el/spec/codegen-js.md

# El JavaScript Backend (codegen-js)

**Status:** scaffolded. Hello-world compiles and runs. ~50% language coverage. Core runtime (~30 builtins) implemented. CGI / DHARMA / LLM / Engram intentionally stubbed.

**Authoritative files**

| File | Role |
|---|---|
| `el-compiler/src/codegen-js.el` | El → JS code generator (mirrors `codegen.el`) |
| `el-compiler/runtime/el_runtime.js` | Browser/Node runtime that compiled programs link against |
| `el-compiler/src/compiler.el` | Adds `compile_js()` and `--target=js` CLI dispatch |
| `spec/codegen-js.md` | This document |

---

## 1. Why a JS backend exists

El compiles to C today. C is the right substrate for the agent runtime, the DHARMA daemon, and Engram. But three first-class consumers of El need to **run in a browser**, where C is not an option:

1. **`el-ui/runtime/`** — the activation-based frontend framework written in JS. The long-term plan is to author components and the runtime itself in El and compile them down to JS.
2. **`cgi-studio`** — the web app for cultivating CGIs. Today it is hand-written JS. Once the JS backend is mature, the studio's UI logic can be authored in El and share types/identifier names with the CGI it cultivates.
3. **Marketplace plugin UIs** — third parties writing browser-side El that runs untrusted in a sandbox. They need a JS target.

A secondary motivation: **El-on-Node**. CLI tooling, build scripts, and tests benefit from a tight `el → js → node` cycle without a `cc` step.

---

## 2. Type representation strategy

The C backend pretends every value is `int64_t`. That is a deliberate runtime trick to avoid dynamic dispatch in generated C. JavaScript already has tagged dynamic values, so the JS backend is **simpler**: every El value is a native JS value, and the tag of `el_val_t` collapses into the JS type system.

| El type | C representation | JS representation |
|---|---|---|
| `Int` | `int64_t` (direct) | `number` (with `Number.isSafeInteger` caveat — see §6) |
| `Float` | `int64_t` bit-cast of `double` via `el_from_float` | `number` (no bit-cast — JS number IS a double) |
| `Bool` | `int64_t`, 0 = false, nonzero = true | `boolean` |
| `String` | `(int64_t)(uintptr_t)cstring` | `string` |
| `Void` | C `void` | `undefined` |
| `[T]` (List) | `el_val_t` pointer to refcounted struct | `Array<any>` |
| `Map<K,V>` | `el_val_t` pointer to refcounted struct | plain object `{[key]: any}` |
| `EL_NULL` (`0`) | `(el_val_t)0` | `null` |
| Any | `el_val_t` | `any` (no compile-time check) |

**Key consequences:**

- `+` on two strings is JS `+` (string concat) — no `el_str_concat()` runtime call needed for the common case. The runtime DOES export `el_str_concat` for the cases where codegen does not know the types.
- `==` on strings is `===` — not `str_eq()`. Same disambiguation logic as the C backend (look at left/right kind, fall back to `str_eq` for identifiers without int annotation).
- `Map` access `m["foo"]` compiles to JS `m["foo"]` (no `el_get_field`). For `Field` access (`m.foo`) we emit `m["foo"]` so it works on plain objects regardless of prototype shape.
- List access `arr[i]` is JS `arr[i]`. No bounds checking — same as C (which segfaults on bad index). Could add `el_list_get` wrapper later for safe access.
- `EL_NULL` becomes JS `null`, not `undefined`. The runtime checks for `=== null` consistently. This avoids the JS undefined/null fork and matches El's single null value.

---

## 3. Builtin runtime layer (`el_runtime.js`)

Same function names as `el_runtime.c` wherever possible, so codegen-js can emit the same call sites. The runtime is a single ES module that exposes every builtin as a named export AND attaches them to a `globalThis.__el` namespace (so generated code can do either `import * as el from './el_runtime.js'` or assume globals).

**The codegen-js generated output uses the global-namespace style:** every emitted file starts with `import './el_runtime.js'` (which side-effects the globals) so call sites stay flat — `println(x)` not `el.println(x)`. This matches the C backend's flat call surface and keeps the generated code grep-compatible across targets.

