Skip to main content

harn_vm/vm/
state.rs

1use std::collections::{BTreeMap, HashMap, HashSet};
2use std::path::PathBuf;
3use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
4use std::sync::{Arc, Mutex};
5use std::time::Instant;
6
7use crate::chunk::{Chunk, ChunkRef, Constant};
8use crate::runtime_limits::RuntimeLimits;
9use crate::value::{
10    ModuleFunctionRegistry, VmAsyncBuiltinFn, VmBuiltinFn, VmClosure, VmEnv, VmError, VmMutex,
11    VmTaskHandle, VmValue,
12};
13use crate::BuiltinId;
14
15use super::debug::DebugHook;
16use super::modules::ModuleCache;
17use super::VmBuiltinMetadata;
18
19/// A lazy callable's resolved export set together with the module graph that
20/// was loaded to produce it.
21///
22/// The exported closures — and every function they can transitively reach
23/// through imports — hold only `Weak`s into their home module's function
24/// registry and module state. The child VM that first loaded the graph is
25/// normally their sole strong owner (via its `module_cache`), and it dies once
26/// the hook fire completes. A later child VM that hits this cache never
27/// re-imports the graph, so a transitively imported callee's sibling `pub fn`
28/// would fall through name resolution to host-bridge dispatch (`Undefined
29/// builtin`). Retaining the complete loaded graph here for the cache entry's
30/// lifetime keeps every transitively reachable registry and module state
31/// upgradeable at call time, for every fire — not just the first.
32pub(crate) struct ResolvedLazyCallable {
33    pub(crate) exports: BTreeMap<String, Arc<VmClosure>>,
34    /// Intentionally unread: retained solely to keep the loaded module graph's
35    /// function registries and module states alive for this cache entry (the
36    /// same liveness role [`crate::value::RetainedModuleScope`] plays for a
37    /// single retained closure, generalized across the whole import graph).
38    #[allow(dead_code)]
39    pub(crate) retained_module_graph: ModuleCache,
40}
41
42pub(crate) type LazyCallableResolution = Arc<ResolvedLazyCallable>;
43pub(crate) struct LazyCallableCacheSlot {
44    pub(crate) execution_guard: Option<Arc<harn_modules::package_execution::PackageExecutionGuard>>,
45    pub(crate) resolution: Arc<tokio::sync::OnceCell<LazyCallableResolution>>,
46}
47pub(crate) type LazyCallableModuleCache =
48    Arc<VmMutex<BTreeMap<PathBuf, Vec<LazyCallableCacheSlot>>>>;
49
50/// RAII guard that starts a tracing span on creation and ends it on drop.
51pub(crate) struct ScopeSpan(u64);
52
53impl ScopeSpan {
54    pub(crate) fn new(kind: crate::tracing::SpanKind, name: String) -> Self {
55        Self(crate::tracing::span_start(kind, name))
56    }
57}
58
59/// Cancellation-safe host deadline state. The guard owns the shared state so
60/// dropping an in-flight `execute_with_timeout` future restores the previous
61/// deadline and poisons the abandoned VM even though the future still holds
62/// `&mut Vm`.
63pub(crate) struct ExecutionDeadlineState {
64    origin: Instant,
65    /// Nanoseconds from `origin`, plus one so zero remains the inactive value.
66    deadline_offset: AtomicU64,
67    /// Set when the host drops a polled execution future. Arbitrary async
68    /// cancellation cannot unwind interpreter state, so subsequent execution
69    /// entries fail loudly instead of resuming a partial frame.
70    abandoned: AtomicBool,
71}
72
73impl ExecutionDeadlineState {
74    pub(crate) fn new(origin: Instant, deadline: Option<Instant>) -> Arc<Self> {
75        Arc::new(Self {
76            origin,
77            deadline_offset: AtomicU64::new(Self::encode(origin, deadline)),
78            abandoned: AtomicBool::new(false),
79        })
80    }
81
82    #[inline]
83    pub(crate) fn is_active(&self) -> bool {
84        self.deadline_offset.load(Ordering::Acquire) != 0
85    }
86
87    #[inline]
88    pub(crate) fn is_abandoned(&self) -> bool {
89        self.abandoned.load(Ordering::Acquire)
90    }
91
92    pub(crate) fn fork(&self) -> Arc<Self> {
93        let state = Self::new(self.origin, self.current());
94        state
95            .abandoned
96            .store(self.is_abandoned(), Ordering::Release);
97        state
98    }
99
100    pub(crate) fn current(&self) -> Option<Instant> {
101        let encoded = self.deadline_offset.load(Ordering::Acquire);
102        (encoded != 0)
103            .then(|| self.origin + std::time::Duration::from_nanos(encoded.saturating_sub(1)))
104    }
105
106    pub(crate) fn install(self: &Arc<Self>, deadline: Instant) -> ExecutionDeadlineGuard {
107        let previous = self.deadline_offset.load(Ordering::Acquire);
108        let requested = Self::encode(self.origin, Some(deadline));
109        let active = if previous == 0 {
110            requested
111        } else {
112            previous.min(requested)
113        };
114        self.deadline_offset.store(active, Ordering::Release);
115        ExecutionDeadlineGuard {
116            state: Arc::clone(self),
117            previous,
118            completed: false,
119        }
120    }
121
122    fn encode(origin: Instant, deadline: Option<Instant>) -> u64 {
123        deadline.map_or(0, |deadline| {
124            let nanos = deadline.saturating_duration_since(origin).as_nanos();
125            u64::try_from(nanos)
126                .unwrap_or(u64::MAX - 1)
127                .saturating_add(1)
128        })
129    }
130}
131
132pub(crate) struct ExecutionDeadlineGuard {
133    state: Arc<ExecutionDeadlineState>,
134    previous: u64,
135    completed: bool,
136}
137
138impl ExecutionDeadlineGuard {
139    /// Mark an awaited execution as terminal before restoring its prior host
140    /// deadline. Dropping without this acknowledgement poisons the VM.
141    pub(crate) fn complete(mut self) {
142        self.completed = true;
143    }
144}
145
146impl Drop for ExecutionDeadlineGuard {
147    fn drop(&mut self) {
148        self.state
149            .deadline_offset
150            .store(self.previous, Ordering::Release);
151        if !self.completed {
152            self.state.abandoned.store(true, Ordering::Release);
153        }
154    }
155}
156
157impl Drop for ScopeSpan {
158    fn drop(&mut self) {
159        crate::tracing::span_end(self.0);
160    }
161}
162
163#[derive(Clone)]
164pub(crate) struct LocalSlot {
165    pub(crate) value: VmValue,
166    pub(crate) initialized: bool,
167    pub(crate) synced: bool,
168}
169
170impl Drop for LocalSlot {
171    fn drop(&mut self) {
172        // Slot locals hold script values directly (e.g. a `let` bound to a
173        // deeply nested list). When a frame is torn down, the default
174        // recursive drop of such a value would overflow the native stack and
175        // abort the process. For the overwhelmingly common scalar slot this is
176        // a single `matches!` check and then the normal trivial drop; only a
177        // nested container is moved out and torn down iteratively, so hot
178        // frame teardown is unaffected.
179        if crate::value::recursion::is_recursive_container(&self.value) {
180            crate::value::recursion::dismantle(std::mem::replace(&mut self.value, VmValue::Nil));
181        }
182    }
183}
184
185#[derive(Clone)]
186pub(crate) struct InterruptHandler {
187    pub(crate) handle: i64,
188    pub(crate) signals: Vec<String>,
189    pub(crate) once: bool,
190    pub(crate) graceful_timeout_ms: Option<u64>,
191    pub(crate) handler: VmValue,
192}
193
194/// Call frame for function execution.
195pub(crate) struct CallFrame {
196    pub(crate) chunk: ChunkRef,
197    /// VM-local inline-cache set for this frame's chunk. Computed once at
198    /// frame entry so hot opcode dispatch can index cache feedback directly
199    /// instead of hashing the chunk id on every cached opcode.
