Sandbox bytecode loading: K.load + load-sxbc, bytecode param, web stack sxbc via loadModule

Bytecode modules now load correctly in sandbox mode. HS .sxbc modules
use K.load('(load-sxbc ...)') which syncs defines to eval env. Web stack
.sxbc modules use K.loadModule with import suspension drive loop.

K.eval used directly for expression eval (not thunk wrapper) so bytecode-
defined symbols are visible. Falls back to callFn thunk on IO suspension.

Sandbox now reproduces the bytecode repeat bug: source gives 6/6
suspensions, bytecode gives 4/6. Bug is in bytecode compilation of
when/do across perform boundaries, not the runtime wrapper.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-08 20:56:54 +00:00
parent c521ff8731
commit 520424954b
2 changed files with 103 additions and 30 deletions

View File

@@ -107,7 +107,9 @@ let register_mcp_jit_hook () =
if !(Sx_vm._jit_compiling) then None
else
(try Some (Sx_vm.call_closure_reuse cl args)
with e ->
with
| Sx_vm.VmSuspended _ as e -> raise e (* let Fix 1 in jit_try_call handle it *)
| e ->
if not (Hashtbl.mem _jit_warned fn_name) then begin
Hashtbl.replace _jit_warned fn_name true;
Printf.eprintf "[mcp-jit] %s runtime fallback to CEK: %s\n%!" fn_name (Printexc.to_string e)
@@ -124,7 +126,9 @@ let register_mcp_jit_hook () =
| Some cl ->
l.l_compiled <- Some cl;
(try Some (Sx_vm.call_closure_reuse cl args)
with e ->
with
| Sx_vm.VmSuspended _ as e -> raise e
| e ->
Printf.eprintf "[mcp-jit] %s first-call fallback: %s\n%!" fn_name (Printexc.to_string e);
Hashtbl.replace _jit_warned fn_name true;
l.l_compiled <- Some Sx_vm.jit_failed_sentinel;
@@ -1572,6 +1576,10 @@ let handle_sx_playwright args =
let island = args |> member "island" |> to_string_option in
let phase = args |> member "phase" |> to_string_option in
let filter = args |> member "filter" |> to_string_option in
let setup = args |> member "setup" |> to_string_option in
let stack = args |> member "stack" |> to_string_option in
let bytecode = try args |> member "bytecode" |> to_bool with _ -> false in
let files_json = try args |> member "files" with _ -> `Null in
(* Determine whether to run specs or the inspector *)
let use_inspector = match mode with
| Some m when m <> "run" -> true
@@ -1631,6 +1639,12 @@ let handle_sx_playwright args =
(match island with Some i -> Some ("island", `String i) | None -> None);
(match phase with Some p -> Some ("phase", `String p) | None -> None);
(match filter with Some f -> Some ("filter", `String f) | None -> None);
(match setup with Some s -> Some ("setup", `String s) | None -> None);
(match stack with Some s -> Some ("stack", `String s) | None -> None);
(if bytecode then Some ("bytecode", `Bool true) else None);
(match files_json with
| `List items -> Some ("files", `List (List.map (fun j -> `String (Yojson.Safe.Util.to_string j)) items))
| _ -> None);
]) in
let args_json = Yojson.Basic.to_string inspector_args in
(* Single-quote shell wrapping — escape any literal single quotes in JSON *)
@@ -1690,10 +1704,26 @@ let handle_sx_harness_eval args =
let session = Sx_ref.cek_call (env_get e "make-harness") mock_arg in
(* Install interceptors *)
ignore (call_sx "install-interceptors" [session; Env e]);
(* IO-aware eval: drives cek_step_loop + cek_resume to handle perform *)
let io_log = ref [] in
let eval_with_io expr =
let state = Sx_ref.make_cek_state expr (Env e) (List []) in
let rec drive st =
let final = Sx_ref.cek_step_loop st in
match Sx_ref.cek_suspended_p final with
| Bool true ->
let request = Sx_runtime.get_val final (String "request") in
io_log := request :: !io_log;
let resumed = Sx_ref.cek_resume final Nil in
drive resumed
| _ -> Sx_ref.cek_value final
in
drive state
in
(* Evaluate the expression *)
let exprs = Sx_parser.parse_all expr_str in
let result = List.fold_left (fun _acc expr ->
try Sx_ref.eval_expr expr (Env e)
try eval_with_io expr
with exn -> String (Printf.sprintf "Error: %s" (Printexc.to_string exn))
