plain language,
no filler, decisions discussed one at a time BEFORE dispatching, agents
dispatched one at a time. No commit watermarks. `dune test` stays fast.

## State of the repo

- Branch `dev-loop`, tip `73ab213` (merge of the x86 dyn + conditions lane).
  All green after merge: dune test 287/0, @x86 137 match / 0 differ,
  @sanitize clean.
- dyn now works on the default x86 dev daemon. handler-bind is now a value
  (body's last form) on both backends.
- `sand.flan` is the author's UNCOMMITTED WIP. Never commit it. Stash around
  merges (`git stash push sand.flan -m wip`, pop after). `game-data.edn`
  untracked, leave it.
- Known open hole (flagged, not fixed): `dune test` exits 0 even when
  test_acceptance prints failures — nothing surfaces them.

## In flight: M2 item 1 — dyn maps + keywords

Agent worktree: `.claude/worktrees/agent-a3009901a950e8daa`
(branch `worktree-agent-a3009901a950e8daa`). NOT committed yet — uncommitted
changes across lib/{ast,check,emit,load,parse,prelude}.ml, runtime/flan_dyn.{h,c},
test/dune; it was writing checker/parse tests in test/test_flan.ml when it hit
the session limit (resets 7:20pm Asia/Bangkok).

To finish: spawn an agent IN THAT WORKTREE (don't reset it) with: "Finish the
in-progress dyn maps + keywords lane. Read the uncommitted diff first to see
what exists. Remaining: tests in test_flan.ml, acceptance rows (LLVM, -O0,
--x86; run programs for real output, never guess wording; no double-quotes in
OCaml comments), then ONCE at the end: dune test --force, dune build @x86
--force, dune build @sanitize --force. Commit in the repo's prose voice."
Then: stash sand.flan, merge, run the three sweeps, pop stash.

## M2 queue (FIX.org "M2 queue" section, line ~408) — dispatch ONE AT A TIME

1. dyn maps + keywords (in flight, above)
2. per-type descriptors — use Opus-class agent (representation change)
3. typed containers into dyn as views (rides on 2)
4. nil↔None at (Option T)
5. typed =/!= grow strings
6. defclass = named dyn map + CLOS/Clojure dispatch (after 1)
7. dyn if truthiness
8. return slot stays mandatory — decision only, no work

Model plan agreed with user: session on Opus, Sonnet agents for items 3–8,
Opus agents for item 2 and for reviews of merges.

## Not yet recorded in FIX.org (do this early, cheap)

Append a section on the syntax-family discussion (2026-09-19):
- User wants an F#-ish indentation/ML surface syntax, side by side with
  s-exprs. Languages that "disappear" for them: Python > Odin > F#.
- Agreed architecture if it happens: one AST (the existing forms), second
  reader; macros stay usable from both; Nim-style quote-block could allow
  writing macros in ML syntax too.
- Middle options discussed: Parinfer (editor-only), wisp/sweet-exprs
  (indentation-implied parens), simplified in-paren syntax. Rhombus = the
  maximal reference.
- DECISION: deferred, no spike queued. User's current hypothesis: their
  Clojure friction may be immutability/planning-ahead, not parens — they
  will write imperative Flan as-is and see if parens still grate. Revisit
  with that evidence.

## Environment notes

- /tmp (7.6G tmpfs) was 100% full; freed ~470M of stale dune test dirs, now
  ~95%. Remaining bulk: 2.6G Refold-Defold Claude scratch + ~850M numeric
  JVM/Defold dirs — user hasn't said whether that session is closed; ask
  before deleting.
- Verify agents' worktree bases: every past agent arrived stale. Prompt must
  demand git log check + reset to dev-loop tip.
- Repeated full sweeps are the main time sink — tell agents to verify
  incrementally, full sweeps once at the end.