200    pub(crate) inline_cache_set: usize,
201    pub(crate) ip: usize,
202    pub(crate) stack_base: usize,
203    pub(crate) saved_env: VmEnv,
204    /// Env snapshot captured at call-time, *after* argument binding. Used
205    /// by the debugger's `restartFrame` to rewind this frame to its
206    /// entry state (re-binding args from the original values) without
207    /// re-entering the call site. Cheap to clone because `VmEnv` is
208    /// already cloned into `saved_env` on every call. `None` for
209    /// scratch frames (evaluate, import init) where restart isn't
210    /// meaningful.
211    pub(crate) initial_env: Option<VmEnv>,
212    pub(crate) initial_local_slots: Option<Vec<LocalSlot>>,
213    /// Iterator stack depth to restore when this frame unwinds.
214    pub(crate) saved_iterator_depth: usize,
215    /// Function name for stack traces (empty for top-level pipeline).
216    pub(crate) fn_name: String,
217    /// Number of arguments actually passed by the caller (for default arg support).
218    pub(crate) argc: usize,
219    /// Saved VM_SOURCE_DIR to restore when this frame is popped.
220    /// Set when entering a closure that originated from an imported module.
221    pub(crate) saved_source_dir: Option<std::path::PathBuf>,
222    /// Module-local named functions available to symbolic calls within this frame.
223    pub(crate) module_functions: Option<ModuleFunctionRegistry>,
224    /// Shared module-level env for top-level `let` / `const` bindings of
225    /// this frame's originating module. Looked up after `self.env` and
226    /// before `self.globals` by `GetVar` / `SetVar`, giving each module
227    /// its own live static state that persists across calls. See the
228    /// `module_state` field on `VmClosure` for the full rationale.
229    pub(crate) module_state: Option<crate::value::ModuleState>,
230    /// Slot-indexed locals for compiler-resolved names in this frame.
231    pub(crate) local_slots: Vec<LocalSlot>,
232    /// Env scope index that corresponds to compiler local scope depth 0.
233    pub(crate) local_scope_base: usize,
234    /// Current compiler local scope depth, updated by PushScope/PopScope.
235    pub(crate) local_scope_depth: usize,
236}
237
238pub(crate) struct InlineCacheSite {
239    pub(crate) cache_set: usize,
240    pub(crate) slot_count: usize,
241    pub(crate) slot: Option<usize>,
242}
243
244impl CallFrame {
245    #[inline]
246    pub(crate) fn inline_cache_site_for_previous_op(&self) -> InlineCacheSite {
247        let op_offset = self.ip.saturating_sub(1);
248        InlineCacheSite {
249            cache_set: self.inline_cache_set,
250            slot_count: self.chunk.inline_cache_slot_count(),
251            slot: self.chunk.inline_cache_slot(op_offset),
252        }
253    }
254}
255
256/// Exception handler for try/catch.
257pub(crate) struct ExceptionHandler {
258    pub(crate) catch_ip: usize,
259    pub(crate) stack_depth: usize,
260    pub(crate) frame_depth: usize,
261    pub(crate) env_scope_depth: usize,
262    /// When present, this catch only handles errors whose enum_name matches.
263    pub(crate) error_type: Option<crate::value::HarnStr>,
264}
265
266/// A structured-concurrency nursery (`scope { }`). Tasks spawned while this
267/// scope is innermost record their id here; `TaskScopeExit` joins them.
268pub(crate) struct TaskScope {
269    /// Ids of tasks spawned in this scope that have not been explicitly
270    /// `await`ed away. Joined (normal exit) or cancelled (unwind) on close.
271    pub(crate) task_ids: Vec<String>,
272    /// Frame depth at which the scope was opened, for unwind pruning.
273    pub(crate) frame_depth: usize,
274    /// Env scope depth at open, for unwind pruning.
275    pub(crate) env_scope_depth: usize,
276}
277
278/// Terminal exit requested by any VM in one execution tree. A process exit is
279/// global control flow, so child VMs share this latch with their parent rather
280/// than relying on a task being explicitly awaited.
281pub(crate) struct ProcessExitRequest {
282    code: Mutex<Option<i32>>,
283    requested: AtomicBool,
284}
285
286impl ProcessExitRequest {
287    fn new() -> Self {
288        Self {
289            code: Mutex::new(None),
290            requested: AtomicBool::new(false),
291        }
292    }
293
294    fn request(&self, code: i32) {
295        let mut recorded = self
296            .code
297            .lock()
298            .expect("process exit request lock poisoned");
299        if recorded.is_none() {
300            *recorded = Some(code);
301            self.requested.store(true, Ordering::Release);
302        }
303    }
304
305    fn code(&self) -> Option<i32> {
306        if !self.requested.load(Ordering::Acquire) {
307            return None;
308        }
309        *self
310            .code
311            .lock()
312            .expect("process exit request lock poisoned")
313    }
314}
315
316/// Iterator state for for-in loops.
317pub(crate) enum IterState {
318    Vec {
319        items: Arc<Vec<VmValue>>,
320        idx: usize,
321    },
322    Dict {
323        entries: Arc<crate::value::DictMap>,
324        keys: Vec<String>,
325        idx: usize,
326    },
327    Channel {
328        receiver: std::sync::Arc<tokio::sync::Mutex<tokio::sync::mpsc::Receiver<VmValue>>>,
329        close: std::sync::Arc<crate::value::VmChannelCloseState>,
330    },
331    Generator {
332        gen: Arc<crate::value::VmGenerator>,
333    },
334    Stream {
335        stream: Arc<crate::value::VmStream>,
336    },
337    /// Step through a lazy range without materializing a Vec.
338    /// Inclusive ranges keep `end` as an actual value so `i64::MAX to i64::MAX`
339    /// still yields one item instead of overflowing a one-past-end sentinel.
340    Range {
341        next: i64,
342        end: i64,
343        inclusive: bool,
344        done: bool,
345    },
346    VmIter {
347        handle: crate::vm::iter::VmIterHandle,
348    },
349}
350
351#[derive(Clone)]
352pub(crate) enum VmBuiltinDispatch {
353    Sync(VmBuiltinFn),
354    Async(VmAsyncBuiltinFn),
355}
356
357#[derive(Clone)]
358pub(crate) struct VmBuiltinEntry {
359    pub(crate) name: Arc<str>,
360    pub(crate) dispatch: VmBuiltinDispatch,
361}
362
363/// The Harn bytecode virtual machine.
364pub struct Vm {
365    pub(crate) stack: Vec<VmValue>,
366    pub(crate) env: VmEnv,
367    pub(crate) output: String,
368    pub(crate) builtins: Arc<BTreeMap<String, VmBuiltinFn>>,
369    pub(crate) async_builtins: Arc<BTreeMap<String, VmAsyncBuiltinFn>>,
370    pub(crate) builtin_metadata: Arc<BTreeMap<String, VmBuiltinMetadata>>,
371    /// Numeric side index for builtins. Name-keyed maps remain authoritative;
372    /// this index is the hot path for direct builtin bytecode and callback refs.
373    pub(crate) builtins_by_id: Arc<HashMap<BuiltinId, VmBuiltinEntry>>,
374    /// IDs with detected name collisions. Collided names safely fall back to
375    /// the authoritative name-keyed lookup path.
376    pub(crate) builtin_id_collisions: Arc<HashSet<BuiltinId>>,
377    /// Iterator state for for-in loops.
378    pub(crate) iterators: Vec<IterState>,
379    /// Call frame stack.
380    pub(crate) frames: Vec<CallFrame>,
381    /// Exception handler stack.
382    pub(crate) exception_handlers: Vec<ExceptionHandler>,
383    /// Spawned async task handles.
384    pub(crate) spawned_tasks: BTreeMap<String, VmTaskHandle>,
385    /// Shared terminal process-exit latch for this execution tree.
386    pub(crate) process_exit_request: Arc<ProcessExitRequest>,
387    /// Shared process-local synchronization primitives inherited by child VMs.
388    pub(crate) sync_runtime: Arc<crate::synchronization::VmSyncRuntime>,
389    /// Shared process-local cells, maps, and mailboxes inherited by child VMs.
390    pub(crate) shared_state_runtime: Arc<crate::shared_state::VmSharedStateRuntime>,
391    /// Per-isolate inline cache entries. `inline_cache_set_by_chunk` maps a
392    /// compiled chunk identity to an index in this vector at frame entry; the
393    /// dispatch loop uses the frame-local index for per-op reads/writes.
