GitHub AI Daily · 2026-11-02 (Monday)

1. Today's picks (5)

1. Alice39s/agent-stash: a "skill drawer" for DSH — it distils reusable workflows out of finished sessions and saves them as Skills after review, which is exactly the missing piece in my expert-package workflow

URL: https://github.com/Alice39s/agent-stash ｜ ★16 ｜ Apache-2.0 ｜ TypeScript ｜ personal Agent Skills collection plus plugins for DSH (DeepSeek Harness), Pi, Cline, Grok Build, goose and Aider

What happened

Three commits inside the window (11-02 01:13 ~ 01:19Z). The volume is small, but every one lands on the question of how plugins get installed and used.

• The substantive change today is demoting the DSH requirement from a hard dependency to an optional peer: feat(dsh): declare the DSH requirement as an optional peer. In a repo that ships plugins for six harnesses, DSH is only one of them — downgrading it to an optional peer means Pi / Cline / goose / Aider users no longer have to install DeepSeek Harness first.
• The plugin index page gained categories and labels: ci(pages): categorize and label the plugins in the index — the six DSH plugins are no longer a flat list but grouped and annotated.
• A version release shipped alongside: Version Packages (changesets flow), which says this is a project with release discipline rather than a script tossed onto GitHub.
• The plugin list is itself a product roadmap: the README lists six DSH plugins — /init (analyse a project with another coding agent's /init preset and write an AGENTS.md), message edit (edit a message, regenerate a reply or retry a turn, with changes saved as switchable versions), skill scout (distil reusable workflows out of finished sessions and store them as Skills after review), prompt recall (page back through prompts sent in the current session with ↑↓), prompt autosuggest (complete input from the prompt history of all sessions), and prompt enhancer (one-click polish, structure and extraction of the core content of a prompt draft). On the skills side there is also code-styleguide (synced daily from the Google, Angular and Android Kotlin upstream) and self-use-<model>-subagent (delegate a task to a chosen model's subagent inside DSH).

Why it matters

• The "skill scout" plugin fills the thinnest part of my expert-package QA workflow. I build a standardised expert-generation workflow with WorkBuddy's expert-manager skill, and the most time-consuming step is always "turn a successful session into a reusable workflow". My current approach is manual review plus a hand-written SKILL.md; this plugin does the opposite — it distils candidate workflows out of finished sessions automatically, and a human reviews them before they become Skills. That "automatic distil, human review" boundary is drawn exactly where I need it.
• "Save the change as a switchable version" is a detail worth copying. When the message-edit plugin rewrites an agent's reply it does not overwrite in place — it stores a switchable version. That fixes one of my own bad habits: when I tune a prompt I overwrite the original, and afterwards I cannot say which version produced a given conclusion. Keeping edits as a rollbackable version chain is cheap, but it is worth a great deal when you need to trace a conclusion back.
• Apache-2.0: readable, modifiable, commercially usable — no copyleft burden for a one-person company that intends to ship a product. Skills install with npx -y skills add Alice39s/agent-stash and are compatible with the skills.sh ecosystem.
• And this is still the ecosystem already on my machine. The DSH plugin featured in the previous issue showed the community building on DSH; here comes another one — this is the second issue in a row with a DSH-ecosystem project, which is enough to conclude that DSH's plugin mechanism is not a private toy but genuinely used.

2. MarissaFamularo/Calibrate: turning "should the doctor trust the AI" into a measurable 2×2 matrix — over-reliance and reflexive dismissal name the exact risk I carry when producing supplement plans

URL: https://github.com/MarissaFamularo/Calibrate ｜ ★0 ｜ unstated ｜ Python ｜ clinical simulator for AI dependence in surgical residents (Meta Global AI Developer Hackathon, 2026-10-27 ~ 11-04; solo entry by a vascular surgery resident programme director)

What happened

Ten commits inside the window (11-01 19:28Z ~ 11-02 02:14Z). The main thread is making the simulator survive live playtesting.

