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 v2.1.222 tightens auto-mode safety, improves diff accuracy, and restricts Remote Control auto-start to user scope
└──▷ GET THIS VERSION
$ git clone --branch v2.1.222 https://github.com/anthropics/claude-code.git
# already have the repo? check out this version:$ git checkout v2.1.222
›Remote Control auto-start can no longer be enabled via repo-local .claude/settings.json or .claude/settings.local.json; enable it at user scope via /config (those files can still turn it off).
›Auto mode now routes messages sent to other agent sessions via SendMessage through the permission classifier before dispatch, preventing unreviewed cross-agent actions.
›The /diff view, Remote Control workspace diff, and file-edit diffs in Claude Code web sessions now use raw git blob content, bypassing workspace-configured diff drivers and textconv for cleaner output.
›Worktree-isolated sessions and their subagents can no longer run destructive git commands against the main checkout; isolation now covers file edits and Bash across every session type.
›The disable-model-invocation refusal now instructs Claude to ask the user to run the skill instead of attempting to replicate its workflow.
+1 moreshow less
›Removes the ultraplan feature.
└──▷ BREAKING ON UPGRADE
!Remote Control auto-start can no longer be enabled by repo-local settings (.claude/settings.json or .claude/settings.local.json); repos relying on those files to turn Remote Control on must switch to user-scope configuration via /config.
Validate a plugin before publishing to catch marketplace name rejections that would break managed sync.
$ claude plugin validate ./my-plugin
›Adds mode: "mask" for sandbox credential files on Linux and WSL — sandboxed commands read a sentinel copy (the whole file, or just spans captured by an extract regex) while the sandbox proxy substitutes the real value on egress; macOS falls back to deny.
›Adds a prompt-audit subcommand to the claude-api skill for auditing prompts and tool descriptions for patterns written for older models.
›Adds warnings to claude plugin validate when a marketplace or plugin name would be rejected by Claude Desktop's managed marketplace sync.
›Changes /status to show the session kind: interactive, or a background job that is attached or unattended.
›Changes sessions forked with /fork to create a new worktree of their own instead of working in the original session's checkout.
+15 moreshow less
›Changes plugins to accept "." as a skills path, with the root-level SKILL.md validation error now suggesting the plugin root.
›Changes /plugin install to refresh a stale marketplace catalog and retry before reporting a plugin not found.
›Changes plugins installed from /plugin to activate immediately when safe, instead of always requiring /reload-plugins.
›Changes the Gateway model field validation to reject non-string values with a 400 instead of forwarding them.
›Adds Focus view in VSCode: a chat-menu toggle that hides tool activity behind an expandable per-turn summary with a live running-tool indicator, toggled with Ctrl+Alt+F or the 'Claude Code: Toggle Focus view' command.
›Improves Stats panel to count cache tokens in its token totals, with a breakdown by input, output, cache read, and cache write.
›Re-enables tool search on Google Vertex AI for Claude 4.5-generation and newer models.
›Changes background sessions to commit and push to preserve work, open a draft PR only when the task calls for one, follow CLAUDE.md git instructions, and always report where the work lives.
›Changes emoji autocomplete to accept alternate shortcodes like :thumbsup:, :thumbsdown:, and :love:.
›Changes Claude in Chrome to close browser tabs it opens once it no longer needs them.
›Changes fast mode to report on the stream when usage credits run out mid-session, instead of failing silently.
›Changes Monitor: a watch that exits without producing output now says so instead of reporting 'stream ended'.
›Improves Windows startup to read process creation times via a native kernel32 call instead of spawning PowerShell, so endpoint security tools that gate powershell.exe no longer prompt.
›Improves /ultrareview error messages when a repo shares no history with its base: a checkout with no branches is now refused up front with advice to create one, and refusal hints no longer suggest git fetch --unshallow on already-complete clones.
›Reduces prompt-cache costs for auto-mode permission checks by reusing the cached conversation prefix across decisions.
Lightweight coding agent that runs in your terminal
Codex CLI now enables cached web search for Amazon Bedrock, falling back from unsupported live/indexed modes automatically.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.147.0-alpha.6.4 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.147.0-alpha.6.4
›Adds hosted web search support for Amazon Bedrock, advertising cached web search while marking external live and indexed web access as unsupported — unsupported modes resolve to cached search, or the tool is disabled when cached search is prohibited by managed requirements.
›Normalizes built-in and configured Bedrock model catalogs to text-only web search payloads, removing the search_content_types field that Bedrock rejects for multimodal search.
›Retains the runtime provider in session configuration so per-turn setup can apply provider-specific capabilities (e.g. text-only search for Bedrock).
Codex CLI raises MCP namespace tool-spec description limit from 1,000 bytes to 512 KiB, preserving complete server instructions.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.147.0-alpha.6.3 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.147.0-alpha.6.3
›Raises the MCP namespace tool-spec description limit from 1,000 bytes to 512 KiB, truncating only at a UTF-8 character boundary when the new limit is exceeded — preserving complete server instructions and preventing tools from being hidden.
›Keeps complete MCP namespace descriptions in tool-search source metadata.