394    pub(crate) inline_cache_sets: Vec<Vec<crate::chunk::InlineCacheEntry>>,
395    pub(crate) inline_cache_set_by_chunk: HashMap<u64, usize>,
396    /// VM-scoped pool registry inherited by child VMs and scoped into Tokio tasks.
397    pub(crate) pool_registry: Arc<crate::stdlib::pool::PoolRegistry>,
398    /// Inline LLM fixtures and observations shared by this VM execution tree.
399    pub(crate) llm_mock_context: crate::llm::mock::LlmMockContext,
400    /// Reader leases opened by package_snapshot_open in this execution tree.
401    /// Child VMs share the registry; the final VM drop releases abandoned
402    /// leases without touching concurrent executions.
403    pub(crate) package_snapshot_registry: Arc<crate::stdlib::PackageSnapshotRegistry>,
404    /// Shared task/channel wait graph for this VM execution tree.
405    pub(crate) wait_for_graph: Arc<crate::wait_for_graph::VmWaitForGraph>,
406    /// Permits acquired by lexical synchronization blocks in this VM.
407    pub(crate) held_sync_guards: Vec<crate::synchronization::VmSyncHeldGuard>,
408    /// Locks held by an ancestor VM that is *suspended on this VM's execution*:
409    /// an inline async-builtin child runs while its parent is parked
410    /// mid-instruction still holding these permits. Re-acquiring more permits
411    /// than the primitive can grant is a provably-unresolvable self-deadlock, so
412    /// HARN-ORC-011 fires across the child boundary. Empty for new concurrent
413    /// tasks (`spawn`/`parallel`/triggers), where the parent keeps running and
414    /// blocking can be legitimately resolvable.
415    pub(crate) inherited_held_keys: Arc<Vec<crate::synchronization::VmSyncHeldKey>>,
416    /// Structured-concurrency nursery stack. Each `scope { }` block pushes a
417    /// `TaskScope`; tasks spawned while it is innermost register their id here.
418    /// On normal exit (`TaskScopeExit`) the scope's tasks are joined and the
419    /// first error propagates; on unwind they are cancelled. Modeled on
420    /// `held_sync_guards` (push on enter, prune/cancel on frame/handler exit).
421    pub(crate) task_scopes: Vec<TaskScope>,
422    /// Counter for generating unique task IDs.
423    pub(crate) task_counter: u64,
424    /// Counter for logical runtime-context task groups.
425    pub(crate) runtime_context_counter: u64,
426    /// Logical runtime task context visible through `runtime_context()`.
427    pub(crate) runtime_context: crate::runtime_context::RuntimeContext,
428    /// Active deadline stack: (deadline_instant, frame_depth).
429    pub(crate) deadlines: Vec<(Instant, usize)>,
430    /// Uncatchable wall-clock deadline imposed by the embedding host.
431    pub(crate) execution_deadline: Arc<ExecutionDeadlineState>,
432    /// Breakpoints, keyed by source-file path so a breakpoint at line N
433    /// in `auto.harn` doesn't also fire when execution hits line N in an
434    /// imported lib. The empty-string key is a wildcard used by callers
435    /// that don't track source paths (legacy `set_breakpoints` API).
436    pub(crate) breakpoints: BTreeMap<String, std::collections::BTreeSet<usize>>,
437    /// Function-name breakpoints. Any closure call whose
438    /// `CompiledFunction.name` matches an entry here raises a stop on
439    /// entry, regardless of the call site's file or line. Lets the IDE
440    /// break on `llm_call` / `host_run_pipeline` / any user pipeline
441    /// function without pinning down a source location first.
442    pub(crate) function_breakpoints: std::collections::BTreeSet<String>,
443    /// Latched on `push_closure_frame` when the callee's name matches
444    /// `function_breakpoints`; consumed by the next step so the stop is
445    /// reported with reason="function breakpoint" and the breakpoint
446    /// name available for the DAP `stopped` event.
447    pub(crate) pending_function_bp: Option<String>,
448    /// Whether the VM is in step mode.
449    pub(crate) step_mode: bool,
450    /// The frame depth at which stepping started (for step-over).
451    pub(crate) step_frame_depth: usize,
452    /// Whether the VM is currently stopped at a debug point.
453    pub(crate) stopped: bool,
454    /// Last source line executed (to detect line changes).
455    pub(crate) last_line: usize,
456    /// Source directory for resolving imports.
457    pub(crate) source_dir: Option<std::path::PathBuf>,
458    /// Installed-package identity retained for the active lazy pipeline.
459    pub(crate) package_execution_guard:
460        Option<Arc<harn_modules::package_execution::PackageExecutionGuard>>,
461    /// Modules currently being imported (cycle prevention).
462    pub(crate) imported_paths: Vec<std::path::PathBuf>,
463    /// Imports that hit an in-progress module (an import cycle) and so could
464    /// not bind inline. Drained by `flush_deferred_cyclic_imports` once the
465    /// involved modules finish loading.
466    pub(crate) deferred_cyclic_imports: Vec<super::modules::DeferredCyclicImport>,
467    /// Loaded module cache keyed by canonical or synthetic module path.
468    pub(crate) module_cache: ModuleCache,
469    /// Immutable hydrated module bytecode shared across fresh VM isolates.
470    /// Runtime closures, registries, state, and init execution are not cached.
471    pub(crate) prepared_module_cache: crate::PreparedModuleCache,
472    /// Optional timing recorder shared by this VM execution tree.
473    pub(crate) module_phase_recorder: Option<super::ModulePhaseRecorder>,
474    /// Lazy manifest modules initialized by any child in this execution tree.
475    /// The complete export set is retained so handlers and predicates from the
476    /// same module share one module state across repeated child invocations.
477    pub(crate) lazy_callable_modules: LazyCallableModuleCache,
478    /// Source text keyed by canonical or synthetic module path for debugger
479    /// retrieval. Entries share the bytes owned by
480    /// [`crate::module_source`] rather than copying them.
481    pub(crate) source_cache: Arc<BTreeMap<std::path::PathBuf, Arc<str>>>,
482    /// Source file path for error reporting.
483    pub(crate) source_file: Option<String>,
484    /// Source text for error reporting.
485    pub(crate) source_text: Option<String>,
486    /// Line-coverage accumulator. `Some` only while a coverage session is
487    /// active (see [`crate::coverage`]); folded into the global report on drop.
488    pub(crate) coverage: Option<crate::coverage::Coverage>,
489    /// Optional bridge for delegating unknown builtins in bridge mode.
490    pub(crate) bridge: Option<Arc<crate::bridge::HostBridge>>,
491    /// Builtins denied by sandbox mode (`--deny` / `--allow` flags).
492    pub(crate) denied_builtins: Arc<HashSet<String>>,
493    /// Cancellation token for cooperative graceful shutdown (set by parent).
494    pub(crate) cancel_token: Option<std::sync::Arc<std::sync::atomic::AtomicBool>>,
495    pub(crate) interrupt_signal_token: Option<std::sync::Arc<std::sync::Mutex<Option<String>>>>,
496    /// Remaining instruction-boundary checks before a requested host
497    /// cancellation is forcefully raised. This gives `is_cancelled()` loops a
498    /// deterministic chance to return cleanly without letting non-cooperative
499    /// CPU-bound code run forever.
500    pub(crate) cancel_grace_instructions_remaining: Option<usize>,
501    /// User-visible interrupt handlers registered through `std/signal`.
502    pub(crate) interrupt_handlers: Vec<InterruptHandler>,
503    pub(crate) next_interrupt_handle: i64,
504    pub(crate) pending_interrupt_signal: Option<String>,
505    pub(crate) interrupted: bool,
506    pub(crate) dispatching_interrupt: bool,
507    pub(crate) interrupt_handler_deadline: Option<Instant>,
508    /// Captured stack trace from the most recent error (fn_name, line, col).
509    pub(crate) error_stack_trace: Vec<(String, usize, usize, Option<String>)>,
510    /// Yield channel sender for generator execution. When set, `Op::Yield`
511    /// sends values through this channel instead of being a no-op.
512    pub(crate) yield_sender: Option<tokio::sync::mpsc::Sender<Result<VmValue, VmError>>>,
513    /// Project root directory (detected via harn.toml).
514    /// Used as base directory for metadata, store, and checkpoint operations.
