The AI Toolchain — issue 012, August 31, 2026

A new XALGORIX_NOTIFY_SCAN_COMPLETE environment variable (default false) opts in to Discord and Telegram notifications when a scan completes, separate from the existing per-vulnerability alerts.

New tg batches command group added to the Together CLI, including tg batches submit (upload a local JSONL file or pass an existing file ID to create a batch job), tg batches list, tg batches get (alias ls), tg batches cancel, and tg batches download (streams results to stdout or writes output/error files to disk with --output).

Connecting a Vercel project via Integrations settings now creates a dedicated API key per linked Vercel project, automatically set as the TOGETHER_API_KEY environment variable in that Vercel project.

API key expiration is now configurable at creation time in the console, with options for 1 hour, 1 day, 7 days, 30 days, or a custom date via Set an expiration date.

Adds SEMANTICA_API_KEY enforcement via Depends(require_auth) on X-API-Key across all 11 Explorer API routers; the API now fails closed with HTTP 503 when the key is unset. SEMANTICA_ALLOW_ANONYMOUS=true is an explicit opt-out for local/dev use, replacing the previous silent anonymous-allow behavior.

$ # production
export SEMANTICA_API_KEY=your-secret-key

# local dev only — explicit opt-out of auth
export SEMANTICA_ALLOW_ANONYMOUS=true

Adds full PROV-O completeness to ProvenanceManager: invalidation-instead-of-delete tombstoning, hash-chained integrity via verify_chain(), typed Agent/Activity records, downstream lineage traversal, W3C PROV-O qualified relations, bitemporal and Bundle support, and a configurable namespace. New CLI commands semantica provenance invalidate, semantica provenance verify-chain, and semantica provenance descendants expose these operations, and every ingestion module now persists to a global default SQLite-backed audit trail instead of silently falling back to in-memory storage.

Adds an embedded Oxigraph backend for TripletStore via pip install 'semantica[tripletstore-oxigraph]' — runs in-process with pyoxigraph, requires no external SPARQL server, and persists to a local directory or runs in memory.

$ pip install 'semantica[tripletstore-oxigraph]'

Adds a shared validate_url_for_request() SSRF guard wired into WebIngestor, SitemapCrawler, and RESTIngestor; an allow_private_ips opt-in remains available on seed ingestion for trusted internal API sources that need to bypass the default guard.

Adds shared semantica/graph_store/query_sanitize.py to validate every Cypher label, relationship-type, and property-key interpolation site across the Neptune, Neo4j, and FalkorDB backends, and adds sparql_escaping.validate_uri() at every IRI interpolation site across the Blazegraph, RDF4J, and Jena backends to block SPARQL injection.

Adds a 50,000 node/edge cap on SPARQL graph materialization to prevent unbounded DoS, plus a 10,000-character defense-in-depth query length cap on the SPARQL route alongside a rewritten _PREFIX_DECL regex to close a polynomial-time ReDoS path.

Replaces pickle.load() in VectorStore.save()/load() with JSON serialization; legacy .pkl files are now refused rather than deserialized, so old pickle files must be regenerated.

Adds run_shacl_validation as a public API for programmatic SHACL constraint validation against a graph, letting violations be surfaced before an audit export.

from semantica.ontology import run_shacl_validation

results = run_shacl_validation(graph)
for violation in results:
    print(violation)

Adds Altair Anzo as a fourth triplet store backend — a peer to Blazegraph, RDF4J, and Jena — speaking plain SPARQL 1.1 over HTTP with no new dependency required.

Adds unsloth studio --secure to expose Studio via a free Cloudflare HTTPS link, unsloth studio -H 0.0.0.0 -p <port> to bind to a custom host/port, unsloth studio reset-password, the UNSLOTH_STUDIO_PASSWORD environment variable for headless starts, and --disable-tools to block server-side tool execution when exposed publicly; also adds LAN access via Settings > API keys > LAN access. Later replaced/extended by a dedicated Settings > Remote & LAN Access (Preview) section that can be toggled without restarting, shows connection addresses and QR codes, supports optional auto-start (disabled by default), requires changing the generated admin password, and supports keyless/password-less LAN API access.