### Implemented today (~30 builtins)

| Category | Functions |
|---|---|
| I/O | `println`, `print` |
| String | `el_str_concat`, `str_concat`, `str_eq`, `str_starts_with`, `str_ends_with`, `str_len`, `int_to_str`, `str_to_int`, `str_slice`, `str_contains`, `str_replace`, `str_to_upper`, `str_to_lower`, `str_trim`, `str_index_of`, `str_split`, `str_char_at`, `str_char_code`, `str_lower`, `str_upper` |
| Math | `el_abs`, `el_max`, `el_min` |
| List | `el_list_new`, `el_list_len`, `el_list_get`, `el_list_append`, `el_list_empty`, `el_list_clone`, `list_push`, `list_join`, `list_range` |
| Map | `el_map_new`, `el_get_field`, `el_map_get`, `el_map_set` |
| HTTP | `http_get`, `http_post`, `http_post_json` (via `fetch()`, returns `Promise<string>` — see §5 async caveat) |
| FS | `fs_read`, `fs_write`, `fs_list` (Node-only, throw in browser) |
| JSON | `json_parse`, `json_stringify`, `json_get`, `json_get_string`, `json_get_int` |
| Time | `time_now`, `time_now_utc`, `sleep_secs` (Node), `sleep_ms` |
| Bool | `bool_to_str` |
| Process | `exit_program` (Node `process.exit`, throw in browser) |
| Refcount | `el_retain`, `el_release` (no-ops — JS has GC) |
| ARC method-call shortforms | `append`, `len`, `get`, `map_get`, `map_set` |
| Native VM aliases | `native_list_get`, `native_list_len`, `native_list_append`, `native_list_empty`, `native_list_clone`, `native_string_chars`, `native_int_to_str` |
| `args` | `args()` returns `process.argv.slice(2)` in Node, `[]` in browser |
| `state_*` | In-memory `Map` keyed by string |
| `env` | `process.env[k]` in Node, throws in browser |

### Stubbed (throw at runtime)

Every function in this list compiles successfully but throws `Error("not supported in JS target — needs server-side delegation: <name>")` when called. This is a **runtime** error, not a compile error, so it doesn't block compilation of code that has dead-code paths through these functions.

- All `dharma_*` (membership in DHARMA network requires the daemon)
- All `engram_*` (needs the embedded SQLite + activation engine — could be reimplemented in JS later)
- All `llm_*` (CORS + API key handling — must go through a server-side proxy)
- `http_serve` (browsers don't host servers; Node could, but that's a separate runtime mode)
- `el_cgi_init` (CGI identity is a server-side concept)
- Crypto: `sha256_*`, `hmac_sha256_*`, `base64*` (deferred — can use `crypto.subtle` later)

### Browser-side specific behavior

When running in a browser:
- `println` / `print` map to `console.log` (no stdout in browsers)
- `http_get` / `http_post` use `fetch()` (CORS applies)
- `fs_*` throws (browsers have no fs)
- `args()` returns `[]`
- `env(k)` throws (or could read from a global config object — TBD)

When running in Node:
- `println` / `print` map to `console.log` and `process.stdout.write`
- `fs_*` use `node:fs/promises` (sync versions for the simple cases)
- `args()` returns `process.argv.slice(2)`
- `env(k)` returns `process.env[k] ?? null`

The runtime auto-detects via `typeof window === 'undefined'`.