515    pub(crate) project_root: Option<std::path::PathBuf>,
516    /// Global constants (e.g. `pi`, `e`). Checked as a fallback in `GetVar`
517    /// after the environment, so user-defined variables can shadow them.
518    pub(crate) globals: Arc<crate::value::DictMap>,
519    /// Optional debugger hook invoked when execution advances to a new source line.
520    pub(crate) debug_hook: Option<parking_lot::Mutex<Box<DebugHook>>>,
521    /// Effective runtime ceilings for this VM execution.
522    pub(crate) runtime_limits: RuntimeLimits,
523}
524
525/// Reusable VM baseline for hosts that need many clean executions with the
526/// same stable builtin/source setup.
527///
528/// The baseline intentionally does not snapshot execution state. Each
529/// instantiation gets fresh stacks, frames, tasks, cancellation fields, sync
530/// primitives, shared cells/maps/mailboxes, and debug state. Builtin tables are
531/// shared through `Arc` until a per-execution rebind needs copy-on-write.
532#[derive(Clone)]
533pub struct VmBaseline {
534    builtins: Arc<BTreeMap<String, VmBuiltinFn>>,
535    async_builtins: Arc<BTreeMap<String, VmAsyncBuiltinFn>>,
536    builtin_metadata: Arc<BTreeMap<String, VmBuiltinMetadata>>,
537    builtins_by_id: Arc<HashMap<BuiltinId, VmBuiltinEntry>>,
538    builtin_id_collisions: Arc<HashSet<BuiltinId>>,
539    source_dir: Option<std::path::PathBuf>,
540    source_file: Option<String>,
541    source_text: Option<String>,
542    project_root: Option<std::path::PathBuf>,
543    globals: Arc<crate::value::DictMap>,
544    denied_builtins: Arc<HashSet<String>>,
545    prepared_module_cache: crate::PreparedModuleCache,
546    runtime_limits: RuntimeLimits,
547}
548
549impl VmBaseline {
550    pub fn from_vm(vm: &Vm) -> Self {
551        Self {
552            builtins: Arc::clone(&vm.builtins),
553            async_builtins: Arc::clone(&vm.async_builtins),
554            builtin_metadata: Arc::clone(&vm.builtin_metadata),
555            builtins_by_id: Arc::clone(&vm.builtins_by_id),
556            builtin_id_collisions: Arc::clone(&vm.builtin_id_collisions),
557            source_dir: vm.source_dir.clone(),
558            source_file: vm.source_file.clone(),
559            source_text: vm.source_text.clone(),
560            project_root: vm.project_root.clone(),
561            globals: Arc::clone(&vm.globals),
562            denied_builtins: Arc::clone(&vm.denied_builtins),
563            prepared_module_cache: vm.prepared_module_cache.clone(),
564            runtime_limits: vm.runtime_limits,
565        }
566    }
567
568    pub fn instantiate(&self) -> Vm {
569        crate::initialize_runtime_assets();
570        let mut source_cache = BTreeMap::new();
571        if let (Some(file), Some(text)) = (&self.source_file, &self.source_text) {
572            source_cache.insert(std::path::PathBuf::from(file), Arc::from(text.as_str()));
573        }
574        if let Some(dir) = &self.source_dir {
575            crate::stdlib::set_thread_source_dir(dir);
576        }
577
578        let mut vm = Vm {
579            stack: Vec::with_capacity(256),
580            env: VmEnv::new(),
581            output: String::new(),
582            builtins: Arc::clone(&self.builtins),
583            async_builtins: Arc::clone(&self.async_builtins),
584            builtin_metadata: Arc::clone(&self.builtin_metadata),
585            builtins_by_id: Arc::clone(&self.builtins_by_id),
586            builtin_id_collisions: Arc::clone(&self.builtin_id_collisions),
587            iterators: Vec::new(),
588            frames: Vec::new(),
589            exception_handlers: Vec::new(),
590            spawned_tasks: BTreeMap::new(),
591            process_exit_request: Arc::new(ProcessExitRequest::new()),
592            sync_runtime: Arc::new(crate::synchronization::VmSyncRuntime::new()),
593            shared_state_runtime: Arc::new(crate::shared_state::VmSharedStateRuntime::new()),
594            inline_cache_sets: Vec::new(),
595            inline_cache_set_by_chunk: HashMap::new(),
596            pool_registry: crate::stdlib::pool::new_pool_registry(),
597            llm_mock_context: crate::llm::mock::LlmMockContext::for_new_vm(),
598            package_snapshot_registry: Arc::new(Default::default()),
599            wait_for_graph: Arc::new(crate::wait_for_graph::VmWaitForGraph::new()),
600            held_sync_guards: Vec::new(),
601            inherited_held_keys: Arc::new(Vec::new()),
602            task_scopes: Vec::new(),
603            task_counter: 0,
604            runtime_context_counter: 0,
605            runtime_context: crate::runtime_context::RuntimeContext::root(),
606            deadlines: Vec::new(),
607            execution_deadline: super::execution::new_execution_deadline_state(None),
608            breakpoints: BTreeMap::new(),
609            function_breakpoints: std::collections::BTreeSet::new(),
610            pending_function_bp: None,
611            step_mode: false,
612            step_frame_depth: 0,
613            stopped: false,
614            last_line: 0,
615            source_dir: self.source_dir.clone(),
616            package_execution_guard: None,
617            imported_paths: Vec::new(),
618            deferred_cyclic_imports: Vec::new(),
619            module_cache: Arc::new(BTreeMap::new()),
620            prepared_module_cache: self.prepared_module_cache.clone(),
621            module_phase_recorder: None,
622            lazy_callable_modules: Arc::new(crate::value::VmMutex::new(BTreeMap::new())),
623            source_cache: Arc::new(source_cache),
624            source_file: self.source_file.clone(),
625            source_text: self.source_text.clone(),
626            coverage: crate::coverage::for_primary(self.source_file.as_deref()),
627            bridge: None,
628            denied_builtins: Arc::clone(&self.denied_builtins),
629            cancel_token: None,
630            interrupt_signal_token: None,
631            cancel_grace_instructions_remaining: None,
632            interrupt_handlers: Vec::new(),
633            next_interrupt_handle: 1,
634            pending_interrupt_signal: None,
635            interrupted: false,
636            dispatching_interrupt: false,
637            interrupt_handler_deadline: None,
638            error_stack_trace: Vec::new(),
639            yield_sender: None,
640            project_root: self.project_root.clone(),
641            globals: Arc::clone(&self.globals),
642            debug_hook: None,
643            runtime_limits: self.runtime_limits,
644        };
645
646        crate::stdlib::rebind_execution_state_builtins(&mut vm);
647        vm
648    }
649}
650
651impl Vm {
652    pub(crate) fn ensure_execution_available(&self) -> Result<(), VmError> {
653        if self.execution_deadline.is_abandoned() {
654            return Err(VmError::AbandonedExecution);
655        }
656        Ok(())
657    }
658
659    pub(crate) fn fresh_local_slots(chunk: &Chunk) -> Vec<LocalSlot> {
660        chunk
661            .local_slots
662            .iter()
663            .map(|_| LocalSlot {
664                value: VmValue::Nil,
665                initialized: false,
666                synced: false,
667            })
668            .collect()
669    }
670
671    pub(crate) fn bind_param_slots(
672        slots: &mut [LocalSlot],
673        func: &crate::chunk::CompiledFunction,
674        args: &[VmValue],
675        synced: bool,
676    ) {
677        Self::bind_param_slots_args(slots, func, &super::CallArgs::Slice(args), synced);
678    }
679
680    pub(crate) fn bind_param_slots_args(
681        slots: &mut [LocalSlot],
682        func: &crate::chunk::CompiledFunction,
683        args: &super::CallArgs<'_>,
684        synced: bool,
685    ) {
686        let param_count = func.params.len();
687        for (i, _param) in func.params.iter().enumerate() {
688            if i >= slots.len() {
689                break;
690            }
691            if func.has_rest_param && i == param_count - 1 {
692                let rest_args = args.to_vec_from(i);
693                slots[i].value = VmValue::List(std::sync::Arc::new(rest_args));
694                slots[i].initialized = true;
695                slots[i].synced = synced;
696            } else if let Some(arg) = args.get(i) {
697                slots[i].value = arg.clone();
698                slots[i].initialized = true;
699                slots[i].synced = synced;
700            }
701        }
702    }
703
704    pub(crate) fn visible_variables(&self) -> crate::value::DictMap {
705        let mut vars = self.env.all_variables();
706        let Some(frame) = self.frames.last() else {
707            return vars;
708        };
709        for (slot, info) in frame.local_slots.iter().zip(frame.chunk.local_slots.iter()) {
710            if slot.initialized && info.scope_depth <= frame.local_scope_depth {
711                vars.insert(crate::value::intern_key(&info.name), slot.value.clone());
712            }
713        }
714        vars
715    }
716
717    pub(crate) fn sync_current_frame_locals_to_env(&mut self) {
718        let frames = &mut self.frames;
719        let env = &mut self.env;
720        let Some(frame) = frames.last_mut() else {
721            return;
722        };
723        let local_scope_base = frame.local_scope_base;
724        let local_scope_depth = frame.local_scope_depth;
725        for (slot, info) in frame
726            .local_slots
727            .iter_mut()
728            .zip(frame.chunk.local_slots.iter())
729        {
730            if slot.initialized && !slot.synced && info.scope_depth <= local_scope_depth {
731                slot.synced = true;
732                let scope_idx = local_scope_base + info.scope_depth;
733                while env.scopes.len() <= scope_idx {
734                    env.push_scope();
735                }
736                // Slot-backed locals are never boxed captures (those resolve
737                // to `None` in the compiler and are defined straight into the
738                // env as `Cell`s), so syncing a slot always yields a `Value`.