) Nil exprs in
(* Get the IO log *)
@@ -1705,7 +1735,13 @@ let handle_sx_harness_eval args =
let args_val = call_sx "get" [entry; String "args"] in
Printf.sprintf " %s(%s)" op (Sx_types.inspect args_val)
) items)
| _ -> "\n\n(no IO calls)"
| _ ->
if !io_log <> [] then
"\n\nIO suspensions (" ^ string_of_int (List.length !io_log) ^ "):\n" ^
String.concat "\n" (List.rev_map (fun req ->
" (perform " ^ Sx_types.inspect req ^ ")"
) !io_log)
else "\n\n(no IO calls)"
in
let warn_str = if !warnings = [] then "" else
"\n\nWarnings:\n" ^ String.concat "\n" (List.rev !warnings)
@@ -2767,14 +2803,18 @@ let tool_definitions = `List [
[];
tool "sx_playwright" "Run Playwright browser tests or inspect SX pages interactively. Modes: run (spec files), inspect (page/island report with leak detection and handler audit), diff (full SSR vs hydrated DOM), hydrate (lake-focused SSR vs hydrated comparison — detects clobbering), eval (JS expression), interact (action sequence), screenshot, listeners (CDP event listener inspection), trace (click + capture console/network/pushState), cdp (raw CDP command), trace-boot (full console capture during boot — ALL prefixes), hydrate-debug (re-run island hydration with full env/state tracing), eval-at (inject eval at a specific boot phase)."
[("spec", `Assoc [("type", `String "string"); ("description", `String "Spec file to run (run mode). e.g. stepper.spec.js")]);
("mode", `Assoc [("type", `String "string"); ("description", `String "Mode: run, inspect, diff, hydrate, eval, interact, screenshot, listeners, trace, cdp, trace-boot, hydrate-debug, eval-at")]);
("mode", `Assoc [("type", `String "string"); ("description", `String "Mode: run, inspect, diff, hydrate, eval, interact, screenshot, listeners, trace, cdp, trace-boot, hydrate-debug, eval-at, sandbox (offline WASM kernel — no server needed)")]);
("phase", `Assoc [("type", `String "string"); ("description", `String "Boot phase for eval-at mode: before-modules, after-modules, before-pages, after-pages, before-components, after-components, before-hydrate, after-hydrate, after-boot")]);
("filter", `Assoc [("type", `String "string"); ("description", `String "Filter prefix for trace-boot mode (e.g. '[sx-platform]')")]);
("url", `Assoc [("type", `String "string"); ("description", `String "URL path to navigate to (default: /)")]);
("island", `Assoc [("type", `String "string"); ("description", `String "Filter inspect to a specific island by name (e.g. home/stepper)")]);
("selector", `Assoc [("type", `String "string"); ("description", `String "CSS selector for screenshot/listeners/trace modes")]);
("expr", `Assoc [("type", `String "string"); ("description", `String "JS expression (eval mode), selector (listeners/trace), or CDP command (cdp mode)")]);
("actions", `Assoc [("type", `String "string"); ("description", `String "Semicolon-separated action sequence (interact mode). Actions: click:sel, fill:sel:val, wait:ms, text:sel, html:sel, attrs:sel, screenshot, screenshot:sel, count:sel, visible:sel")])]
("actions", `Assoc [("type", `String "string"); ("description", `String "Semicolon-separated action sequence (interact mode). Actions: click:sel, fill:sel:val, wait:ms, text:sel, html:sel, attrs:sel, screenshot, screenshot:sel, count:sel, visible:sel")]);
("files", `Assoc [("type", `String "array"); ("items", `Assoc [("type", `String "string")]); ("description", `String ".sx files to load (sandbox mode)")]);
("setup", `Assoc [("type", `String "string"); ("description", `String "SX setup expression (sandbox mode)")]);
("stack", `Assoc [("type", `String "string"); ("description", `String "Module stack for sandbox: core (default), web (full web stack), hs (web + hyperscript), test (web + test framework)")]);
("bytecode", `Assoc [("type", `String "boolean"); ("description", `String "Load .sxbc bytecode instead of .sx source in sandbox (default: false)")])]
[];
]

