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 v2.1.233 adds GitLab MR support, per-user identity forwarding, memory cgroup limits, and WebFetch cache TTL control
└──▷ GET THIS VERSION
$ git clone --branch v2.1.233 https://github.com/anthropics/claude-code.git
# already have the repo? check out this version:$ git checkout v2.1.233
└──▷ TRY IT
Cap memory usage of Bash tool subprocesses on Linux so a runaway build cannot exhaust the host and stall the session.
$ export CLAUDE_CODE_TOOL_MEMORY_LIMIT=2G
claude
Extend the WebFetch URL cache lifetime to avoid redundant fetches during long-running research sessions.
$ export CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS=3600000
claude
Re-enable todo/task-tracking tools when working with a newer model (e.g. Sonnet 5) that disables them by default.
$ export CLAUDE_CODE_ENABLE_TODO_TOOLS=1
claude --model claude-sonnet-5
›Adds CLAUDE_CODE_TOOL_MEMORY_LIMIT environment variable (opt-in) to enforce memory cgroup limits on Bash tool commands on Linux, preventing runaway builds from stalling a session.
›Adds CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS environment variable to configure the WebFetch session URL cache TTL (default: 15 minutes).
›Adds opt-in forward_user_identity setting on Anthropic apps gateway upstreams, sending the signed-in user's identity as headers so a proxy can attribute spend per user.
›Adds CLAUDE_CODE_ENABLE_TODO_TOOLS=1 environment variable to re-enable TodoWrite/TaskCreate/Get/Update/List tools on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models (now disabled by default on those models).
›Extends the --worktree flag and claude agents view to support GitLab merge request URLs, displaying MRs as !N.
+3 moreshow less
›Improves claude plugin validate to check bare .claude/skills directories and report SKILL.md files with unparseable frontmatter.
›Adds a [claude-code:unrecognized_model] diagnostic line to stderr in print mode when a request targets an unrecognized model ID; silence it by mapping the model with modelOverrides.
›Improves claude self-hosted-runner session start time by creating the session branch without rewriting the working tree and removing two blocking server round trips.
└──▷ BREAKING ON UPGRADE
!Todo/task-tracking tools (TaskCreate, TaskGet, TaskUpdate, TaskList, TodoWrite) are no longer available by default on Opus 4.8, Sonnet 5, Fable 5, Mythos 5, and newer models; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to restore them.
Cline v4.1.10 adds web search during tasks and Crusoe as a new model provider (SDK bundle only).
└──▷ GET THIS VERSION
$ git clone --branch v4.1.10 https://github.com/cline/cline.git
# already have the repo? check out this version:$ git checkout v4.1.10
›Adds web search capability for supported models, enabled via a toggle in Feature Settings; search calls and results appear in the conversation and persist across reloads.
›Adds Crusoe as a new model provider in the model catalog, alongside updated model lists and per-provider default models.
Cline CLI 3.0.55 adds Crusoe as a model provider, improves cline doctor fix reporting, and renders web search results in the transcript.
└──▷ GET THIS VERSION
$ git clone --branch cli-v3.0.55 https://github.com/cline/cline.git
# already have the repo? check out this version:$ git checkout cli-v3.0.55
└──▷ TRY IT
Check the health of your Cline Hub and get an accurate breakdown of process states after attempting repairs.
$ cline doctor fix
›The cline update command now installs immediately and explicitly reports that the update applies on next start, rather than swapping the package under a live session.
›Adds Crusoe as a provider in the model catalog, along with updated model lists and per-provider default models across all providers.
›Web search calls and their results from models that run search natively now render in the transcript.
›Auto-updates no longer install while a CLI is attached to the Hub — the update is recorded at startup and applied on exit once the Hub confirms no other clients are attached.
›Streaming assistant markdown no longer flashes back to raw text — settled headings, links, and code stay rendered as new chunks arrive.
+1 moreshow less
›Idle plugin sandbox processes are now reclaimed instead of lingering for the life of the session.