739                Arc::make_mut(&mut env.scopes[scope_idx].vars).insert(
740                    info.name.clone(),
741                    crate::value::Binding::Value {
742                        value: slot.value.clone(),
743                        mutable: info.mutable,
744                    },
745                );
746            }
747        }
748    }
749
750    pub(crate) fn closure_call_env_for_current_frame(
751        &self,
752        closure: &crate::value::VmClosure,
753    ) -> VmEnv {
754        if closure.module_state().is_some() {
755            return closure.env.cloned_for_call();
756        }
757        let call_env = Self::closure_call_env(&self.env, closure);
758        // Same compile-time short-circuit as the env walk in
759        // `closure_call_env`: when the callee body never resolves an
760        // outer name through the env, injecting closure-typed *slot*
761        // locals from the caller's frame is wasted work too.
762        if !closure.func.chunk.references_outer_names {
763            return call_env;
764        }
765        let mut call_env = call_env;
766        let Some(frame) = self.frames.last() else {
767            return call_env;
768        };
769        for (slot, info) in frame
770            .local_slots
771            .iter()
772            .zip(frame.chunk.local_slots.iter())
773            .filter(|(slot, info)| slot.initialized && info.scope_depth <= frame.local_scope_depth)
774        {
775            if matches!(slot.value, VmValue::Closure(_)) && !call_env.contains(&info.name) {
776                let _ = call_env.define(&info.name, slot.value.clone(), info.mutable);
777            }
778        }
779        call_env
780    }
781
782    pub(crate) fn active_local_slot_value(&self, name: &str) -> Option<VmValue> {
783        let frame = self.frames.last()?;
784        let idx = self.active_local_slot_index(name)?;
785        frame.local_slots.get(idx).map(|slot| slot.value.clone())
786    }
787
788    /// Returns the slot index of an initialized active local with the given
789    /// name, walking from innermost to outermost scope. Used by legacy by-name
790    /// hot paths that still want to mutate the slot value in place without
791    /// paying a defensive `VmValue::clone` first.
792    pub(crate) fn active_local_slot_index(&self, name: &str) -> Option<usize> {
793        let frame = self.frames.last()?;
794        for (idx, info) in frame.chunk.local_slots.iter().enumerate().rev() {
795            if info.name == name && info.scope_depth <= frame.local_scope_depth {
796                if let Some(slot) = frame.local_slots.get(idx) {
797                    if slot.initialized {
798                        return Some(idx);
799                    }
800                }
801            }
802        }
803        None
804    }
805
806    pub(crate) fn assign_active_local_slot(
807        &mut self,
808        name: &str,
809        value: VmValue,
810        debug: bool,
811    ) -> Result<bool, VmError> {
812        let Some(frame) = self.frames.last_mut() else {
813            return Ok(false);
814        };
815        for (idx, info) in frame.chunk.local_slots.iter().enumerate().rev() {
816            if info.name == name && info.scope_depth <= frame.local_scope_depth {
817                if !debug && !info.mutable {
818                    return Err(VmError::ImmutableAssignment(name.to_string()));
819                }
820                if let Some(slot) = frame.local_slots.get_mut(idx) {
821                    crate::value::recursion::dismantle(std::mem::replace(&mut slot.value, value));
822                    slot.initialized = true;
823                    slot.synced = false;
824                    return Ok(true);
825                }
826            }
827        }
828        Ok(false)
829    }
830
831    pub fn new() -> Self {
832        crate::initialize_runtime_assets();
833        Self {
834            stack: Vec::with_capacity(256),
835            env: VmEnv::new(),
836            output: String::new(),
837            builtins: Arc::new(BTreeMap::new()),
838            async_builtins: Arc::new(BTreeMap::new()),
839            builtin_metadata: Arc::new(BTreeMap::new()),
840            builtins_by_id: Arc::new(HashMap::new()),
841            builtin_id_collisions: Arc::new(HashSet::new()),
842            iterators: Vec::new(),
843            frames: Vec::new(),
844            exception_handlers: Vec::new(),
845            spawned_tasks: BTreeMap::new(),
846            process_exit_request: Arc::new(ProcessExitRequest::new()),
847            sync_runtime: Arc::new(crate::synchronization::VmSyncRuntime::new()),
848            shared_state_runtime: Arc::new(crate::shared_state::VmSharedStateRuntime::new()),
849            inline_cache_sets: Vec::new(),
850            inline_cache_set_by_chunk: HashMap::new(),
851            pool_registry: crate::stdlib::pool::new_pool_registry(),
852            llm_mock_context: crate::llm::mock::LlmMockContext::for_new_vm(),
853            package_snapshot_registry: Arc::new(Default::default()),
854            wait_for_graph: Arc::new(crate::wait_for_graph::VmWaitForGraph::new()),
855            held_sync_guards: Vec::new(),
856            inherited_held_keys: Arc::new(Vec::new()),
857            task_scopes: Vec::new(),
858            task_counter: 0,
859            runtime_context_counter: 0,
860            runtime_context: crate::runtime_context::RuntimeContext::root(),
861            deadlines: Vec::new(),
862            execution_deadline: super::execution::new_execution_deadline_state(None),
863            breakpoints: BTreeMap::new(),
864            function_breakpoints: std::collections::BTreeSet::new(),
865            pending_function_bp: None,
866            step_mode: false,
867            step_frame_depth: 0,
868            stopped: false,
869            last_line: 0,
870            source_dir: None,
871            package_execution_guard: None,
872            imported_paths: Vec::new(),
873            deferred_cyclic_imports: Vec::new(),
874            module_cache: Arc::new(BTreeMap::new()),
875            prepared_module_cache: crate::PreparedModuleCache::default(),
876            module_phase_recorder: None,
877            lazy_callable_modules: Arc::new(crate::value::VmMutex::new(BTreeMap::new())),
878            source_cache: Arc::new(BTreeMap::new()),
879            source_file: None,
880            source_text: None,
881            coverage: crate::coverage::for_primary(None),
882            bridge: None,
883            denied_builtins: Arc::new(HashSet::new()),
884            cancel_token: None,
885            interrupt_signal_token: None,
886            cancel_grace_instructions_remaining: None,
887            interrupt_handlers: Vec::new(),
888            next_interrupt_handle: 1,
889            pending_interrupt_signal: None,
890            interrupted: false,
891            dispatching_interrupt: false,
892            interrupt_handler_deadline: None,
893            error_stack_trace: Vec::new(),
894            yield_sender: None,
895            project_root: None,
896            globals: Arc::new(crate::value::DictMap::new()),
897            debug_hook: None,
898            runtime_limits: RuntimeLimits::default(),
899        }
900    }
901
902    pub fn baseline(&self) -> VmBaseline {
903        VmBaseline::from_vm(self)
904    }
905
906    /// Replace the scoped immutable module-template cache used by this VM.
