GitHub AI Daily · 2026-10-24 (Saturday)

1. Today's picks (5)

1. Omnigent: an 11k-star "meta-harness" that pulls Claude Code, Codex and Cursor under one roof — and starts governing them

URL: https://github.com/omnigent-ai/omnigent ｜ ★11,013 ｜ Apache-2.0 ｜ Python, published on PyPI plus a macOS desktop app

What happened
Omnigent positions itself as a meta-harness: a common orchestration layer above Claude Code, Codex, Cursor, OpenCode, Hermes, Pi and agents you write yourself, so you can swap or combine harnesses without rewriting. The capability list is unusually complete. Sessions follow you across devices — start in the terminal, continue in the browser, pick it up on the phone, with messages, sub-agents, terminals and files staying in sync. Multiple agents run in one session, so you can ask one agent to review another's work. On models it accepts a first-party API key, a Claude/ChatGPT subscription, or any compatible gateway. Policy governance can be scoped to the whole server, a single agent, or one chat: pause for human approval before risky actions, cap spend, limit which tools an agent can reach. Execution supports 12 sandbox backends (Modal, Daytona, Blaxel, Islo, E2B, Gensee, CoreWeave, Kubernetes, OpenShell, Boxlite, microsandbox, Databricks), plus team collaboration and session forking.

All four commits today are main-line stability fixes, and every one of them is an "orchestration-layer-only" problem: #11469 makes the server say when a held turn's host came back without its runner; #11463 leaves a Claude error report's failed edge to the StopFailure hook; #11458 shows the model-unavailable card on a model_not_found stop; #11436 stops the canvas from listing other people's sessions.

Why it matters

• Yesterday's theme was that harness competition had moved to cost structure and lifecycle management (hermes-agent, openhuman, deer-flow). Today's is one layer up: not building a harness, but governing harnesses. The practical consequence for enterprises is that you no longer have to bet on Claude Code versus Codex first — put policy and audit at the orchestration layer, keep the bottom replaceable.
• Where it fixes bugs is telling: "the host came back but the runner didn't", "which hook owns this error edge" are distributed-session state-reconciliation problems. Once this layer becomes the entry point, its reliability bar is infrastructure-grade, not developer-tool-grade.
• It maps unusually well onto healthcare work: one policy layer can express "no PHI leaves", "daily cap on paid API calls", "these tools are intranet-only", and it applies to every harness underneath — exactly the "declare once, enforce everywhere" shape a compliance audit wants, instead of every agent re-writing its own prompt.

2. Last EHR turns "an agent writing to the chart" into an approvable protocol — and publishes its own coverage ceiling

URL: https://github.com/cbetz/last-ehr ｜ ★4 ｜ Apache-2.0 ｜ TypeScript, Web + MCP surfaces

What happened
Last EHR is an agent layer on top of a headless FHIR EHR — not an EHR itself, and it never forks the backend. Three things are specified in unusual detail.

Reads broadly, and is willing to publish the number. It covers 25 of the 27 readable resource types in US Core 9.0.0, across 23 patient-scoped sections, following references so that authors, encounters and provenance stop being dead pointers. The coverage doc states its denominator and generation date (2026-05-31) and explicitly refuses to quote a percentage of R4 (with a section on why not).

The agent cannot fake an absence. A read reports when its window was capped, when a reference lookup was refused, and when a code filter matched nothing in a section that does hold records. Filters a section cannot apply are refused with the legal values instead of being silently dropped. The README calls these safety-boundary tests, not best-effort behaviour.

Every write is a proposal. It open-sourced Approval-Gated Agent Writes on FHIR v0.1: Proposal → Decision → Commit → Audit, with two implementations (a web approval card and MCP elicitation-gated write tools) and a conformance suite that checks implementations using its own FHIR reads (npx @lastehr/agent-write-conformance). The MCP surface is read-only by default with four bounded read tools; writes require per-action human approval. Five FHIR backends sit behind one interface (Medplum, HAPI, Firely, Aidbox, Oystehr).

Today's commits (PR #200, "durable work lab") connect a real Claude assistant to a reviewed, durable FHIR workflow, and record acceptance and persistence evidence for that real assistant.

Why it matters

• Clinical agents have long been stuck at "can read, can't write". This project's answer is not "make the model more reliable" but demoting write access to a proposal and backing it with a protocol plus conformance tests — one of the few practices that turns human-in-the-loop from product copy into a checkable interface.
• The line "it cannot report an absence it never checked for" belongs in every clinical/health agent design spec: the most dangerous model error is not a wrong answer, it is turning "I didn't find it" into "there isn't any".
• It self-labels alpha, restricts itself to synthetic data, and gives HIPAA deployment groundwork its own document — the "declaring your own boundaries is a maturity signal" pattern we have tracked for several issues. Four stars does not weaken the point; the protocol matters more than the star count.