• At its centre is a 2×2 dependence matrix that classifies every AI-assisted decision: AI right and the doctor followed = appropriate reliance; AI wrong and the doctor followed = over-reliance; AI right and the doctor overrode = reflexive dismissal; AI wrong and the doctor overrode = appropriate override. The report also counts separately the outputs the AI produced that the doctor never consulted, the alerts opened but never read, and the low-value alerts that were opened.
• The AI assistant is deliberately built to be right most of the time and wrong at specific points — sometimes confidently wrong, sometimes citing a real trial in the wrong context. The key design intent: the resident is not hunting for the AI's errors. The AI is a colleague to be used well, and the final decision remains the doctor's.
• Today pushed "unscripting" and detail alignment: the practice script became an unscripted sweep, workup orders from earlier notes carry into the workup step, ABI-alert scoring was scoped to visit 1, and outputs contradicting the case spec (such as "PVR waveforms provided") were removed.
• The reporting criteria were tightened too: the overall debrief now gives counts and named outputs only, with no speculation about why; the demo script and the live rehearsal script ship as a pair; and the attending never sees internal IDs or which items were planted.

Why it matters

• "Over-reliance" and "reflexive dismissal" name the real risk of what I am building. The AI assistant on the 4P supplement platform is, in essence, using an LLM to substitute for part of a nutritionist's judgement. Calibrate's whole evaluation framework can be moved across directly: when the AI's supplement plan conflicts with the lab markers, which cell should that decision be filed under? My platform currently records "what the AI said" but not "did the human trust the AI or overrule it" — and that is precisely the variable that decides whether a plan is safe.
• "Grade alerts by value" is the other thing the 4P platform should copy. The report counts alerts opened but not read, and low-value alerts that were read — my intervention report has the same problem: if all 24 risk dimensions are presented at the same weight and the same visual priority, a clinician will ignore all of them, which in effect is the same as showing no risk prompt at all. Counting the low-value alerts explicitly is the precondition for an alert system being usable at all.
• "The AI is a colleague, not an authority, and responsibility stays with the human" is a more accurate framing for medicine than "the AI suggests". For the supplement content and training I produce, this framework can become teaching language directly.
• One honest caveat: the repository has no LICENSE file and is a solo project. But its evaluation methodology (the 2×2 matrix plus per-item reconciliation) is not subject to copyright, so the thinking can be borrowed freely.

3. richmondgoh8/macro-longevity: writing a "coverage regression" into a health app — 140 lab references each mapped to a usable field, the single most transferable piece for the 4P supplement platform

URL: https://github.com/richmondgoh8/macro-longevity ｜ ★0 ｜ MIT ｜ JavaScript ｜ local-first Singapore health screening and longevity guide (semantic HTML, shared CSS, browser-native ES modules; no framework, no backend, no runtime packages, no build step)

What happened

Thirteen commits inside the window (11-01 12:22Z ~ 11-02 02:01Z). Every one of them circles the same goal: making the lab-reference side of the product trustworthy.

• The core of today is an "honest range coverage audit": Add sourced numerical laboratory intervals and honest range coverage audit and Record numerical interval deployment verification and remaining review scope. Translated: attach a sourced reference interval to every lab value, then state plainly which ones are not covered yet — rather than presenting a list that merely looks complete.
• The release gate is versioned: Complete versioned protocol measurement coverage and release gates and Close versioned coverage checklist with verified live release evidence — every release requires a checked coverage checklist plus verified evidence from the live site.
• The data scope is declared in layers: the Singapore baseline workspace carries 13 readily accessible markers (naming hs-CRP, ApoB and Lp(a) explicitly), plus 32 optional tracker fields covering full lipids, insulin, liver, kidney, iron, thyroid, chemistry, blood count and vitamin D; the Blueprint holds 140 published baseline measurements, 3 male additions, 1 female addition and 86 matched follow-up entries.
• One easily missed but clever design: a coverage regression maps every existing core-panel reference onto the usable tracker fields — the reference interval and the field you can actually enter into have to be mapped; if the mapping breaks, that reference effectively does not exist. There is also risk-stratified lipid targets and report-unit selection for ApoB, Lp(a) and urine ACR.

Why it matters

• The "coverage regression" concept is precisely the gap in my 4P supplement platform. I noted on the claimseed entry last issue that the lab-OCR extraction pipeline needs a ground-truth answer key plus a manifest SHA-256 regression evaluation. This repository offers a different answer to the same problem — not answer keys checked against answers, but a regression built on whether each reference has a corresponding enterable field. Both approaches hold, but this one fits my content side far more directly: what I actually have is 140 reference intervals, not a batch of PDFs waiting to be extracted.
• "Stating honestly which parts are not covered" matters more than "covering more". My intervention report carries 116 evidence-graded rules, each with its own reference interval. But if I never check which references in a given report are actually missing, then the number 116 is hollow. That is exactly what this repository did today — writing the remaining gap out explicitly instead of letting the numbers look good.
• Local-first, no framework, no build step — directly usable. MIT licensed with no commercial obstacle. It already runs a reachable HTTPS site with PWA install, offline support and Apple touch icons — the "I have to maintain this alone" constraint is satisfied at the architectural level, exactly in step with my hard limit of eight hours a week.
• It has already put ApoB, Lp(a) and hs-CRP into the baseline workspace. Those three are the hard markers in the atherosclerosis direction I care about, and someone else has already built and open-sourced this structure — which saves me the scarcest resource I have, namely time.

4. yantrikos/yantrikdb: the memory engine starts managing "contradictions" rather than just "storage" — and it puts the SQLite footgun that silently corrupts data on the first screen of its README

URL: https://github.com/yantrikos/yantrikdb ｜ ★71 ｜ Apache-2.0 ｜ Rust with Python bindings ｜ cognitive memory engine for AI agents: temporal decay, autonomous consolidation, contradiction detection and a knowledge graph

What happened

Eighteen commits inside the window (11-01 ~21:xx to 11-02 01:51Z). Today's thread is CI, cross-platform work and configurability.

• The concrete work today is engineering-side: ci: apply the CPU-only torch install to the lint and Linux test jobs, ci(benchmark): install CPU-only torch to stop filling the runner disk, ci: move macOS jobs from macos-14 to macos-15; redundancy_sim_threshold was made reachable from Python and enabled by default on deserialisation; contributors and reviewers are credited by name.
• The product boundary is stated very clearly: the README opens with "your agent starts every session as a stranger. Bolting a vector store onto it doesn't fix that — nothing is ever forgotten, near-duplicate memories pile up, and two contradictory facts come back ranked side by side with no signal that they disagree." What it builds is memory management, not just memory storage.
• The capability list: temporal decay, autonomous consolidation, contradiction detection and a knowledge graph, embedded in a Rust library with Python bindings.
• Two deployment shapes: as a library, pip install yantrikdb then call record_text() / recall_text() — with no sentence-transformers install and no ONNX runtime; or as a service, pip install yantrikdb-mcp dropped into the MCP config for Claude, Cursor, Windsurf or Copilot, where the agent auto-recalls context, auto-remembers decisions and auto-detects contradictions with no prompting required.
• The first screen of the README carries a footgun that silently corrupts data: never open the store with a second SQLite library in the same process — POSIX locks are per process, so two writers interleaving WAL commits corrupt the store silently. On Linux the engine detects the second instance and refuses to write (ForeignSqliteInstance); on every platform a commit that did not come through the engine queues an integrity check, and a failed check refuses writes with IntegrityCheckFailed.

Why it matters

• "Contradiction detection" is table stakes for my 4P platform, not a nice-to-have. A client's time-series data in the 4P supplement platform conflicts by nature: March LDL and September LDL are not the same number; three months of magnesium supplementation left serum magnesium flat; last round's dose and this round's dose differ. Today those conflicts are papered over by "take the newest row" — but taking the newest row is not the same as taking the most defensible one. A memory layer that can say "these two records contradict each other" directly decides whether an intervention report should stop and ask.
• "Nothing is ever forgotten" is a trap I keep stepping in. Every time I revise a client's protocol I add new information and never clean out the old — near-duplicate plans pile up until nobody can say which version is actually in force. "Temporal decay plus autonomous consolidation" targets exactly that: old information fades automatically while being merged into higher-level conclusions.
• That SQLite footgun is a direct warning about my own architecture. The 4P precision nutrition platform is Flask + SQLite. The failure mode the README describes — a second SQLite connection in the same process interleaving WAL writes and silently corrupting the data — happens easily in a Flask app (one request thread connecting directly while another goes through the ORM: a textbook double-write). The protection pattern it offers (engine-owned locking plus an integrity check plus refusing writes on failure) is worth auditing my own persistence code against.
• Apache-2.0 with no dependency on sentence-transformers or ONNX means installing it does not drag in a whole model runtime. For a one-person company working eight hours a week, that dependency bill is easy to justify.

