neuron-technologies/el
3
0
Fork 0
Code Issues 1 Pull Requests Actions Packages Projects Releases 5 Wiki Activity

feat(el-runtime): native Windows port of el_runtime.c (+ engram_node_full corruption fix) #55

Merged
will.anderson merged 4 commits from feat/windows-el-runtime into stage 2026-06-20 00:06:09 +00:00
Conversation 1 Commits 4 Files Changed 3
+190 -19
tim.lingo commented 2026-06-15 23:19:26 +00:00
Member
Copy Link
Copy Source

Native Windows port of el_runtime.c plus the engram_node_full corruption fixes the port was based on (2026-06-15).

Commits (not yet in main):

  • Fix engram_node_full wrapper field corruption + add node_type/tier validation
  • allow SessionSummary node_type in validation allowlist
  • native Windows port of el_runtime.c (winsock/dlsym/CreateProcess)
  • promote http_handler typedefs to el_runtime.h (cross-module + Windows)

The first two are engram data-integrity fixes; the last two are the Windows runtime port. Opening for Will to review.

🤖 Generated with Claude Code

Native Windows port of el_runtime.c plus the engram_node_full corruption fixes the port was based on (2026-06-15). Commits (not yet in main): - Fix engram_node_full wrapper field corruption + add node_type/tier validation - allow SessionSummary node_type in validation allowlist - native Windows port of el_runtime.c (winsock/dlsym/CreateProcess) - promote http_handler typedefs to el_runtime.h (cross-module + Windows) The first two are engram data-integrity fixes; the last two are the Windows runtime port. Opening for Will to review. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
tim.lingo changed target branch from main to stage 2026-06-17 18:32:34 +00:00
tim.lingo added 2 commits 2026-06-17 18:32:35 +00:00
feat(el-runtime): native Windows port of el_runtime.c (winsock/dlsym/CreateProcess) 28ef43264a 
Compiles for Windows x64 via mingw-w64 and still compiles clean on POSIX
(darwin/linux) — all Windows code is behind #ifdef _WIN32, POSIX path unchanged.

- el_platform_win.h (new): winsock2 + auto WSAStartup, el_closesocket(),
  dlsym->GetProcAddress, popen/_popen, mkdir/_mkdir, setenv/_putenv_s,
  timegm/_mkgmtime, localtime_r/gmtime_r. Threading unchanged — mingw
  winpthreads supplies <pthread.h> + -lpthread.
- el_runtime.c: include block guarded; 10 socket-close sites -> el_closesocket();
  setsockopt arg4 cast; tm_zone guarded; exec_bg fork/exec -> CreateProcess.

Part of feat/windows-port. Core-el change, for Will's review.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
fix(el-runtime): promote http_handler typedefs to el_runtime.h (cross-module + Windows)
El SDK Release / build-and-release (pull_request) Failing after 13m0s
Details
a36a62ca14 
http_handler_fn / http_handler4_fn were defined only inside el_runtime.c, so soul
modules (routes/chat/...) that reference them via cross-module forward declarations
couldn't see the types — which broke the Windows link of every module. Moving the
public function-pointer types to the shared header is the correct home and unblocks
the build on all platforms (identical typedef, C11-safe redefinition in el_runtime.c).

With this, the soul links into a native Windows neuron.exe (mingw, static) that boots
and serves HTTP on :7770 — verified /health → 200 {"status":"alive",...} in a Win11 VM.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
will.anderson commented 2026-06-19 19:53:53 +00:00
Owner
Copy Link
Copy Source

Review: feat(el-runtime): native Windows port of el_runtime.c (+ engram)

Thanks for the port Tim — the overall structure is solid (consistent _WIN32 guards, correct Winsock include ordering, WSAStartup via __attribute__((constructor)), winpthreads strategy). That said there are 5 blockers across two categories that need to land before this can merge.

Blockers — will not compile or link

1. nanosleep has no Windows shim
nanosleep() is called in four places in el_runtime.c with no shim in el_platform_win.h. On mingw-w64 this is a linker error. Needs a Sleep()-based replacement (e.g. timeBeginPeriod(1) + Sleep(ms) for sub-millisecond precision, or just Sleep(ns/1e6)).

2. Duplicate http_handler_fn / http_handler4_fn typedefs
These typedefs are now defined in both el_runtime.h (newly promoted) and as local typedefs inside el_runtime.c. C11 doesn't allow duplicate typedef declarations of the same type — this is a compile error under -Wpedantic/-Werror on all platforms, not just Windows.

3. el_mem_check declared but definition deleted
The function body was removed in this PR but the declaration remains in el_runtime.h. Any caller gets an undefined symbol at link time on both Windows and Unix.