3. An HL7 official repository grows a molecular tumour board from patient banner to AI synthesis in one day

URL: https://github.com/FHIR/genomics-operations ｜ ★34 ｜ Apache-2.0 ｜ public reference implementation of HL7 FHIR Genomics Operations

What happened
FHIR/genomics-operations is the official reference implementation of HL7 FHIR Genomics Operations (with an accompanying JAMIA paper). It provides operations such as find-subject-variants and find-population-tx-implications, standardizing "which variants does this patient have, and what phenotype or medication implications follow" as FHIR interfaces. Its commit log over the past two days is tightly clustered around a newly added MTB (Molecular Tumor Board) application, and it builds up in layers:

• Oct 23 18:41 — use clinical data to create a patient banner
• Oct 23 20:59 — use clinical data to create a chart summary
• Oct 23 23:47 — construct a treatment timeline
• Oct 24 00:51 — add AI-based synthesis

In other words, an interface reference implementation maintained by a standards body grew an AI application skeleton aimed at a clinical decision conference — banner, summary, timeline, synthesis — in about six hours.

Why it matters

• This is the most important directional signal of the day: the AI application layer is being written directly into a standards reference implementation. Standards repos used to own interfaces and tests; now they also demonstrate how AI should sit on top. For health-IT teams this means official examples will gradually become the default answer for product shape, and both procurement and compliance review will align to them.
• MTB is a well-chosen scenario: it is where genomic data actually lands (somatic variants → available trials → therapy), and it is the archetypal "multi-source evidence + timeline + synthesized recommendation" problem — precisely what agents are good at.
• One caveat: a reference implementation is not a deployable clinical system. It gives you data shapes and a process skeleton; the judgement logic and accountability boundaries are still yours. Today's value is that the standards side has acknowledged the AI-synthesis step, not that this is ready for clinical use.

4. A 240-run ablation on a pharma evidence agent: the skills layer is the cost-performance knee; Sonnet costs 16× more for 2 points

URL: https://github.com/Akhilesh-Vangala/healthcare-evidence-agent ｜ ★1 ｜ MIT ｜ Python 3.10+, Claude Code as runtime

What happened
This is a governed evidence agent for biotech commercial, medical and government-affairs questions, pinned to three public sources: FDA labels, ClinicalTrials.gov, and Medicare NCDs. The engineering shape is clear: Claude Code is the runtime, tools are a 6-tool MCP server, the procedure ships as 4 Agent Skills, and it states plainly that it uses public data only, contains no PHI, and is not medical advice.

The most valuable part is that the ablation is part of the body copy (240 runs, with 95% CIs):

The second hard constraint is a programmatic verifier: a claim is accepted only if its source_id was retrieved in that run and the quote is a verbatim substring of that source. Today's commits split the repo into frontend, backend and agent workspaces and remove local files that would leak deploy tokens.

Why it matters

• These numbers give a rare quantitative answer to "should we invest in a skills layer": closed book 23% → tools 80% → skills 95%, while a stronger model only moves 95% to 97% at 16× the cost. For budget-sensitive business (health management, nutrition coaching — high frequency, low unit price), the implication is direct: fund tools and skills before you fund a bigger model.
• The "verbatim quote + retrieved this run" check is an order of magnitude stronger than "prompt the model to cite": it turns hallucinated citations into an executable assertion rather than a spot check. Same lineage as yesterday's Odyssey "a test per claim".
• Governance sits in the code: a source allowlist, no PHI, pass rate and failure cases (all three misses in config C were "review-flag over-escalation") are all in the report. Classifying failures beats reporting a single score.

5. Nutrition as hard-constraint solving: budget, protein floor, sodium cap and allergens at once — reconciled against official pricing

URL: https://github.com/harry1729/mcd-nutrition-planner ｜ ★1 ｜ MIT ｜ JavaScript, zero dependencies

What happened
A decision assistant for group and meeting catering. You say "20 people, ¥40 per head, Shenzhen Science Park"; it queries real store menus through the McDonald's MCP, prices them with the official pricing API, and then solves under four simultaneous constraints — per-head budget as a hard cap, a protein floor, a sodium ceiling, and allergen avoidance — using a two-dimensional 0/1 knapsack DP to allocate coupons optimally. Output is an orderable list (quantity, unit price, total, per-head cost, plus per-head calories/protein/sodium estimates), finished with a consistency check between MCP official pricing and the local estimate (the sample output shows ¥630.00 on both sides).

Engineering: zero dependencies, zero build, and npm test verifies the entire solving logic offline (45 tests passing) without an MCP token. The README states outright that "store menus and prices rotate daily, so re-runs differ — that is normal, not mock data".

Today's commits are a README first-screen rewrite and a WeChat publish package (image design, compliance checks, one-click export); the solver itself landed on Oct 22. It is a McDonald's 1024 Developer Day entry: small, one star — read it as a method sample, not a production tool.