Codex CLI alpha.7 adds dual-WebSocket transport for code mode, MCP conformance regression gates, and per-environment login shell policy.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.147.0-alpha.7 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.147.0-alpha.7
›Adds exec resume --last performance improvement by querying the state database first, avoiding a full rollout scan; falls back to scanning only on a complete database miss.
›Supports per-environment allow_login_shell policy, exposing the login argument for shell tools when any selected environment permits login shells and validating each command against its own environment's policy.
›Adds a dual-WebSocket transport for code mode, routing frames to the appropriate transport lane via transport_lane().
›Adds MCP client conformance regression gates covering HTTP and stdio transports, OAuth scenarios, shipping/intermediate/modern protocol versions, and an app-server regression matrix for transport, security, schema, pagination, SSE, multi-round, and catalog-boundary behavior.
›Adds codex-utils-audio workspace crate for canonicalizing audio inputs and estimating token usage.
+4 moreshow less
›Supports deferred loading for freeform tools and custom tools in namespaces.
›Adds timeout handling for stalled code-mode host requests.
›Identifies agents by name in token budget context.
›Prefers SQLite-backed names when resolving local archive, delete, and unarchive session targets before falling back to rollout scanning.
Qwen Code v0.21.5 adds a finding-centric Web Shell review view, Qwen 3.8 reasoning effort support, and safer MCP replay controls.
└──▷ GET THIS VERSION
$ git clone --branch v0.21.5 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:$ git checkout v0.21.5
›Enables Qwen 3.8 reasoning effort levels by mapping existing effort flags to the provider-specific reasoning_effort parameter.
›Introduces a structured, finding-centric view in Web Shell for review results, displaying severity, confidence, and suggested fixes alongside durable reports.
›Adds an opt-in one-time update bridge for macOS users to migrate from the Electron desktop app to the new Tauri shell.
›Introduces detailed execution-specific outcome tracking for tool calls to distinguish between invocation success, failure, and cancellation.
Run a cloud-only Kimi K3 model that has no default local tag without hitting a failure.
$ ollama run kimi-k3:cloud
›Adds stream_options.include_usage support to /v1/chat/completions streaming, now matching OpenAI's wire format: role only on the first chunk, finish_reason on its own trailing chunk, and usage in a separate chunk.
›Adds kimi-k3:cloud tag so ollama run kimi-k3 can pull cloud-only models that publish no default tag, instead of failing.
›Qwen3.5 on Apple GPUs now automatically uses the model's MTP head for speculative decoding via the MLX engine, improving inference speed.
└──▷ BREAKING ON UPGRADE
!Experimental image generation has been temporarily removed; users must continue using v0.32.5 for image generation support.
PydanticAI v2.23.0 adds cost tracking with cost and cost_limit, Bedrock extra_headers, and dynamic tool availability parts.
└──▷ GET THIS VERSION
$ git clone --branch v2.23.0 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:$ git checkout v2.23.0
└──▷ USE IT
Cap how much an agent run can spend by setting a cost_limit on UsageLimits alongside an existing token budget.
python
from pydantic_ai import Agent
from pydantic_ai.usage import UsageLimits
agent = Agent('openai:gpt-4o')
result = await agent.run(
'Summarize this document...',
usage_limits=UsageLimits(request_limit=10, cost_limit=0.05),
)
print(result.usage().cost)
Pass custom headers (e.g. for cost allocation tagging) to every Bedrock request via ModelSettings.extra_headers.
python
from pydantic_ai import Agent
from pydantic_ai.settings import ModelSettings
agent = Agent('bedrock:anthropic.claude-3-5-sonnet-20241022-v2:0')
result = await agent.run(
'Explain zero-trust networking.',
model_settings=ModelSettings(extra_headers={'x-amzn-bedrock-workload-name': 'sec-review'}),
)
print(result.output)
›Adds cost field to RunUsage and cost_limit field to UsageLimits to track and cap monetary spend per agent run.
›Adds extra_headers support in ModelSettings for Amazon Bedrock requests.
›Adds ToolAvailabilityDeltaPart with native tool_addition and additional_tools rendering to represent dynamic tool availability changes in agent message streams.
llama.cpp b10271 adds per-conversation working directory picker with path navigation to the server UI agent.
└──▷ GET THIS VERSION
$ git clone --branch b10271 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10271
›Adds per-conversation working directory picker to the server UI, supporting path navigation via POSIX (/, ~), Windows drive roots (C:), and UNC shares (//host/share) in addition to glob-matching.
›Extends file_glob_search on the server with hardened directory listing including kind enum, timeout, symlink guard, and absolute base enforcement.
›Persists per-call tool cwd on tool result messages in the database, so working directory context survives across conversation turns.
›Adds get_info tool call block rendering in the UI.
llama.cpp b10270 adds Qwen3-TTS support via a revamped llama-tts binary and new mtmd_helper_gen_audio API.
└──▷ GET THIS VERSION
$ git clone --branch b10270 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10270
›Adds mtmd_helper_gen_audio API for multimodal audio generation, enabling programmatic TTS workflows.
