feat(runtime): make connection pooling configurable - #3975
Merged
sxlijin merged 4 commits intoAug 14, 2026
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
|
@dexhunter is attempting to deploy a commit to the Boundary Team on Vercel. A member of the Team first needs to authorize it. |
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
🚧 Files skipped from review as they are similar to previous changes (1)
📝 WalkthroughWalkthroughAdds opt-in HTTP connection pooling for supported BAML clients. Provider-aware validation rejects unsupported providers and invalid types. Runtime tests verify connection reuse. Documentation describes configuration, defaults, and limitations. ChangesHTTP Connection Pooling
Estimated code review effort: 3 (Moderate) | ~25 minutes Merge Risk: ⚪ Minimal · up to This change makes HTTP connection pooling opt-in while preserving existing behavior by default; no actionable merge-blocking risk remains after normal checks and review. Sequence Diagram(s)sequenceDiagram
participant BAMLClient
participant create_http_client
participant ReqwestTCPServer
BAMLClient->>create_http_client: Configure enable_connection_pooling
create_http_client->>ReqwestTCPServer: Send three requests
ReqwestTCPServer-->>create_http_client: Return HTTP responses
create_http_client-->>BAMLClient: Reuse one connection or open three connections
Poem
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches 💡 1🛠️ Fix failing CI checks 💡
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
🧹 Nitpick comments (1)
fern/03-reference/baml/clients/connection-pooling.mdx (1)
34-36: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winAdd links to related documentation pages.
The coding guidelines state that documentation should include well-designed links to guide readers to relevant pages. This page mentions fallback, round-robin, and AWS Bedrock clients but does not link to their respective documentation pages.
🔗 Suggested links
-Connection pooling is not available for fallback or round-robin clients. Set it -on their leaf clients instead. It is also not available for AWS Bedrock clients, -whose HTTP transport is managed by the AWS SDK. +Connection pooling is not available for [fallback](/reference/baml/clients/strategy/fallback) or +[round-robin](/reference/baml/clients/strategy/round-robin) clients. Set it on their leaf +clients instead. It is also not available for [AWS Bedrock](/reference/baml/clients/providers/aws-bedrock) +clients, whose HTTP transport is managed by the AWS SDK.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@fern/03-reference/baml/clients/connection-pooling.mdx` around lines 34 - 36, Add inline links to the documentation pages for fallback clients, round-robin clients, and AWS Bedrock clients in the connection-pooling limitation paragraph. Preserve the existing wording while linking each referenced client type to its corresponding documentation page.Source: Coding guidelines
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@fern/03-reference/baml/clients/connection-pooling.mdx`:
- Around line 1-3: Add the required subtitle field to the frontmatter of the
“Connection Pooling” document, preserving the existing title and matching the
specified frontmatter format used by other fern/**/*.mdx files.
