2026-11-13 · GitHub AI Daily

WindowBeijing time 2026-11-13 00:00 – 10:05 (equivalent to GitHub committer time from 2026-11-12 16:00 UTC). Judged by real commit timestamps on each repository's default branch — `pushed_at` is never used as a proxy — supplemented by same-day releases and in-repo official announcements.

Coverage① AI Coding / coding-agent & skill ecosystem ② Medical / clinical research AI & agent skill libraries ③ Health management / nutrition & food AI ④ Model layer & agent infrastructure. Four tracks in total.

De-duplicationChecked one by one against every published issue from 2026-10-16 to 2026-11-12 (1,821 historical repositories in total). All 83 repositories in this issue appear for the first time. There is no cross-table duplication, and index entries = table rows = unique URLs, all three equal. Four repositories that topped this week's trending list — anthropics/skills (★193,723), mattpocock/skills (★323,052), addyosmani/agent-skills (★116,712) and cloudflare/security-audit-skill (★71,253) — recorded no commit on their default branch today, so they are excluded from the tables and noted only under "Trends today".

VerificationStars and licenses come from the GitHub REST API (`repos` and `license` endpoints, with the LICENSE file read back to confirm the actual text); commits come from the `commits` endpoint filtered by `committer.date`; releases come from the `releases` endpoint; every README conclusion quoted here was checked against the source text. No repository name, star count, commit message or evaluation figure in this issue is invented.

1. Today's picks (5)

1. Evaluating the evaluators: BLEU and BERTScore cannot catch planted clinical errors — structured fact extraction plus image localization can

URL: https://github.com/FavianaNoronha/ultrasound-report-metaeval ｜ ★0 ｜ MIT ｜ On breast, thyroid and fetal-head ultrasound, plants one clinical error into an AI-written report and asks whether four layers of scorers rank the erroneous report below a harmless rewording

What happened

3 commits today completed Steps 31-38 of the roadmap: a cross-dataset segmenter, a stricter fact extractor, pre-registered fact-vs-wording and adversarial tests, metamorphic and limit-of-detection checks, RadGraph-F1, and responsible-AI checks on the image check.
Metamorphic checks (Step 34) measured unit, order and synonym invariance violation rates: BLEU 60%, ROUGE-L 48%, chrF 36%, RaTEScore 35%, AI judge 30% (subset), finding_f1 6%, BERTScore 5%, llm_finding_f1 5%.
Limit of detection (Step 35), measured against matched synonyms with an AUROC of at least 0.80 required: BLEU, ROUGE-L, chrF, BERTScore and RaTEScore never reach it. The AI judge reaches it only at 5% size error and 1 mm head circumference; the MedSAM oracle-box image check reaches 25%, the label-free U-Net check 50%.
The pre-registered adversarial rewrite attack on the judge (Step 36): the winning transform, a reorder, moved the judge score by −0.13 on average across 100 wrong reports, yet 6% of them overtook their correct counterpart (9% before the attack). RadGraph-F1 turned out to be wording-driven (fact 0.00 vs wording 0.17 across 17,444 blocks) and never reached AUROC 0.80.
Responsible-AI checks (Step 38): abstaining on low U-Net confidence lowers the position-check error from 4.8% to 3.3% at 80% coverage and 2.2% at 60% coverage, and mask failure from 12.1% to 3.8% and 1.7%. GREEN's official 13 GB checkpoint did not finish scoring two pairs in 36 minutes on a 16 GB machine, and the author states plainly that no GREEN number exists and nothing was substituted.

Why it matters

This lands exactly on the thing this site and my business fear most: a report may claim evidence, zero hallucination and cited references, but if the review metric is BLEU or BERTScore, it cannot even tell a swapped number from a flipped negation — in 93% of ROUGE-L scenarios the planted error ranks above a harmless synonym swap.
The finding has a direct engineering consequence. The combination that actually works is structured fact extraction (finding_f1 / llm_finding_f1) plus image localization (mask comparison) plus an abstain mechanism — not a longer prompt asking the model to judge itself. That belongs directly in my 4P intervention-report generation spec.
The figure that 6% of erroneous reports overtake their correct counterpart after a rewrite is a hard reminder for my content production: once text is drafted by AI, wording is itself an attack surface. Every conclusion in my 4P platform marked "must be human-confirmed" should be a required step in the flow, not a line in a prompt.
Writing "there is no GREEN number" instead of substituting a weaker metric is a rule I should copy verbatim into my own spec documents, for the same reason as last issue's openfit publishing the verification status of all 95 of its rules in four tiers.

2. Skills need quality gates too: NVIDIA now separates a failed evaluation arm from an arm that never ran

URL: https://github.com/NVIDIA/SkillEvaluator ｜ ★888 ｜ Apache-2.0 ｜ A tiered quality-gate framework with semantic overlap detection, synthetic evaluation dataset generation and live agent evaluation, measuring what a skill actually changes about task outcomes

What happened

24 commits today, making it one of the densest-moving mature frameworks in this issue. The project sits inside the NVIDIA Verified Skills pipeline and ships with an arXiv paper (2608.20614) and official documentation.
Three commits expose the tightening of evaluation semantics directly: fix(reporting): describe a failed Tier 3 arm as incomplete, not as not run (#407), fix(tier1): report dependency advisories as structured findings (#421) and fix(tier3): stop per-trial apt installs in Docker agent setup (#413).
It works in two stages. First static quality checks, wired into GitHub or GitLab CI via ready-made templates with no evaluation code to import; then live agent runs with the skill and without it, compared to measure the skill's effect on the task result.
skillevaluator ci init bootstraps the configuration inside your own repository and opens a setup pull request where you pick checks and thresholds. The most recent release is v0.8.0 (2026-11-06).

Why it matters

The homework I set myself in the previous issue was exactly "build a 30–50 question pairwise comparison evaluation for the 4P expert package". This framework's shape — skill-on versus skill-off live runs, tiered gates, explicit thresholds — can be copied directly, which removes the need to invent an evaluation protocol from scratch.
The distinction between a failed arm and an arm that never ran is the piece I most want to borrow. Last issue's clinmesh enforced human sign-off as a property of the system rather than of the text; here it is enforced by semantics in the report. Both approaches prevent "unverified" from being reported as "passed".
Reporting dependency advisories as structured findings rather than log lines turns skill review from reading logs into reading conclusions — that can move straight into my expert package's compliance checklist.
A practical caution: NVIDIA is now pushing skill quality gates, meaning the skill ecosystem is turning from prompt fragments into engineering assets that CI is expected to police. If my 4P expert package stays a plain Markdown list, it will fall a generation behind.

3. Label photos into the app: food-label recognition is split into "AI reads only, values normalized per 100 g, energy cross-checked against macros"

URL: https://github.com/oscardanielnc/kawsay ｜ ★0 ｜ No license declared (no LICENSE file in the repository) ｜ An offline-first journal for food, activity, weight and mood where the AI only picks catalog items and portions while every number is computed from a table

What happened

4 commits today, three of which land the same batch of functionality: feat(api): read a food label from a photo with Gemini Flash-Lite, checked and retried once (ADR 0029), feat(core): label values per 100 g, energy check against macronutrients; own label foods rank first and feat(mobile): own food from a label, by photo or typed, confirmed number by number.
The project calls its design principle "Measured AI": the AI emits only catalog IDs and portions, and every number on the intake card is computed by a deterministic table.
The nutrition catalog comes from Peru's INS 2023 composition tables with household units. Daily targets are recalculated from the day's actual activity, plus five liver-health gauges for added sugar, sweet drinks, movement, alcohol and sleep.
Private by default: intimate entries are locked on the phone and never reach the AI (ADR 0013). Target recalculation and AI boundaries are recorded as ADRs 0009, 0008 and 0020. The author reports using it daily since October 2026.