Cline sdk/sdk/v0.0.75 adds provider-executed web search, a dedicated Cline gateway provider, and improved cline doctor fix diagnostics.
└──▷ GET THIS VERSION
$ git clone --branch sdk/sdk/v0.0.75 https://github.com/cline/cline.git
# already have the repo? check out this version:$ git checkout sdk/sdk/v0.0.75
└──▷ HOW TO FIND IT
Enable provider-executed web search so a model can fetch live results mid-turn, with searches persisted in session history.
📍In Settings, locate the model tools section and enable the web_search model tool.
›Adds provider-executed web search via the web_search model tool in settings; search calls and results are persisted in session history and replay on reload. Off by default.
›Adds a dedicated Cline provider for the Cline gateway, replacing the generic OpenAI-compatible path, so extended thinking budgets and other gateway options now apply correctly to both cline and cline-pass.
›Adds Crusoe as a provider in the refreshed model catalog, with updated model lists and per-provider default models.
Lightweight coding agent that runs in your terminal
Codex CLI alpha.18 enables unified exec on Windows by default, makes Guardian v2 risk classification configurable, and removes the gRPC code-mode session limit.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.148.0-alpha.18 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.18
›Makes features.guardianv2 in config accept either a boolean toggle or a structured object specifying classifier instructions, review threshold, reasoning effort, action/instruction token limits, transcript source controls, per-entry and total token budgets, and the number of recent non-user entries.
›Adds an override to skip project configuration at startup.
›Enables unified_exec by default on Windows, exposing exec_command and write_stdin instead of shell_command on that platform.
›Removes the gRPC code-mode open session limit, allowing more than MAX_IN_FLIGHT_REQUESTS concurrent sessions.
›Adds MCP protocol discovery metrics: records a counter and duration per discovery attempt, tagged with legacy or auto mode and classified as modern, legacy, or failure.
+1 moreshow less
›Routes 'None of the above' selection in request input prompts to the notes editor on Enter (in addition to Tab), preventing accidental submission.
Codex CLI alpha.16 adds Guardian V2 risk scoring, parallel tool calls for all models, and stricter MCP auto-review enforcement.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.148.0-alpha.16 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.16
›Enforces strict auto-review for MCP tool calls even when the approval policy, tool annotations, or a remembered session decision would otherwise skip review.
›Adds Guardian V2 risk scoring: initializes per-thread samplers, records risk scores as snapshots on threads, preserves transcript boundaries in sampling input, and requires automatic review for high-risk actions.
›Installs Guardian V2 in the app server and isolates Guardian reviewer sessions from parent extensions.
›Reuses pending MCP server connections during reconciliation when identity, catalog limit, and protocol mode still match, avoiding duplicate startups.
›Adds agent_name field to Responses API turn metadata, falling back to /root when no canonical agent path exists, and reserves the key from client metadata overrides.
+3 moreshow less
›Instructs the model to use tool_search to discover relevant tools when an explicitly selected plugin has apps available, before falling back to built-in tools.
›Restricts filesystem helper sandbox access for tighter isolation.
›Honors cloud-managed requirements in feature listings.
Codex CLI adds an Amazon Bedrock Runtime provider and skill-level model annotations with Luna delegation support.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.148.0-alpha.15 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.15
›Adds the built-in amazon-bedrock-runtime provider for regional OpenAI-compatible Bedrock endpoints, with SigV4 auth, per-provider AWS profile/region/transport overrides, global and US cross-region model variants, and web search disabled.
›Adds an optional model field to skill frontmatter (e.g. model: luna), enabling per-skill model annotations while silently ignoring unsupported values.
›Adds SkillModelDelegationInstruction so skills requesting Luna can delegate to it when Luna is available in the current provider namespace, with bounded/validated model identifiers and instruction size limits.
$ git clone --branch rust-v0.148.0-alpha.14 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.148.0-alpha.14
└──▷ USE IT
Pin a specific MCP server to its own OAuth callback port so it does not conflict with other servers sharing the global port.
yaml
oauth:
callback_port: 9876
›Adds oauth.callback_port to MCP server configuration, letting each server override the global mcp_oauth_callback_port for CLI login, app-server, plugin installation, executor, and skill dependency OAuth flows.