Why it matters

• Nutrition products have been stuck at "log + advise". This one changes the question: not "what should you eat" but "is there a feasible solution under these hard constraints". Turning nutrition from soft guidance into a feasibility problem is a shape you can drop straight into corporate catering, canteen menu planning or chronic-disease meal planning.
• The price reconciliation step is the key move: a plan is only emitted if the local estimate matches official pricing. That makes data trust a runtime assertion rather than a disclaimer.
• The continuity is worth noting. We covered FoodGuard (offline national-standard ingredient grading) on Oct 19, and Edamam's official skills plus CKD six-axis measurement on Oct 22; today's is the third route — constraint solving. Together they show the competitive edge in nutrition AI has moved from "recognizing food" to "producing an executable, checkable, bounded optimum".

2. Category roundup

① AI Coding / Programming agents and the skills ecosystem

② Medical / Clinical research AI and agent skill libraries

③ Health management / Nutrition and food AI

④ Model layer and agent infrastructure

3. Trends today

1. Write permission has become the dividing line for clinical agents. Yesterday's clinical agents were still competing on "answers correctly, explains clearly"; last-ehr rewrites the question as: who approved this write, was what got committed exactly what was reviewed, and can it be traced afterwards. The four-stage Proposal → Decision → Commit → Audit protocol plus a conformance suite is likely to become the default skeleton for the next wave of in-hospital agents; writes without an audit trail will struggle to pass review.
2. Evaluation is shifting from a leaderboard to an ablation table plus checkable citations. The 240 runs in healthcare-evidence-agent decompose the gain cleanly: +60 points from tools, +15 more from skills, and only +2 from a stronger model at 16× the cost. That gives a reusable framework for where to spend: fund tools and skills before models; and "verbatim quote + retrieved this run" turns hallucination from a spot-check problem into an assertion problem. Same discipline as Odyssey's "a test per claim".
3. A layer that governs harnesses is growing on top of the orchestration layer. Omnigent at 11k stars pulls Claude Code, Codex, Cursor, Hermes and Pi into one surface, then stacks policy and sandboxes on top; everything it fixed today was distributed-state work — hosts losing runners, which hook owns an error edge. The implication: enterprises no longer need to bet on one harness; they can unify governance and audit at the orchestration layer and keep the bottom replaceable.
4. Standards-body repositories have started demonstrating the AI application layer directly. HL7's FHIR/genomics-operations went from patient banner to AI synthesis in a day. That changes procurement and review expectations more than any startup demo could — the pressure of "the official reference implementation looks like this, so the product should too" will flow through to health-IT vendors.
5. The third route for nutrition AI: treat nutrition as hard-constraint solving. After the offline national-standard rule engine (FoodGuard) and official skills plus specialty measurement (Edamam / CKD six-axis), today brings multi-constraint optima reconciled against official pricing. All three routes share one trait: they are no longer satisfied with recognizing food, and demand something executable, checkable and explicitly bounded. Teams in health management should treat "are the constraints hard-coded, is the data reconciled" as a selection criterion.
6. Health-security watch (issue 2): the php2shell fix window is closing. PHP 8.5.12 / 8.4.27 / 8.3.36 shipped on Oct 21, yet php-src still climbed to No. 2 on the trending page today with +716★, which means many teams are still at the "which version do I need" stage. If the self-hosted health/nutrition/MCP projects we have covered (HealthLog, SparkyFitness, plate, medicationtrackerbot and others) run on a PHP stack, check the version today — this pre-auth RCE needs no application-layer bug, so it will not wait for next week's sprint.

4. Repo index

AI Coding

• https://github.com/omnigent-ai/omnigent
• https://github.com/codewhale-hq/Codewhale
• https://github.com/esengine/DeepSeek-Reasonix
• https://github.com/kirodotdev/KiroCrew
• https://github.com/agentconnect-md/agentconnect
• https://github.com/modu-ai/moai-adk
• https://github.com/SnailSploit/Claude-Red

Medical / Clinical research AI

• https://github.com/cbetz/last-ehr
• https://github.com/FHIR/genomics-operations
• https://github.com/Akhilesh-Vangala/healthcare-evidence-agent
• https://github.com/aurelianware/cloudhealthoffice
• https://github.com/MrCosta77/FHIR-to-OMOP
• https://github.com/open-recruit/recruit
• https://github.com/microsoft/HealthAgentBench

Health management / Nutrition and food AI

• https://github.com/harry1729/mcd-nutrition-planner
• https://github.com/January-ai/january-sdk-ios
• https://github.com/BasedHardware/omi
• https://github.com/Mentra-Community/MentraOS
• https://github.com/Delubear/GlucoseTray
• https://github.com/Growstuff/growstuff
• https://github.com/castefeudal/gymdo

Model layer and agent infrastructure

• https://github.com/opensearch-project/agent-health
• https://github.com/ai-dynamo/dynamo
• https://github.com/raullenchai/Rapid-MLX
• https://github.com/algorithmicsuperintelligence/openevolve
• https://github.com/cloudgeni-ai/opengeni
• https://github.com/holon-run/holon
• https://github.com/php/php-src
• https://github.com/OpenBMB/VoxCPM