Why it matters

This is close to a working reference for the lab-report-photo OCR into AI into app import path I have queued next, and it faces the identical hard parts: values from a photo must be normalized per 100 g, energy must be cross-checked against macronutrients, and recognition output must be validated and retried once rather than written straight to the database.
The "AI picks but never computes" division of labour is the rule my GLP-1 nutrition companion has followed all along, and here it is implemented far more cleanly: even the portion is chosen by the AI, but every number comes from a table — far more robust than telling a model in a prompt not to make things up.
Recording each design decision as its own ADR is worth copying. Every algorithm module in my 4P platform should have an ADR explaining why it computes the way it does, otherwise three months from now I will not be able to say where a given threshold came from.
A self-hosted tool that states clearly that it is not a medical device matches the "no clinical credentials claimed" boundary I have set for myself. The author puts that sentence at the top of the README instead of burying it, which is itself a trust design.

4. Authoritative nutrition and terminology sources packaged as MCP: USDA composition tables, OBO/FoodOn ontologies and taxonomy in one tool surface

URL: https://github.com/Knuckles-Team/world-reference-mcp ｜ ★0 ｜ MIT ｜ Species taxonomy, ontology reference terms, food nutrition composition, observations and Wikidata alignment turned into a bounded, checkpointed MCP tool surface

What happened

3 commits today finished the ecosystem rename from agent-utilities to meristem and adopted scion-sdk 0.10.1; the latest release is v0.4.0 (2026-11-13). The notable line in it is the addition of tool-use contracts, links, locations, roles and agent prompts to the surface (0.3.9).
Nine MCP tools map onto five source ledgers: NCBI Taxonomy plus the GBIF backbone (taxonomy), EBI OLS4 covering GO, UBERON, ENVO, PATO, ChEBI and FoodOn (reference terms), USDA FoodData Central (nutrition composition), GBIF occurrences, iNaturalist, Open-Meteo and NOAA GHCND (observations), and Wikidata SPARQL (alignment).
Ingestion runs through scion-sdk's declarative mcp_tool source adapter into the medulla world_model ontology, characterized as bounded, checkpointed ingestion streams rather than one-shot full fetches.
The credential policy is explicit: every tool is keyless except fdc_food_search and noaa_ghcnd_daily, whose credentials are resolved server-side via WORLD_REFERENCE_FDC_API_KEY and WORLD_REFERENCE_NOAA_TOKEN and are never accepted as tool arguments.

Why it matters

This closes a gap in my 4P precision-nutrition stack. I maintain 153 ingredient templates by hand, but the source of the nutrition values and terminology has always been manual transcription. The USDA FoodData Central plus FoodOn pairing is an existing authoritative source that can be attached over MCP — one order of magnitude more reliable than asking a model to report calories from memory.
The bounded, checkpointed ingestion design mirrors what the 4P thousand-practitioner pilot needs for data collection: a large file cannot be pulled in one shot, it has to be segmentable, resumable and progress-verifiable.
Never passing credentials as tool arguments matches the rule I have already set about never exposing tokens; this can be written straight into my MCP connector spec.
If supplement and ingredient terminology can be joined this way, the long-standing question of whether "magnesium" means the element, the ion or a particular supplement form finally gets a machine-readable answer rather than a debate.

5. DSH moves from personal plugin to enterprise: MOVO wraps it into a self-hosted platform with approvals, governance and a rollback baseline

URL: https://github.com/himovo/movo ｜ ★665 ｜ Apache-2.0 plus additional terms (MOVO Community License) ｜ A self-hosted enterprise agent platform built on DeepSeek Harness, adding knowledge, identity and access control, governance, admin console and file delivery

What happened

7 commits today concentrate on making the boundary between "DSH runs the agent" and "MOVO brings the agent into enterprise production" concrete: test: align DSH contract checks with MOVO approval policy and fix: retain one released DSH rollback baseline.
The remaining commits address run legibility: showing activity icons by tool capability, fixing the placement of commentary in completed runs, registering user questions with agent identity, and preserving intermediate DSH replies before the final answer.
Deployment is reduced to a single command: git clone followed by ./movo up, where the launcher pulls official prebuilt images, waits for the services to become healthy and prints the setup address — no .env file, no local image build.
The repository ships an English README, a Simplified Chinese one and an online documentation site. The most recent release is v0.2.6 (2026-10-30).

Why it matters

I have been building native agents on dsh locally but always treated it as a framework. MOVO demonstrates that the same DSH Runtime, Skills, Tools and MCP stack can directly carry the enterprise side of approvals, governance and delivery — a path I do not have to walk from scratch.
"Retain one released DSH rollback baseline" maps precisely onto the operational reality of this site: the server auto-pulls and resets every 15 minutes, and my own 4P platform and AI daily pipeline need an explicit, revertible release baseline rather than "ship it and see".
"DSH contract checks aligned with the MOVO approval policy" shows the two layers aligning through a contract rather than through convention. If the "zero hallucination plus human confirmation" rule in my 4P platform only lives inside a prompt, it is effectively not a contract.
The rest of the DSH ecosystem moved today too: dsh-session-insights (local-first, evidence-backed DSH workflow retrospectives), dsh-session-hub (cross-agent session import and search) and DSH-Bot (personal agents with memory and workspace group chats). From this issue on I track DSH as its own line rather than scattering it across the four tracks.

2. Category roundup

① AI Coding / coding-agent & skill ecosystem

Project	Stars	One-liner
https://github.com/NVIDIA/SkillEvaluator	888	Tiered quality gates plus live agent A/B runs, measuring what a skill actually changes
https://github.com/qybaihe/bla	5468	Adversarial defense shield for blind LLM agents, protecting binaries and Electron apps
https://github.com/hashgraph-online/awesome-ai-plugins	1040	Curated AI plugin list, with an Agent Bridge added today
https://github.com/vercel-labs/fx	3824	A Unix-like coding agent that turns terminal interaction into composable tools
https://github.com/mixpeek/amux	559	Agent control plane running Claude Code, Codex and Gemini CLI in parallel
https://github.com/Th0rgal/orb	715	All model providers and agent harnesses behind one local entry point, with a managed machine pool
https://github.com/activeing123/mcptoon	429	Zero-dependency MCP gateway CLI unifying every MCP server and skill, with token savings
https://github.com/ZaxbyHub/opencode-swarm	526	Hub-and-spoke orchestration plugin for OpenCode, v7.204.2
https://github.com/intent-hq/intent	428	Open-source workspace for coding agents where specialists plan before Claude Code, Codex and Cursor build
https://github.com/gy15901580825/Argus	208	Red-team testing for agents with 266 bundled probes mapped to OWASP LLM Top 10 and MITRE ATLAS
https://github.com/coddy-project/coddy-agent	205	General-purpose agent compiled into a single Go binary, 1.4.12
https://github.com/nbn07/agent-tracedown	16	The trace says "success" but the answer is wrong — locates the offending step and replays with a fix
https://github.com/Innocent-children/TaskBelay	13	Task control for coding agents: explicit scope, bounded verification, durable state
https://github.com/writfold-dev/writfold	1	Evidence-and-approval harness: exact evidence, explicit approval, reproducible delivery
https://github.com/rschuft/ai-engineering-harness	2	Open-source AI engineering harness turning human intent into verified software
https://github.com/a-littlecat/ai-dev-flow	1	AI-assisted development workflow Skill for solo developers, with approval and land-check steps
https://github.com/MyDear1ove/reflection-log	0	Compiles one reflection into checkable rules plus a trigger-word index, escalating to resident rules on repeat failure
https://github.com/Jenoyrex/ctx-receipt	0	A token receipt for agent context, so you can see what it loaded before you type a word
https://github.com/yy/scholar-kit	40	Renamed to scholar-kit, academic search skills now work with any agent harness
https://github.com/localixai/tokendoctor	70	A Claude Code plugin that cuts per-call cost by up to 38%

② Medical / clinical research AI & agent skill libraries

Project	Stars	One-liner
https://github.com/FavianaNoronha/ultrasound-report-metaeval	0	Meta-evaluation of ultrasound report scorers, quantifying who catches planted clinical errors
https://github.com/microsoft/fhir-server	1393	FHIR reference implementation, today rejecting SMART scopes constrained by non-filtering parameters
https://github.com/cornerstonejs/cornerstone3D	1168	JavaScript medical imaging libraries, v5.18.1 records frame delivery before volume load
https://github.com/Zachsteuer98/medical_assistant_agent	0	Clinical RAG assistant with page-level citations, adding a public-corpus eval and an answer-correctness judge
https://github.com/felipertoledo/clinical-llm-guardrails	0	Deterministic release gate for LLM-drafted clinical documents
https://github.com/MichaelEnny/agentic_gateway	0	Checks every tool call a clinical agent proposes and allows, defers or blocks it by calibrated confidence
https://github.com/P1Manav/SugenEngTest	0	Offline, zero-LLM ICD-11 coding for outpatient clinical notes
https://github.com/capstone-group-l/clinical-evidence-rag	1	Clinical decision support RAG with verifiable citations, comparing retrieval setups today
https://github.com/vinay3254/medicluster	1	Clinical decision support combining LSTM/Prophet forecasting with a DenseNet
https://github.com/burakdon/neurocue	1	A clinically grounded synthetic aphasia patient plus an RL therapy trainer with clinician corrections
https://github.com/rogeriorrodrigues/nursia-research-lab	0	Clinical AI and FHIR research demos: zero cloud, FHIR R4, local LLMs
https://github.com/Luosanj/VOICE_HE_ST_Pred	5	A vision-omics foundation model predicting transcriptome-wide single-cell expression from H&E images (KDD 2027 Oral)
https://github.com/natera-open-source/vep-rs	48	Rust reimplementation of Ensembl VEP, same flags and outputs at 100x+ the speed
https://github.com/Deliverome-Project/accessible-surfaceome	6	Open atlas of which human proteins reach the cell surface, and in what context
https://github.com/OpenwaterHealth/openlifu-transmitter-fw	4	open-LIFU low-intensity focused ultrasound transmitter firmware, including FDA preset generation
https://github.com/taryev/diag-llm	0	Extracting structured diagnoses from unstructured clinical notes with LLMs
https://github.com/josuexpositoacosta/SEGUIMORBI	1	Epidemiological case management supporting clinical assessment and medical decision-making
https://github.com/openshiporg/openfront-hospital	2	Open-source hospital information alternative on Next.js, handling refunds and settlement scope

③ Health management / nutrition & food AI

Project	Stars	One-liner
https://github.com/Knuckles-Team/world-reference-mcp	0	USDA composition tables, OBO/FoodOn ontologies and taxonomy packaged as a bounded, resumable MCP surface
https://github.com/oscardanielnc/kawsay	0	Offline-first food journal with "Measured AI": the AI picks catalog items, the table computes every number
https://github.com/grocy/grocy	9675	Self-hosted household inventory and purchasing, today optimizing the OpenAPI schema and table column types
https://github.com/TandoorRecipes/recipes	8733	Recipe, meal plan and shopping list management
https://github.com/bioconda/bioconda-recipes	1884	Conda recipes for the bioconda channel, with 11 package updates today
https://github.com/omicverse/omicverse	1396	Unified multi-omics framework adding lineage markers and a specificity/contrast conflict rule, storage back to memory by default
https://github.com/openfoodfoundation/openfoodnetwork	1302	Connects local producers, distributors and consumers to trade local produce
https://github.com/scverse/PyDESeq2	772	A Python implementation of the DESeq2 pipeline for bulk RNA-seq differential expression
https://github.com/GMOD/jbrowse-components	298	Source for JBrowse 2, a modern React genome browser, today gating ring-view fetches
https://github.com/merenlab/anvio	546	An analysis and visualization platform for 'omics data
https://github.com/stjude-rust-labs/sprocket	222	Bioinformatics workflow engine on top of WDL, today pinpointing the first out-of-order element in WDL lint
https://github.com/cogent3/cogent3	142	Comparative genomics toolkit, release 2026.11.13
https://github.com/sidhant947/Path	57	Privacy-focused lightweight step tracking app, V2.0.8
https://github.com/bardesss/haelan	43	Self-hosted dashboard and local mirror for household Google Health data, in one SQLite file
https://github.com/tarantila/CaloGraph	3	Self-hosted way to explore calorie and nutrition data from Apple Health and Google Health
https://github.com/nschoch/macrolog	0	Self-hosted AI photo calorie and macro tracker with adaptive targets
https://github.com/SparkSidd/IngredientInsightAI	0	AI-native food label nutrition co-pilot
https://github.com/sedoglia/diabetes-m-mcp	0	A TypeScript MCP server connecting Claude securely to Diabetes:M Analytics
https://github.com/BeBernos/Mesa-Business	0	Family nutrition PWA: per-person meals and targets plus one shared shopping list
https://github.com/caferkarabulut/bite-routine	0	Food diary and portion-based nutrition tracking with Open Food Facts packaged products
https://github.com/timmylich/wellness	1	Private, offline-first Android health tracker for food and 12 nutrients, v0.7.0 adds a shared pantry

④ Model layer & agent infrastructure

Project	Stars	One-liner
https://github.com/ddalcu/mlx-serve	2441	Native LLM inference server for Apple Silicon, compatible with both OpenAI and Anthropic APIs
https://github.com/vercel/workflow	2507	TypeScript workflow SDK, today storing the local world in a single SQLite database
https://github.com/theopenco/llmgateway	1736	Route, manage and analyze LLM requests across multiple providers behind one gateway
https://github.com/durable-workflow/workflow	1280	Core package for durable workflows and activities, v2.27.0 fixes operations still open at close
https://github.com/vllm-project/vllm-ascend	3078	Community hardware plugin for vLLM on Huawei Ascend, expanding Qwen3.8-Flash-Next A3 PD deployment guidance
https://github.com/himovo/movo	665	Self-hosted enterprise agent platform on DSH, with approvals, governance and a rollback baseline
https://github.com/basecompute/baseRT	250	LLM inference runtime for Apple Silicon, v0.5.0 open-source sync
https://github.com/salesforce/agentscript	277	An open, schema-driven language for configuring agent orchestration, synced from internal today
https://github.com/DecapodLabs/decapod	238	Metacognition for AI agents as a repo-native governance kernel, v0.141.0
https://github.com/NVIDIA-NeMo/nemo-helix	145	Ship agents faster, more accurately and more safely, adding a datasets list route today
https://github.com/cloudrift-ai/emmy	123	Optimized GPU compiler for LLM inference, v0.3.25 reads the cache once per verify step in native serving
https://github.com/flyteorg/flyte-sdk	129	Type-safe, distributed orchestration of agents, ML pipelines and real-time inference
https://github.com/alexsergivan/kwant	554	Your own team of AI bots in one self-hosted binary, with apps and accounts consolidated
https://github.com/stacklok/mecatl	309	Open-source agent harness built for cloud-native production, today wiring the embedded gateway's authorization server as AS-2
https://github.com/Fergana-Labs/stash	343	Open-source infrastructure for agents that learn from experience
https://github.com/slow-stack/mneme	223	Cross-session memory for DeepSeek Harness, v0.14.0
https://github.com/bigfish1913/pi-rust	260	Rust-native, library-first coding-agent runtime and terminal CLI
https://github.com/infercrane/infercrane	40	Open-source infrastructure for the full inference lifecycle
https://github.com/default-cybe/conclave	3	Self-hosted agent harness and hub, coding agents with any model on your own machines
https://github.com/MJ-CJM/DSH-Bot	28	Personal agents with memory and workspace group chats for DeepSeek Harness
https://github.com/GreenLv/dsh-session-insights	2	Local-first, evidence-backed workflow retrospectives for DSH, v0.8.1
https://github.com/zyf2492313716-cloud/dsh-session-hub	1	Cross-agent session hub for DSH, importing and searching OpenCode and other sessions
https://github.com/echohello-dev/pidex	0	Yet another coding agent harness with Pi, carrying session history context
https://github.com/SylphxAI/anyagents	0	Run any coding agent harness as a long-lived agent with durable mailboxes and real-time steering