View File

@@ -1331,7 +1331,7 @@ const SANDBOX_STACKS = {
],
};
async function modeSandbox(page, expr, files, setup, stack) {
async function modeSandbox(page, expr, files, setup, stack, bytecode) {
const fs = require('fs');
const path = require('path');
const PROJECT_ROOT = path.resolve(__dirname, '../..');
@@ -1489,25 +1489,47 @@ async function modeSandbox(page, expr, files, setup, stack) {
if (window.SxKernel.beginModuleLoad) window.SxKernel.beginModuleLoad();
});
for (const mod of modules) {
// Try .sx file in wasm/sx/ dir
const sxPath = path.join(SX_DIR, mod + '.sx');
// Also try lib/hyperscript/ for hs- prefixed modules
const libPath = path.join(PROJECT_ROOT, 'lib/hyperscript', mod.replace(/^hs-/, '') + '.sx');
let src;
try {
src = fs.existsSync(sxPath)
? fs.readFileSync(sxPath, 'utf8')
: fs.readFileSync(libPath, 'utf8');
} catch(e) {
loadErrors.push(mod + ': file not found');
continue;
let loaded = false;
// Try bytecode first if requested
if (bytecode) {
const bcPath = path.join(SX_DIR, mod + '.sxbc');
if (fs.existsSync(bcPath)) {
const bcSrc = fs.readFileSync(bcPath, 'utf8');
const err = await page.evaluate(src => {
try {
const K = window.SxKernel;
// Use K.load with (load-sxbc ...) wrapper — handles IO suspension in CEK
window.__sxbcText = src;
const r = K.load('(load-sxbc (first (parse (host-global "__sxbcText"))))');
delete window.__sxbcText;
if (typeof r === 'string' && r.startsWith('Error')) return r;
return null;
} catch(e) { delete window.__sxbcText; return 'CATCH: ' + e.message; }
}, bcSrc);
if (err) loadErrors.push(mod + '.sxbc: ' + err);
else { loaded = true; loadedModules.push(mod + '.sxbc'); }
}
}
// Fall back to .sx source
if (!loaded) {
const sxPath = path.join(SX_DIR, mod + '.sx');
const libPath = path.join(PROJECT_ROOT, 'lib/hyperscript', mod.replace(/^hs-/, '') + '.sx');
let src;
try {
src = fs.existsSync(sxPath)
? fs.readFileSync(sxPath, 'utf8')
: fs.readFileSync(libPath, 'utf8');
} catch(e) {
loadErrors.push(mod + ': file not found');
continue;
}
const err = await page.evaluate(src => {
try { window.SxKernel.load(src); return null; }
catch(e) { return e.message; }
}, src);
if (err) loadErrors.push(mod + ': ' + err);
else loadedModules.push(mod);
}
const err = await page.evaluate(src => {
try { window.SxKernel.load(src); return null; }
catch(e) { return e.message; }
}, src);
if (err) loadErrors.push(mod + ': ' + err);
else loadedModules.push(mod);
}
await page.evaluate(() => {
if (window.SxKernel.endModuleLoad) window.SxKernel.endModuleLoad();
@@ -1549,11 +1571,22 @@ async function modeSandbox(page, expr, files, setup, stack) {
window._asyncPending = 0;
let syncResult;
try {
K.eval('(define _sandbox_thunk (fn () ' + expr + '))');
syncResult = K.callFn(K.eval('_sandbox_thunk'), []);
// First try K.eval directly (handles most expressions)
syncResult = K.eval(expr);
} catch(e) {
resolve({ result: 'Error: ' + e.message, ioTrace: window._ioTrace });
return;
// If eval fails with IO suspension, retry via callFn thunk
if (e.message && e.message.includes('suspension')) {
try {
K.eval('(define _sandbox_thunk (fn () ' + expr + '))');
syncResult = K.callFn(K.eval('_sandbox_thunk'), []);
} catch(e2) {
resolve({ result: 'Error: ' + e2.message, ioTrace: window._ioTrace });
return;
}
} else {
resolve({ result: 'Error: ' + e.message, ioTrace: window._ioTrace });
return;
}
}
// If the thunk itself suspended (direct perform), drive it inline
@@ -1688,7 +1721,7 @@ async function main() {
result = await modeEvalAt(browser, url, args.phase || 'before-hydrate', args.expr || '(type-of ~cssx/tw)');
break;
case 'sandbox':
result = await modeSandbox(page, args.expr || '"hello"', args.files || [], args.setup || '', args.stack || '');
result = await modeSandbox(page, args.expr || '"hello"', args.files || [], args.setup || '', args.stack || '', !!args.bytecode);
break;
default:
result = { error: `Unknown mode: ${mode}` };