›Adds experimental thread queue APIs to the app server.
›Adds a running-task exit menu to local daemon sessions: pressing Ctrl-C with an empty composer while a task runs now offers choices to cancel and stay, exit while leaving the task running, or stop and exit.
›Exposes model upgrade retirement times in the model context.
›Preserves client-authored developer messages across context compaction when retain_client_developer_messages is enabled.
+5 moreshow less
›Provides Guardian V2 with full tool action context — conversation transcript and a structured planned-action object containing tool name and arguments — for more accurate risk classification.
›Wraps injected current-time reminders in <current_time_reminder> tags in model context.
›Deduplicates current-time reminders when spawning full-history subagents, replacing inherited parent reminders with a single fresh reminder.
›Adds structured telemetry for response retries.
›Includes node_repl images in Guardian review evidence.
Pi v0.84.2 adds fullscreen transcript search, configurable default tools, a per-run theme flag, and new extension/gateway APIs.
└──▷ GET THIS VERSION
$ git clone --branch v0.84.2 https://github.com/earendil-works/pi.git
# already have the repo? check out this version:$ git checkout v0.84.2
└──▷ TRY IT
Apply a specific TUI theme for a single run without overwriting your saved theme preference.
$ pi --use-theme dracula
Enable experimental strict JSON-schema constrained sampling for built-in tools to tighten model output conformance.
$ PI_EXPERIMENTAL=1 pi
Set the default tools available at startup for every project, then override for a specific project.
json
{ "defaultTools": ["read", "bash", "edit"] }
›Adds --use-theme <name[/name]> flag to choose a per-run interactive theme without changing saved settings.
›Adds defaultTools setting for configuring the initial built-in tool selection globally or per project.
›Adds expandPromptTemplates option to the extension pi.sendUserMessage() API for explicitly dispatching commands and expanding skills and prompt templates.
›Adds createGatewayBindingFetch() for routing Cloudflare AI Gateway requests through a Workers AI binding without an API token.
›Adds AssistantMessage.endTurn field to preserve OpenAI Codex's terminal end_turn signal for diagnostics.
+6 moreshow less
›Adds PI_EXPERIMENTAL=1 environment variable to enable experimental strict JSON-schema constrained sampling for the default read, bash, edit, and write tools.
›Adds PI_TUI_ESC_TIMEOUT environment variable to tune Escape-input timeout for high-latency terminals (e.g. over SSH).
›Adds fullscreen transcript search with Ctrl+Shift+F, incremental match highlighting, configurable search match theme colors, and next/previous navigation via Enter/Ctrl+G and Shift+Enter/Ctrl+Shift+G.
›Adds a fullscreen exit output setting to choose between printing the final transcript and printing only a session resume hint.
›Adds unbound single-line transcript scrolling actions for fullscreen mode.
›Documents the AI_AGENT=pi process marker and how it differs from PI_CODING_AGENT=true.
$ git clone --branch v0.21.12 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:$ git checkout v0.21.12
›Adds a guard to block cross-worktree Git mutations from model-issued commands that target directories outside the current session.
›Adds optional OpenTelemetry trace and span IDs to daemon logs for correlation and debugging of sampled requests.
›Adds OpenTelemetry-compliant traces for main agent invocations with stable identities and correct status semantics for success, cancellation, and errors.
›Enables per-agent JSONL transcripts for all workflow agent dispatches, recording prompts, tool calls, and results consistently.
›Implements a diff growth brake in autofix reviews to limit source and test line increases per window using configurable budgets.
+11 moreshow less
›Adds automatic escalation to maintainers when autofix diffs exceed growth budgets across multiple rounds.
›Confirmed Critical findings now require an executed witness with observed output, automatically demoting unverified claims to low confidence and hiding them from PR posts.
›Review ledger markers now include the anchor commit SHA to ensure accurate incremental diff scoping across environments.