---
Nitpick comments:
In `@fern/03-reference/baml/clients/connection-pooling.mdx`:
- Around line 34-36: Add inline links to the documentation pages for fallback
clients, round-robin clients, and AWS Bedrock clients in the connection-pooling
limitation paragraph. Preserve the existing wording while linking each
referenced client type to its corresponding documentation page.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository UI
Review profile: CHILL
Plan: Pro
Run ID: e4f337d3-f5ba-4fe0-853d-918adcd9d41b
📒 Files selected for processing (8)
engine/baml-lib/baml/tests/validation_files/client/http_config_aws_pooling.bamlengine/baml-lib/baml/tests/validation_files/client/http_config_invalid_pooling.bamlengine/baml-lib/baml/tests/validation_files/client/http_config_regular_with_total.bamlengine/baml-lib/baml/tests/validation_files/client/http_config_valid.bamlengine/baml-lib/llm-client/src/clients/helpers.rsengine/baml-runtime/src/request/mod.rsfern/03-reference/baml/clients/connection-pooling.mdxfern/docs.yml
dexhunter
had a problem deploying
to
boundary-tools-dev
July 10, 2026 21:40 — with 
GitHub Actions
Failure
GitHub Actions
Failure
dexhunter
had a problem deploying
to
boundary-tools-dev
July 10, 2026 21:40 — with
GitHub Actions
Failure
GitHub Actions
Failure
dexhunter
had a problem deploying
to
boundary-tools-dev
July 10, 2026 21:40 — with
GitHub Actions
Failure
GitHub Actions
Failure
Contributor
|
are you using AWS provider with baml? |
dexhunter
force-pushed
the
feat/opt-in-connection-pooling
branch
from
July 10, 2026 23:22
b6265f2 to
7c4a2fd
Compare
Contributor
Author
|
No, I am not using AWS Bedrock for this path. The target is the regular HTTP-based providers. I kept AWS explicitly unsupported because its HTTP transport is owned by the AWS SDK, so exposing this reqwest pooling option there would be misleading. The motivation is #3072 plus the reproduced connection-count result: 10 requests use 10 connections by default and 1 connection when pooling is explicitly enabled. I have also rebased onto current canary and addressed the documentation review. |
Contributor
Author
|
One status note, because the red ✗ on this PR are misleading rather than informative. The two failing checks are Vercel deploy statuses, and both read "Authorization required to deploy" — a fork can't deploy to the Boundary team. The four checks that would actually tell you something ( Re-checked against If the surface is the concern rather than the mechanism, I'm happy to drop the docs page and ship just the flag plus validation, or to narrow it further. |
Contributor
|
Note GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer. |
sxlijin
had a problem deploying
to
boundary-tools-dev
August 14, 2026 21:54 — with
GitHub Actions
Failure
GitHub Actions
Failure
meefs
pushed a commit
to meefs/baml
that referenced
this pull request
Aug 15, 2026
## Summary - add comprehensive release notes for the independent `baml_language` 0.17.0 release line - separate user-facing features, breaking and compatibility changes, fixes, performance work, and internal changes - synchronize the canary version, SDK packages, bridge runtime identities, generated Node loader guards, and VS Code extension to `0.17.0` ## Verified comparison range - previous release: `baml-language-0.16.0` at `2e021b3429db8769656e9b8646657c1863789169` - reviewed product target: `baml-language-source-c0153b20108bc428b77a53128b9a59ddeb4e2b42` at `c0153b20108bc428b77a53128b9a59ddeb4e2b42` - the reviewed product target was also `origin/canary` when the updated notes were prepared - full product comparison: BoundaryML/baml@baml-language-0.16.0...baml-language-source-c0153b20108bc428b77a53128b9a59ddeb4e2b42 The product range contains 43 monorepo commits. The notes cover every merged PR in the range that touched `baml_language/`, plus the baml_language-specific macOS test suppression and binary-size baseline workflow update. The post-BoundaryML#4424 review added BoundaryML#4427, BoundaryML#4431, BoundaryML#4432, BoundaryML#4433, BoundaryML#4434, BoundaryML#4103, and BoundaryML#4435. BoundaryML#3975 and BoundaryML#3510 are in the monorepo comparison but are excluded because they change only the main engine/docs release line. Other unrelated documentation, website, JetBrains, on-call, and general monorepo changes are also excluded. The five commits in this PR add and refine the changelog and mechanically prepare release metadata; they do not add product behavior beyond the