type User = { name: String } was silently broken: the parser consumed
the type name then called expect(LBrace) while sitting on the = token.
expect() advances unconditionally on mismatch, so it consumed = and
treated { as the first field name, producing a corrupt TypeDef node.
The FnDef following the broken TypeDef was then parsed incorrectly or
lost entirely -- causing greet() and similar functions to vanish from
JS/C output with no error.
Fix: detect and skip the optional Eq token before expecting LBrace.
Both targets benefit; rebuild elc to pick up the fix.
Parser now handles `SomeEnum::Variant` in match arm patterns, emitting
a Variant pattern node with enum_name and variant fields. Previously
these fell through to Binding, producing broken codegen.
JS codegen: emit str_eq check against the variant name string (El enums
are plain strings at runtime). C codegen: same, via EL_STR + str_eq.
Rebuild elc to pick up the parser change.
Replace accumulate-by-concatenation loops with native_list_append + str_join.
Eliminates quadratic memory growth when processing large source files.
This is the v2 compiler state — what produced /tmp/elc-v2.
Replaces the need for product-level denylist sanitizers. Small
state-machine parser; tag-and-attribute allowlist passed as JSON;
URL scheme validation on href/src attrs (http, https, mailto,
fragment, relative); whole-subtree drop for script/style/iframe/
object/embed/form (plus rarer media containers). No comment-
wrapping (was fragile to comment-injection bypass via a literal
--> inside an attacker-supplied attribute value).
Also picks up the codegen and parser changes for first-class
Instant/Duration types (postfix-literal time values, typed binop
dispatch) that were sitting in tree alongside this work.
Test corpus at tests/html_sanitizer/ covers the live attacker
probes (script, iframe, form, javascript:, about:, data:, img
onerror, onclick) plus structural attacks (comment-injection
bypass, tab-in-scheme bypass, encoded payloads, malformed input,
empty input, plain text). 29 cases, all green.
Self-host fixed point holds at 5720 lines via the canonical
el-compiler/src/compiler.el entry. Snapshot tagged at
dist/platform/elc.20260502-1249-self-host.
Backlog: bl-dc55ae07
1. Parser+codegen: bare reassignment `x = expr` inside an if-body
was compiling to three orphan expressions with no store. Now
emits a real assignment.
2. Runtime json_get: dot-path segments that are all digits now
correctly traverse array indices. `json_get(s, "0.field")` works.
3. Runtime HTTP writer: response bodies starting with
`{"__status__":<int>,...}` now set the HTTP status header to
that value and strip the marker from the served body. Existing
404/401/503 paths in product code now produce real status codes
instead of HTTP 200 with the status hidden in the body.
Self-host fixed point holds: gen2 == gen3 byte-identical.
Snapshot tagged at dist/platform/elc.20260502-1231-self-host.
Backlog: bl-c121edda
Capability becomes a compile-time structural property, not a runtime
convention. A program's top-level block determines what runtime
primitives it may call; the codegen rejects forbidden calls with
#error directives so cc fails with a clear message.
Three kinds:
cgi — full self-formation. All primitives.
service — bounded. Cannot call self-formation primitives:
llm_call_agentic, llm_register_tool, dharma_emit,
dharma_field. Single-turn LLM calls allowed.
utility — default (no top-level block). No DHARMA, no LLM.
Pure compute + I/O.
Deep claim: the binary either CAN or CANNOT do a thing. There is no
runtime check, no opt-in, no override. A weather service compiled
with `service { ... }` is structurally incapable of becoming Neuron.
Sponsors of services know exactly what they're vouching for.
Implementation
- Lexer: `service` keyword.
- Parser: parse_service_block parallels parse_cgi_block. Produces
ServiceBlock AST with name/sponsor/domain.
- Codegen entry: scans top-level for cgi/service blocks, sets
__program_kind state ("cgi" / "service" / "utility"). Rejects
programs declaring both kinds.
- cg_expr Call: cap_check_call(fn_name) per emission. Records
violations in __cap_violations CSV. emit_cap_violations() writes
one #error per violation at end of generated C.
- Helpers: is_self_formation_call, is_dharma_call, is_llm_call.