Adds unsloth start claude --model <model>, unsloth start codex, unsloth start hermes, unsloth start openclaw, and unsloth start opencode subcommands to connect Claude Code, OpenAI Codex, and other agents to local models with a single command; later extended with OpenCode V2 support via unsloth start opencode and a --reasoning-effort flag documented on the Agents tab for controlling reasoning in Studio.

Adds DeepSeek V4 Flash 0731 Dynamic GGUF support with quantization tiers UD-IQ1_S (83 GB), UD-IQ3_XXS (104 GB), UD-Q4_K_XL (155 GB), and UD-Q8_K_XL (162 GB); enables DSpark inference by default for DeepSeek V4 Flash, delivering 2x faster inference. Adds Kimi K3 Dynamic GGUF support for Moonshot AI's 2.8T-parameter MoE model (104B active parameters, native vision, 1M context window) with tiers UD-IQ1_S (595 GB), UD-Q4_K_XL (1.51 TB), and UD-Q8_K_XL (1.56 TB), with low/high/max reasoning effort levels for this thinking-only model; both models are also listed among the local models runnable in Unsloth Desktop alongside Muse Glimmer 30B, Qwen3.8, and Gemma 4.

Supports Qwen3.8-Flash-Next (125B) locally on 75 GB RAM via 1-bit Unsloth Dynamic GGUF, with up to 262K context and switchable reasoning levels (None, Low, Medium, Extra High), plus 'Preserved Thinking' to keep reasoning consistent across longer conversations; supports GLM-5.3-Flash (320B, 18B active) locally on 102 GB combined RAM + VRAM via 1-bit GGUF, with up to 1M context and Low/High/Max reasoning modes.

New Auto Compaction (Experimental) feature lets long local chats continue past a model's context limit by rolling older turns out of active context, indexing evicted turns into a per-thread searchable archive via search_conversation, using the existing RAG pipeline (store, chunker, embedder, retrieval) without summarization.

Training now checks whether a model requires a newer Transformers version and prompts before installing it, and warns when a model must use 16-bit instead of 4-bit. On Kaggle, large saves can use the bigger temporary drive when the working folder is too small. Bounds dataset_num_proc by available memory and treats a value of 1 as genuine single-process mode rather than no multiprocessing. Fetches sentence-transformers model subfolder weights directly instead of pruning them. Answers GGUF variant requests from the specific named cache snapshot and skips unreadable entries. Settles deferred compile-mode switches between training steps rather than mid-step. Auto-falls back to HTTP download if XET is stuck, improving reliability on Colab, low-memory, and high-memory/CPU systems. Revamped training page in Unsloth Studio, with the Train page now linked across to image training.

Adds export of trained models to NVFP4, GGUF, and FP8 formats after no-code training. GGUF exports now check available disk space before starting a long merge and show how much space is needed, and Unsloth now reports the real cause when a model save or GGUF export fails instead of a generic error.

Adds support for running and fine-tuning Qwen3.8-27B and Qwen3.8-2.4T locally, including Dynamic GGUF quantizations runnable in 17 GB RAM and NVFP4 quants.

$ unsloth start claude --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL

Model settings now expose an 'Extra Arguments' box for custom llama-server flags, validated against the installed build and saved per model for normal, startup, and API load modes; invalid or security-breaking flags are rejected with a clear message.

Supports multi-GPU inference and training across NVIDIA, AMD, Intel, and Mac hardware, plus CPU-only operation; adds Intel XPU support for inference and hardware acceleration; adds improved support for AMD RDNA 3, RDNA 4, and Strix Halo systems with safer multi-GPU ROCm device matching; automatically detects multi-GPU setups and offloads model layers to system memory for DeepSeek V4 Flash and Kimi K3.

Projects feature lets chats, files, and workspaces be organized together, with attached files shared across a project; chats remember composer settings, system prompts, and model sampling options. Adds drag-and-drop support for files (including .ods) onto project sources and knowledge bases, plus an attachment preview modal for all attachment types. A full chat or individual reply can be saved into a project's sources, preserving reasoning, tool calls, and citations.

