hephaestus

A personal context management system — a single, self-hosted Rust application that fuses a wiki-style knowledge base (Zettelkasten) with task management. Notes and tasks are first-class, cross-linkable entities in one database, so a task like "Fix the roof leak" stays coupled to the home-repair log and contractor-call notes that give it context.

It is offline-first: fully useful on a laptop with no network, and (when the distributed layer lands) auto-syncing to a central instance hosted in blumeops. The primary surface will be a Neovim plugin (heph.nvim, an obsidian.nvim replacement); a CLI (heph) is the utility/scripting surface.

Why "what is next?" is the flagship. heph is organized around concise, honest answers to recurring questions — above all "what do I do right now?" — built from ~10 years of the owner's lived prioritization discipline: projects-as-contexts; attention colors (white/orange/red/blue, where red = a consequence exists if late, not importance); do-dates, not due-dates (earliest-actionable, never a deadline alarm); and working-set tensions surfaced honestly (no "fake happy", no overwhelm).

See docs/explanation/design.md for the vision and rationale, and docs/reference/v1-prototype-tech-spec.md for the implementation-facing specification.

Status

Phase 1 (v1 prototype) — at Todoist feature-parity (2026-06-03) on branch feature/v1-prototype. The Rust backend is feature-complete and all three surfaces are installed daily-drivers: the heph CLI (capture/scripting + the complete daemon API), heph-tui (the agenda/triage surface — attention-colored views, fzf move-to-project, undo/redo, an Inbox), and heph.nvim (the context/knowledge base — journals, wiki-links by id with conceal, a [[ picker, YAML-frontmatter task editing). A global ⌘' quick-capture popover (heph-quickadd) rounds out capture. Under the hood: all three runtime modes work, replicas sync through a hub over HTTP, authenticated end-to-end with OIDC (Authentik; JWKS/RS256, single-tenant), the offline-first config (local + hub_url) converges with a yrs text-CRDT, and the daemon runs as an OS service. CI is green, fully through Dagger; built test-first.

heph now self-hosts its roadmap. With parity reached, remaining and future work is tracked in heph itself — as tasks in the Hephaestus project (heph view ondeck, or the On Deck view in heph-tui) — rather than in a document. The v1-prototype tech-spec is now the historical build record (its §14 tracker); the design doc remains the living rationale.

Area	State
Data model, markdown extraction, wiki-links, export	✅ done
Tasks, links, "what is next?" ranking, recurrence, per-task logs	✅ done
hephd daemon — local mode (file lock + JSON-RPC over a unix socket)	✅ done
heph CLI; list / health / journal / full-text search (FTS5)	✅ done
Sync engine — HLC, op-log, converging merge + conflict queue (no network yet)	✅ done
yrs text-CRDT for body merge	✅ done
server (hub) mode + spoke push/pull sync over HTTP (axum)	✅ done
client mode + RemoteStore (online-only, no replica)	✅ done
OIDC hub auth — bearer-token verification + owner gate	✅ done
OIDC client — device-code login, keyring token cache	✅ done
heph.nvim (primary surface) — RPC client, buffer-backed editing, wiki-link follow, journal (slice 11a)	✅ done
heph.nvim — Tactical/Organizational task views, capture, attention, done/drop, log (slice 11b)	✅ done
heph.nvim — context-item promotion + Dagger headless-nvim CI (slice 11c)	✅ done
heph.nvim — managed daemon (plug-and-play, self-heal), follow-or-create, home/index, dailies picker, interactive views; installed + dev-isolated	✅ done

Architecture

A Cargo workspace, layered so the same core runs from a laptop to a hub:

• crates/heph-core — the library: data model, the Store trait + SQLite store, markdown parsing/extraction, recurrence, the "what is next?" engine, and the sync engine (op-log, hybrid logical clocks, CRDT/LWW merge, conflict detection). Synchronous and clock-injected (no ambient wall-clock reads) so ranking and merge are deterministic.
• crates/hephd — the per-device daemon. One binary, three modes — local (own SQLite replica; a syncing spoke when given --hub-url), server (also the sync hub: an HTTP endpoint others sync against), client (thin, remote, no replica — proxies to a --server-url) — selected by configuration via a targetable Store backend. Surfaces connect to it over a unix socket; it owns the DB handle and background sync.
• crates/heph — the CLI: a thin client of the daemon (no direct DB access).
• hephaestus.nvim (its own forge repo, eblume/hephaestus.nvim) — the Neovim editing/knowledge-base surface, a thin client of the daemon; see docs/reference/heph-nvim.md and docs/how-to/install-heph.md.

Storage: SQLite is the source of truth; a node's body is markdown; export materializes the whole store as a directory of .md files. Sync: each device holds a full replica + an append-only op-log; devices reconcile through a hub with automatic merge (text-CRDT bodies, last-writer-wins scalars, OR-set links) and a conflict queue for the ambiguous remainder. Auth: the hub verifies an OIDC bearer token (Authentik) on every op exchange — RS256/JWKS verification + a single-tenant owner gate — and clients obtain tokens via the OAuth 2.0 device-code flow (heph auth login), cached in the OS keyring. Local-only instances need no auth.

Build & run

Requires a Rust toolchain (stable). The build is a standard Cargo workspace:

cargo build              # build all crates
cargo test --all         # run the full test suite
cargo clippy --all-targets -- -D warnings

Run the daemon in local mode, then drive it with the CLI:

# Terminal 1 — start the daemon (creates ~/.local/share/heph/heph.db and a socket)
cargo run -p hephd

# Terminal 2 — talk to it
cargo run -p heph -- task "Fix the roof leak" --attention red
cargo run -p heph -- next                 # the "what is next?" ranking
cargo run -p heph -- doc "Roof log" --body "Called the contractor."
cargo run -p heph -- search roof          # full-text search
cargo run -p heph -- journal 2026-06-01   # open/create today's journal
cargo run -p heph -- export ./snapshot    # write the store as a .md tree

The daemon takes an exclusive lock on its DB file; --db and --socket override the defaults.

Development

• Test-driven. Every feature has tests at the appropriate layer(s) — unit, property tests, real-socket daemon integration, and CLI process tests. No feature is "done" without them.
• Change process. Changes are classified C0 / C1 / C2 (quick fix / human-review PR / Mikado chain). The v1 prototype is a single long-lived C1. See docs/how-to/agent-change-process.md.
• Git hooks & tooling. prek runs formatting, linting, and secret detection; mise runs repo automation; CI (Forgejo) runs prek plus cargo fmt/clippy/test via .forgejo/scripts/build.

prek install && prek install --hook-type commit-msg   # set up hooks
prek run --all-files                                  # run all checks
mise tasks                                            # list automation tasks
mise run ai-docs                                      # docs AI agents read first

• Documentation is Diataxis-structured and built with Quartz; changelog fragments use Towncrier under docs/changelog.d/. Working agents should read AGENTS.md first.

Repository layout

./Cargo.toml            # workspace manifest
./crates/heph-core/     # core library: model, store, extraction, recurrence, ranking, sync
./crates/hephd/         # daemon: local/server/client modes — unix-socket RPC + HTTP sync/rpc
./crates/heph/          # CLI: thin client of the daemon
./docs/                 # Diataxis docs (design, v1-prototype-tech-spec, how-to), Quartz config
./.forgejo/             # CI build + release workflows and hooks
./.dagger/              # Dagger module backing docs builds/releases
./mise-tasks/           # repo automation via `mise run`

License

All rights reserved. This is a personal, private project — not licensed for use, copying, modification, or distribution. Open-sourcing may be considered in the future. See LICENSE.