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.
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
Claude Code v2.1.225 adds gateway spend-limit warnings, workspace trust for agents, and cross-machine Remote Control messaging.
└──▷ GET THIS VERSION
$ git clone --branch v2.1.225 https://github.com/anthropics/claude-code.git
# already have the repo? check out this version:$ git checkout v2.1.225
›Adds gateway spend-limit support to the usage warning: the limit-reached message now names the cap, its reset time, and the operator's message (requires gateway on 2.1.225).
›SendMessage can now start a conversation with Remote Control sessions on other machines by name, as shown by ListAgents in the name [ref] format — previously only reply-after-first-contact was possible.
›Adds a workspace trust prompt to claude agents for untrusted directories, matching the existing behavior of claude.
›Remote Control: photos attached from the Claude app are now passed directly to Claude instead of being read from disk via a separate tool call.
Lightweight coding agent that runs in your terminal
Codex CLI adds migrate-rollouts subcommand, Markdown conversation export, session forking in codex exec, and a Guardian circuit-breaker for cyber models.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.148.0-alpha.5 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.5
└──▷ TRY IT
Inspect or migrate legacy local sessions to paginated thread history before resuming old conversations.
$ codex migrate-rollouts
›Adds migrate-rollouts subcommand to inspect or migrate legacy local sessions to paginated thread history.
›Adds Markdown conversation export to the TUI.
›Adds session forking to codex exec.
›Adds session archiving and restore to the resume picker.
›Adds MCP event discovery and subscriptions.
+10 moreshow less
›Mounts a minimal /dev in full-filesystem Bubblewrap sandboxes.
›Tracks context windows per agent.
›Exposes multi-agent versions in model/list.
›Exposes app-server diagnostics through the experimental API.
›Adds process diagnostics snapshots.
›Fetches remote installed plugins across all scopes.
›Recovers MCP servers after OAuth reauthentication.
Codex CLI alpha adds multi-agent version metadata in model/list, MCP event discovery, process diagnostics, and a shared project-instruction byte budget.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.148.0-alpha.4 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.4
›Adds nullable multiAgentVersion field to v2model/list responses, with supported values disabled, v1, or v2, exposing multi-agent capability metadata for each model.
›Adds codex-diagnostics crate that snapshots process ID, resident-memory measurements, and registered process-wide gauges; tracks live CodexThread instances via the core.threads.live gauge.
›Adds MCP event discovery and subscriptions, enabling clients to subscribe to server-side events over MCP.
›Includes call IDs in MCP requests for end-to-end traceability.
›Treats project_doc_max_bytes as a single shared byte budget across all selected environments, truncating at the limit and skipping later environments once exhausted, capping total project instruction payload size.
+7 moreshow less
›Exposes app-server and runtime activity diagnostics through the experimental API.
›Includes sandbox mode in response metadata.
›Keeps response streams alive through connection failures, improving resilience for long-running agent turns.
›Disables Nagle's algorithm for code-mode WebSockets, reducing latency for interactive sessions.
›Reports temporary directories in exec-server environment info.
›Includes tool namespace inventory in turn metadata.
›Generalizes skill locator aliases across providers and aliases resource-backed skill locators under context pressure.
Pause a running Dynamic Workflow from the TUI without aborting it, so you can inspect state and resume later.
$ /workflows p
›Adds activeWork, activeWorkReporting, and activeWorkStaleMs fields to the GET /health?deep=1 endpoint to expose current work state.
›Documents the --test-delta flag for severity adjustments when running /review, with a warning if the executing bundle is older than the current code.
›Adds optional Mem0 memory write support to the Direct External Context integration, enabled explicitly in configuration.
›Adds a repository context manifest to guide the /review command with bounded domains, related paths, and recommended tests for specific repository areas.
›Enables compression cache sharing for all OpenAI-compatible providers, including stable session keys for the official OpenAI API, to reuse conversation prefixes and reduce redundant input processing.
+11 moreshow less
›Enables compression cache sharing for Gemini and Vertex AI to reuse conversation prefixes and reduce redundant input processing.
