repositories · 2026-10-17

the evidence behind d7 repository shape: what the organization's 118 repositories are, who visits them, what couples them, and what the prior art says. measured on 2026-10-17 with gh, the crates.io api and the trees in ~/cyber.

the inventory

the ladder already distinguishes crate, part (a named part that is not its own crate: six parts of trident live in one repository, nine parts of cy in cyb), product, network, ops. the registry in .github exists because 118 doors need a map.

who comes

stars: go-cyber 360 · cyb 174 · chaingear 82 · cyber 53 · space-pussy 12 · trident 10 · honeycrisp 7 · erga 5 · cyber-valley 3 · soft3 2; every other repository has one or none. the ~60 repositories of the new stack hold about 30 stars together.

traffic, 14 days, all 72 public repositories: 446 views by 188 unique visitors. cyb 106 uniques, cyber 23, trident 12, soft3 6; 68 of 72 repositories had fewer than five visitors. referrers are github.com itself, t.co and a few Google hits on trident; the most visited paths are cyb/issues, cyb/releases, cyber/pulls. the 20,736 clones are our own agents and runners (soft3 2,529, zheng 1,104, radio 869).

reading: on github the organization has three doors — cyb, cyber, trident — and the rest is invisible. a visitor who lands on lens sees one star and no context.

crates.io

21 of the 47 main crates checked are published under cyberia-to; ~7,000 downloads in total, mostly our own ci. publishing already happens from folders: unimem and rane from honeycrisp, strata-core from strata, cyber-tok from plumb, three crates (cyber-mudra, cyber-rune, inf-cli) published without a repository link. six repository names are taken on crates.io by others — cy, mir, optica, particle, trop-cli, wasmi — so repository name and crate name are already different things. crates.io does not care how many repositories hold the crates.

what couples the repositories

• path dependencies into siblings: honeycrisp 107 references, rs 51, strata 48, cyber-bao 39, evy 27, hemera 24; git dependencies on radio 18, hemera 7.
• soft3/release/phase1.toml pins eight sibling revisions for the release train; the rule "a bump pull request, never a path edit" exists only because the parts live apart.
• the component-boundary change of 2026-10-07 took 16 pull requests across 17 repositories, several stacked on other open pull requests because origin defaults did not build against local siblings (soft3/roadmap/component-boundaries.md).
• continuation slices wait for the row's first pull request to merge in another repository (launch rows).
• open pull requests today: 113 across 38 repositories; 23 repositories hold one or two.

prior art

• polkadot-sdk (Substrate + Polkadot + Cumulus, merged 2023-08): open-to-merge time fell about 25 % a month after the merge, the share of issues resolved rose from 39–48 % to 50–55 %, and the up-to-30 monthly "companion" pull requests disappeared; merged pull-request volume fell and the authors call that a misleading metric (forum, article, announcement).
• the rust norm: one workspace per product, many crates published from its folders — bevy (~50 crates), nushell (42 crates; our nu fork carries all 42 in one repository), iroh, wasmi, risc0. separate repositories appear where cadence, ownership or audience differ (tokio keeps mio and tracing apart).
• github mechanics (rename, transfer): a rename or a transfer keeps stars, issues and watchers and redirects the old address; folding repository B into a folder of A keeps nothing of B's — stars and watchers do not merge, old links die unless B stays as an archived pointer. issues move between repositories of one organization; pull requests do not. an archived repository keeps its stars and its page.

the terms of cyberia, applied

from dev § the objects: a product is a goal kept alive, one per rung, its page the door of its repository; a project is a task tree that closes; a part (a chip) is a part of one product, never a dependency, and no component depends on a rung above its own; a warrior is one per target family. github knows one durable unit, the repository. so the mapping that follows from the ontology, not from taste: a repository is the door of something that stays — a product, a warrior, a network, the organization; parts are folders of their product; projects are pages and branches, never repositories.

against that rule the inventory reads:

costs of folding, measured where possible

• history: git subtree add or git filter-repo keeps blame; a day of agent work per rung with the build green at the end.
• links: the old repository stays, archived, with a one-line README pointing at the folder; nothing breaks; crates.io repository fields update on the next publish.
• ci: a workspace builds everything unless workflows filter by path; the Mac runner is the bottleneck, so path filters are a precondition, not a nicety.
• size: cyb 797 MB and lytics 547 MB carry media that belongs in particles anyway; radio 296 MB (the iroh fork) is the heaviest honest part.
• queues: 113 pull requests over 38 repositories become ~10 queues; one is easier to read than thirty-eight, and it is also longer.
• permissions: one owner today; folder-level CODEOWNERS covers the rest.