Tests verified:
cgi + llm_call_agentic → compiles ✓
service + llm_call_agentic → cc fails with capability violation
for 'service' on 'llm_call_agentic'
service + llm_call (1-turn) → compiles ✓
utility + dharma_send → cc fails with capability violation
for 'utility' on 'dharma_send'
utility + http/json/state → compiles + runs ✓ ("got: world")
cgi + dharma_emit (manager) → compiles ✓ (VBD also enforced)
cgi + dharma_emit (engine) → cc fails with VBD violation
Three-stage closure: stage1.c == stage2.c (byte-identical).
Engram rebuilt against new compiler — daemon on :8742 healthy,
{"node_count":0,"edge_count":0}.
A bug found and fixed during testing: cap_record_violation had
`csv = ","` (bare assignment, not valid in El) instead of
`let csv = ","`. Without the let, the leading comma never made
it into the accumulator, off-by-one'ing the kind extraction so
"service" appeared as "ervice" in error messages. Pattern
fixed; this confirms once more that El requires `let X = ...`
for all rebindings (codegen converts to assignment when X is
already declared).
Two parallel agent sweeps closing the remaining structural gaps.
== Compiler completions ==
- match codegen: lowers Match into GCC/Clang statement-expression
({ ... }). Patterns: Wildcard, Binding, LitInt (==), LitStr
(str_eq), LitBool. Per-match unique label via state counter.
Verified: classify(0)→"zero", classify(1)→"one", classify(7)→"other".
- cgi block parsing: `cgi "name" { dharma_id, principal, network,
engram }` → CgiBlock AST node → el_cgi_init() emitted as the first
call in main() after el_runtime_init_args. Multiple cgi blocks per
program emit a #error directive. Missing optional fields → EL_NULL.
- VBD compile-time enforcement: parser attaches `decorator: <name>`
to FnDef. Codegen recursively walks fn bodies (Call/BinOp/Not/Neg/
Field/Index/Try/Array/Map/If/For/Match plus Let/Return/Expr/While/
For). If a non-@manager function calls dharma_emit or dharma_field,
emit `#error "VBD violation: ... fn '<name>'"` before the function
body. Verified: @engine fn calling dharma_emit → cc fails with the
message. @manager fn calling dharma_emit → compiles clean.
Three-stage closure: stage1.c == stage2.c == stage3.c (2791 lines
each, byte-identical). dist/platform/elc rebuilt at 165 KB; .prev5
preserved.
== Runtime completions ==
- Real dharma_* primitives, no more stubs. Channel registry,
request/response over HTTP, network-wide spreading activation,
fire-and-forget event emission, blocking dharma_field with
pthread_cond_timedwait (30s default), Hebbian relationship
weights stored as Engram edges between dharma:self and
dharma:peer:<id>, sorted-by-weight peer list. URL/ID arrays
snapshotted before network I/O so mutexes never block on socket.
- New public C contract: el_runtime_dharma_event_arrive(type, payload,
source) — application HTTP handler calls this when /dharma/event
arrives, runtime broadcasts on _dharma_event_cv. Keeps the HTTP
server generic; events flow through the application's router.
- llm_call_agentic real multi-turn loop. Tool registry (mutex-
protected, dlsym-resolved, mirroring http_set_handler). Loop:
build request with tools+messages → POST → dispatch on stop_reason.
end_turn → return text. max_tokens → text + "[truncated]". tool_use
→ walk content[], call registered handler per block, build
tool_result message, append to conversation, loop. Iteration cap
10. Tools not registered return {"error":"tool not registered: X"}
with is_error: true.
- New builtin: llm_register_tool(name, handler_fn_name).
Compile clean: cc -std=c11 -Wall -Wextra -c → zero warnings, zero
errors. Smoke test exercises every new dharma_* primitive +
llm_register_tool round-trip.
Runtime grew 3309→4079 lines (.c, ~155 KB), 312→342 lines (.h).
== Integration ==
Engram rebuilt against the new runtime: 130 KB binary, daemon
swapped on :8742 cleanly, /health and /api/stats both returning
correctly under launchd. No regressions.