›Supports installing Web Shell Extensions directly from local .zip or .tar.gz archive files via the Extension manager.
›Dynamic Workflows now support cooperative pause and resume via the p shortcut in Background Tasks or the /workflows p command in the TUI.
›Restores real-time autofix support for pull requests opened from forks by bridging review events to credentialed workflows.
›Allows ACP agent tool calls to run concurrently at the configured limit, removing artificial serialization that slowed fan-out operations.
›Group chats can now be approved once via a stable chat ID using the new group pairing policy, granting access to all members.
›Enriches Feishu contact observations with sender display names and group names to replace opaque IDs with recognizable labels.
›Feishu users can now answer structured single-select and multi-select questions directly via native Card V2 forms instead of text replies.
›Enhances DingTalk integration to include stable identifiers of mentioned members in inbound context when the bot is mentioned alongside other users.
›The daemon now observes and reports memory pressure levels and ratios using real denominators for both the root process and child heaps.
›Improves telemetry attribution to distinguish daemon-spawned sessions from direct CLI launches in usage statistics.
Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.
CrewAI 1.15.14 splits runtime context from the coding agent and introduces project ID support.
└──▷ GET THIS VERSION
$ git clone --branch 1.15.14 https://github.com/crewAIInc/crewAI.git
# already have the repo? check out this version:$ git checkout 1.15.14
›Separates runtime context from the coding agent, allowing project ID to be tracked independently per execution.
$ git clone --branch v2026.6.33 https://github.com/openclaw/openclaw.git
# already have the repo? check out this version:$ git checkout v2026.6.33
›Adds extended-stable channel support so package installations can select, update from, and receive availability notices for the extended-stable release line without silently falling back to another channel.
›Scopes external MCP loopback clients to short-lived session-bound attach grants instead of inheriting mutable child-process authority, limiting lateral movement from compromised MCP sessions.
›Restricts narrow tool allowlists to the factory that constructs them, preventing privilege escalation through allowlist inheritance.
›Ensures Gateway message actions retain trusted requester provenance and reject untrusted callers, closing an authorization bypass in action bridges.
›Improves long-running agent reliability by distinguishing genuine stalls from active long model calls and wedged backends in run-release, liveness-check, and watchdog logic.
+1 moreshow less
›Preserves SecretRef-backed Telegram credentials across service restarts, preventing credential loss on restart.
Switch DeepSeek-MLA DCP decode to the FlashInfer MNNVL kernel on GB200 and skip the per-layer Q all-gather to reduce cross-rank communication overhead.
Accelerate large MoE prefill by using DWDP to prefetch expert weights over NVLink P2P instead of dispatching tokens all-to-all across expert-parallel ranks.
›Adds --dcp-comm-backend {ag_rs, a2a, fi_a2a} to select pluggable communication backends for the DeepSeek-MLA decode context-parallel path; a2a exchanges packed attention output plus fp32 LSE in a single NCCL collective per layer with fp8 KV as uint8 byte transport, and fi_a2a delegates cross-rank exchange to the FlashInfer MNNVL kernel on GB200.
›Adds --dcp-replicate-q-proj flag to project full-head Q locally and skip the per-layer Q head-dim all-gather in DCP decode.
›Adds --dwdp-size to enable DWDP (Download-While-Dispatch-Prefill) parallelism for MoE prefill, which prefetches peer expert weights over NVLink P2P and computes all experts locally, removing EP all-to-all token dispatch; achieves 1.92x over DEP4 at MNT 32K / ISL 32K on 4x B200 with gpt-oss-120b.
›Adds --enable-session-radix-cache flag and /close_session endpoint for session-reference-aware Unified Radix Cache, allowing requests to carry a stable session_id so eviction respects prefixes still referenced by active sessions — targeted at agentic and RL-rollout workloads.
›Adds SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE=1 environment variable to enable the DeepGEMM MegaMoE A2A path on SM90 for DeepSeek-V4-Flash/Pro FP8, including the pre-dispatch JIT kernel and FP8 expert weight preparation.
