Skip to content

Boundary with tamias and peregrin ​

tarsi owns release semantics — changelog → bump → tag → publish. Its neighbors own the layers around it:

neighborownstarsi's relation
tamiascontent distribution — fetch, store, cache, dispense; type: content driversrelease profiles are type: tarsi packs; tamias-driver-tarsi validates/installs them; the tarsi-serve binary itself ships as a type: files pack (libraries/tarsi)
peregrinorchestration — campaigns, multi-machine workflowslibraries/release bundle drives tarsi-serve release --yes; campaigns call the same ndjson ops humans and agents use
lasiusgit & workspace expertise — lasius-vcs portall history/tag/push goes through lasius-vcs + lasius-vcs-git; tarsi-core never touches git
sepiasession lifecycleoptional changelog polish — the caller may prompt a sepia session with rendered notes; no in-repo LLM path
cuorasecrets brokertoken_env names resolve through env/cuora; values never enter config, manifests, or logs
scarabconfig/template expertisetarsi.toml discovery/layering is scarab-config's FileSource/LayeredFileSource; tarsi-config is only the wiring
hyphadistributed substratehypha-serve-cli is the shared one-shot arg→params shim every *-serve bin uses

What tarsi is not ​

  • Not a CLI. No tarsi binary, no verb tree — the surface is ops. (The legacy TypeScript changelog CLI under the README's bottom half is pre-rewrite reference only.)
  • Not an orchestrator. tarsi-serve composes ports and answers ops; scheduling, approval flows, and UI live in peregrin/pavo.
  • Not a publisher of record for the org. Org repos release tag-only; crates.io publishing is tamias-only and GitHub Releases are unused (ADR-0004). The tarsi-publish-* crates exist for external consumers who opt back in per-repo.
  • Not a registry. It reads git history and writes tags/manifests; it does not host anything.

The dogfooding loop ​

tarsi's own tarsi.toml configures its release — units resolve from the cargo workspace, tarsi-xtask is skipped, both publish targets are off. Releasing tarsi itself is plan F of PL-014: a human or campaign decision, never an agent's.

MIT licensed — part of the 9aia GAIA stack.