Adds support for custom llama.cpp builds with toggles for Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, and Vision On/Off.

Desktop update controls moved to Settings > General; adds Check for Updates and Update Now options to Settings, plus the app version shown in Settings > About; Desktop warns before quitting during an in-progress update and skips the installer fallback when the user cancels; Desktop draft publishing is enabled without a full rebuild.

Adds MiniMax-H3 for local video generation. Adds Fast FP8 MiniMax-H3 inference, delivering approximately 10x faster throughput (approximately 3 minutes vs. 30); MiniMax-H3 now splits large model parts to fit smaller supported GPUs, and the model picker hides H3 options the current hardware cannot run.

Adds a Settings > Debugging panel showing main server and model-runner logs with live, 3-second, and manual refresh modes; passwords, tokens, and API keys are redacted in the viewer.

Enables 5x faster inference for RAM offloading, with large GGUFs automatically split across GPU and system RAM via the offload planner, now on by default.

Adds support for running and fine-tuning Meta's Muse Glimmer 30B dense model locally on 20 GB RAM/VRAM (Mac and GPU) via Unsloth Dynamic quants, under the Apache 2.0 license; also listed among models runnable locally in Unsloth Desktop.

Adds prompt queueing with Cmd/Ctrl+Enter and drag-to-reorder, plus customizable keyboard shortcuts.

Adds Qwen3-ASR dictation to Unsloth Studio, with cancellable model downloads and a download prompt before starting; adds stop and send buttons on the Studio chat dictation bar; adds custom STT (speech-to-text) connections in Studio; loads text-to-speech models on demand rather than at startup; supports searching and downloading embedding models directly from Hugging Face.

Studio image and video pickers now populate from a curated catalog rather than raw model listings; Studio detects stalled Deep Research output streams and surfaces the condition to the user; Studio now applies saved per-model settings on API loads, adds an API monitor, and reports active training status in the UI; serves gated diffusion base models from ungated Unsloth mirrors, removing the need for user tokens; Desktop supports dropping documents into a chat (non-GGUF files accepted); Chat now provides downloadable artifacts from generated responses; pinned models can be reordered by dragging them in the model list; Chat now shows live prompt and generation speeds during streaming.

The new /spec run command opens a dedicated full-screen task execution view for V3 spec runs, with real-time progress tracking and the ability to select task scope before execution begins.

$ /spec run

Adds a new superset scripts add subcommand to create terminal scripts from the CLI, and renames the former 'terminal presets' feature to 'terminal scripts'.

$ superset scripts add

Adds locale persistence via a Postgres user.locale field and a desktop Language setting (Settings > Language), a language switcher on marketing and docs pages, and automatic translation of new strings on PRs; ships 14 new UI languages.

Raises MIN_HOST_SERVICE_VERSION so pre-1.22 host services hit the version gate instead of displaying a blank terminal pane.

Adds tools.update_plan.enabled config key, which now defaults to false; set it to true to explicitly expose the update_plan tool. When disabled, bundled planning guidance is removed from prompts, so workflows relying on update_plan being available by default will stop receiving it until the key is set explicitly.

tools:
  update_plan:
    enabled: true

MCP server names may now contain :, @, /, and . in addition to letters, numbers, -, and _, enabling package-style names like npm:@modelcontextprotocol/server-sequential.thinking across mcp add, get, list, and remove commands, runtime tool namespaces, and OAuth credential lookup.

$ codex mcp add npm:@modelcontextprotocol/server-sequential.thinking

The TUI now renders actionable rate-limit banners above the composer, with actions covering usage, credits, resets, owner notifications, plan management, and automatic fallback-model switching when instructed by the banner.

Sets history_ingest_requested to true in Responses turn metadata when the history-notes token-budget extension is enabled, and reserves the key so caller-provided metadata cannot override it.