›Adds llama_model_get_tok_embd function to the llama library API.
›Adds mtmd_helper_model_can_chat helper to query whether a loaded model supports chat-style interaction.
›Adds step_prompt() method to the mtmd generation API for incremental prompt stepping.
›Supports Qwen3-TTS model in the mtmd multimodal framework, including text model, speaker encoder, code predictor, and GEN_WAV (audio generation) subgraph.
+2 moreshow less
›Revamps llama-tts binary with updated CLI interface and Qwen3-TTS support, including voice clone demo capability.
›Adds suppress_tokens support to the TTS backbone vocabulary pipeline.
└──▷ BREAKING ON UPGRADE
!The llama-tts binary interface has breaking changes as part of the Qwen3-TTS integration revamp — existing invocations will need to be updated.
llama.cpp b10255 extends SYCL oneDNN SDPA to Q4_0–Q8_0 and FP32 KV caches on Intel hardware.
└──▷ GET THIS VERSION
$ git clone --branch b10255 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10255
›Extends the oneDNN Scaled Dot-Product Attention (SDPA) path on SYCL to support quantized KV cache types Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, and F32 in addition to the existing F16 path, by dequantizing K/V to dense FP16 on-device before the fused systolic kernel runs.
›Non-F16 KV cache SDPA is gated to prefill scenarios only: requires sequence length K >= 1024 and query count Q >= 32; F16 KV continues to run at any length.
llama.cpp b10254 adds a DeepSeek V4 Flash 0731 chat template with reasoning-effort mapping and structured output support.
└──▷ GET THIS VERSION
$ git clone --branch b10254 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10254
›Adds a separate chat template for DeepSeek-V4-Flash-0731 with updated high and max reasoning-effort mappings.
›Adds structured output response-format instructions to DeepSeek V4 templates, passing the schema into template rendering.
›Defaults drop_thinking for DeepSeek V4 conversation history so prior reasoning is omitted unless preserve_reasoning is requested or tools are present.
›Adds DSV4_ANSWER_RESERVE environment variable to control how many tokens are reserved for the answer portion of DeepSeek V4 Flash output — set to 0 to disable the budget split entirely.
›DeepSeek V4 Flash now automatically reserves part of the output budget for the answer when thinking mode is active, using a bounded reserve (up to 2048 tokens) rather than a percentage — so long reasoning is not penalized at large token budgets.
Arize Phoenix v19.17.0 adds AI-powered query assistance for filter fields in the UI.
└──▷ GET THIS VERSION
$ git clone --branch arize-phoenix-v19.17.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:$ git checkout arize-phoenix-v19.17.0
›Adds AI Query assistance for filter fields in the UI, enabling natural-language construction of trace/span filters.
Phoenix v19.16.0 adds an OpenAI-compatible /v1/chat/completions proxy with server-side credentials and parallelized span export downloads.
└──▷ GET THIS VERSION
$ git clone --branch arize-phoenix-v19.16.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:$ git checkout arize-phoenix-v19.16.0
›Adds an OpenAI-compatible v1/chat/completions proxy endpoint to the Phoenix server, enabling server-side credential management for LLM requests.
›Parallelizes span export fetches and streams downloads to disk, unlocking faster large-scale trace exports from the UI.
›Adds preview card contents while collapsed in the UI, letting practitioners skim trace details without expanding each card.
Langfuse v4.4.0 adds in-app agent background execution with client/web support and expands session sidebar with span list.
└──▷ GET THIS VERSION
$ git clone --branch v4.4.0 https://github.com/langfuse/langfuse.git
# already have the repo? check out this version:$ git checkout v4.4.0
›Adds client/web support for 'in-app agent' background execution with a canonical hydration contract, enabling agents to run in the background from browser contexts.
›Increases the maximum number of steps available to in-app agents.
›Adds a span list to the modern session sidebar, giving visibility into individual spans directly from the session view.
›Forwards semantic root filters to v4 tables in dashboards, making dashboard filter behavior consistent across surfaces.
›Tracks association origin for media, recording how media attachments are linked.
›Adds POST /v1/search/{collection}/near-text REST endpoint for near-text search queries, with responses enveloped as {id, properties, references, metadata} and camelCase payload fields with a nested rerank object.
›Introduces /grpc-web endpoint, enabling gRPC-web protocol access to Weaviate.
›Adds GA resource-oriented index endpoints for the Alter Schema reindex feature (v1.39 RFC rework).
›Adds namespace suspend endpoints, RAFT state management, and DB-user status checks for namespace lifecycle control on shared clusters.
›Returns the first letters of API keys to admins on namespaced clusters.
+9 moreshow less
›Adds namespace graduation via backup/restore.
›Adds namespace-local roles for per-namespace RBAC isolation.
›Introduces drop-vector-index capability: supports removing a vector index from an existing collection property to reclaim disk space, with RBAC integration, multi-tenancy support, and cold-tenant completion.
›Adds cross-property AND matching in BM25 search, allowing queries to require term matches across multiple properties simultaneously.
THE AI TOOLCHAIN