5. bhargavchintam/aureproof: adding an "integrity" layer to AI evaluation — prove the grader itself works before letting its scores count

URL: https://github.com/bhargavchintam/aureproof ｜ ★0 ｜ Apache-2.0 ｜ Python ｜ open-source integrity layer for AI-agent evaluations: per-task health records, signed re-scorable evidence bundles and a pre-registered gate

What happened

Twelve commits inside the window (11-01 23:29Z ~ 11-02 02:14Z), mixing engineering progress with outward-facing material.

• It certifies three things about an evaluation run on top of the runners teams already use (Inspect AI first, Harbor next): ① the instrument — a per-task health record proving the grader works (its reference-positive, repaired-negative, no-op and anti-cheat controls all pass) before a task is allowed to count; ② the record — a signed, re-scorable evidence bundle built from content-addressed artefacts, a hash-chained journal and an in-toto statement; ③ the gate — a pre-registered verdict returning pass, fail or inconclusive.
• Today's specific changes: writing agent-explore in canonical JSON (fix(protocols): write agent-explore in canonical JSON so it can be signed) — because a protocol that cannot be signed is a protocol that cannot be verified; adding single-shot probes on OpenAI-compatible endpoints (feat(probe): single-shot probes on OpenAI-compatible endpoints); publishing an exploratory agent comparison on AkashML (feat(bundles): the exploratory agent comparison on AkashML); and adding the speedrun deck with its HyperFrames source and diagrams.
• It states its own status plainly as pre-release (0.0.x): the integrity core, the Inspect runner, signing, the comparison gate and one live signed evidence chain are built, but no benchmark result or performance claim is asserted yet. What remains before v0.1.0, and who it waits on, is written down in the release acceptance matrix.
• The author plants a hard rule first: Every claim, backed by evidence. The repo carries an SPDX header, an OpenSSF Scorecard badge and a REUSE status badge.

Why it matters

• "Prove the grader is good before letting its scores count" is worth an issue of this daily on its own. Two things I am building right now sit on this fault line: the lab-trend judgement in the GLPL nutrition companion, and the question-bank engine of the AI patient supplement simulation agent. My current evaluation method is "look at whether the agent's report is right" — which is the same mistake as "an evaluation that only looks at model output systematically overstates reliability". If the person doing the grading (me) is themselves fallible, the score means nothing. Aureproof runs four control tests against the grader (reference positive / repaired negative / no-op / anti-cheat), so the grader passes a medical check-up before it grades anyone else.
• The hash-chained journal plus offline re-scoring speaks directly to the persistence work on my 4P platform. The engineering conclusion I recorded last issue — an "intervention report persistence checklist" needing exact-value numeric comparison, None matching a stored null, and rejection of literal placeholders — lacks the layer Aureproof adds: the evidence bundle has to be re-scorable offline by a third party. For "AI assistant plus evidence-backed reports", the missing piece in my setup is exactly this: after the report is stored, can somebody verify it independently?
• A gate returning three states (pass / fail / inconclusive) is far more honest than binary pass/fail. It applies directly to my case: in evaluating generated intervention plans, "insufficient evidence to determine" has to be a first-class conclusion rather than being folded into "pass". A report that quietly mixes in items that were really inconclusive is more dangerous than one that labels them.
• Apache-2.0 is commercially usable, and it explicitly marks itself pre-release with no benchmark claims asserted — that kind of non-overclaiming status statement is itself a signal I trust when selecting a dependency, more than the many projects whose READMEs are full of benchmark numbers.

2. Category roundup

① AI Coding / programming agents and the skill ecosystem

② Medical / clinical research AI and agent skill libraries

③ Health management / nutrition and food AI

④ Model foundation and agent infrastructure

3. Trends today

Across the 102 entries today there is a single through-line: the agent ecosystem is starting to show layered self-doubt — not making models smarter, but adding verification layers to the evaluation, memory and installation plumbing that surrounds them. This is happening across all four directions at once.