907    /// Fresh runtime module state is still instantiated for every VM.
908    pub fn set_prepared_module_cache(&mut self, cache: crate::PreparedModuleCache) {
909        self.prepared_module_cache = cache;
910    }
911
912    /// Return the effective runtime limit profile for this VM.
913    pub fn runtime_limits(&self) -> RuntimeLimits {
914        self.runtime_limits
915    }
916
917    /// Return a host/debug report describing the VM's effective runtime limits.
918    pub fn runtime_limit_report(&self) -> crate::RuntimeLimitsReport {
919        self.runtime_limits.report()
920    }
921
922    /// Returns true if any debugging affordance is active — DAP hook,
923    /// line breakpoints, or function breakpoints. Call-site code uses
924    /// this to decide whether to capture per-frame restart snapshots
925    /// (`initial_env`, `initial_local_slots`); without a debugger those
926    /// snapshots are dead weight, so skipping them removes two
927    /// allocations from every function call hot path.
928    ///
929    /// All three signals are stable across a function call's lifetime
930    /// (they're set before pipeline execution starts), so the gate is
931    /// consistent between frame creation and any later `restart_frame`
932    /// invocation. The three `is_empty` checks compile to a handful of
933    /// branch-predicted memory probes — cheaper than a single
934    /// `BTreeMap` clone, which is what we're avoiding.
935    #[inline]
936    pub(crate) fn debugger_attached(&self) -> bool {
937        self.debug_hook.is_some()
938            || !self.breakpoints.is_empty()
939            || !self.function_breakpoints.is_empty()
940    }
941
942    /// Set the bridge for delegating unknown builtins in bridge mode.
943    pub fn set_bridge(&mut self, bridge: Arc<crate::bridge::HostBridge>) {
944        self.bridge = Some(bridge);
945    }
946
947    /// Set builtins that are denied in sandbox mode.
948    /// When called, the given builtin names will produce a permission error.
949    pub fn set_denied_builtins(&mut self, denied: HashSet<String>) {
950        self.denied_builtins = Arc::new(denied);
951    }
952
953    /// Set source info for error reporting (file path and source text).
954    pub fn set_source_info(&mut self, file: &str, text: &str) {
955        self.source_file = Some(file.to_string());
956        self.source_text = Some(text.to_string());
957        if let Some(cov) = self.coverage.as_mut() {
958            cov.set_primary_file(file);
959        }
960        Arc::make_mut(&mut self.source_cache)
961            .insert(std::path::PathBuf::from(file), Arc::from(text));
962    }
963
964    /// Initialize execution (push the initial frame).
965    pub fn start(&mut self, chunk: &Chunk) -> Result<(), VmError> {
966        self.ensure_execution_available()?;
967        // The top-level pipeline frame captures env at start so
968        // restartFrame on the outermost frame rewinds to the
969        // pre-pipeline state — basically "restart session" in
970        // debugger terms. Skipped when no debugger is attached:
971        // the snapshot is dead weight in that case and dominates
972        // call-overhead bench numbers (~5-10%).
973        let debugger = self.debugger_attached();
974        let initial_env = if debugger {
975            Some(self.env.clone())
976        } else {
977            None
978        };
979        let initial_local_slots = if debugger {
980            Some(Self::fresh_local_slots(chunk))
981        } else {
982            None
983        };
984        let chunk = Arc::new(chunk.clone());
985        let local_slots = Self::fresh_local_slots(&chunk);
986        let inline_cache_set = self.inline_cache_set_index_for_chunk(&chunk);
987        self.frames.push(CallFrame {
988            chunk,
989            inline_cache_set,
990            ip: 0,
991            stack_base: self.stack.len(),
992            saved_env: self.env.clone(),
993            initial_env,
994            initial_local_slots,
995            saved_iterator_depth: self.iterators.len(),
996            fn_name: String::new(),
997            argc: 0,
998            saved_source_dir: None,
999            module_functions: None,
1000            module_state: None,
1001            local_slots,
1002            local_scope_base: self.env.scope_depth().saturating_sub(1),
1003            local_scope_depth: 0,
1004        });
1005        Ok(())
1006    }
1007
1008    /// Create a child VM that shares builtins and env but has fresh execution state.
1009    /// Used for parallel/spawn to fork the VM for concurrent tasks.
1010    pub(crate) fn child_vm(&self) -> Vm {
1011        Vm {
1012            stack: Vec::with_capacity(64),
1013            env: self.env.clone(),
1014            output: String::new(),
1015            builtins: Arc::clone(&self.builtins),
1016            async_builtins: Arc::clone(&self.async_builtins),
1017            builtin_metadata: Arc::clone(&self.builtin_metadata),
1018            builtins_by_id: Arc::clone(&self.builtins_by_id),
1019            builtin_id_collisions: Arc::clone(&self.builtin_id_collisions),
1020            iterators: Vec::new(),
1021            frames: Vec::new(),
1022            exception_handlers: Vec::new(),
1023            spawned_tasks: BTreeMap::new(),
1024            process_exit_request: Arc::clone(&self.process_exit_request),
1025            sync_runtime: self.sync_runtime.clone(),
1026            shared_state_runtime: self.shared_state_runtime.clone(),
1027            inline_cache_sets: Vec::new(),
1028            inline_cache_set_by_chunk: HashMap::new(),
1029            pool_registry: self.pool_registry.clone(),
1030            llm_mock_context: self.llm_mock_context.clone(),
1031            package_snapshot_registry: self.package_snapshot_registry.clone(),
1032            wait_for_graph: self.wait_for_graph.clone(),
1033            held_sync_guards: Vec::new(),
1034            inherited_held_keys: Arc::new(Vec::new()),
1035            task_scopes: Vec::new(),
1036            task_counter: 0,
1037            runtime_context_counter: self.runtime_context_counter,
1038            runtime_context: self.runtime_context.clone(),
1039            deadlines: self.deadlines.clone(),
1040            execution_deadline: self.execution_deadline.fork(),
1041            breakpoints: BTreeMap::new(),
1042            function_breakpoints: std::collections::BTreeSet::new(),
1043            pending_function_bp: None,
1044            step_mode: false,
1045            step_frame_depth: 0,
1046            stopped: false,
1047            last_line: 0,
1048            source_dir: self.source_dir.clone(),
1049            package_execution_guard: self.package_execution_guard.clone(),
1050            imported_paths: Vec::new(),
1051            deferred_cyclic_imports: Vec::new(),
1052            module_cache: Arc::clone(&self.module_cache),
1053            prepared_module_cache: self.prepared_module_cache.clone(),
1054            module_phase_recorder: self.module_phase_recorder.clone(),
1055            lazy_callable_modules: Arc::clone(&self.lazy_callable_modules),
1056            source_cache: Arc::clone(&self.source_cache),
1057            source_file: self.source_file.clone(),
1058            source_text: self.source_text.clone(),
1059            coverage: crate::coverage::for_primary(self.source_file.as_deref()),
1060            bridge: self.bridge.clone(),
1061            denied_builtins: Arc::clone(&self.denied_builtins),
1062            cancel_token: self.cancel_token.clone(),
1063            interrupt_signal_token: self.interrupt_signal_token.clone(),
1064            cancel_grace_instructions_remaining: None,
1065            interrupt_handlers: Vec::new(),
1066            next_interrupt_handle: 1,
1067            pending_interrupt_signal: None,
1068            interrupted: self.interrupted,
1069            dispatching_interrupt: false,
1070            interrupt_handler_deadline: None,
1071            error_stack_trace: Vec::new(),
1072            yield_sender: None,
1073            project_root: self.project_root.clone(),
1074            globals: Arc::clone(&self.globals),
1075            debug_hook: None,
1076            runtime_limits: self.runtime_limits,
1077        }
1078    }
1079
1080    /// Create a child VM for external adapters that need to invoke Harn
1081    /// closures while sharing the parent's builtins, globals, and module state.
