Automated flow to bump version${COMMIT_MSG} by hellovai · Pull Request #2 · BoundaryML/baml

hellovai · 2023-10-10T05:11:42Z

No description provided.

Complete the BEX garbage collector by closing 5 identified gaps: - Gap #1: Parked VM roots now collected via VM pointer registry - Gap #2: VM stacks updated with forwarding pointers after GC - Gap #3: TLABs invalidated after GC space swap - Gap #6: Handle race condition fixed with EpochGuard Key changes: - Add parked_vms registry to EpochState for tracking VMs at safepoints - Collect roots from all parked VM stacks during GC - Update parked VM stacks with forwarding pointers after copying - Remove cached ObjectIndex from Handle (always resolve through table) - Add gc_in_progress flag to synchronize new calls with GC - Add EpochGuard token for compile-time epoch protection enforcement - Add heap accessor API for safe external code access (read_string, etc.) - Add GC coordination documentation to module docs

…ements Orchestration: - Replace flat build_plan + execution loop with recursive execute_client / execute_client_once — RR counters now advance only at execution time, including nested RR inside fallback branches (fully fixes Codex #2) - Last retry attempt no longer applies local retry sleep (legacy parity) - Extract resolve_primitive helper to deduplicate client resolution Codex review fixes: - #3: Compile-time EmptyStrategy diagnostic for composite clients with no sub-clients (prevents runtime modulo-by-zero crash) - #4: Cross-file UnknownRetryPolicy diagnostic when a client references a retry policy that doesn't exist - #5: Synthetic error HTTP responses now store error message so text() returns error details instead of "already consumed" Parity improvements: - Finish-reason filtering enforced in parse path (sys_llm) - Raw provider finish reason preserved for allow/deny matching - Round-robin start index plumbed through HIR → emit → engine metadata - Cancellation parity TODO added All tests pass: 375 baml_tests, 17 orchestration, 13 llm_render, 79 sys_llm.

…yML#3185) ## Summary Fixes three concurrency bugs in the Go CFFI callback layer that cause deadlocks and process crashes under sustained concurrent load (~100 streaming/call requests in parallel). Discovered during load testing. All changes are in `engine/language_client_go/pkg/` (Go only — no Rust changes). ## Changes ### 1. Replace random callback IDs with atomic counter (`callbacks.go`) `create_unique_id` uses `rand.Intn(1000000)` in a retry loop while holding an exclusive lock. When callback entries leak (see bug BoundaryML#2 below), the ID space fills up, the loop spins longer, and all goroutines needing `callbackMutex.RLock()` starve — full deadlock. Fix: `sync/atomic.Uint32` counter. IDs are unique by construction, no retry loop, lock acquired only after ID is chosen. ### 2. Prevent panics from concurrent callback channel operations (`callbacks.go`, `runtime.go`) `error_callback` and `trigger_callback` can race on the same callback ID — one closes the channel while the other sends. `send on closed channel` panics crash the entire plugin process and all in-flight requests. Go has no atomic check-and-send for channels (`select { case ch <- v: default: }` still panics on closed channels). Fix: `safeSend()`/`safeClose()` helpers that wrap operations with `defer`/`recover()`. Zero overhead in the non-panic path. Also makes callback channels buffered (64 for stream, 1 for object-handle) so that post-cancellation sends don't block forever, which was causing the callback entry leaks that feed bug BoundaryML#1. ### 3. Recover from panics in `on_tick_callback` (`callbacks.go`) `on_tick_callback` invokes user-provided handlers that send on application-level output channels. When the stream consumer closes its channel (context cancelled, iteration done), a concurrent tick panics. This is a different code path from BoundaryML#2 — it's the user's channel, not the internal callback channel. Fix: `defer func() { recover() }()` at the top of `on_tick_callback`. ## Testing Each fix was validated independently using a "leave one out" methodology against a 100k-request stress test (40k call + 20k stream + 40k parse, concurrency 100, mock LLM): | Fix removed | Result | |---|---| | Atomic IDs | Deadlock at ~23k requests | | safeSend/safeClose | 77% success, repeated panics | | on_tick recover | 99.9% success, 22 failures | | All fixes applied | **100% success, 1500+ req/s** |  ## Summary by CodeRabbit * **Bug Fixes** * Replaced non-deterministic callback IDs with deterministic sequencing to improve reliability. * Introduced buffered callback channels and safer send/close handling to prevent panics, races, and deadlocks. * Added panic recovery and safer cleanup for error paths to ensure callbacks are closed reliably and reduce crashes.  --------- Co-authored-by: hellovai <vbv@boundaryml.com>

Add test projects and LSP tests covering confirmed issues from PR review: - captured_field_chain: lambda capture roots (#1) and map fallback (#2) - generic_field_chain: generic type substitution (#3, deferred to separate PR) - deep_method_call: chained method call throw target (#5) - completions tests: chained field-access (#7) and lambda param (#8) - semantic token test: type-aware final segment classification (#9) Also accept lambda_advanced MIR snapshot (void→int type fix) and net.rs bytecode updates (call→dispatch_future). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

BamlRuntime#spawn_and_wait and FunctionResultStream#done called queue.pop with no timeout, so a hung Tokio task or a dropped future would wedge the calling Ruby thread indefinitely. Add a `timeout:` kwarg (default nil = infinite) to both paths; on expiry the pending call-id is removed and BamlError is raised. For streams the timeout is applied per chunk rather than as a total budget so long-running legitimate streams keep working as long as chunks keep arriving within the window. Implemented via stdlib Timeout rather than Queue#pop(timeout:), which is Ruby 3.2+ only (the gemspec still targets >= 3.1). Flagged in thoughts/ryan/reviews/2026-04-20-ruby-cffi-migration-review.md (High BoundaryML#2).