Blockers — runtime correctness on Windows

4. close(fd) in http_worker / http_worker_v2 (line ~1576, ~1826)
http_worker and http_worker_v2 call close(fd) in their teardown paths. On Windows, close() is the CRT file-descriptor closer — it does nothing to Winsock sockets. Every HTTP request leaks one SOCKET handle. The accept()-loop error paths were correctly migrated to el_closesocket() but the normal teardown paths were not. A long-running server exhausts system handle space and starts rejecting connections.

5. accept() → int cfd truncates 64-bit SOCKET handle (line ~1613, ~1863)
accept() returns SOCKET (UINT_PTR, 64-bit unsigned). Storing that in int cfd silently truncates the upper 32 bits. HttpWorkerArg.fd is also typedef'd as int (line ~1539), so the corrupted handle is passed through to the worker thread. Any socket handle above 0x7FFFFFFF causes WSAENOTSOCK or closes the wrong descriptor. GCC -Wconversion fires on mingw-w64 for this narrowing. Fix: HttpWorkerArg.fd should be SOCKET (or uintptr_t) behind #ifdef _WIN32.

Significant — fix before merge

6. unsetenv("TZ") shim sets TZ="" instead of removing it
_el_apply_zone() calls unsetenv("TZ") + tzset() to restore the OS local timezone. POSIX: TZ removed → tzset reads /etc/localtime → correct local time. Your shim calls _putenv_s("TZ", "") which sets TZ to an empty string, not removes it — _tzset() sees TZ="" and silently falls back to UTC. All El programs doing local-timezone date operations on Windows return UTC. Fix: use _putenv_s("TZ=", "") (the environment deletion form) or check SetEnvironmentVariable("TZ", NULL).

7. IANA timezone names silently broken on Windows CRT
_el_apply_zone() passes IANA names (e.g. America/New_York) to setenv("TZ"). The Windows CRT only understands POSIX timezone specs (e.g. EST5EDT,M3.2.0,M11.1.0). On Windows, IANA names produce no error — the CRT just defaults to UTC. Either require callers to pass POSIX specs, or bundle a minimal IANA→POSIX lookup table.

8. el_closesocket(int s) parameter type
SOCKET is UINT_PTR (64-bit). The int parameter means handles above 0x7FFFFFFF are sign-extended before being passed to closesocket(). Same root issue as finding #5. Parameter should be SOCKET or uintptr_t behind #ifdef _WIN32.

9. EINTR retry after accept() never fires on Windows
The accept() error path checks errno == EINTR. On Windows, Winsock errors come from WSAGetLastError(), not errno. Any transient socket error permanently exits the server loop instead of retrying.

Minor

  • setenv shim ignores overwrite=0 — no current callers use it but the contract is violated
  • WSAStartup return not checked; inited set to 1 even on failure (silently breaks all networking with no diagnostic)
  • No CI changes / no cross-compile verification included — at minimum confirm a local x86_64-w64-mingw32-gcc build succeeds

What's working

Platform detection is clean and consistent, WSAStartup/include ordering is correct, threading via winpthreads needs no changes, exec_bg/CreateProcess path is properly guarded, setsockopt cast to const char* is correct.

Do not merge until the 5 blockers above are resolved.