---

## 4. Tradeoffs vs the C backend

| Concern | C backend | JS backend |
|---|---|---|
| **Static types** | El's `Int` becomes `int64_t`, real arithmetic | El's `Int` becomes `number` — loses precision past 2^53 |
| **Linking model** | Static link against `el_runtime.c` + libcurl + libpthread | ES module import of `el_runtime.js` |
| **Dynamic dispatch** | `dlsym` for `http_set_handler` / `llm_register_tool` (requires `-rdynamic`) | JS function value lookup via `globalThis[name]` — no compiler flag |
| **Tool registry** | dlsym walks symbol table; tool fns must be top-level C symbols | Tool fns live as exports of the generated module; trivially callable |
| **Memory model** | Refcounted lists/maps with `el_retain`/`el_release` to avoid leaks | JS GC handles all of it; `el_retain`/`el_release` are no-ops |
| **`+` overload** | Has to dispatch in codegen between `el_str_concat` and integer `+` because at C level both are `int64_t` | JS `+` is already overloaded: `"a" + "b"` → `"ab"`, `1 + 2` → `3`. Codegen still preserves the existing dispatch for safety, but the runtime fallback is correct |
| **Concurrency** | `pthread`-backed `http_serve` | Single-threaded event loop; `http_serve` not supported in this target |
| **HTTP client** | libcurl, blocking, returns body string | `fetch()` is async — see §5 |
| **CGI identity** | `el_cgi_init` runs at start of `main()` | Not supported; UI code is not a CGI principal |
| **DHARMA / LLM** | Native, blocking, libcurl-backed | Not supported — all such calls throw and the program is expected to delegate to a server-side El daemon via plain HTTP |
| **Compile speed** | El → C → cc → binary (cc is the slow step) | El → JS → done. Faster iteration |
| **Output size** | Static binary ~2MB | Source `.js` + ~10kb runtime |

---

## 5. The async problem (the big deferred decision)

`fetch()` is async. The C backend's `http_get(url)` is synchronous and returns the body string directly. El source was written assuming sync. Three options:

1. **Pretend it's sync from El's POV; use synchronous XHR (browser) or `child_process.execSync('curl …')` (Node).** Bad: synchronous XHR is deprecated and frozen on the main thread; `execSync` is a hack.
2. **Make every `http_*` builtin in the JS runtime return a `Promise`, and rewrite codegen-js to insert `await` everywhere.** This requires turning every El function that transitively calls a network builtin into an `async fn` in JS. Doable, but invasive — the El AST does not currently mark async-ness.
3. **Compile El's call sites with implicit await; compile-time taint tracking marks every fn that transitively calls a network builtin as `async`. Generated JS uses `async function` and `await`.** This is the right answer long-term.

**Decision for this scaffold:** option 3, but only the runtime side is implemented. `http_get` in `el_runtime.js` returns a `Promise<string>`. `codegen-js.el` does NOT yet emit `async`/`await`. Calling `http_get` from compiled El will return a Promise that the El program will treat as a string (which produces `"[object Promise]"`). This is documented and accepted for the scaffold; the compile-time taint pass is a follow-up.

For now, programs that don't touch HTTP work correctly. That covers `el-ui/runtime` (which only manipulates the DOM and a graph), most of cgi-studio's pure UI components, and all hello-world style programs.

---

## 6. Number precision

JS `number` is IEEE 754 double — only 53 bits of integer precision. El `Int` is `int64_t` and the runtime sometimes uses the full 64 bits (e.g. `time_now_utc` returns nanoseconds-since-epoch, which exceeds 2^53 in practice).

**Decision for this scaffold:** accept the precision loss. Document it. UI code does not use 64-bit timestamps. If/when a use case demands it, `time_now_utc` can return a `BigInt` and we can introduce a `BigInt` sub-mode. That's a follow-up.

---

## 7. What's NOT supported in JS target initially

This is the canonical list. Programs that use any of these compile (no `#error`-style fail-fast like the C backend's capability check) but throw at runtime or behave as documented.

| Feature | Status | Notes |
|---|---|---|
| `cgi {}` block | Compiled to a no-op + warning comment | CGI identity is server-side. UI code is not a CGI. |
| `service {}` block | Compiled to a no-op + warning comment | Same. |
| All `dharma_*` | Stub throws | Programs needing DHARMA must call a server-side daemon over HTTP |
| All `engram_*` | Stub throws | Could be ported to in-browser (IndexedDB-backed) later |
| All `llm_*` | Stub throws | Browser cannot hold API keys; route through server |
| `llm_register_tool` | Stub throws | Same |
| `http_serve` | Stub throws | Browsers cannot serve. Node-mode could, deferred |
| `http_set_handler` | Stub throws | Same |
| `match` expressions | Compiled (basic) | LitInt/LitStr/LitBool/Wildcard/Binding all work via `if/else` chain. Tagged-union match deferred |
| `type` (struct) defs | Skipped at codegen | Treated as documentation; structs are plain JS objects. `t["field"]` works |
| `enum` defs | Skipped at codegen | Same — enum values are bare strings or ints |
| `?` postfix (nil-prop) | No-op | Same as C backend's current state |
| `try` postfix | Stripped to inner | Same as C backend |
| Capability enforcement | Not enforced | The C backend uses `#error` directives; the JS backend lets the runtime stubs throw. Future: emit `throw new Error('capability violation')` at compile time |
| VBD role check | Not enforced | Same |
| Float bit-cast | Not needed | JS number is already a double |
| Crypto primitives | Stub throws | Easy to add via `crypto.subtle` later |
| `state_*` | In-memory only | No persistence; resets on page reload |
| `args()` | Node-only | Browser returns `[]` |
| `fs_*` | Node-only | Browser throws |

