Changelog
Track our rapid iteration. We ship updates constantly to bring you the best local AI developer experience.
Stabilized Windows and Unix CI tests by replacing flaky shell wrappers (`pause`, `timeout`, `sh -c`) with a native Go `blocker` binary.
Bypassed `syscall.SIGTERM` signal usage on Windows to fix a persistent 5-second process kill timeout delay.
Core Engine Rebuild & Runtime System Redesign
Unified all inference backends under a common `Engine` interface (`llama.cpp`, `vLLM`, and `SGLang`), allowing seamless plug-and-play capability. Standardized core CLI execution via a stage-based execution pipeline (`Validate` ➔ `Download` ➔ `Resolve` ➔ `Probe` ➔ `MapFlags` ➔ `Launch` ➔ `Await` ➔ `Present`). Implemented the `process.Supervisor` shared execution layer to unify container/subprocess lifecycles, signal handling, watchdog-monitored startups, active health polling (`/health`), and concurrent persistent logging under `~/.cache/bloc/logs/`. Added capability-first flag mapping, dynamically resolving flag naming and syntax at runtime (e.g., `-fa` vs `--flash-attn on`) via help-probing instead of hardcoded string arrays. Security Hardening Removed dangerous shell execution patterns (`sh -c`) in pre-run script stages, replacing them with a secure allowlist-filtered system execution. Mitigated path traversal risks by introducing strict containment checking on all user-supplied paths via `filepath.Abs` and prefix validation. Enhanced API token privacy by preventing Hugging Face tokens from escaping into error messages or system logs. Implemented environment filtering to scrub sensitive environment variables (such as `LD_PRELOAD`) from child subprocesses. Performance Tuning & Correctness Cached CPU and VRAM capabilities probing via `sync.Once` to prevent expensive redundant python execution loops. Gated CUDA-based Docker testing behind a host-level `nvidia-smi` presence check, preventing automatic ~100MB downloads on CPU-only machines. Resolved nil-map panic vectors in local download operations and fixed deprecated flag compatibility issues with vLLM 0.10.0+.
Speculative Decoding & vLLM Reasoning Flag Updates
Updated speculative decoding flags to use the modern `llama-server` flags (`--spec-draft-model`, `--spec-draft-n-max`, `--spec-draft-p-min`) instead of deprecated/legacy aliases, preventing capability probe and startup failures on modern builds. Updated vLLM flag builder to automatically append the `--enable-reasoning` boolean flag when a reasoning parser is specified (such as `deepseek_r1`), enabling chain-of-thought extraction natively. Updated corresponding unit test suites to verify modern flags are built and probed correctly.
Flash Attention Flag Fix
Fixed a crash affecting all recipes with `flash_attn: true`. A breaking change in `llama.cpp` renamed `--flash-attn` from a boolean toggle (`-fa`) to a value-required flag (`--flash-attn [on|off|auto]`). The CLI was emitting bare `-fa`, causing `llama-server` to consume the next flag (`-b`) as the value and abort on startup. `BuildFlags()` now correctly emits `--flash-attn on`, and `RequiredFlags()` probes for the long-form flag. Tests updated to reject bare `-fa` emission.
Sync Tab State
Fixed an aggressive caching bug in the TUI. The Chat and History tabs now seamlessly synchronize session states dynamically.
Dynamic TUI Shortcuts & Recent History
Replaced hardcoded CLI tips in the Studio with context-aware shortcuts and dynamically linked the Recent Chats list to actual session history.
Introduced the Bloc Studio
A major update bringing a full-fledged Terminal User Interface (TUI) to Bloc for managing models, chatting, and viewing history.
CLI Self-Update Fallback
Implemented robust Homebrew upgrade detection and a seamless self-update fallback mechanism for the update command.
SGLang Engine Support
Added native integration for the highly optimized SGLang engine, enabling massive throughput improvements for production deployments.
Transitioned Deploy to Run
Refactored 'bloc deploy' to 'bloc run' for better ergonomics, while maintaining complete backward compatibility.
Expanded Security Blocklist
Migrated security rules to an embedded JSON deny-list and expanded the blocklist to proactively intercept 36 potentially dangerous engine flags.
Deadlock Fixes
Resolved a critical sync.RWMutex deadlock encountered during the DeleteCachedModel operation.
Update Pointer Migration
Successfully migrated the repository pointer in the update module to the new Bloc-ai/Bloc organization.
Cert Pinning Updates
Hardened security by updating Vercel GTS cert pinning hashes to the new WR1 intermediate certificates.
Windows Cross-Compilation Support
Resolved cross-compilation toolchain errors, stabilizing the Windows build pipeline.
Initial Release
The first stable release of Bloc, featuring the core CLI, dynamic recipe system, and native llama.cpp execution capabilities.