reviewed target. ## Release safety This PR remains draft. Its `baml_language/release.toml` change is the documented canary release request: merging it to `canary` after green CI would dispatch the production 0.17.0 release workflow. Do not merge or mark ready until Sam explicitly approves the release. No package was published, no release tag was created or pushed, no artifact was uploaded, and no production release workflow was triggered during preparation. ## Validation - `scripts/baml-language-version sync` - `scripts/baml-language-version check` - `scripts/baml-language-version show` → `0.17.0` - `scripts/baml-language-version compute --channel canary` → `0.17.0` - `python3 -m unittest scripts.tests.test_baml_language_version scripts.tests.test_release_pipeline_contract` — 39 passed - `cargo test --manifest-path baml_language/Cargo.toml -p baml_version` - `mise run validate-markdown` from `baml_language/` - changed-file `prek` hooks, including Cargo formatting, Clippy, C++ formatting, TOML checks, and Markdown validation - `git diff --check` <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **New Features** - Added system and filesystem capabilities, cryptography utilities, time arithmetic, assertions, bitwise operations, improved indexing, toolchain pinning, and LSP code actions. - Added stable merge sorting for array helpers. - **Bug Fixes** - Improved language behavior, diagnostics, CLI evaluation, playground workflows, and language-server synchronization. - **Breaking Changes** - Updated string APIs, function field requirements, prompt syntax, and type-checking behavior. - **Documentation** - Added release notes for version 0.17.0. - **Release** - Published version 0.17.0 across supported SDKs and tooling. <!-- end of auto-generated comment: release notes by coderabbit.ai -->
meefs
pushed a commit
to meefs/baml
that referenced
this pull request
Aug 15, 2026
# BEP-066: Evaluation, type construction, and reflection ## Summary This PR implements BEP-066 across the language, compiler, VM, package format, and host SDK boundary. BAML programs can inspect values and types, construct runtime types, compile and mount runtime packages, invoke reflected callables, and isolate dynamic work in sessions without weakening the static type system. The resulting model has four properties: - Reflection uses one sealed, canonical type algebra for all nine public reflection kinds. Equivalent types normalize to the same representation and stable digest. - Runtime generic arguments retain ordered static/runtime provenance through HIR, MIR, bytecode, serialization, and invocation. Runtime occurrences never become unconstrained solver variables. - Compiled and mounted packages preserve declaration identity, interface metadata, bounds, associated defaults, `requires` clauses, implementations, and callable targets without fabricated source locations. - Dynamic type information is lexically scoped. It can guide checking and execution inside its owning body, but values escaping that scope are erased back to their static occurrence type. ## BEP-066 scenarios | Scenario | Delivered behavior | Primary evidence | | --- | --- | --- | | 1. Runtime enums through an LLM | An enum obtained at runtime can be supplied to an `ai`/provider call and reflected back with the correct definition identity. | `baml_language/crates/baml_tests/tests/reflect_call_any.rs` | | 2. Saved forms become runtime classes | Stored rows can be converted into runtime class definitions; returned objects retain those definitions for extraction and field lookup. | `baml_language/crates/baml_tests/tests/runtime_classes_and_composites.rs` | | 3. Runtime tool unions | Runtime class definitions can be composed into a tool union, passed through model/tool dispatch, and recovered without losing member identity. | `baml_language/crates/baml_tests/tests/runtime_classes_and_composites.rs` | | 4a. Bounded generics over runtime classes | Runtime class witnesses participate in generic bounds and interface dispatch using the same conformance rules as static types. | `baml_language/crates/baml_tests/tests/runtime_interface_witnesses.rs` | | 4b. Package views over runtime definitions | Runtime package reflection exposes local declarations and their relationships with package-local identity. | `baml_language/crates/baml_tests/tests/runtime_type_bindings.rs`, `runtime_package_api_consistency.rs` | | 5. Model-written schema compilation | A schema produced as data can be compiled