---

## 8. CLI dispatch — `--target=js`

The compiler entry point `compiler.el` adds a `compile_js(source: String) -> String` alongside the existing `compile()`. The CLI behavior:

```
elc <source.el> <output>          # default — emit C
elc --target=c <source.el> <out>  # explicit — emit C
elc --target=js <source.el> <out> # emit JS

elc --target=js source.el         # write JS to stdout (no out path)
```

The argv parser scans for a `--target=<lang>` token; remaining positional args are `<source>` and optional `<out>`. The dispatch logic stays in El: a `compile_dispatch(target, source) -> String` switch.

---

## 9. The path to compiling el-ui/runtime through this backend

This is the real-world test. `el-ui/runtime/src/` is currently 5 hand-written `.js` files. The path to authoring them in El:

1. **Phase 1 — Hello-world** (this scaffold). Done.
2. **Phase 2 — language coverage.** Get codegen-js to ~95% parity with codegen.el for non-network features. Specifically: `match`, struct/enum field access, `?`-propagation, full `for`-over-list, complete unary/binary operators, lexical closures (the C backend doesn't have these but we'll need them for el-ui's component model).
3. **Phase 3 — DOM bridge.** Add `dom_*` builtins to el_runtime.js: `dom_create_element`, `dom_set_text`, `dom_append_child`, `dom_query`, `dom_listen`, etc. These are Node-as-El builtins for the browser; the C backend will add a stub set that errors. Source-shareable El UI code becomes possible.
4. **Phase 4 — Component class lowering.** El doesn't have classes; el-ui's `Component` is a JS class. Decide: extend El with a `component` keyword that compiles to JS class + C struct? Or have el-ui authors define components as `fn render_<name>(state) -> String` and provide a small bootstrap. The latter is the lower-impact path.
5. **Phase 5 — Async taint pass.** Implement compile-time async tracking so `http_get` and friends produce `await fetch()` correctly. Required before authoring code that fetches data.
6. **Phase 6 — Port `el-ui/runtime/`.** Translate the 5 JS files to El, compile to JS, swap in. Run el-ui's existing tests. Iterate.
7. **Phase 7 — Port cgi-studio UI.** Larger surface area; same pattern.
8. **Phase 8 — Marketplace plugins.** Open the door for third-party UI El.

The blocking item between phase 1 and phase 2 is incremental — every El construct used by el-ui's source needs codegen-js coverage. Phase 5 (async) is the architectural decision that needs explicit user buy-in, because it changes the language's effective semantics on the JS target.

---

## 10. Test

```bash
echo 'fn main() -> Void { println("hello from el-js") }' > /tmp/hello.el
elc --target=js /tmp/hello.el > /tmp/hello.js
node /tmp/hello.js
# → hello from el-js
```

This should pass after the bootstrap rebuild. See §11.

---

## 11. Bootstrap status

Adding `--target=js` to `compile()` requires regenerating the shipped `elc` binary at `dist/platform/elc`. The rebuild path is:

1. Existing `elc` binary compiles updated `elc-combined.el` (which now includes `codegen-js.el` and the `--target=js` dispatch) → `elc.c`.
2. `cc` compiles `elc.c` → new `elc` binary.
3. New `elc` binary supports `--target=js`.

The scaffold checks all four scaffold files in. The bootstrap rebuild happens as a follow-up step, gated on review of this design doc.