The daily firehose — everything the toolchain shipped today, already filtered.
// HOW THIS ISSUE IS MADE
We read every release from the 174 tools on our watchlist at the source — GitHub and GitLab release notes, vendor release pages and changelogs, project blogs and feeds, vendor press releases, and the source code behind the tag. Bug-fix-only releases and non-product newsroom noise are dropped; what's left is summarized down to the new capability, how to try it, and any screenshots or videos the release itself published. Every entry links to the sources it was built from.
Cline CLI v3.0.20 names installed plugin wrappers from their source instead of opaque hashes.
└──▷ GET THIS VERSION
$ git clone --branch cli-v3.0.20 https://github.com/cline/cline.git
# already have the repo? check out this version:$ git checkout cli-v3.0.20
›Installed plugin wrappers are now named after their source (npm package name, git repo, remote filename, official slug, or local directory) instead of an opaque hash, making plugins easier to identify.
›Adds builtInAgents.rubberDuckAutoInvoke setting to control automatic rubber duck agent invocation (disabled by default).
›LSP server config now accepts bash, powershell, and cwd keys; cwd supports plugin vars like PLUGIN_ROOT, and command launch default cwd stays project-root unless cwd is overridden.
›Adds -r as a shorthand for --resume when reopening sessions.
›Adds --available-tools, --excluded-tools, and --reasoning-effort flags support in ACP mode.
›Adds maximum reasoning effort level for Anthropic models and makes all effort levels available on every plan.
+19 moreshow less
›Adds ACP allow_all config option to apply unrestricted permissions for tools, paths, and URLs.
›Adds vim-style navigation keys (g, G, Ctrl+D, Ctrl+U) to the /diff view.
›Adds billing help topic with an overview of AI credit usage features.
›Repurposes ctrl+s to stash and pop the current prompt (Claude Code parity); the slash-command picker remains available by typing /.
›/context now separates Custom Instructions from the system prompt and cross-references per-server MCP tool token costs with /mcp.
›/env command now shows hook counts and source provenance for active hooks.
›Shows cache write tokens alongside cache read tokens in the /usage display.
›Shows Mission Control sharing status of synced sessions in the /session info view.
›Shows session link in /session info table when remote exporting or steering.
›MCP tools added or removed by a server mid-turn are now available immediately in the same turn.
›Prompts for approval before running commands that set dynamic-loader or git-config environment variables (e.g. LD_PRELOAD, GIT_EXTERNAL_DIFF).
›BYOK file attachments larger than 5 MiB now send successfully via the OpenAI Responses provider.
›Rewind picker now shows working-tree diff stats (+added −removed) at each checkpoint.
›Enables creating a git worktree for a pull request directly from the pull requests screen.
›Auto-links bare #number issue and PR references to the current git repository.
›Enforces bypass permissions policy for --allow-all-tools, --allow-all-paths, and --allow-all-urls flags.
›Tab-completes .. parent traversal in slash-command path arguments instead of switching tabs.
›IDE picker appends a process id to entries sharing the same editor and folder, making git worktrees of the same repo distinguishable.
└──▷ BREAKING ON UPGRADE
!On Windows, executables are no longer discovered in the working directory when invoking by bare name (e.g. git). Add the working directory to PATH to restore previous discovery behavior.
CUDA PDL enrollment for MoE matrix-vector ops boosts MTP speculative decoding throughput by ~5–6% on Blackwell GPUs.
└──▷ GET THIS VERSION
$ git clone --branch b9521 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b9521
›Enrolls mul_mat_vec_q_moe into CUDA PDL (Programmatic Dependent Launch), improving Multi-Token Prediction (MTP) speculative decoding throughput by roughly 5–6% on Blackwell-architecture GPUs (e.g. B4500), with tok/s gains observed across code generation, summarization, QA, translation, and math workloads.
vLLM v0.22.1 adds Mellum v2 model support and zentorch-accelerated quantized inference on AMD Zen CPUs.
└──▷ GET THIS VERSION
$ git clone --branch v0.22.1 https://github.com/vllm-project/vllm.git
# already have the repo? check out this version:$ git checkout v0.22.1
›Adds W8A8 (int8 dynamic-symmetric) and W4A16 (GPTQ) linear inference through zentorch kernels on AMD Zen CPUs, registered ahead of generic oneDNN CPU kernels, with transparent fallback on non-Zen CPUs, GPUs, and XPU.
›Adds support for JetBrains' Mellum v2, an open-weights Mixture-of-Experts code-generation model.
Milvus v2.6.18 adds nullable vector fields, element-level Struct Array search, and HTTP/2 for the proxy REST server.
└──▷ GET THIS VERSION
$ git clone --branch v2.6.18 https://github.com/milvus-io/milvus.git
# already have the repo? check out this version:$ git checkout v2.6.18
›Adds HTTP/2 support for the proxy REST server, including h2c and ALPN-based TLS listeners.
›Adds support for importing Arrow FixedSizeList data from Parquet into non-nullable array and dense vector fields.
›Extends Arrow IO thread pool configuration to DataNode to improve compaction and import throughput.
›Introduces nullable vector fields: vector fields can now be declared nullable so entities with missing or not-yet-generated embeddings can be inserted without a placeholder; NULL vectors consume no extra storage and are skipped automatically during search.
›Enables element-level vector search on Struct Array fields, returning each result's matched element offset within the array rather than scoring the whole row.
+6 moreshow less
›Improves QueryNode read-task scheduling under heavy load with deadline-aware admission, cleanup, grouping, and metrics.
›Limits QueryNode delegator post-load concurrency to reduce CPU spikes during segment loading.
›Optimizes QueryCoord collection filtering in ChannelDistManager and reduces temporary allocations in distribution lookups.
›Optimizes ReplicaManager locking to reduce cross-collection contention in QueryCoord.
›Optimizes garbage collection for dropped segment index files and metadata.
›Improves REST timeout handling to safely discard late handler writes after request timeouts.
Weaviate v1.38 adds Namespaces (Preview), Nested Object Filtering (Preview), Runtime Property Reindex (Preview), and promotes HFresh to GA.
└──▷ GET THIS VERSION
$ git clone --branch v1.38.0 https://github.com/weaviate/weaviate.git
# already have the repo? check out this version:$ git checkout v1.38.0
›Introduces Namespaces (Preview) — control-plane and data isolation between users on a shared cluster, with RBAC/OIDC wiring, per-namespace collection limits, object limits, cascading delete, user management, alias endpoints, and audit log entries.
›Adds Nested Object Filtering (Preview) — enables search and filtering within indexed JSON properties, supporting IsNull, positional arr[N] filtering, Contains* operators, correlated AND resolution, scope-aware NOT, and gRPC + GraphQL ingress for nested filter paths.
›Adds Runtime Property Reindex (Preview) — allows changing a property's index type at runtime without recreating the collection, with two-phase RAFT swap barrier for semantic migrations and graceful-restart resilience for in-flight reindex units.
›Promotes HFresh index to GA, with asymmetric distance computation, query-vector normalization before rescoring, and reduced posting-map memory usage.
THE AI TOOLCHAIN