== Status of "planned" items in language.md ==
- match codegen → IMPLEMENTED
- cgi block parsing → IMPLEMENTED
- VBD enforcement → IMPLEMENTED
- % operator → IMPLEMENTED (earlier today)
- vessel keyword → lexed (codegen uses package compatible)
- activate construct → still planned (low priority; engram_activate
builtin covers the use case for now)
- sealed block → still planned
- dharma_emit fanout parallelization → potential future work, current
serial behavior matches spec
Closes the known limitation from the self-host commit: `fn add(a:Int,
b:Int) { a + b }` now compiles to integer addition, not string concat.
Previously the codegen heuristic guessed string concat whenever both
operands were Idents with no literal anchor.
Mechanism
- parser captures the leading type identifier from `let x: T = ...`
bindings (new "type" field on Let) and from function parameter
annotations (new "type" field on each param).
- codegen maintains a per-function int-name set in process state via
state_set("__int_names", csv). cg_fn seeds it from typed parameters;
cg_stmt extends it from typed `let` bindings and from `let x = <Int
literal>` (literal inference).
- BinOp Plus: when both sides are Idents and both names are in the
int-name set, emit arithmetic; otherwise the existing literal-anchor
heuristic applies, with string concat as the fallback.
This is the first compiler change made entirely through the self-
hosting workflow — no Python bootstrap. Edit el source, run existing
elc on elc-combined.el, cc the output, test. Closure holds at the
new binary.
Tests
- add(40, 2) → 42
- count_to(10) → 45 (let i: Int / let total: Int rebinding)
- Regression suite (tiny/implret/whiletest/lextest) unchanged.
dist/platform/elc updated; .prev preserved.
Today's milestone: dist/platform/elc compiles itself byte-for-byte to
itself (stage1 == stage2 == stage3 verified). The compiler is now a
real binary in the world.
What landed
- Spec rewrite (language.md) to truth — every feature marked
implemented / planned / not-in-this-language with no fiction.
- C runtime extension: 51 new builtins. JSON parser + accessors,
time, UUID, env, in-process state K/V, float formatting + math,
string ops (index_of, split, char_at, char_code, pad_left/right,
format), list ops (push, push_front, join, range), bool_to_str.
Runtime grew 631 → 1611 lines, header 171 → 247.
- Codegen fix: transform_implicit_return lifts a function's bare
trailing expression into an explicit return. Without it, lex(),
parse(), and every other implicit-return function returned 0/nil
and the whole pipeline produced empty C output.
- Codegen fix: index expressions dispatch on AST kind. obj["literal"]
→ el_get_field (map), arr[i] → el_list_get (list). Same Index node
in the parser, two different runtime calls.
- Codegen fix: skip emitting fn main() (collides with C main()) and
honor parsed return-type annotations so Void functions don't get
return-wrapped (return println(x) is a C type error).
- Parser: capture return-type identifier from -> Ret annotations.
- Lexer: + vessel keyword, + % operator, + \r escape.
- Runtime fix: el_list_append now allocates a fresh list rather than
realloc'ing the input. Realloc moved blocks made caller pointers
dangle, which was inserting garbage values into declared lists and
causing strcmp segfaults. Persistent allocation eliminates the
whole class of use-after-free at modest memory cost.
Bootstrap path
- One-shot Python helper translated elc-combined.el to C and
produced stage1. Helper is disposable; not committed.
- stage1 compiles elc-combined.el → stage2.c which cc compiles to
stage2; stage2 compiles elc-combined.el → stage3.c. stage2.c and
stage3.c are byte-identical. Closure proven.
- New elc installed at dist/platform/elc; old broken binary
preserved as dist/platform/elc.legacy.
- dist/platform/elc.c is the canonical generated source.
- elvm and the bytecode pipeline are no longer on the critical path.
Known gap
- The `+` operator's heuristic dispatch still picks string concat
when both operands are Idents with no literal anchor. Self-hosting
works because the compiler source is careful, but `fn add(a:Int,
b:Int) { a + b }` will not do arithmetic until codegen reads the
parsed type annotations to dispatch. Fix is wiring; not done here.
Tested
- tiny / lextest / whiletest / map+field / array build all run.
- cgi-studio (1037 lines real El) compiles to C cleanly. Link fails
only because runtime is missing fs_list, json_encode, llm_*; those
are scheduled batches.
- Three-stage closure (stage1 vs stage2 vs stage3) byte-identical.
Will Anderson