Architecture
Tarsi is the org's release plane: one pure domain core, a config loader, an ops wire, a publish port — and nothing else. It answers questions about releases; it never decides when to release or who approves.
caller (human, agent, peregrin campaign, scripts/release.sh)
│ ndjson {"op": …, "params": …} or one-shot argv
▼
tarsi-serve ────────── hypha-serve-cli (shared arg→params shim)
│ │
│ plans · bumps · │ tarsi.toml / type: tarsi profiles
│ changelogs · tags │ (scarab-config FileSource/Layered)
▼ ▼
tarsi-core (pure) tarsi-manifest (Cargo.toml scan/patch)
│
▼
lasius-vcs ── git history, tags, push (tarsi-core never touches git)
│
▼
tarsi-publish port ── crates / ghreleases / dryrun implsThe pieces
tarsi-core— pure release domain: plans, units, semver bumps, commit classification, changelog rendering. No I/O.tarsi-manifest— Cargo manifest scanning: workspace members →ReleaseUnitSpec, version reads/writes viatoml_edit.tarsi-config—tarsi.toml/.ymlloading (scarab-config) plustype: tarsiprofile-pack resolution.tarsi-serve— the ops surface:manifest/profiles/plan/changelog/notes/check/schema/bump/tag/publish/release. NotarsiCLI exists.tarsi-publish+-crates/-ghreleases/-dryrun— the publish port and per-forge impls.tarsi-driver—tamias-driver-tarsi: thetype: tarsipack driver, plus a read-only pavo view.tarsi-testkit— conformance fixtures + golden corpus.
The ops wire
Two modes over the same ops: no args = ndjson JSON-lines on stdio; <op> [--dir D] [--set k=v] [--yes|--dry-run] = one-shot argv via the shared hypha-serve-cli shim. release-family ops are dry-run by default — mutation requires dry_run: false (--yes). Humans, agents, and peregrin campaigns speak identical wire; there is deliberately no tarsi-llm.
Profiles are packs
Org release conventions ship as type: tarsi tamias packs (libraries/9aia-release) — layered config, versioned by the tamias store. tamias-driver-tarsi validates and installs them; tarsi-config resolves the layers through scarab-config's FileSource/LayeredFileSource.
Where it sits
In the org's layering — tamias (package & content manager) → engines (myco records, kuda code maps, mimus proofs, sepia sessions, sturn coordination, hypha substrate, cicada schedules, cypera docs, scarab config, peregrin orchestration) → surfaces (pavo views, MCP fronts).
Tarsi sits at the release plane beside the engines: every repo's scripts/release.sh is a thin shim over tarsi-serve, org conventions ride tamias profile packs, and peregrin's libraries/release campaign drives the same ops — tarsi computes plans and mutations; orchestration decides when to run them.
Deeper reads
- Release semantics — the load-bearing rules (dry-run default, per-unit tags, ordering, rollback).
- Boundaries — owns / doesn't-own, neighbors named.
- Concepts — units, strategies, profiles, the wire.
- Decisions — ADR-lite.