- pull_semantics: debug_assert on the default `make_closure_with_type_args` so a sink that drops `ntypeargs` fails fast instead of silently modeling the wrong stack shape (#1). - mir/lower: `enclosing_generic_params()` now prepends class-level generics (matching TIR's `enclosing_class_generic_params ++ user` convention), so `reflect.type_of<T>()` and explicit `<T>` calls inside a method on `class Box<T>` lower to `TypeArgRef(N)` instead of silently emitting `Concrete(Void)` (#2, MIR side only — runtime ABI for class type-arg threading via the receiver is still TODO; documented on the helper). Two stdlib snapshots regenerated to reflect the now-correct lowering for `Stream<T,S>::final()` / `__sap_parse_partial<T,S>`. - tir/builder: add `Ty::Type { .. }` arm to `try_resolve_member_on_ty` so unions like `(cond ? type_of<A>() : type_of<B>()).to_string()` resolve via the `TypeValue` companion class (#4). - vm: gate the post-`Call` `type_args` writeback on `frames.len() > frames_before` so completed Native callees don't clobber the caller's frame, and `extend` rather than replace so a closure callee's `captured_type_args` are preserved with call-site args appended (#6, critical). - tests/load_type: tighten the concrete `LoadType` tests to match `Object::Type(ty)` against `Ty::int()`/`Ty::string()` and assert they differ, instead of accepting any `Value::Object(_)` (#7). - type_class / reflect_type_of: document `TypeValue.to_string()` as a stable identity key (package names are unique within a workspace, so the existing `Display` form is unambiguous) and add an end-to-end `map<string, V>` keyed-by-`type_of<T>().to_string()` test as the user-reachable analogue of `map<type, V>` (#8). - conversion: document the runtime gap that class-instance matching in `value_matches_type` and `find_matching_union_member` ignores `Ty::Class` type args because `Object::Instance` carries no resolved class type args — needed for proper union dispatch on parametric classes (#9, runtime side bundled with #2 follow-up).

Address CodeRabbit review on PR #3615: - baml_rustgen_check::walk now panics on read_dir/read errors instead of silently skipping them — a partial/unreadable source-of-truth must abort the build/regen, never hash a partial tree that could match stale output. (#1) - tools_rustgen status message: "wrote N generated file(s)" instead of the misleading "up to date" (write_all always rewrites). (#4) - Document normalize()'s known whitespace-in-string-literal limitation; the up_to_date regen test is the backstop, so a quote-aware tokenizer isn't worth the complexity. (#2) Not changed: the build.rs "sys_types/src/..." paths flagged critical are correct — assert_generated_matches_baml_std resolves via crates_dir() (CARGO_MANIFEST_DIR/..), not the build script cwd, and CI is green. (#3) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Three describe cleanups from the audit: - Resolve-once (#2): `describe_top_level` now resolves the symbol's `Definition` a single time and threads it to build_shape / find_dependencies / resolve_type_for_item / class methods+fqn / builtin-signature, replacing ~5 independent `resolve_name_at` calls. Pure refactor — output is byte-identical. - Class dependencies (#1): `collect_ty_deps` matched the canonical `pkg.Name` qtn string against the flat outline's short names, so a class's user-type dependencies never resolved. It now keys on the short name and only surfaces local (user-package) types — builtin/dependency types are well-known noise. Types used only in method signatures are now collected too (e.g. `WrapperMarker` in `-> T | WrapperMarker`). - Harden method pairing (#3): the HIR-method <-> package-interface signature pairing is index-based (the interface is built 1:1 with `class.methods`); `exported_method` now verifies the name at that index so a future reordering can't silently mispair signatures — it falls back to unresolved source types instead. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- Generic static methods: bind own TypeVars via _types= (named wire), no class args; engine slots each binding by name into the static frame (class params harmlessly unknown). Fixtures: GenericBox.new<V> + NamedStatic.make<D,E> (distinct names) prove name-keyed slotting. - Instance methods: class TypeVars recovered from the parameterized receiver, method's own via _types=. - Engine: scope the declared-generic-param completeness check to free functions (a static's display_type_params include enclosing class params it never binds; is_class_method gates it); check #2 (no unbound TypeVar in substituted sig) still covers statics. - .pyi: required keyword-only _types for generic methods (own params). - Tests: static new/make, required-ness, unparameterized-receiver negative. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Automated flow to bump version${COMMIT_MSG}

769d924

hellovai closed this Oct 10, 2023

hellovai deleted the auto-bump-version branch October 10, 2023 05:13

aaronvg mentioned this pull request Dec 4, 2024

'baml-cli generate' supports ES Module #1213

Open

hellovai mentioned this pull request Jan 25, 2025

Is there support for 'unlimited' output tokens? #1353

Open

hellovai mentioned this pull request Jan 20, 2026

feat(bex): complete garbage collector implementation #3003

Merged

6 tasks

invakid404 mentioned this pull request Feb 26, 2026

fix(go-cffi): fix deadlock and panics under high concurrency #3185

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Automated flow to bump version${COMMIT_MSG} - #2

Automated flow to bump version${COMMIT_MSG}#2
hellovai wants to merge 1 commit into
canaryfrom
auto-bump-version

hellovai commented Oct 10, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

hellovai commented Oct 10, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant