fgcz-context

FGCZ Institutional Context

This skill is a substrate, not a workflow. It does not perform an action — it tells the agent what world the FGCZ skills operate in. Load it once per session when working on FGCZ tasks, and the institutional facts below become available to subsequent skill activations.

When to use

• At the start of any session involving FGCZ infrastructure (SUSHI, B-Fabric, gStore, Draugr, FGCZ cluster).
• Before invoking any other plugin in this marketplace, especially for first-time users or external collaborators unfamiliar with FGCZ conventions.
• When a downstream skill references a path (/srv/gstore/projects/, /srv/GT/analysis/), a host (fgcz-r-029, fgcz-c-051), a module (Dev/R/4.5.0), or a conda environment (gi_py3.11.5) and the agent needs to know what these are.
• When generating a report or recommendation that mentions FGCZ-specific entities and the audience may be unfamiliar with them.

Do not use as a primary how-to source. This skill describes the world; operational skills (fgcz-infrastructure, bfabric-lims, etc.) describe how to act in it.

What is FGCZ

Functional Genomics Center Zurich (FGCZ) — joint genomics, transcriptomics, proteomics, and metabolomics core facility of the University of Zurich (UZH) + ETH Zurich, located in Zürich, Switzerland.

• Scientific scope: NGS (bulk RNA-seq, WGS/WES, ATAC-seq), single-cell (10x Chromium ecosystem, scVI/Seurat), spatial transcriptomics (Visium, VisiumHD, Xenium, CytAssist), immune repertoire (TCR/BCR via MiXCR + scRepertoire), proteomics, LC-MS metabolomics. Strong 10x Genomics emphasis.
• Service shape: Academic core-facility model. Per-project accounting via pXXXXX (project) / oXXXXX (order) IDs in B-Fabric.
• User base: Primarily internal FGCZ staff, secondarily UZH/ETH researchers submitting projects via B-Fabric. External collaborators are minority.
• Web entry points:
  • https://fgcz-bfabric.uzh.ch — LIMS
  • https://fgcz-sushi.uzh.ch — workflow framework (production)
  • https://shiny.fgcz.uzh.ch — public ShinyProxy
  • https://gitlab.bfabric.org — self-hosted GitLab

Tech / operational stack

Layer	Component	Notes
LIMS / metadata	B-Fabric	fgcz-bfabric.uzh.ch. Python: bfabric + btools. Credentials at ~/.bfabricpy.yml. Conda env gi_py3.11.5.
Workflow framework	SUSHI	Ruby on Rails + Python job manager + MySQL. 4 instances: prod / demo / course / test. App pattern = Ruby parameter class + R/Rmd backend via ezRun.
Analysis backend	ezRun R package	Source under ~/git/ezRun / ~/git/ezRun-scmultiomics.
Demultiplexing	Draugr v2.5.5+	FGCZ-internal Python tool. Cron-driven (populate_queue.py + monitor_runs.py), wraps bcl2fastq / bases2fastq.
Compute	SLURM on Debian 12 + Lmod	Named hosts: fgcz-r-029 (reference builds), fgcz-c-041/043/044/051/053/054/176 (compute). GPU: NVIDIA L40S, Blackwell. Common partition: employee.
Environment management	Pixi + Lmod modules + conda envs	Pixi for project-local rapidsai/conda-forge. module load Dev/R/4.5.0 style. Production envs (gi_* prefix) are read-only; use _dev variants for installable.
Storage	gStore	Long-term delivery: /srv/gstore/projects/pXXXXX/ (read-only after copy). Working: /srv/GT/analysis/pXXXXX/. References: /srv/GT/reference/. Databases: /srv/GT/databases/. Raw BCL: /export/local/data/. Copy semantics: g-req copynow / gstore-list.
SCM / CI	Self-hosted GitLab	gitlab.bfabric.org. PAT-only auth (OAuth disabled). Manage via glab CLI.
App delivery	ShinyProxy	Two instances: private (LDAP against B-Fabric users) + public. Containerized via fgcz_base_image Docker base.
Pipelines	Nextflow / nf-core	rnaseq, sarek, atacseq. Coexists with SUSHI.
10x tools	CellRanger / -ARC / -ATAC / multi, SpaceRanger, XeniumRanger	FGCZ-specific Sample2Barcode CSV format (differs from upstream 10x).
Single-cell stack	Seurat v5, scVI/scANVI/totalVI/MultiVI, RAPIDS singlecell, scRepertoire, RCTD (R + GPU py), CellChat, Slingshot, SCEVAN, SComatic, CytoTRACE 2, clusterProfiler, infercnvpy, STalign	Multi-language (R + Python), qs2 serialization standard.
Reporting	R Markdown + ezRun templates	Institutional style: 300 DPI, polychrome palette, coord_fixed UMAPs, ezInteractiveTableRmd.