into a package, inspected, mounted, and used for typed extraction. | `baml_language/crates/baml_tests/tests/runtime_package_compile.rs` | | 6. Current and packed package lookup | `Package.current`, package enumeration, and `get_function` work for live and packed programs with loc-free callable targets. | `baml_language/crates/baml_tests/tests/runtime_package_compile.rs`, `baml_cli/tests/pack_e2e.rs` | | 7. Session isolation | Sessions provide isolated runtime definitions and calls while preserving the lexical lifetime and escape-erasure rules. | `baml_language/crates/baml_tests/tests/runtime_session.rs` | ## Consistency and audit hardening The implementation closes the consistency gaps found while auditing the end-to-end feature: - Constructor lookup has one behavior across parser, checker, and runtime: reserved words are rejected as constructor names, enum constructors retain their precise kind, and removed reader spellings resolve as ordinary missing names. See `baml_language/crates/baml_tests/tests/constructor_consistency.rs`. - Runtime rendering keys definitions by semantic identity, not display name. Non-equivalent same-name definitions fail before rendering with E0162; equivalent and recursive definitions remain renderable. See `runtime_render_identity.rs`. - Runtime and static failures share diagnostic codes and messages for bare generics, failed bounds, and duplicate serialized keys. Runtime-only diagnostics are structured and carry null spans instead of fake locations. See `runtime_diagnostic_consistency.rs`. - Package reflection includes `get_enum`, `get_interface`, kind-precise enumeration, null for wrong-kind or missing lookups, package-local identity, and distinct identity across packages. See `runtime_package_api_consistency.rs`. - Every declaration in a compiled package shares one identity across direct lookup, enumeration, and reflected function signatures. See `compiled_package_identity.rs`. - Runtime definition builders validate atomically, support recursive groups, and produce definitions identical to map-based construction. See `builder_witness_parity.rs` and `to_baml_witness_roundtrip.rs`. ## Canary reconciliations - [BoundaryML#4301](BoundaryML#4301) made `hir_ty` the production inference engine, removed TIR, and established `TYPE_SYSTEM.md` as the semantic authority. BEP-066 now lowers, checks, and reports diagnostics through that path only. - [BoundaryML#4352](BoundaryML#4352) reorganized the LLM surface around `ai`, provider namespaces, `@spec`, and agents. Reflection and model-call coverage targets that current surface rather than the superseded stdlib layout. - [BoundaryML#4367](BoundaryML#4367) removed Jinja prompt interpolation. BEP-066 prompts and fixtures use backtick strings and `${...}` interpolation, and obsolete Jinja metadata is not carried into runtime packages. ## `hir_ty` port contract The `hir_ty` implementation preserves the following invariants from syntax to execution: - Each explicit type-argument occurrence is recorded in source order as either a static type or a runtime operand. Static occurrences are solved normally; runtime occurrences are checked against their static occurrence type and never enter the solver as unknown variables. - The authoritative call metadata carries the resolved callable target, ordered type-argument slots, realized bindings, deferred dependent checks, and runtime definition identity through MIR and bytecode. Free functions, methods, and interface calls use symbolic, source-location-free targets. - Only checks that depend on runtime definitions are deferred. Static arity, ordinary argument types, and independent bounds remain compile-time errors. - Runtime type refinements live in an overlay keyed by body owner and statement identity, respect lexical shadowing, and are erased at scope escape. - Mounted package metadata is sufficient for normal generic realization and interface selection: bounds, associated defaults, `requires`, implementation registrations, and method resolution are not reconstructed heuristically at runtime. - `_` remains an exact contextual hole. Let annotations and constructors solve permitted holes; declaration signatures and explicit call/upcast type arguments reject them with E0147. A top-level `throws T | _` is open and exposes `T` plus inferred throws to callers; plain `throws T` is closed. Intentional differences from the retired TIR behavior are part of the contract: - Multiple generic bounds are conjunctive