+12 moreshow less
›Adds SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D environment variable to copy oversized or non-contiguous CPU weight views into contiguous storage before H2D transfer, cutting DeepSeek-V4-Pro TP8 load time from ~35 minutes to 6m20s (5.6x) and GPT-OSS-20B BF16 from 545s to 70s (7.8x); off by default.
›Adds day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE with 1M-token context, served with DCP, DSpark speculative decoding, chunked-prefill PP with TP decode, KDA-aware prefix caching, HiCache L2 over DCP, LoRA on quantized weights, and OpenAI-compatible serving; verified on NVIDIA GB300 and AMD MI35x.
›Adds day-0 support for MiniMax-H3 video-and-audio generation on SGLang-Diffusion across task profiles t2va (text-to-video-and-audio), fl2va (first/last-frame conditioning), and ref2va (image/video/audio reference conditioning); verified on B200 (TP2 + Ulysses4), H100, AMD MI300X/MI355X, and 2x RTX 5090 with layerwise offload.
›Adds an initial Rust frontend that replaces the Python server path from network ingress through tokenization and request hand-off to the GPU scheduler with a multi-threaded Rust implementation, including tokenizer manager, ingress validation, egress, OpenAI-compatible API server, and PD disaggregation support; shipped as prebuilt release artifacts.
›Adds a weight-cache daemon that holds GPU weights so a restarting engine can recover from cache rather than reloading from storage and recapturing CUDA graphs, reducing restart time for large models (e.g., Qwen3-235B FP8 on 4 GPUs from ~6.5 minutes).
›Adds support for EmbeddingGemma and LFM2.5 embedding models, nvidia/MiniMax-M3-NVFP4, and cookbook recipes for Poolside Laguna-S-2.1 and Inkling-Small.
›Adds trtllm_mha support for Gemma 4 MTP draft attention backend in speculative decoding.
›Adds grammar-constrained decoding support (including tool_choice=auto) for DSpark speculative decoding.
›Adds grammar-constrained decoding support in DFLASH speculative verify.
›Adds grammar overlap scheduling for standalone speculative decoding.
›Removes unnecessary expert padding on AMD, dropping MI355X FP4 MoE model weights from 159.07 GB to 112.36 GB and cutting colocated RL training-phase footprint from ~143 GiB to ~87 GiB per GPU.
›Updates dependencies: flashinfer 0.6.15.post1, sgl-deep-gemm 0.1.5.post1, helion 1.4, mooncake 0.3.12.post1, dynamo-tokenizers 1.7.0; PyTorch remains at 2.11.0 and CUDA base image at 13.0.1.
llama.cpp server tools now declare working-directory support, so the WebUI /cwd control only appears when a relevant tool is active.
└──▷ GET THIS VERSION
$ git clone --branch b10329 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10329
›Built-in tools now publish a working-directory flag in the /tools listing, indicating whether they resolve paths against the working directory — the WebUI shows the working-directory chip and enables the /cwd command only when at least one such tool is both served and enabled.
llama.cpp server gains Docker-backed tool isolation via x-tool-runtime for sandboxed LLM tool execution.
└──▷ GET THIS VERSION
$ git clone --branch b10328 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10328
›Adds initial tool isolation support to the server via Docker, separating tool execution into tools_io_sandbox and tools_io_docker modes, exposed through the x-tool-runtime request header.
vMLX 1.6.25 delivers a new indexed-attention prefill kernel for DeepSeek V4 Flash, live KV cache rewind on retry, and honest 413 errors for over-ceiling context requests.
└──▷ GET THIS VERSION
$ git clone --branch v1.6.25 https://github.com/jjang-ai/vmlx.git
# already have the repo? check out this version:$ git checkout v1.6.25
›New indexed-attention prefill kernel (on by default) boosts DeepSeek V4 Flash prompt processing from 347 to ~449 tokens/s at 15k context (+29%) and from 199 to ~316 tokens/s at 40k context (+58%), unlocking practical use of very long prompts.
›Aborting and retrying a long prompt now rewinds the live KV cache to the shared prefix instead of re-prefilling from scratch, cutting retry time-to-first-token at 15k context from 46s to 2.3s.
›Two-pass answer path now reuses the first pass's KV cache, eliminating full prompt re-prefill overhead on second-pass generation.