## Review: `feat(el-runtime): native Windows port of el_runtime.c (+ engram)` Thanks for the port Tim — the overall structure is solid (consistent `_WIN32` guards, correct Winsock include ordering, `WSAStartup` via `__attribute__((constructor))`, winpthreads strategy). That said there are 5 blockers across two categories that need to land before this can merge. --- ### Blockers — will not compile or link **1. `nanosleep` has no Windows shim** `nanosleep()` is called in four places in `el_runtime.c` with no shim in `el_platform_win.h`. On mingw-w64 this is a linker error. Needs a `Sleep()`-based replacement (e.g. `timeBeginPeriod(1)` + `Sleep(ms)` for sub-millisecond precision, or just `Sleep(ns/1e6)`). **2. Duplicate `http_handler_fn` / `http_handler4_fn` typedefs** These typedefs are now defined in both `el_runtime.h` (newly promoted) and as local `typedef`s inside `el_runtime.c`. C11 doesn't allow duplicate typedef declarations of the same type — this is a compile error under `-Wpedantic`/`-Werror` on all platforms, not just Windows. **3. `el_mem_check` declared but definition deleted** The function body was removed in this PR but the declaration remains in `el_runtime.h`. Any caller gets an undefined symbol at link time on both Windows and Unix. --- ### Blockers — runtime correctness on Windows **4. `close(fd)` in `http_worker` / `http_worker_v2` (line ~1576, ~1826)** `http_worker` and `http_worker_v2` call `close(fd)` in their teardown paths. On Windows, `close()` is the CRT file-descriptor closer — it does nothing to Winsock sockets. Every HTTP request leaks one `SOCKET` handle. The `accept()`-loop error paths were correctly migrated to `el_closesocket()` but the normal teardown paths were not. A long-running server exhausts system handle space and starts rejecting connections. **5. `accept()` → `int cfd` truncates 64-bit `SOCKET` handle (line ~1613, ~1863)** `accept()` returns `SOCKET` (`UINT_PTR`, 64-bit unsigned). Storing that in `int cfd` silently truncates the upper 32 bits. `HttpWorkerArg.fd` is also `typedef`'d as `int` (line ~1539), so the corrupted handle is passed through to the worker thread. Any socket handle above `0x7FFFFFFF` causes `WSAENOTSOCK` or closes the wrong descriptor. GCC `-Wconversion` fires on mingw-w64 for this narrowing. Fix: `HttpWorkerArg.fd` should be `SOCKET` (or `uintptr_t`) behind `#ifdef _WIN32`. --- ### Significant — fix before merge **6. `unsetenv("TZ")` shim sets `TZ=""` instead of removing it** `_el_apply_zone()` calls `unsetenv("TZ") + tzset()` to restore the OS local timezone. POSIX: TZ removed → tzset reads `/etc/localtime` → correct local time. Your shim calls `_putenv_s("TZ", "")` which sets TZ to an *empty string*, not removes it — `_tzset()` sees `TZ=""` and silently falls back to UTC. All El programs doing local-timezone date operations on Windows return UTC. Fix: use `_putenv_s("TZ=", "")` (the environment deletion form) or check `SetEnvironmentVariable("TZ", NULL)`. **7. IANA timezone names silently broken on Windows CRT** `_el_apply_zone()` passes IANA names (e.g. `America/New_York`) to `setenv("TZ")`. The Windows CRT only understands POSIX timezone specs (e.g. `EST5EDT,M3.2.0,M11.1.0`). On Windows, IANA names produce no error — the CRT just defaults to UTC. Either require callers to pass POSIX specs, or bundle a minimal IANA→POSIX lookup table. **8. `el_closesocket(int s)` parameter type** `SOCKET` is `UINT_PTR` (64-bit). The `int` parameter means handles above `0x7FFFFFFF` are sign-extended before being passed to `closesocket()`. Same root issue as finding #5. Parameter should be `SOCKET` or `uintptr_t` behind `#ifdef _WIN32`. **9. `EINTR` retry after `accept()` never fires on Windows** The `accept()` error path checks `errno == EINTR`. On Windows, Winsock errors come from `WSAGetLastError()`, not `errno`. Any transient socket error permanently exits the server loop instead of retrying. --- ### Minor - `setenv` shim ignores `overwrite=0` — no current callers use it but the contract is violated - `WSAStartup` return not checked; `inited` set to 1 even on failure (silently breaks all networking with no diagnostic) - No CI changes / no cross-compile verification included — at minimum confirm a local `x86_64-w64-mingw32-gcc` build succeeds --- ### What's working Platform detection is clean and consistent, WSAStartup/include ordering is correct, threading via winpthreads needs no changes, `exec_bg`/`CreateProcess` path is properly guarded, `setsockopt` cast to `const char*` is correct. **Do not merge until the 5 blockers above are resolved.**
will.anderson added 1 commit 2026-06-19 23:59:35 +00:00
fix(windows): resolve PR blockers — nanosleep shim, unsetenv, duplicate typedefs, SOCKET type, el_closesocket
El SDK CI - stage / build-and-test (pull_request) Failing after 22s
Details
c087b97093
will.anderson added 1 commit 2026-06-20 00:06:00 +00:00
Merge branch 'stage' into feat/windows-el-runtime
El SDK CI - stage / build-and-test (pull_request) Failing after 15s
Details
99b113ea9d 
Resolve el_runtime.c conflict: include both sys/resource.h (from stage)
and el_closesocket POSIX shim (from Windows port) within the #else block.
will.anderson merged commit 2d751890ea into stage 2026-06-20 00:06:09 +00:00
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
No items
No labels
Milestone
No items
No Milestone
Projects
Clear projects
No projects
Assignees
Clear assignees
andre.almeida (Andre Almeida) tim.lingo (Tim Lingo) will.anderson (Will Anderson)
No Assignees
2 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: neuron-technologies/el#55
Delete Branch "feat/windows-el-runtime"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?