3. Trends today

The single most important thing today is that evaluators themselves now have to be evaluated. The ultrasound report meta-evaluation produced the hardest numbers in this issue: BLEU, ROUGE-L, chrF, BERTScore and RaTEScore never reached AUROC 0.80 on the "planted clinical error versus harmless rewording" benchmark, while structured fact extraction and image localization did. That means every quality metric for AI reports built on text similarity over the past few years has probably been measuring how similar the writing looks rather than whether it is right. For me this is a concrete change to my own spec: the review metric for 4P intervention reports should move from similarity to fact extraction plus an abstain mechanism.
The division of labour between deterministic parts and probabilistic parts was written into code today in four different ecosystems. kawsay lets the AI pick only catalog IDs and portions while every number comes from a table; deterministic nutrition engines treat AI as an optional assistant; the clinical side pairs LLM drafting with a deterministic release gate; NVIDIA's SkillEvaluator splits static quality checks from live agent comparison. The common point is not "use less model" but fixing the parts that do not need a model as tables and thresholds first, and letting the model handle only the small slice that genuinely requires judgement. This matches exactly what I ran into these months building the GLP-1 nutrition companion — the sooner numeric computation leaves the prompt, the less rework follows.
Four DSH-ecosystem repositories moved on the same day, a first. movo (enterprise platform), mneme (cross-session memory), dsh-session-insights (evidence-backed retrospectives) and dsh-session-hub (cross-agent session hub), plus DSH-Bot. In one week this line went from personal plugin to industry adaptation to enterprise delivery boundary. From this issue on I track DSH as an independent item rather than an ordinary entry under track ④.
The skill ecosystem is turning from prompt fragments into engineering assets that CI is expected to police. NVIDIA is pushing skill quality gates and live agent comparison, google/skills and Anthropic's skill repository maintain skills as product lines, and NVIDIA/SkillSpector updated 7 times today purely on skill security scanning. Together with Argus's 266 red-team probes, this says skills now carry supply-chain risk and need versioned quality gates. If my 4P expert package stays a Markdown checklist, it will be a generation behind.
Local inference is forming its own track on Apple Silicon. mlx-serve (★2,441, compatible with both the OpenAI and Anthropic APIs), baseRT (an Apple Silicon runtime), emmy (a GPU compiler) and vllm-ascend (the Ascend plugin) all moved today. For this site's practical purposes this is direct: beyond the local Ollama qwen2.5:7b there is now a native alternative speaking the Anthropic API, so the 4P platform's local inference layer need not lock onto a single runtime.
Heat and activity are two different things and should be read separately. This week's trending leaders — anthropics/skills (★193,723, +6,055 for the week), mattpocock/skills (★323,052), addyosmani/agent-skills (★116,712) and cloudflare/security-audit-skill (★71,253, +10,700 for the week) — recorded no commit on their default branch today. The tables in this issue include only repositories with actual movement today; trading away hot-but-idle projects is preferable to passing star growth off as engineering progress.

4. Repo index

GitHub AI Daily · 2026-11-13 (Friday)

1. Today's picks (5)

1. Evaluating the evaluators: BLEU and BERTScore cannot catch planted clinical errors — structured fact extraction plus image localization can

2. Skills need quality gates too: NVIDIA now separates a failed evaluation arm from an arm that never ran

3. Label photos into the app: food-label recognition is split into "AI reads only, values normalized per 100 g, energy cross-checked against macros"

4. Authoritative nutrition and terminology sources packaged as MCP: USDA composition tables, OBO/FoodOn ontologies and taxonomy in one tool surface

5. DSH moves from personal plugin to enterprise: MOVO wraps it into a self-hosted platform with approvals, governance and a rollback baseline

2. Category roundup

① AI Coding / coding-agent & skill ecosystem

② Medical / clinical research AI & agent skill libraries

③ Health management / nutrition & food AI

④ Model layer & agent infrastructure

3. Trends today

4. Repo index