1082    pub(crate) fn child_vm_for_host(&self) -> Vm {
1083        self.child_vm()
1084    }
1085
1086    pub(crate) fn request_process_exit(&self, code: i32) {
1087        self.process_exit_request.request(code);
1088    }
1089
1090    pub(crate) fn requested_process_exit(&self) -> Option<i32> {
1091        self.process_exit_request.code()
1092    }
1093
1094    /// Request cancellation for every outstanding child task owned by this VM
1095    /// and then abort the join handles. This prevents un-awaited spawned tasks
1096    /// from outliving their parent execution scope.
1097    pub(crate) fn cancel_spawned_tasks(&mut self) {
1098        for (_, task) in std::mem::take(&mut self.spawned_tasks) {
1099            task.cancel_token
1100                .store(true, std::sync::atomic::Ordering::SeqCst);
1101            task.handle.abort();
1102        }
1103    }
1104
1105    /// Set the source directory for import resolution and introspection.
1106    /// Also auto-detects the project root if not already set.
1107    pub fn set_source_dir(&mut self, dir: &std::path::Path) {
1108        let dir = crate::stdlib::process::normalize_context_path(dir);
1109        self.source_dir = Some(dir.clone());
1110        crate::stdlib::set_thread_source_dir(&dir);
1111        // Auto-detect project root if not explicitly set.
1112        if self.project_root.is_none() {
1113            self.project_root = crate::stdlib::process::find_project_root(&dir);
1114        }
1115    }
1116
1117    /// Explicitly set the project root directory.
1118    /// Used by ACP/CLI to override auto-detection.
1119    pub fn set_project_root(&mut self, root: &std::path::Path) {
1120        self.project_root = Some(root.to_path_buf());
1121    }
1122
1123    /// Get only the explicit or auto-detected project root, without falling
1124    /// back to `source_dir`.
1125    pub(crate) fn explicit_project_root(&self) -> Option<&std::path::Path> {
1126        self.project_root.as_deref()
1127    }
1128
1129    /// Get the project root directory, falling back to source_dir.
1130    pub fn project_root(&self) -> Option<&std::path::Path> {
1131        self.project_root.as_deref().or(self.source_dir.as_deref())
1132    }
1133
1134    /// Return all registered builtin names (sync + async).
1135    pub fn builtin_names(&self) -> Vec<String> {
1136        let mut names: Vec<String> = self.builtins.keys().cloned().collect();
1137        names.extend(self.async_builtins.keys().cloned());
1138        names
1139    }
1140
1141    /// Return discoverable metadata for registered builtins.
1142    pub fn builtin_metadata(&self) -> Vec<VmBuiltinMetadata> {
1143        self.builtin_metadata.values().cloned().collect()
1144    }
1145
1146    /// Return discoverable metadata for a registered builtin name.
1147    pub fn builtin_metadata_for(&self, name: &str) -> Option<&VmBuiltinMetadata> {
1148        self.builtin_metadata.get(name)
1149    }
1150
1151    /// Set a global constant (e.g. `pi`, `e`).
1152    /// Stored separately from the environment so user-defined variables can shadow them.
1153    pub fn set_global(&mut self, name: &str, value: VmValue) {
1154        Arc::make_mut(&mut self.globals).insert(crate::value::intern_key(name), value);
1155    }
1156
1157    /// Read a previously-installed global (the value `set_global` /
1158    /// `set_harness` recorded). Returns `None` for unknown names.
1159    /// Hosts use this to look up runtime-installed capability handles
1160    /// (e.g. the `harness` slot) without having to track them
1161    /// separately.
1162    pub fn global(&self, name: &str) -> Option<&VmValue> {
1163        self.globals.get(name)
1164    }
1165
1166    /// Install the script's `Harness` capability handle as the `harness`
1167    /// global so the auto-call emitted by `Compiler::compile()` (for
1168    /// `fn main(harness: Harness)` entrypoints) can read it. Hosts that
1169    /// drive the VM directly (CLI, MCP server, composition runtime) call
1170    /// this once before `execute()`.
1171    pub fn set_harness(&mut self, harness: crate::harness::Harness) {
1172        self.set_global("harness", harness.into_vm_value());
1173    }
1174
1175    /// Get the captured output.
1176    pub fn output(&self) -> &str {
1177        &self.output
1178    }
1179
1180    /// Drain and return the captured output, leaving the buffer empty.
1181    /// Used by the async-builtin dispatch path to forward closure output
1182    /// from a child VM back to its parent.
1183    pub fn take_output(&mut self) -> String {
1184        std::mem::take(&mut self.output)
1185    }
1186
1187    /// Append text to this VM's captured output. Used to forward output
1188    /// from child VMs (e.g. closures invoked via `call_closure_pub`)
1189    /// back into the parent stream.
1190    pub fn append_output(&mut self, text: &str) {
1191        self.output.push_str(text);
1192    }
1193
1194    pub(crate) fn pop(&mut self) -> Result<VmValue, VmError> {
1195        self.stack.pop().ok_or(VmError::StackUnderflow)
1196    }
1197
1198    pub(crate) fn peek(&self) -> Result<&VmValue, VmError> {
1199        self.stack.last().ok_or(VmError::StackUnderflow)
1200    }
1201
1202    pub(crate) fn const_str(c: &Constant) -> Result<&str, VmError> {
1203        match c {
1204            Constant::String(s) => Ok(s.as_str()),
1205            _ => Err(VmError::TypeError("expected string constant".into())),
1206        }
1207    }
1208
1209    pub(crate) fn release_sync_guards_for_current_scope(&mut self) {
1210        let depth = self.env.scope_depth();
1211        self.held_sync_guards
1212            .retain(|guard| guard.env_scope_depth < depth);
1213        // A `scope { }` torn down without a normal `TaskScopeExit` (break /
1214        // continue out of it) leaves a dangling nursery — cancel its tasks.
1215        self.cancel_task_scopes_where(|s| s.env_scope_depth >= depth);
1216    }
1217
1218    pub(crate) fn release_sync_guards_after_unwind(
1219        &mut self,
1220        frame_depth: usize,
1221        env_scope_depth: usize,
1222    ) {
1223        self.held_sync_guards.retain(|guard| {
1224            guard.frame_depth <= frame_depth && guard.env_scope_depth <= env_scope_depth
1225        });
1226        // Cancel nurseries opened above the catch handler (a `throw` unwound
1227        // past their `TaskScopeExit`).
1228        self.cancel_task_scopes_where(|s| {
1229            !(s.frame_depth <= frame_depth && s.env_scope_depth <= env_scope_depth)
1230        });
1231    }
1232
1233    pub(crate) fn release_sync_guards_for_frame(&mut self, frame_depth: usize) {
1234        self.held_sync_guards
1235            .retain(|guard| guard.frame_depth != frame_depth);
1236        // Cancel any nursery whose `scope {}` block belonged to the frame being
1237        // torn down (e.g. a `return` jumped past its `TaskScopeExit`).
1238        self.cancel_task_scopes_where(|s| s.frame_depth == frame_depth);
1239    }
1240
1241    pub(crate) fn adopt_sync_permit_for_current_scope(
1242        &mut self,
1243        permit: crate::value::VmSyncPermitHandle,
1244    ) {
1245        if permit.is_released()
1246            || self
1247                .held_sync_guards
1248                .iter()
1249                .any(|guard| guard._permit.same_lease(&permit))
1250        {
1251            return;
1252        }
1253        self.held_sync_guards
1254            .push(crate::synchronization::VmSyncHeldGuard {
1255                _permit: permit,
1256                frame_depth: self.frames.len(),
1257                env_scope_depth: self.env.scope_depth(),
1258            });
1259    }
1260
1261    /// Deregister a task id from every open nursery (it was explicitly
1262    /// `await`ed, so it must not be double-joined or cancelled at scope exit).
1263    pub(crate) fn deregister_task_from_scopes(&mut self, id: &str) {
1264        for scope in &mut self.task_scopes {
1265            scope.task_ids.retain(|t| t != id);
1266        }
1267    }
1268
1269    /// Cancel and remove every task scope matching `doomed`, aborting its bound
1270    /// tasks (used when a `scope {}` is torn down without a normal join).