Tool packages (FGCZ-specific)

These are the institutional tool packages an agent should distinguish. Generic tools (SLURM, Lmod, Pixi, Nextflow, GitLab) are not listed here.

B-Fabric — LIMS web app + SOAP API

The central LIMS for FGCZ. Web UI at https://fgcz-bfabric.uzh.ch. Exposes a SOAP API for programmatic access. Stores projects (pXXXXX), orders (oXXXXX), samples, workunits, datasets, resources, users, lineage. Owned by: IT group.

bfabricPy — Python library wrapping B-Fabric SOAP API

The canonical Python interface to B-Fabric. Provides Bfabric class for direct API calls. Credentials in ~/.bfabricpy.yml (sections like PRODUCTION, TEST). Used by all higher-level Python tools and by user scripts. When to use: any Python code that needs B-Fabric read or write access.

btools — CLI tools for B-Fabric operations

A set of command-line tools (e.g. register_sushi_dataset_into_bfabric, update_resource_status, delete_workunit_by_dataset_id, get_project_id_from_order, check_dataset_bfabric) that internally use bfabricPy to talk to the B-Fabric SOAP API. Source: ~/git/btools. When to use: shell-level B-Fabric operations, especially registration of analysis outputs back into B-Fabric lineage. Relationship to bfabricPy: btools = CLI surface; bfabricPy = library. They are layered, not parallel.

SUSHI — workflow framework

FGCZ-internal pipeline submission framework. Ruby on Rails web frontend + Python job manager daemon (the 2024-era replacement of legacy workflow_manager) + MySQL DB. Apps are defined as Ruby parameter class + R/Rmd backend via ezRun. Jobs dispatch to SLURM. Four instances: prod / demo / course / test. Web entry: https://fgcz-sushi.uzh.ch CLI: sushi_fabric Source: ~/git/sushi/master Owned by: Genome Informatics group.

ezRun — R analysis backend

R package providing the analysis layer that SUSHI apps call into. Encodes the FGCZ report templates (300 DPI, polychrome, ezInteractiveTableRmd), helper functions (ezLoadPackage, ezWrite.table, setwdNew, ezSystem), and per-method analysis modules. Source: ~/git/ezRun (and forks like ~/git/ezRun-scmultiomics). Owned by: Genome Informatics group.

Draugr — sequencing demultiplexing automation

FGCZ-internal Python tool (v2.5.5+, Python 3.12+) that wraps bcl2fastq / bases2fastq, polls B-Fabric for new runs, demuxes, and archives FASTQ to gStore. Cron-driven via populate_queue.py + monitor_runs.py (screen-based daemon). Owned by: Genome Informatics group.

gtools — gStore file server CLI

A package of CLI tools (commands include g-req, gstore-list, …) for interacting with the gStore file server (/srv/gstore/). gStore is read-only at the POSIX level for safety (prevents accidental deletion of long-term project deliverables). The only sanctioned way to write to gStore is via gtools, which logs operations and enforces permissions.

• g-req copynow <src> <dst> — copy a file or directory into gStore. The source must be on a working filesystem (typically /srv/GT/analysis/...). gStore-side dst is created by the daemon.
• gstore-list — list / poll gStore contents.

Why this design exists: gStore holds the canonical, billable deliverables that are referenced by B-Fabric lineage. A rm -rf mistake there breaks audit chains. gtools mediation prevents both accidental deletes and silent overwrite of audit-tracked files. Agent rule: never propose rm, mv, or direct cp into /srv/gstore/projects/. Always route writes through g-req.

ShinyProxy — containerized app delivery

Open-source app server (not FGCZ-built, but configured with FGCZ-specific topology). Two instances: private (LDAP-authenticated against B-Fabric users, https://shiny.fgcz.uzh.ch) + public. Hosts R Shiny / Python (Gradio, Streamlit) / Quarto apps in Docker containers based on fgcz_base_image. Used to deliver interactive viewers (e.g. exploreSC, BD Rhapsody dashboards). Owned by: Genome Informatics group (deployment) + IT group (LDAP / cert / infrastructure).