• The evaluation layer is adding "the grader needs verifying too". bhargavchintam/aureproof requires the grader to pass four control tests before it is allowed to score any other task; cmisaico/healoi-ai-clinical-simulator runs two independent graders in parallel (a gap of two points or more goes to human review), verifies every citation (a score without a verifiable citation is discarded) and aligns against human raters with weighted kappa; MarissaFamularo/Calibrate splits whether a doctor trusts AI into a 2×2 matrix. Three projects, three languages, three domains, all pointing at the same thing: the credibility of model output cannot be guaranteed by it merely looking right.
• The memory layer is moving from "store" to "manage". yantrikos/yantrikdb defines the problem explicitly — nothing is forgotten, near-duplicate memories pile up, two contradictory facts come back side by side with no signal that they disagree. Pseudogiant-xr/Pseudolife-MCP emphasises a canonical-fact cortex and "honest forgetting"; fjwood69/mori separates curation (what to keep) from provenance (where it is valid). None of the three treats memory as a retrieval problem any more — it is a governance problem.
• Skill distribution has started treating "installation is an attack surface" as a topic. All three commits in hussi9/skill-router today are hardening: skills.sh registry strings are treated as untrusted input (they may not reach a command or the model before validation), entries are checked by origin file and anchor identifier with fullmatch, and when the audit board is unavailable the router fails closed rather than letting things through. Once a skill ecosystem has a central registry, "installing from the registry" is equivalent to "executing third-party code" — this repo is the only project in the window that confronts that head-on.
• Health and nutrition projects are increasingly writing down the parts they do not cover. richmondgoh8/macro-longevity records coverage gaps and the remaining review scope; iashahussain/rounds-dashboard promises 100% client-side operation, that patient data never leaves the device and that the repository contains no patient data; Razzleberrytt/hippie-scientist-site treats an Excel master sheet as the single source of truth with the public data directory as a publish target only. This is the same demand as the "coverage check across my 116 rules" note I made last issue.
• The medical line has a new shape: turning uncertainty into an executable branch rather than a single number. The assistant in MarissaFamularo/Calibrate is "right most of the time, wrong at specific predefined points"; cmisaico/healoi-ai-clinical-simulator has an explicit attack case where grader A, after prompt injection, invents citations — there to demonstrate that code-level controls catch it; bhargavchintam/aureproof lets the gate return inconclusive. All three refuse to compress "uncertain" into a number.

4. Repo index

Featured

• Alice39s/agent-stash — https://github.com/Alice39s/agent-stash
• MarissaFamularo/Calibrate — https://github.com/MarissaFamularo/Calibrate
• richmondgoh8/macro-longevity — https://github.com/richmondgoh8/macro-longevity
• yantrikos/yantrikdb — https://github.com/yantrikos/yantrikdb
• bhargavchintam/aureproof — https://github.com/bhargavchintam/aureproof

AI Coding

• https://github.com/hesreallyhim/awesome-claude-code
• https://github.com/backnotprop/plannotator
• https://github.com/pacifio/atlas
• https://github.com/Gentleman-Programming/gentle-ai
• https://github.com/rlaope/oh-my-hermes
• https://github.com/ginlix-ai/LangAlpha
• https://github.com/LeoYeAI/agent-master-skills
• https://github.com/zhnt/loushang
• https://github.com/awizemann/scarf
• https://github.com/23blocks-OS/ai-maestro
• https://github.com/FeatherHunter/dsh-mattpocock-skills-deck
• https://github.com/initializ/forge
• https://github.com/orbi-build/orbi
• https://github.com/AnastasiyaW/codex-claude-code-config
• https://github.com/dcc-mcp/dcc-mcp-core
• https://github.com/Fmarzochi/EGC
• https://github.com/dgalarza/agent-skills
• https://github.com/midasmatterhorn/artcraft-skills
• https://github.com/jkitchin/knowhow
• https://github.com/transcendr/slopware-skills
• https://github.com/hussi9/skill-router
• https://github.com/melodic-software/claude-code-plugins
• https://github.com/eugenelim/agent-ready-repo
• https://github.com/randommonicle/claude-skills
• https://github.com/trustabl/agent-reliability-skills
• https://github.com/bobmatnyc/trusty-tools
• https://github.com/paulnsorensen/easy-cheese
• https://github.com/event4u-app/agent-config
• https://github.com/dotcommander/cclint
• https://github.com/Alice39s/agent-stash
• https://github.com/KyaniteLabs/tastecheck
• https://github.com/ellmos-ai/bach
• https://github.com/fxylabs/spfn
• https://github.com/cplieger/marotte
• https://github.com/Pigbibi/TelegramAgentBot
• https://github.com/LindsayRidgeway/llm-symposium
• https://github.com/game4837/island-terminal-console