end to end. Runtime and static checking require every bound; the prior single-bound asymmetry is not preserved. - Existential interface dispatch permits exactly one `Self` witness. Ambiguous multi-`Self` shapes are rejected instead of being guessed. - Associated defaults, `Self` substitution, and `requires` realization follow the current spec and `hir_ty` rules where TIR behavior differed. - Current diagnostic codes and normalized types are authoritative; ignored TIR tests and legacy wording do not override them. The legacy root `engine/` is not an authority for this work. ## Test evidence map | Contract surface | Evidence | | --- | --- | | Canonical reflection algebra and stable identity | `baml_type` normalization/type-kind unit tests; type-spec tables for reflection kinds and canonicalization | | Syntax, formatting, AST, and HIR preservation | lexer/parser/formatter tests; HIR tests for ordered static/runtime type-argument occurrences | | Wildcard and open-throws semantics | `wildcard_hole_in_let_annotation.baml`, `wildcard_hole_in_constructor_generic_arg.baml`, `partial_throws_clause.baml`, `wildcard_type_inference.rs`, `wildcard_expression_holes.rs` | | Mounted-package and source-package parity | `hir_ty_package_interface`, `mounted_package_calls`, `mounted_package_parity` | | Runtime reflection scenarios | `reflect_call_any.rs`, `runtime_classes_and_composites.rs`, `runtime_interface_witnesses.rs`, `runtime_type_bindings.rs`, `runtime_package_compile.rs`, `runtime_session.rs` | | Identity, diagnostics, and API consistency | `compiled_package_identity.rs`, `constructor_consistency.rs`, `runtime_diagnostic_consistency.rs`, `runtime_package_api_consistency.rs`, `runtime_render_identity.rs` | | Builder and round-trip parity | `builder_witness_parity.rs`, `to_baml_witness_roundtrip.rs` | | Host and packed-program boundary | host reflection SDK fixtures and `baml_cli/tests/pack_e2e.rs` | ## Verification state - Full pinned gate on the gated head (`rustup run 1.93.0 cargo insta test --test-runner nextest -p baml_tests -p baml_cli -p baml_lsp2_actions -p baml_lsp2_actions_tests -p baml_surface --all-features --unreferenced=reject`): **3,265 passed, 0 failed** (23 skipped), all doctests passed or were intentionally ignored, no unreferenced snapshots, and no pending `.snap.new`. - Whole-stdlib type census: 121 files, **62,910 typed nodes, 0 error-channel entries, 0 panics**. - Session soak: the 500-evaluation pin runs with flat latency (eval BoundaryML#10: 3.32s → eval BoundaryML#500: 2.39s) and 0 KiB measured RSS growth. - The head commit additionally absorbs a final 4-commit canary drift (BoundaryML#4431 crypto, BoundaryML#3975 connection pooling, BoundaryML#4432 quote prompts, BoundaryML#4433 string standardization) verified by focused validation (targeted runtime/stdlib/compiler suites + no-update snapshot replays); the authoritative final signal is this PR's CI on the head commit. ## Handoff notes (for the next agent or human working on this) **State:** everything above is on the head commit. No temporary port constructs remain in the tree — the Phase-A stub (`E_BEP066_PORT_IN_PROGRESS`), fixture exclusions, and port-era test ignores are all gone (verified by grep). **Authoritative documents:** - `CONTRACTS.md` (repo root, this branch) — the binding data-model contracts (runtime type slots / `CallPlan`, loc-free `SymbolicCallableTarget`, scoped generic overlay) plus the 35-fact acceptance checklist with per-fact conditions. - TIR→hir_ty migration guide: branch `antoniosarosi/hir-ty-research` (`TIR_HIRTY_MIGRATION_GUIDE.md`) — the architecture map, worked examples, and risk register the port followed. - Wildcard adjudication table: branch `antoniosarosi/wildcard-adjudication` (`WILDCARD_VERDICTS.md`) — per-test rulings for the formerly ignored B-230/B-247 cases. - The full port audit trail (per-slice log with every checkpoint, decision, and accepted snapshot delta) is posted as a comment on this PR. **Deep review findings (resolved):** the adversarial review of the six port commits (`462dad20b..3c22664`) is complete — [full report in this comment](BoundaryML#4325 (comment)). All four ratified blockers landed in `f239a032f` and passed the widened pinned gate: 1. **Union interface dispatch (report finding 1):** virtual dispatch now requires every union member to resolve to the same declaring-interface view; heterogeneous unions take the guarded per-member switch, with executable coverage. 