How they compose

B-Fabric (LIMS + SOAP API)
    ↑ ↑
    │ └── bfabricPy ──── btools (CLI)
    │                       │
    │                       └─ used by Draugr, SUSHI, custom scripts
    │
    └── SUSHI (Rails + job manager + MySQL)
            │
            ├── dispatches → SLURM ──── runs ezRun (R) on compute nodes
            │
            └── inputs from gStore, outputs back to gStore via gtools

Draugr ── reads B-Fabric runs ── writes FASTQ to gStore via gtools

ShinyProxy ── reads gStore (read-only mount) ── authenticates via B-Fabric LDAP

Canonical data flow

B-Fabric sample/order metadata
   ↓
sequencing → BCL on /export/local/data/
   ↓
Draugr demux + B-Fabric integration
   ↓
FASTQ archived to gStore /srv/gstore/projects/pXXXXX/
   ↓
SUSHI app submission (CellRanger / SpaceRanger / nf-core / custom)
   ↓
analysis under /srv/GT/analysis/pXXXXX/
   ↓
Rmd/HTML reports + qs2 objects
   ↓
g-req copy back to gStore
   ↓
B-Fabric dataset registration (with lineage)
   ↓
(optional) custom Rmd → SUSHI-shaped artefact (Hubert blueprint)

FGCZ organizational structure

FGCZ is organized into 5 groups:

Group	Role	Skills user?
Genomics & Transcriptomics group	Wet-lab — sample prep, library construction, sequencing runs	Indirect (sees B-Fabric / SUSHI web UI)
Genome Informatics group	Manages and develops genomics infrastructure (ezRun, SUSHI, Draugr) + analyzes user data	Primary skills user
Proteomics group	Wet-lab + bioinformatics for proteomics	Primary skills user (bioinformaticians)
Metabolomics group	Wet-lab + bioinformatics for LC-MS metabolomics	Primary skills user (bioinformaticians)
IT group	B-Fabric and infrastructure operations	Uses bfabric-lims, fgcz-infrastructure

Primary skills users (the core audience): bioinformaticians in Genome Informatics + Proteomics + Metabolomics groups. When in doubt about the intended audience of a skill, assume one of these three groups.

User personas (task roles within the primary groups)

These are task shapes — a single person typically wears several of these hats depending on the project.

Persona	Typical skill use	Task shape	Primary group
Pipeline / infrastructure developer	sushi-framework, fgcz-sushi-app-dev, fgcz-modules, fgcz-gitlab, shinyproxy-deployment, scmultiomics-analysis	Build / maintain SUSHI apps, ezRun backends, deployment	Genome Informatics
Demux / sequencing operations	draugr-demultiplexing, sample2barcode-generation	Routine demux monitoring, base-mask config, queue triage	Genome Informatics
Single-cell / spatial analyst	seurat-analysis, cellranger-fgcz, scvi-tools, cellchat, slingshot, cytotrace2, clusterprofiler, rctd, scevan, scomatic, screpertoire	Per-project analysis pipeline, repeatable template, FGCZ-styled Rmd	Genome Informatics
Spatial specialist	spaceranger-fgcz, xenium-ccf-registration, insitucnv, split-purification	Method-heavy Visium/VisiumHD/Xenium work	Genome Informatics
ML / GPU specialist	rapids-gpu-analysis, scvi-tools, rctd-py	>500K cells, Pixi envs, GPU SBATCH	Genome Informatics
Custom analysis / project delivery	autonomous-render, fgcz-custom-analysis-register, bfabric-tools	One-off Rmd → SUSHI-shaped artefact (Hubert blueprint); register back to lineage	Genome Informatics
Metabolomics analyst	metabolomics-curation	LC-MS feature QC (100-300 features)	Metabolomics
LIMS / dataset admin	bfabric-query, bfabric-tools	Registration, lineage tracing, metadata export	IT + Genome Informatics
External / GEO reanalysis (occasional)	nextflow-development	Self-service nf-core on public data — aspirational; actual primary user is staff doing the same	Any

Why these skills exist (gaps they fill)