Medical / clinical research AI

• https://github.com/Builder106/med-core
• https://github.com/lyc0603/clinical_bench
• https://github.com/petergraffy/CLIF-LungCx-Epi
• https://github.com/Beijing-Haola-Tech/haola-checkup-mcp
• https://github.com/claimseed/claimseed
• https://github.com/iqbal-rahmatullah/tapal-ai
• https://github.com/MedYoussef-Abdelhedi/Medical-Agentic-GraphRAG
• https://github.com/eschraufnagel/trial-prescreening-agent
• https://github.com/not162/record-medical-development
• https://github.com/bandlatejesh18-alt/medagent-ai
• https://github.com/Sagark7825/Causal-Multimodal-Diagnostic-Agent-for-Medical-Imaging
• https://github.com/ldelvillar/veritrust
• https://github.com/raniyaaa/MedAgent-Radiology-Orchestrator
• https://github.com/datalife-ehealth/datalife-clinical-agents
• https://github.com/anaslari23/Multi-Modal-Health-AI-Agent-Guide
• https://github.com/southcul/international-business-for-medical-devices
• https://github.com/ichirio/tflplanner
• https://github.com/Cmy666688/medical-agent-workspace
• https://github.com/senseiewok/cf-research
• https://github.com/cmisaico/healoi-ai-clinical-simulator
• https://github.com/midnightnow/vetnotes-me
• https://github.com/bryanlandia/synthetic-medical-risk-agent
• https://github.com/ArjunaFransesco/clinical-ehr-mortality-risk-predictor
• https://github.com/zhaochenggang2026/medical-ai-hub
• https://github.com/yukiouma/aegis
• https://github.com/MarissaFamularo/Calibrate
• https://github.com/drliuhao/mh-liver-pathways
• https://github.com/abasuprab-png/MLR-Genie
• https://github.com/oloveluck/cql
• https://github.com/NytelAI/nytel.ai
• https://github.com/rgt47/zzedc

Health / nutrition

• https://github.com/ilodezis/vitals
• https://github.com/dhyyan/AI-Health-Nutrition
• https://github.com/PurnaJear06/Tracend
• https://github.com/SachinVenugopalan30/cupful
• https://github.com/robinchristopherheering/stride_app
• https://github.com/dexiadigi-cloud/trophe-api
• https://github.com/hammad-masood26/AI-CareCompanion
• https://github.com/Devinshen123/forma-health
• https://github.com/maxgfr/calorie-ai-tracker
• https://github.com/richmondgoh8/macro-longevity
• https://github.com/iashahussain/rounds-dashboard
• https://github.com/pipeworx-io/mcp-fruityvice
• https://github.com/courtneymuscat/Prokol
• https://github.com/yehsohimself/health
• https://github.com/Razzleberrytt/hippie-scientist-site
• https://github.com/akuttruff/cairn
• https://github.com/davehrsh/SaarAI

Model foundation / agent infrastructure

• https://github.com/rocketride-org/rocketride-server
• https://github.com/yantrikos/yantrikdb
• https://github.com/Amio242/mcp-memory-server
• https://github.com/fjwood69/mori
• https://github.com/Astrix-Labs/papez
• https://github.com/LindaHaviv/second-brain
• https://github.com/kagura-ai/memory-cloud
• https://github.com/axisvela/praxovela-open
• https://github.com/MrPeppersDev/agent-infrastructure-landscape
• https://github.com/cirwel/unitares
• https://github.com/Pseudogiant-xr/Pseudolife-MCP
• https://github.com/Cryremis/aether-core
• https://github.com/joaoh82/flotta
• https://github.com/Heinsithukyaw/Agent-Harness
• https://github.com/jihua3445/multi-model-dual-brain-memory-system
• https://github.com/go-llmflow/gateway
• https://github.com/bhargavchintam/aureproof