Adds four new integrations for investigation and alert triage: Cloudflare (HTTP traffic, firewall events, Zero Trust Gateway activity, DNS, zones, and devices), Auth0 (identity configurations, users, audit events, and sign-ins), 1Password (sensitive item activity and audit events), and Tailscale (devices, users, and configuration changes across tailnets).

Adds extraction of .robot and .resource files through the optional [robot] extra, capturing suites, test cases, user keywords, keyword-call edges, and resource/library imports, with case/space/underscore-insensitive keyword resolution. Install with uv tool install 'graphifyy[robot]'.

$ uv tool install 'graphifyy[robot]'

The graphify extract --force --code-only flag now fully rescans all code files instead of skipping unchanged files and keeping stale import/alias resolution, while preserving the existing document/semantic tier.

$ graphify extract --force --code-only

New OPENCLAW_SUPERVISOR_MODE=external environment variable lets an external supervisor own Gateway restarts, service lifecycle, and updates without the native service competing for those changes.

New openclaw telemetry show command inspects the outgoing usage-statistics payload before deciding whether to opt in; feature statistics are off by default.

$ openclaw telemetry show

New openclaw update repair and openclaw doctor --fix commands recover missing configured provider packages after failed installs or upgrades.

$ openclaw update repair

The bundled OpenProse plugin and /prose command are removed; run openclaw doctor --fix to clean stale configuration and follow the upstream Agent Skill migration, keeping existing .prose source files.

All shipped codex/* and openai-codex/* model refs, provider config, stored sessions, and automation routes must be migrated to openai/*; run openclaw doctor --fix to perform the migration and flag conflicts for operator repair.

New --agent-name flag during interactive onboarding names the first agent and migrates legacy main-session history to its clear owner; onboarding now sets up named agents and reuses main as an ordinary agent ID after required doctor repairs.

New shared credential store manages team-scoped secrets and environment entries in SQLite, accessible through the CLI and Settings, with write-only secret values, protected egress bound to declared hosts, and a distinction between readable environment values and secrets.

New modelPolicy.allow configuration key defines explicit per-agent model allowlists and provider wildcards, kept separate from aliases and per-model settings.

Plugin uninstall now cleans up plugins.load.paths, removing exact recorded install paths for marketplace, npm, and other managed installs while preserving parent, child, prefix, and unrelated paths.

New ClickClack integration supports one-use expiring setup codes, immediate account verification, and native command menus published at Gateway startup when the token has commands:write.

Plugin installs from arbitrary executable sources via CLI or chat now require --force; trusted ClawHub, bundled, official-catalog, and tracked-update sources skip the warning.

New conversation-bound automations and an owner-only /loop command support fixed-interval or self-paced checks, defaulting new agent-turn automations to their originating conversation.

Optional 1Password broker resolves SecretRefs or reads selected 1Password fields using service-account auth, with per-secret approval and value-free audit.

New recoverable database backups schedule recorded backups, keep versioned snapshots in an operator-owned Git repository, and can restore a verified full archive into a fresh staging directory.

New official provider packages for BytePlus, ComfyUI, Mistral, NovitaAI, OpenCode, Synthetic, Volcengine, Vydra, and Xiaomi are installable through onboarding; separately installable official plugins add Cohere and Meta providers, DuckDuckGo search, Voyage embeddings, and iMessage messaging.

Adds SHA256SUMS.txt with per-archive checksums computed in CI from the tagged commit, verifiable via shasum -a 256 --ignore-missing -c SHA256SUMS.txt.

$ shasum -a 256 --ignore-missing -c SHA256SUMS.txt

Adds signed, self-contained desktop app builds (window + tray around a local server) for Windows x64 (exxperts-desktop-0.11.0-win-x64.zip, exxperts-setup-0.11.0.exe) and macOS Apple Silicon (exxperts-desktop-0.11.0-mac-arm64.dmg), with the macOS build Apple-notarized and Windows binaries publisher-signed by Exxeta AG to eliminate unknown-publisher prompts on first launch. Also adds self-contained exxperts server archives for Windows x64 (exxperts-0.11.0-win-x64.zip), macOS Apple Silicon (exxperts-0.11.0-darwin-arm64.tar.gz), and Linux x64 (exxperts-0.11.0-linux-x64.tar.gz) with a vendored Node runtime, requiring no Node, npm, or Git installation on target machines.