›Adds user settings to control review attribution footers, default effort levels, and default comments, preventing repository files from overriding these policies.
›The /review command now identifies and reports entire classes of unbounded defects prospectively instead of listing individual instances.
›Compact mode (Ctrl+O) now displays model reasoning as a foldable 'Thinking...' summary instead of hiding it entirely.
›Supports uploading workspace files to the Web Shell composer via drag-and-drop or the @ file panel with progress tracking.
›Web Shell Channel management now supports full policy configuration, workspace binding, and a redesigned management interface.
›Background shells are now tracked in activeWork, enabling explicit negotiation of shell categories and preventing premature automatic cleanup.
›Requests routed through Alibaba Cloud API Gateway domains now correctly include metadata fields for session tracing and log correlation.
›Updates the review loop to validate feedback based on content accuracy rather than relying solely on the author's identity.
Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.
CrewAI 1.15.16 adds execution context management with UUID support and richer observability for flows and deployments.
└──▷ GET THIS VERSION
$ git clone --branch 1.15.16 https://github.com/crewAIInc/crewAI.git
# already have the repo? check out this version:$ git checkout 1.15.16
›Adds execution context management with UUID support for tracking individual execution runs.
›Records the type of exception that ended a flow, enabling richer post-mortem observability.
›Records when a trace batch is shared with AMP, improving telemetry auditability.
›Counts deployments from any origin and records where they started, broadening deployment provenance tracking.
langchain-openrouter now surfaces provider identity in response metadata for every API call.
└──▷ GET THIS VERSION
$ git clone --branch langchain-openrouter==0.2.8 https://github.com/langchain-ai/langchain.git
# already have the repo? check out this version:$ git checkout langchain-openrouter==0.2.8
›Preserves the upstream provider in response metadata, giving callers visibility into which OpenRouter provider served each request.
PydanticAI v2.30.0 adds allowed_hosts for the local web UI, OpenRouter web search, Gemini 3.7 Flash, and gRPC metadata on XaiProvider.
└──▷ GET THIS VERSION
$ git clone --branch v2.30.0 https://github.com/pydantic/pydantic-ai.git
# already have the repo? check out this version:$ git checkout v2.30.0
└──▷ USE IT
Allow the local web UI to be reached under a custom hostname in a non-loopback deployment.
python
agent.to_web(allowed_hosts=['mydevbox.internal'])
Run a web search via OpenRouter directly from a PydanticAI agent.
python
from pydantic_ai import Agent
agent = Agent(model='openrouter:web_search')
result = agent.run_sync('What are the latest CVEs in OpenSSL?')
print(result.data)
›Adds allowed_hosts setting to Agent.to_web() and clai web to explicitly permit non-loopback hostnames when deploying the local dev web chat UI under a real hostname.
›Adds support for openrouter:web_search as a web search model via the OpenRouter provider.
›Adds gemini-3.7-flash to the supported Gemini model catalog.
›Exposes gRPC metadata on XaiProvider for passing custom gRPC metadata to xAI endpoints.
└──▷ BREAKING ON UPGRADE
!The local dev web chat UI (Agent.to_web(), clai web) now validates the Host header against localhost/loopback/LAN addresses by default; deployments reached under a real hostname will be blocked unless allowed_hosts is explicitly configured.
Get up and running with Kimi-K2.6, GLM-5.2, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.
Ollama v0.32.13 adds Qwen3.8 model support including developer instructions and OpenAI-compatible agent compatibility.
└──▷ GET THIS VERSION
$ git clone --branch v0.32.13 https://github.com/ollama/ollama.git
# already have the repo? check out this version:$ git checkout v0.32.13
›Adds Qwen3.8 model support with a dedicated renderer handling reasoning-effort and preserved-thinking chat template semantics.
›Supports developer-role instructions for Qwen3.8, folding leading system/developer prefixes into a single system turn so OpenAI-compatible coding agents (which commonly send developer role messages) work without modification.