›KV pool quantization keeps cache RAM under 8 GB even near the model's maximum context length.
›Engine now advertises the true hardware context ceiling at startup, shrinks prefill chunks adaptively under Metal memory pressure, and returns a clean 413 instead of a Metal OOM for over-ceiling requests — eliminating the flat request timeout that killed healthy very-long prefills.
An MCP that lets AI tools securely connect to your infrastructure, write IaaS code, debug issues, and assist during incidents - without risking production stability. Built for security teams to approve and infrastructure teams to experience like magic.
Emisar v0.38.0 adds policy override validation warnings and fixes session disconnect and runbook size reporting.
└──▷ GET THIS VERSION
$ git clone --branch v0.38.0 https://github.com/AndrewDryga/emisar.git
# already have the repo? check out this version:$ git checkout v0.38.0
›Adds a live validation warning when a policy override glob pattern cannot match any action ID — for example, regex-style patterns like cassandra\.drop_* — so deny rules that silently protect nothing are caught while writing, not after deployment.
›Session termination now disconnects the active console session immediately, not only the backing cookie, so administrators ending sessions during an incident actually cut access.
›Runbook size failures now report the actual size rather than appearing as a missing record, and character-limit bounds on title and description fields are now derived from the byte budget rather than assumed, preventing multibyte-encoded descriptions from silently vanishing from list_runbooks and get_runbook.
›New EMISAR_GROUP and EMISAR_RUNNER_ID environment variables let operators relabel a fleet without editing configs on each host.
›Official multi-architecture container image published at ghcr.io/andrewdryga/emisar-runner with build provenance and an SBOM.
›Runner tokens now carry a 90-day bounded lifetime and self-refresh two-thirds of the way through over the existing connection, with no host access required; expired tokens are refused at connect.
›Action children now start with no_new_privs, preventing execve inside a pack from picking up setuid or file-capability privileges the runner does not already hold.
›Reads that can return credentials are now approval-gated; sensitive run values are masked in a single pass so one match cannot rewrite another's marker.
+8 moreshow less
›MFA enrollment and recovery-code regeneration now require proof of the current inbox; credential step-up codes are rate-limited across the cluster rather than per node.
›Every pack's curl is confined to an explicit protocol with globbing off, preventing URLs arriving in API responses from expanding into extra transfers or exfiltrating credentials.
›Signed MCP dispatch now signs the narrative a human approver actually reads (attestation v5), and the bridge verifies that narrative rather than a reconstruction of it.
›MCP bridge signing key now comes from a pinned credential directory instead of the environment.
›Runbook targets can name a runner group in the model contract, and an agent can revise and test a runbook draft before a human publishes it.
›Catalog now carries 95 packs and 1,498 actions, adding Apache Airflow, Spark, Google Cloud billing, and BunnyCDN.
›Console runbook target selection adds a stable trigger naming chosen targets, a searchable roster of dense one-line rows, and scope icons encoding cardinality.
›Backlogged run output ships as one frame instead of one per line, and the connection lease renews at half its life instead of on every heartbeat.
└──▷ BREAKING ON UPGRADE
!Action children now start with no_new_privs, so a setuid or setgid helper in a runner's process tree no longer elevates. For example, postqueue (setgid postdrop) can no longer grant access to the Postfix queue — the runner user must own the queue directly or be a postdrop member.
arize-phoenix-otel v0.17.0 adds support for a Phoenix environment file for configuration.
└──▷ GET THIS VERSION
$ git clone --branch arize-phoenix-otel-v0.17.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:$ git checkout arize-phoenix-otel-v0.17.0
›Supports a Phoenix env file for configuring arize-phoenix-otel settings via a file-based environment variable approach.
Arize Phoenix Evals 3.4.0 adds a built-in hallucination evaluator.
└──▷ GET THIS VERSION
$ git clone --branch arize-phoenix-evals-v3.4.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:$ git checkout arize-phoenix-evals-v3.4.0
›Adds a hallucination evaluator to the evals library for detecting hallucinated content in LLM outputs.
THE AI TOOLCHAIN