1. Institutional tacit knowledge — paths, hostnames, queue partitions, module names, conda env names, SUSHI deploy instance choice. Previously tribal.
2. Cross-system glue — B-Fabric ↔ SUSHI ↔ gStore ↔ GitLab handoffs are non-obvious; many failure modes at boundaries.
3. FGCZ-specific formats — Sample2Barcode CSV ≠ upstream 10x, dataset.tsv lineage shape, ezRun report templates, Hubert blueprint for SUSHI lineage capture of custom analyses.
4. Self-hosted quirks — gitlab.bfabric.org PAT-only, ShinyProxy LDAP, FGCZ-internal Draugr tool, gStore read-only-after-copy semantics.
5. Operational gotchas — "CellRanger multi CSV is the #1 error source" (per cellranger-fgcz); CellChat 21 lessons; Draugr base masks; Seurat v5 object changes.
6. Autonomy hooks — autonomous-render collapses iterative render-debug cycles into one agent invocation. AI-shaped, not human-shaped.
7. Standardized report identity — 300 DPI, polychrome, coord_fixed UMAPs codify FGCZ visual brand.

Conventions agents should default to

• Paths: when a user mentions "project NNNNN", interpret as gStore /srv/gstore/projects/pNNNNN/. Working space is /srv/GT/analysis/pNNNNN/. gStore is POSIX-read-only by design (safety against accidental deletion of long-term deliverables). Never propose rm, mv, or direct cp to /srv/gstore/. Write to /srv/GT/analysis/, then deliver via g-req copynow (the gtools-mediated path is the only sanctioned write route).
• Order ID oNNNNN is distinct from project ID pNNNNN. Both are B-Fabric concepts.
• Personal paths in examples: this marketplace contains example paths like /home/{username}/, Analyses_{Name}/, ps_{name} conda envs. Treat these as templates, not literal paths — substitute the current user's equivalents.
• gi_ conda environments are production / read-only* — do not pip install into them. Use _dev variants or Pixi.
• GPU jobs: prefer --constraint=L40S (or Blackwell) over --nodelist=fgcz-r-NNN to avoid queue-forever when a specific node is busy.
• Reports: R Markdown via ezRun, 300 DPI, polychrome palette, coord_fixed() for UMAP, qs2 (not qs) for serialization.
• Credentials: never embed in SKILL.md or scripts. B-Fabric: ~/.bfabricpy.yml. MySQL: --defaults-file=~/.my.cnf or MYSQL_PWD env var from secret. GitLab: PAT via glab auth login (not from agent session).

Known limitations / open questions

The following are not encoded in this skill and should be treated as ambiguous until clarified for a given task:

• UZH vs ETH governance details and exact billing model.
• RBAC: who is permitted to run Draugr / SUSHI admin commands vs only submit via web. Not documented in this marketplace.
• Course/demo SUSHI vs production data boundaries — which instance is appropriate for what data is staff-knowledge.
• CyteType / Nygen Analytics API — institutional license vs per-user keys. Not stated.
• Skill maturity variance — some skills (seurat, cellranger, draugr, spaceranger) are production-hardened; others (cytotrace2, cytetype, scevan, scomatic) are newer. Frontmatter does not flag this.
• Nextflow ↔ SUSHI division of labor — which analyses route through SUSHI vs nf-core is convention, not enforced.
• BD Rhapsody and non-10x platforms — mentioned only in scmultiomics-analysis. Coverage uneven.
• Cross-group skill adoption — Genomics & Transcriptomics (wet lab) and IT groups are not primary skills users today; whether this will shift is open.

Confidence and provenance

This context was constructed on 2026-05-22 by independent skim + integration from 4 LLMs (Claude Opus 4.7 Agent persona, Claude Opus 4.6, Codex GPT-5.5, Cursor Composer 2.5) reviewing the fgcz-skills marketplace at v1.3.2. 4/4 reviewers agreed on all items above with no factual conflicts. Tech stack items are evidenced from the marketplace itself. Governance/funding items marked above are inferred from public domain evidence (*.uzh.ch, [email protected]).

This skill is a snapshot. FGCZ infrastructure evolves (Job Manager 2024 architecture shift, new GPU nodes). When this skill's content conflicts with current operational reality, current reality wins — and the conflict should be flagged for skill update.

See also

• meta-skills/multi-llm-review — workflow to refresh / verify this context against the current state of the marketplace.
• meta-skills/skill-audit — deterministic linter that uses conventions from this skill as baseline.
• fgcz-infrastructure/* — operational skills that assume this context.
• bfabric-lims/* — LIMS operations.