Adds KnowledgeManagementTools with operations ingest_url, ingest_text, ingest_path, list_content, ingest_status, and remove_content (confirmation required by default), supporting scope="shared"|"user", JSON envelopes, and sync/async variants. In the following release the constructor flags were renamed to match the tools they register — ingest_url, ingest_path, ingest_text, remove_content — giving per-tool opt-in control, with ingest_path now defaulting to off because under scope='shared' it exposes loaded content to every agent on the knowledge base; the old enable_ingest/enable_remove flags are removed and silently ignored. The toolkit also moved from agno.tools.knowledge_management to agno.tools.knowledge, alongside the read-only KnowledgeTools, mirroring the agno.tools.mcp and agno.tools.finance layout.

from agno.tools.knowledge import KnowledgeManagementTools

km_tools = KnowledgeManagementTools(
    knowledge_base=my_kb,
    ingest_url=True,
    ingest_text=True,
    ingest_path=True,   # opt-in: exposes any path the server process can read
    remove_content=False,
)

from agno.agent import Agent
from agno.tools.knowledge import KnowledgeManagementTools

agent = Agent(
    tools=[KnowledgeManagementTools(scope="shared")],
)
agent.print_response("Ingest https://docs.agno.com and list all loaded pages.", stream=True)

AgentOS gains GET /knowledge/content?parent_id= to list a site's or folder's content rows with correct totals, and POST /knowledge/content/{id}/refresh to re-run ingest for a URL- or path-sourced row in the background.

$ curl -X POST https://<agentos-host>/knowledge/content/<id>/refresh

$ curl 'https://<agentos-host>/knowledge/content?parent_id=<parent_id>'

Adds HttpxPageFetcher and ParallelPageFetcher as fetch seams below URL readers; ParallelPageFetcher resolves Parallel's keyed SDK, then its keyless MCP endpoint, then plain httpx, honoring retry-after with exponential backoff and recording per-page extractor and attempts provenance. HttpxPageFetcher also routes application/pdf responses and %PDF- bytes served under a wrong content type through PDFReader; a missing pypdf surfaces as a per-page error naming agno[pdf].

Adds pow_workers argument to AtomicMail (default min(4, cpu_count())) to parallelise scrypt nonce search across a bounded thread pool, cutting mean solve time from 25.9 s to 10.4 s at difficulty 10.

Adds SitemapReader that discovers pages via the sitemap protocol (robots.txt Sitemap: lines, /sitemap.xml, /sitemap_index.xml, gzip and nested indexes) with canonical dedup and a max_pages cap; it is auto-selected for bare sitemap*.xml(.gz) URLs and also available in the UI reader dropdown.

Caches AtomicMail auth context (capability JWT, API URL, account and inbox IDs) on the instance until token expiry, reducing warm tool call latency from ~35 s to 0.4–3 s in both sync and async paths.

Per-page website ingestion stores one content row per page with content_id matching its vectors, supporting individual list, refresh, and delete; digest-driven re-ingest skips unchanged pages, replaces only changed pages' vectors, retries failed pages, and prunes removed sitemap entries.

Folder ingestion stores a folder row with one child row per file (nested folders flattened), with byte-digest refresh so unchanged files skip read and embed, per-file failure isolation, pruning of deleted files, and cascade delete.

Makes agno.tools.file and agno.tools.knowledge lazy-loading packages so importing FileTools no longer pulls in reportlab and python-docx — saves 44.8 ms on every FilesystemContextProvider import.

Adds vectorized GLM KDA MTP verification for GLM-family models, adds typed prompt SSD cache support to the MTP benchmark, and makes native MTP restart cache proofs configurable via commit fc88a3ea.

Adds native support for Qwen 3.8 Flash Next (125B), delivering ~60 tok/s on M4 Max and ~78 tok/s with speculative decoding, using ~70 GB RAM with the 4-bit pack ddalcu/Qwen3.8-Flash-Next-MLX-Serve-4bit.