2. **Written union member order (finding 2):** static call slots retain a canonical checking type and a written emission type, preserving runtime coercion order. The audited snapshots restore `audio | image` and the other written forms. 3. **Deferred runtime bounds (finding 3):** bound registration substitutes the call frame and defers checks that depend on active scoped runtime bindings, including no-value-argument calls. 4. **Session top-level-let cycles (finding 5):** inference has an error cycle seed plus RAII in-flight ownership; resolution order preserves functions, exported values, and reserved package roots, and recursive lets diagnose without poisoning the Session. Canary integration and audited follow-ups are recorded in `55f6318f5`, `957d4514b`, `4fef90eff`, and `8de2d10bb`. Report findings 4, 6, and 8 remain agreed fast-follow work; finding 9 remains explicitly pre-existing. The report's clean-area conclusions remain valid. **CI playbook for this branch:** - darwin cargo-test: Blacksmith runner evictions ("No active SSH sessions") are infra flakes; cancelled jobs display as failures — check each job's `.conclusion`, then `gh run rerun <run-id> --failed` (up to 2 attempts). - tsweb-macos: same infra treatment. Its one real failure mode (vitest 5s default timeout on the worker `Package.compile` test) is already fixed with an explicit 30s timeout. - Size Gate: if it fails with the absolute cap **below** the reported baseline, the ceiling is stale — re-bake and bump the pinned literal in `baml_language/crates/tools_size_gate/src/config.rs` to baseline×1.03 as a single-file commit. - Local CI mirror (all three matter: toolchain, features, package set): `rustup run 1.93.0 cargo insta test --test-runner nextest -p baml_tests -p baml_cli -p baml_lsp2_actions -p baml_lsp2_actions_tests -p baml_surface --all-features --unreferenced=reject` **Agreed follow-up PRs (deliberately not in this PR):** 1. Add a nextest serialization group for `sdk_test_typescript_web` (`max-threads = 1`) — prevents concurrent ~8 GiB transient workerd import peaks from stacking (measured; this OOM'd a dev machine). 2. Investigate the 82.4 MB `bridge_web_core_bg.wasm` (wasm-opt is currently disabled in its package metadata); consider splitting compiler-only code out of the runtime web bridge. 3. Build-time precompiled-stdlib artifact + an `emit_units` variant that accepts it, so `Package.compile`'s first call skips the ~4–5 s full 90-file stdlib recompile (measured; a lazy in-process cache is not sufficient for the first-call cost). 4. Address report findings 4, 6, and 8 in focused follow-ups; triage the pre-existing items under finding 9 separately. **Coordination items:** - The two schema decisions living in hir_ty's crate (`SymbolicCallableTarget` in `callable.rs`; the `PackageInterface` Borsh expansion) should get a review from the hir_ty owner — they were frozen unilaterally under time pressure and intentionally invalidate cached package-interface bytes per compiler build. - GC: BEP-066's runtime-constructed type declarations live in the moving heap; the ty-heapptr line of work assumes GC-inert type heads. Alignment needed before that lands (`visit_heads` hook is reserved for this). **Semantic authority:** `TYPE_SYSTEM.md` + current hir_ty behavior. Do not "fix" anything back toward TIR-era snapshots or ignored TIR-era test expectations — several behavior deltas (conjunctive bounds, one-`Self` existential dispatch, hole/E0147 handling) are intentional and contract-pinned. ## Reviewer entry points - Public language and stdlib surface: `baml_language/TYPE_SYSTEM.md`, the BEP-066 specification, and `baml_language/crates/baml_builtins2/baml_std/baml/ns_reflect/`. - Shared type algebra: `baml_language/crates/baml_type/src/type_kind.rs` and `normalize.rs`. - Syntax and HIR preservation: AST/HIR type-argument reference definitions and lowering in `baml_compiler2_hir` and `baml_compiler2_hir_ty`. - Inference and mounted metadata: `baml_language/crates/baml_compiler2_hir_ty/src/infer.rs`, `callable.rs`, `package_interface.rs`, `impls.rs`, and interface method resolution. - MIR and bytecode handoff: the runtime type-argument lowering/provider code in `baml_compiler2_mir` and its bytecode emitter. - Runtime identity, reflection, and builders: the reflection/runtime-definition modules in `baml_language/crates/bex_vm`. - Behavioral review: start with the seven scenario tests above, then the audit-hardening tests, then the host SDK and pack coverage.