Ollama v0.32.12 adds Qwen3.8 27B support with an Apple Silicon MLX-optimized variant.
└──▷ GET THIS VERSION
$ git clone --branch v0.32.12 https://github.com/ollama/ollama.git
# already have the repo? check out this version:$ git checkout v0.32.12
└──▷ TRY IT
Run the Qwen3.8 27B model locally for coding or agentic tasks.
$ ollama run qwen3.8:27b
Run the Apple Silicon-optimized MLX variant of Qwen3.8 27B for faster inference on Mac.
$ ollama run qwen3.8:27b-mlx
›Adds qwen3.8:27b model, runnable via ollama run qwen3.8:27b, targeting coding, professional work, research, and long-horizon agentic tasks.
›Adds qwen3.8:27b-mlx, an Apple Silicon-optimized MLX variant of Qwen3.8 27B, tuned for maximum performance and output quality in repeated tasks and coding agents.
llama.cpp b10434 adds reasoning_effort to chat templates and generation params, enabling model-specific thinking control via the OpenAI-compatible API.
└──▷ GET THIS VERSION
$ git clone --branch b10434 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10434
›Adds reasoning_effort field to common_chat_templates_inputs, making OpenAI Chat Completions reasoning_effort available to Jinja templates with model-specific translations.
›Adds reasoning_effort to generation params, positioned alongside enable_thinking for consistent reasoning control.
›Server now reads reasoning_effort from the request body, with server_chat_convert_responses_to_chatcmpl handling conversion of Responses API reasoning.effort to reasoning_effort.
llama.cpp server now serves /metrics and /slots endpoints without blocking during active llama_decode() calls
└──▷ GET THIS VERSION
$ git clone --branch b10429 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b10429
›Enables concurrent access to /metrics and /slots endpoints while llama_decode() is executing, eliminating observability blind spots during active inference.
vMLX v1.6.28 adds Qwen3.6-27B bundle support with automatic multi-token prediction, delivering a 33% throughput gain.
└──▷ GET THIS VERSION
$ git clone --branch v1.6.28 https://github.com/jjang-ai/vmlx.git
# already have the repo? check out this version:$ git checkout v1.6.28
›Supports the Qwen3.6-27B bundle line (and upcoming Qwen 3.8): the qwen3_coder tool-parser name now resolves to the XML-function parser, capability-only stamps route through the stock loader, and the bundled multi-token prediction head is constructed, quantized per per-module overrides, and engaged automatically at the stamp's trained speculative depth — yielding 23.3 to 31 tokens/sec (+33%) on the 4-bit bundle with no flags required.
Arize Phoenix client 3.2.0 adds a new POST /traces/transfer endpoint for transferring traces server-side.
└──▷ GET THIS VERSION
$ git clone --branch arize-phoenix-client-v3.2.0 https://github.com/Arize-ai/phoenix.git
# already have the repo? check out this version:$ git checkout arize-phoenix-client-v3.2.0
›Adds POST /traces/transfer API endpoint to transfer traces between projects or destinations on the server.
Langfuse v4.11.0 adds configurable markdown render limits, annotation counts in sessions, and surfaced background agent activity.
└──▷ GET THIS VERSION
$ git clone --branch v4.11.0 https://github.com/langfuse/langfuse.git
# already have the repo? check out this version:$ git checkout v4.11.0
›Makes the markdown render character limit configurable via a new web setting, preventing runaway rendering of large LLM outputs.
›Exposes migration observation evidence in the v4 upgrade flow, giving teams visibility into what data is being migrated.
›Shows existing annotation counts directly in the sessions view, so practitioners can gauge coverage at a glance.
›Surfaces background in-app agent conversation activity and toast cards, keeping users informed of async agent work without leaving the current view.
›Introduces a forced v3 experience for projects integrating through third-party providers on legacy SDKs, ensuring consistent behavior during the v4 transition.
+1 moreshow less
›Adds a v4 features pitch with documentation links to the upgrade sidebar header, making migration guidance discoverable in-product.
THE AI TOOLCHAIN