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 v2.1.32 adds Opus 4.6, experimental multi-agent teams, automatic memory, and partial conversation summarization.
└──▷ GET THIS VERSION
$ git clone --branch v2.1.32 https://github.com/anthropics/claude-code.git
# already have the repo? check out this version:$ git checkout v2.1.32
└──▷ TRY IT
Spin up an experimental multi-agent team session to tackle complex, parallelizable tasks.
$ CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude
›Adds Claude Opus 4.6 model support.
›Introduces experimental multi-agent collaboration via agent teams (enable with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1).
›Claude now automatically records and recalls memories as it works across sessions.
›Adds 'Summarize from here' option in the message selector for partial conversation summarization.
›Skills defined in .claude/skills/ inside --add-dir directories are now loaded automatically.
+2 moreshow less
›Skill character budget now scales with context window size (2% of context), allowing more skill descriptions to appear untruncated for users with larger context windows.
›VSCode: Adds a spinner when loading the past conversations list.
Lightweight coding agent that runs in your terminal
Codex CLI v0.98.0 ships GPT-5.3-Codex support and promotes Steer mode to stable-on-by-default.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.98.0 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.98.0
›Adds GPT-5.3-Codex as a supported model.
›Steer mode is now stable and enabled by default: Enter sends immediately during running tasks while Tab explicitly queues follow-up input.
└──▷ BREAKING ON UPGRADE
!Steer mode is now enabled by default, changing keyboard behavior: Enter sends immediately during running tasks instead of queuing input, and Tab is now the explicit queue key — any workflow that relied on the previous Enter-to-queue behavior will behave differently without reconfiguration.
Codex CLI v0.97.0 adds session-scoped tool auto-approval, live skill reloading, /debug-config, and configurable log_dir.
└──▷ GET THIS VERSION
$ git clone --branch rust-v0.97.0 https://github.com/openai/codex.git
# already have the repo? check out this version:$ git checkout rust-v0.97.0
└──▷ TRY IT
Inspect the exact merged configuration Codex is running with — useful when debugging unexpected behavior from config files, env vars, or MDM settings.
$ /debug-config
›Adds session-scoped 'Allow and remember' approval for MCP/App tools, so repeated calls to the same tool are auto-approved without re-prompting during a session.
›Adds live skill update detection so changes to skill files are picked up without restarting the CLI.
›Supports mixed text and image content in dynamic tool outputs for app-server integrations.
›New /debug-config slash command in the TUI lets you inspect the effective configuration at runtime.
›Introduces initial memory plumbing (API client and local persistence) to support thread memory summaries.
+1 moreshow less
›Adds configurable log_dir to redirect logs, overridable via -c flag.
Qwen Code v0.9.0-nightly adds symlink support for the skill manager and preserves UTF-8 BOM during file edits.
└──▷ GET THIS VERSION
$ git clone --branch v0.9.0-nightly.20260205.ce7506d6 https://github.com/QwenLM/qwen-code.git
# already have the repo? check out this version:$ git checkout v0.9.0-nightly.20260205.ce7506d6
›Adds symlink support for the skill manager, allowing skill files to be referenced via symbolic links.
›Preserves UTF-8 BOM markers when editing files, preventing BOM stripping on save.
langchain-core 1.2.9 adds approximate token counting scaled by reported usage.
└──▷ GET THIS VERSION
$ git clone --branch langchain-core==1.2.9 https://github.com/langchain-ai/langchain.git
# already have the repo? check out this version:$ git checkout langchain-core==1.2.9
›Enables scaling of approximate token counts by reported usage, improving token estimation accuracy when exact counts are unavailable.
$ git clone --branch v0.8.0 https://github.com/openai/openai-agents-python.git
# already have the repo? check out this version:$ git checkout v0.8.0
└──▷ USE IT
Gate a sensitive tool behind human approval so an operator can confirm or deny each invocation before the agent proceeds.
python
@function_tool(needs_approval=True)
async def delete_record(record_id: str) -> str:
# Only runs after a human approves
return f"Record {record_id} deleted"
result = await Runner.run(agent, "Delete record 42")
for interruption in result.interruptions:
state = result.to_state()
if await confirm(f"Approve {interruption.name}({interruption.arguments})?"):
state.approve(interruption)
else:
state.reject(interruption)
result = await Runner.run(agent, state)
Restore fail-fast behavior for MCP tool errors on an agent that previously expected the run to abort when an MCP tool failed.
›Adds needs_approval=True parameter to @function_tool to declare that a tool call requires human approval before execution; pending approvals surface as result.interruptions on the run result.
›Adds RunState class with state.approve(interruption) and state.reject(interruption) methods, and result.to_state() to serialize a paused run so it can be resumed via Runner.run(agent, state) after human decisions.
›Adds mcp_config={"failure_error_function": ...} agent-level config key to control MCP tool failure handling; defaults now return model-visible error output instead of failing the whole run; set failure_error_function=None on individual MCP servers to restore fail-fast behavior.
›Adds tool_error_formatter parameter for customizing the error output returned to the model when a tool call fails.
›Adds max_turns run error handlers so callers can supply a callback when the agent hits its turn limit.
+5 moreshow less
›Adds session customization parameters to Runner for controlling session behavior.
›Adds MCP tool meta resolver support, allowing dynamic resolution of MCP tool metadata.
›Supports image responses from MCP servers, enabling MCP tools to return image content.
›Adds CRLF line-ending support for apply_diff, broadening compatibility with Windows-style patch content.
!Synchronous Python function tools now execute on worker threads via asyncio.to_thread(...) instead of the event loop thread; tools that depend on thread-local state or thread-affine resources must migrate to async implementations or make thread affinity explicit.
!Local MCP tool failure handling default behavior changed: failures now return model-visible error output instead of failing the whole run; to restore fail-fast semantics, set mcp_config={"failure_error_function": None} at the agent level and failure_error_function=None on each local MCP server that has an explicit handler.
llama.cpp b7951 adds adaptive CPU/GPU interleave on Metal for improved Apple Silicon inference.
└──▷ GET THIS VERSION
$ git clone --branch b7951 https://github.com/ggml-org/llama.cpp.git
# already have the repo? check out this version:$ git checkout b7951
›Adds adaptive CPU/GPU interleave scheduling on Metal, dynamically balancing compute across CPU and GPU based on the number of graph nodes for Apple Silicon and iOS deployments.
$ git clone --branch v2.6.10 https://github.com/milvus-io/milvus.git
# already have the repo? check out this version:$ git checkout v2.6.10
└──▷ USE IT
Enable map_populate for mmap to pre-fault pages and eliminate random page-fault latency spikes during vector search on memory-mapped segments.
yaml
map_populate: true
›Adds the map_populate flag for mmap to reduce page faults during memory-mapped data access.
›Adds loading timeout and cancellation support for better control of long-running segment load operations.
›Supports alter_collection_field() updating a field's description.
›Adds BM25 search_by_pk support, enabling primary-key lookups through the BM25 path.
›Adds automatic FP32-to-FP16/BF16 conversion during search, reducing memory and compute overhead without manual type casting.
+5 moreshow less
›Updates default auto-index configuration for vector fields.
›Disables storage-version upgrade compaction by default.
›Limits segment load concurrency by submitting loads to a dedicated load pool.
›Improves slow logs by recording average cost per NQ for latency diagnostics.
›Adds StoppingBalancer and extracts assign policy from the balancer for finer query-node load-balancing control.
└──▷ BREAKING ON UPGRADE
!Storage-version upgrade compaction is now disabled by default; deployments relying on automatic storage-version upgrade compaction must re-enable it explicitly.
THE AI TOOLCHAIN