solverforge-cli 3.7.x: Parallel Evaluation and Per-Phase Telemetry

solverforge-cli 3.7.1 was published on October 27, 2026, following 3.7.0, which introduced this line’s runtime and telemetry changes after 3.6.0 scaffolded solverforge 0.22.0. The 3.7.x packages target solverforge 0.23.0; the web UI, maps, and MCP targets remain independent. Existing applications keep their own manifests and generated source.

What Changed

MCP scaffolds use rmcp 3.8.0 and ServerConfig

The 3.7.1 patch updates the MCP shell from rmcp 3.5.1 to rmcp 3.8.0 and replaces the deprecated ServerInfo alias with ServerConfig in generated server code. Existing MCP apps should update their manifest and the import, return type, and constructor in src/mcp/server.rs; installing the CLI does not rewrite existing source.

The generated server remains stateless over Streamable HTTP and makes no server-to-client sampling, elicitation, or roots requests. The twelve tools, task-capable and legacy lifecycle paths, and stdio JSON-RPC contract remain unchanged. Runtime 0.23.0, UI 0.9.0, and maps 2.1.4 targets are unchanged.

The patch also corrects the scaffold test to compare the commented thread-count block line by line, accepting Windows CRLF endings without changing the configuration template or accepted settings.

Parallel local-search evaluation is available, not enabled by default

Fresh scalar and list templates document an opt-in setting:

solverforge config set move_thread_count.count 4
solverforge check

The default auto stays sequential. Parallel local search scores candidates on several evaluators while one coordinator preserves sequential acceptance order. Construction remains sequential. The effective count is capped by the Rayon pool and by 32; concurrent jobs share one process-wide pool controlled by RAYON_NUM_THREADS. Small models can be slower with more evaluators.

The CLI’s check and config set commands accept only "auto", "none", or a table containing one non-negative integer count. Bare integer values, negative counts, extra table fields, and non-integer counts are rejected. Generated apps embed solver.toml and parse it when solving, so rebuild and restart an app after changing the setting. There is no per-job thread override.

Telemetry records every started phase and speculative work

Generated telemetry replaces the single phase object with a phases array in execution order. Run totals and each phase carry movesSpeculative; candidate traces label discarded ahead-of-decision work as discarded_speculative. This is a breaking wire-shape change for consumers of generated status, snapshot, and typed SSE payloads. The web UI target remains solverforge-ui 0.9.0, which reads neither phase key.

movesGenerated includes speculative work. Use movesEvaluated when comparing logical evaluation work. Detailed candidate pulls remain separate from ordinary control-plane payloads and require opt-in candidate tracing.

The verification scenarios cover fixed-work parity

The updated mixed runtime scenario compares seeded, step-bounded sequential and parallel solves: solution, score, logical counters, and ordered consumed candidate pulls must match. Only the parallel local-search run may report speculative work. MCP scenarios enable parallel evaluation on a pinned Rayon pool, including the byte-level check that stdio stdout remains JSON-RPC only. The 3.7.0 artifact’s exact-commit CI failure and the successful 3.7.1 validation are recorded separately below.

Publication and exact-commit validation are separate

The 3.7.1 crate carries source SHA fcaad8e3ef59444cb28f7e4742494e2b0cbb64ca in .cargo_vcs_info.json. Its SHA-256 is ba476ea2862d6e0a6e2f0a9a434428d2887f8863abdd1e932b6ef87598d7d538, matching the crates.io index. The exact-commit publication workflow, GitHub Release workflow, and CI run all completed successfully. CI passed Linux, macOS, Windows, and Integration Tests.

The 3.7.0 crate carries source SHA ab79f437bf7694bbb166d1247cde30c5c5a00e42 in .cargo_vcs_info.json. Its SHA-256 is 41e70042260d7ed88e7f310822f3c0f583dea2d5dd28fa92e48820919d72841e, matching the crates.io index. The publication workflow and GitHub Release workflow succeeded for that commit.

The exact-commit CI run failed: Linux and macOS Rust jobs passed, but Windows failed test_new_creates_neutral_project_files on the commented thread-count block assertion. Windows reported 42 scaffold tests passed and one failed; its network-retry step and the dependent Integration Tests job were skipped. The later main-branch CI run succeeded at 8c9d62d780730e61d53ab7e882d48ec5e77d2886; that is not validation of the original artifact commit.

Published Version Boundaries

Patch History

Upgrade Checklist

• To stay on this line, install with cargo install solverforge-cli --version 3.7.1 --locked --force and inspect solverforge --version.
• Upgrade existing applications deliberately: installing the CLI does not regenerate their telemetry DTOs or change their dependency manifests.
• For an existing MCP app, update rmcp to 3.8.0 and replace ServerInfo with ServerConfig in the server metadata implementation; run cargo check.
• Replace consumer assumptions about a single phase record with the phases array and retain speculative counters separately from logical work.
• Leave parallel evaluation off until a fixed-work comparison and real-instance timing justify it; rebuild after changing solver.toml.
• Run model, configuration, lifecycle, and transport tests after upgrading.

Where to read next

• CLI manual
• Configuration
• MCP shell
• Complete 3.6.0 to 3.7.0 comparison
• Complete 3.7.0 to 3.7.1 comparison
• GitHub Release