sxlijin
added a commit
to ryanmazzolini/baml
that referenced
this pull request
Aug 17, 2026
## Summary - bump all legacy BAML v0 release surfaces from 0.225.0 to 0.226.0 using the repository version-bump configuration - add exactly the two verified external-contributor engine changes: connection pooling from BoundaryML#3975 and OpenAI transcriptions from BoundaryML#4202 - synchronize the new 0.226.0 sections in the top-level and Fern changelogs - regenerate all nine integration clients without test-report churn and update the engine, Rust SDK, and generated Rust integration lockfile versions ## Release boundary - 0.225.0 boundary: release PR BoundaryML#4310 / merge commit da2a673 - BoundaryML#3975 shipped as 521919c - BoundaryML#4202 merged from reviewed head ff96fd4 as squash commit 79b8705; the reviewed-head and squash diffs have the same stable patch ID - BoundaryML#4464 and its byte-perfect revert BoundaryML#4469 have zero net diff and are excluded - rebased onto current canary after BoundaryML#4202 merged ## Validation - TypeScript runtime build and nine-client generation - Python runtime build/install and nine-client generation - deterministic regeneration: unchanged binary diff hash on repeated generation - locked Cargo metadata for the engine and Rust SDK workspaces - generated Rust integration library build: `cargo test --lib --locked` - OpenAI transcription wiremock closure tests: 2 passed - generated Python multipart request test: 1 passed - connection pooling transport test: 1 passed - HTTP config validation fixtures: 8 passed - release metadata and workflow contracts: 50 passed - stale 0.225.0 audit clean across release-owned surfaces; historical junit.xml intentionally unchanged - final release diff matches the 0.225.0 workflow-owned file set, excludes junit.xml test-report churn, and additionally refreshes the two tracked Rust SDK/integration lockfiles proven stale by locked builds ## Status - dependency satisfied; rebased onto current canary with the reviewed BoundaryML#4202 patch present - all required GitHub Actions checks are green; local and GitHub CodeRabbit reviews are green with approval and no review threads or requested changes - merge-ready; awaiting normal human review
Contributor
|
@dexhunter FYI, this has been released in 0.226.0! |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Issue Reference
Changes
Adds an explicit
http.enable_connection_poolingoption for regular HTTP-backed clients.false, preserving the current fork-safe default.fork()caveat, and unsupported client types.In a deterministic local HTTP/1.1 harness, ten sequential requests produced:
falsetrueThis is a 90% reduction in new connections for repeated calls while leaving existing behavior unchanged by default.
Testing
Validation performed:
cargo test -p internal-llm-clientcargo test -p baml-runtimeThe metric harness also verifies response payloads, streaming, retries, cancellation, concurrency, forced connection close, recovery, and a fresh forked child process.
Screenshots
Not applicable.
PR Checklist
Additional Notes
Applications using Python's
fork()start method should create the BAML runtime in each child process instead of inheriting a pooled client from the parent. The new option is disabled by default for this reason.Summary by CodeRabbit
New Features
http.enable_connection_pooling.Documentation
Bug Fixes / Validation
Tests