1271    fn cancel_task_scopes_where<F: Fn(&TaskScope) -> bool>(&mut self, doomed: F) {
1272        let mut i = 0;
1273        while i < self.task_scopes.len() {
1274            if doomed(&self.task_scopes[i]) {
1275                let scope = self.task_scopes.remove(i);
1276                for id in &scope.task_ids {
1277                    if let Some(task) = self.spawned_tasks.remove(id) {
1278                        task.cancel_token
1279                            .store(true, std::sync::atomic::Ordering::SeqCst);
1280                        task.handle.abort();
1281                    }
1282                }
1283            } else {
1284                i += 1;
1285            }
1286        }
1287    }
1288
1289    /// Total live permits this VM already holds for `kind:key`. The held-set is
1290    /// tiny (bounded by lexical nesting and explicit sync acquisitions), so this
1291    /// scan is cheap and only runs on the rare blocking-acquire path.
1292    pub(crate) fn held_permits_for(&self, kind: &str, key: &str) -> u32 {
1293        let own: u32 = self
1294            .held_sync_guards
1295            .iter()
1296            .filter(|guard| {
1297                !guard._permit.is_released()
1298                    && guard._permit.kind() == kind
1299                    && guard._permit.key() == key
1300            })
1301            .map(|guard| guard._permit.permits())
1302            .sum();
1303        let inherited: u32 = self
1304            .inherited_held_keys
1305            .iter()
1306            .filter(|held| held.kind == kind && held.key == key)
1307            .map(|held| held.permits)
1308            .sum();
1309        own + inherited
1310    }
1311
1312    /// Every live sync permit held by this VM *and* its suspended ancestors: the
1313    /// transitive held-set seen by an inline child.
1314    pub(crate) fn combined_held_keys(&self) -> Vec<crate::synchronization::VmSyncHeldKey> {
1315        let mut keys: Vec<crate::synchronization::VmSyncHeldKey> = self
1316            .held_sync_guards
1317            .iter()
1318            .filter_map(|guard| crate::synchronization::VmSyncHeldKey::from_permit(&guard._permit))
1319            .collect();
1320        keys.extend(self.inherited_held_keys.iter().cloned());
1321        keys
1322    }
1323
1324    /// Clone a child VM for an **inline, same-task** execution (an async builtin
1325    /// awaited while this VM is parked, or a user closure that builtin runs and
1326    /// awaits). The child inherits this VM's transitive held-lock keys so a
1327    /// re-acquire of a parent-held lock is caught as a self-deadlock
1328    /// (HARN-ORC-011). Use plain `child_vm()` for new concurrent tasks.
1329    pub(crate) fn child_vm_inline(&self) -> Vm {
1330        let mut child = self.child_vm();
1331        child.inherited_held_keys = Arc::new(self.combined_held_keys());
1332        child
1333    }
1334}
1335
1336impl Drop for Vm {
1337    fn drop(&mut self) {
1338        if let Some(coverage) = self.coverage.take() {
1339            crate::coverage::merge_into_global(coverage);
1340        }
1341        self.cancel_spawned_tasks();
1342    }
1343}
1344
1345impl Default for Vm {
1346    fn default() -> Self {
1347        Self::new()
1348    }
1349}
1350
1351#[cfg(test)]
1352mod tests {
1353
1354    use super::*;
1355
1356    #[test]
1357    fn vm_construction_initializes_shared_secret_patterns() {
1358        let _vm = Vm::new();
1359        assert!(crate::secret_patterns::default_secret_patterns_initialized());
1360    }
1361
1362    fn baseline_with_stdlib(source: &str) -> VmBaseline {
1363        let mut vm = Vm::new();
1364        crate::register_vm_stdlib(&mut vm);
1365        vm.set_source_info("baseline_test.harn", source);
1366        vm.set_global(
1367            "stable_global",
1368            VmValue::String(arcstr::ArcStr::from("baseline")),
1369        );
1370        vm.baseline()
1371    }
1372
1373    #[test]
1374    fn vm_baseline_instantiates_clean_mutable_execution_state() {
1375        let baseline = baseline_with_stdlib("pipeline main() { __io_println(stable_global) }");
1376
1377        let mut dirty = baseline.instantiate();
1378        dirty.stack.push(VmValue::Int(42));
1379        dirty.output.push_str("dirty");
1380        dirty.task_counter = 9;
1381        dirty.runtime_context_counter = 7;
1382        dirty
1383            .error_stack_trace
1384            .push(("main".to_string(), 1, 1, None));
1385
1386        let clean = baseline.instantiate();
1387        assert!(clean.stack.is_empty());
1388        assert!(clean.output.is_empty());
1389        assert!(clean.frames.is_empty());
1390        assert!(clean.exception_handlers.is_empty());
1391        assert!(clean.spawned_tasks.is_empty());
1392        assert!(clean.held_sync_guards.is_empty());
1393        assert_eq!(clean.task_counter, 0);
1394        assert_eq!(clean.runtime_context_counter, 0);
1395        assert!(clean.deadlines.is_empty());
1396        assert!(clean.cancel_token.is_none());
1397        assert!(clean.interrupt_handlers.is_empty());
1398        assert!(clean.error_stack_trace.is_empty());
1399        assert!(clean.bridge.is_none());
1400        assert!(clean
1401            .globals
1402            .get("stable_global")
1403            .is_some_and(|value| value.display() == "baseline"));
1404    }
1405
1406    #[tokio::test]
1407    async fn inline_child_inherits_held_lock_keys_but_concurrent_child_does_not() {
1408        let mut parent = Vm::new();
1409        let permit = parent
1410            .sync_runtime
1411            .acquire("mutex", "v:test", 1, 1, None, None)
1412            .await
1413            .unwrap()
1414            .unwrap();
1415        parent
1416            .held_sync_guards
1417            .push(crate::synchronization::VmSyncHeldGuard {
1418                _permit: permit,
1419                frame_depth: 0,
1420                env_scope_depth: 0,
1421            });
1422        assert_eq!(parent.held_permits_for("mutex", "v:test"), 1);
1423
1424        // An inline child (async builtin awaited while the parent is parked, or
1425        // a closure the builtin runs inline) inherits the held key, so a
1426        // re-acquire is caught as a cross-context self-deadlock (HARN-ORC-011)
1427        // — even transitively through a further inline child.
1428        let inline = parent.child_vm_inline();
1429        assert_eq!(inline.held_permits_for("mutex", "v:test"), 1);
1430        assert_eq!(
1431            inline.child_vm_inline().held_permits_for("mutex", "v:test"),
1432            1
1433        );
1434
1435        // A new concurrent task (spawn / parallel / trigger) does NOT inherit:
1436        // blocking on a parent-held lock there is legitimately resolvable, so
1437        // flagging it would be a false positive.
1438        let concurrent = parent.child_vm();
1439        assert_eq!(concurrent.held_permits_for("mutex", "v:test"), 0);
1440    }
1441
1442    #[test]
1443    fn vm_reports_effective_runtime_limits() {
1444        let vm = Vm::new();
1445
1446        assert_eq!(vm.runtime_limits(), RuntimeLimits::default());
1447        assert_eq!(
1448            vm.runtime_limit_report().entries.len(),
1449            crate::RUNTIME_LIMIT_DESCRIPTIONS.len()
1450        );
1451        assert_eq!(vm.child_vm().runtime_limits(), vm.runtime_limits());
1452        assert_eq!(
1453            vm.baseline().instantiate().runtime_limits(),
1454            vm.runtime_limits()
1455        );
1456    }
1457
1458    #[tokio::test(flavor = "current_thread")]
1459    async fn vm_baseline_rebinds_shared_state_builtins_per_instance() {
1460        let local = tokio::task::LocalSet::new();
1461        local
1462            .run_until(async {
1463                let source = r#"
1464pipeline main() {
1465  const cell = shared_cell({scope: "task_group", key: "turn", initial: 0})
1466  __io_println(shared_get(cell))
1467  shared_set(cell, shared_get(cell) + 1)
1468}"#;
1469                let chunk = crate::compile_source(source).expect("compile");
1470                let baseline = baseline_with_stdlib(source);
1471
1472                let mut first = baseline.instantiate();
1473                first.execute(&chunk).await.expect("first execute");
1474                assert_eq!(first.output(), "0\n");
1475
1476                let mut second = baseline.instantiate();
1477                second.execute(&chunk).await.expect("second execute");
1478                assert_eq!(
1479                    second.output(),
1480                    "0\n",
1481                    "shared state created by the first VM must not leak into the next baseline instance"
1482                );
1483            })
1484            .await;
1485    }
1486}