Adds a --mtp flag (and a MoE toggle in Settings) to opt into speculative decoding on Flash Next, yielding +41% throughput on code generation; speculative decoding is now also applied to image questions.

$ mlx-serve --model ddalcu/Qwen3.8-Flash-Next-MLX-Serve-4bit --mtp

Adds POST /v2/datasets/{dataset_id}/experiment-runs as the supported public API for paginated experiment comparison. Legacy dataset comparison helpers are removed from the public OpenAPI spec and generated SDKs; existing HTTP routes continue to work only for LangSmith UI clients.

OpenTelemetry resource attributes set via OTEL_RESOURCE_ATTRIBUTES now appear on traces as metadata namespaced under otel.resource.*, enabling attachment of fields like user IDs without changing span emission code.

$ export OTEL_RESOURCE_ATTRIBUTES="user.id=u_12345,deployment.environment=production"
# Traces will appear in LangSmith with metadata keys otel.resource.user.id and otel.resource.deployment.environment

The thread evaluator config preview now shows only the thread message formats the evaluator actually maps, and displays a locked 'Trace count ≥ 2' filter for managed thread evaluators. Evaluators can be tested against a real conversation before saving via the validate endpoint.

$ curl -X POST 'https://<your-langsmith-host>/runs/rules/validate' \
  -H 'Content-Type: application/json' \
  -H 'X-Api-Key: <your-api-key>' \
  -d '{"test_thread_id": "<thread-uuid>", "session_id": "<session-uuid>", "evaluator": {"type": "thread", "code": "..."}}'

A reorderable 'Splits (latest)' column in the experiment comparison view shows each example's current dataset split assignments as chips reflecting live membership. Each split chip is interactive, with an 'Edit splits' action that opens the single-example split picker inline.

Value redaction now covers xAI keys (xai-), Groq keys (gsk_), and Google OAuth 2.0 access tokens (ya29.). Anthropic keys (sk-ant-) are now reported under their own named class instead of being lumped into a generic category.

Tables (local and remote) can declare computed columns using SQL expressions, with columns able to read earlier declarations within the same batch for multi-step derivation. refresh_column_async returns a job handle for tracking refresh progress, and refresh operations now return typed refresh job results.

job = await table.refresh_column_async("summary_embedding")
result = await job.wait()

Adds listTables to the Node.js SDK, deprecating tableNames, for paginated table listing driven by the store's own cursor; existing code that constructed or passed page tokens for table listings must be updated.

for await (const name of db.listTables()) {
  console.log(name);
}

Remote Functions gain a gpu flag requiring a GPU for every execution (baked into the immutable Function version), conda_channels and channels options for pinning Conda packages and channel priority, first-class function wire contracts, scalar function authoring, and a catalog client for managing remote functions. Functions can be bound to a group of columns or to specific columns by version, and support large_utf8 return signatures and nested Arrow types in their definitions.

StreamingDataLoader, a PyTorch DataLoader subclass, carries consumer-committed StreamingDataset checkpoints alongside every internal batch, enabling safe mid-epoch resumption with multiple workers.

from lancedb.streaming import StreamingDataset, StreamingDataLoader

dataset = StreamingDataset(
    table,
    num_splits=8,
    pack_sequences=2048,   # block length in tokens
    eos_id=2,
    pad_id=1,
    blocks_per_epoch="auto",
)
loader = StreamingDataLoader(dataset, batch_size=4, num_workers=4)

for batch in loader:
    # batch contains 'input_ids' and 'doc_ids' LongTensors
    train_step(batch)
    checkpoint = dataset.state_dict()  # consumer-committed checkpoint

Adds field metadata convention keys lancedb:description, lancedb:tag:<name>, lancedb:logical-column, and lancedb:status (values: production, candidate, deprecated, archived) for annotating table columns via the metadata API.

Adds materialized view declarations on local tables plus a refresh API to update them, with Python and Node.js bindings included. Materialized view refresh is bound to the view incarnation so stale refreshes against a replaced view are rejected.