Skip to content

Roadmap

Dates assume the likely estimate with a part-time maintainer (15 h/week) supervising 4 agents. See ESTIMATION.md. They will be re-forecast after M1.

Milestone Goal Start Target Weeks Supervision h Agent h Human team (3 devs) weeks
M0 Foundations Repos, CI, ADRs, schemas, threat model, agent workflow 2026-11-02 2026-12-02 4.4 66.0 17.9 1.3
M1 Walking skeleton Pipeline engine, store, journal, dry-run, undo, plugin host v1, CLI 2026-12-03 2027-02-03 9.0 135.3 45.7 3.3
M2 LLM step Skills, backends, schema validation, cache, budgets, de-identification 2027-02-04 2027-04-12 9.8 146.7 44.0 2.6
M3 Dogfood First plugins, API, scheduler, image, Coolify template: a real daily note and reminders 2027-04-13 2027-08-19 18.4 276.4 95.3 6.6
M4 Marketplace Git marketplace, pinned installs, permissions approval and lockfile 2027-08-20 2027-10-10 7.4 110.6 31.4 1.9
M5 Hard isolation Plugin containers, egress allow-list, malicious-plugin suite, MCP server 2027-10-11 2028-01-15 13.9 208.1 34.8 2.5
M6 Existing MCP services brief() contract and plugin-mcp-brief 2028-01-16 2028-02-01 2.4 35.6 12.0 0.7
M7 Mail, school, home IMAP/Proton, Gmail, PRONOTE/ÉcoleDirecte, Home Assistant 2028-02-02 2028-03-25 7.6 113.9 34.2 2.6
M8 Public v0.1 Docs site, release automation, signed artifacts, v0.1 2028-03-26 2028-04-13 2.8 41.7 9.3 1.0

Critical path: M0 → M1 (engine, plugin host) → M2 (LLM step) → M3 (dogfood). M4-M8 build on a working dogfood instance. M6 and M7 plugins can start in parallel once M3 is done if supervisor time allows.

Burning points

The hot topics that can stall the project. Each one is tracked as a Risk issue in horae and visible in the project's "Burning points" view.

Risk Severity When Mitigation
Scope creep: product and personal tool at once — Building a public product and a personal assistant at the same time can stall both. high M0-M8 M3 dogfooding comes before marketplace and isolation; non-goals in PLAN §2.3; every feature must be used by the reference instance.
Supervisor bandwidth is the real bottleneck — Agents produce PRs faster than one part-time human can specify and review them (industry data: review time up 2-5x with agents). high M0-M8 Graft policy auto-approves trivial and small logic changes; specs written as acceptance criteria up front; at most 3-4 agents in parallel; recalibrate estimates after 15 stories.
Security-critical code written by agents — Permissions, sandboxing, signed links and secrets are where AI code fails most (studies: ~45% insecure choices) and mistakes are costly. high M1, M4, M5 These stories are human-led (agent drafts, human designs and reviews), sensitive paths in graft.yaml, malicious-plugin suite in CI.
Claude terms for the claude-cli backend — Anthropic allows personal headless use of claude -p but not third-party products offering claude.ai login. medium M2 Horae never handles OAuth; API backend is the documented default; re-check terms before v0.1.
LiveSync CLI encryption parity and sync conflicts — CLI encryption docs are thin; conflicts between server writes and phone edits are possible. medium M3 Plugin owns separate files/blocks only; parity test on real devices; nightly git snapshot; Obsidian Sync headless as fallback.
iOS behaviour of ntfy action buttons and deep links unverified — ntfy docs do not say whether http actions and obsidian:// links work on iOS. medium M3 Device test task in M3; fallback to view actions that open the note.
Unofficial school APIs break or lock accounts — PRONOTE/ÉcoleDirecte libraries are reverse-engineered; repeated logins can lock accounts. medium M7 Read-only, rate-limited, degraded status, separate plugin; core never depends on it.
Coolify cross-resource networking — Reaching other services by container name across Coolify resources may not be stable across redeploys. low M3, M6 Spike in M3; fallback to an internal route with a bearer token and IP allow-list.
MCP specification churn — MCP SDK and spec changes can break the plugin host. low M1, M5 Pin the SDK; keep the plugin host behind an internal interface.
Trademark: French marks in class 42 — Several French national marks contain HORAE, two in class 42. low M8 Open source, no commercial use or filing; professional clearance before any commercial step (PLAN §10.1).

How the work flows (AI-first)

  1. Ready: a story has acceptance criteria, an estimate, an executor (Agent, Agent + human check, Human-led, Human-only) and its dependencies done.
  2. Agent works: one agent per story, on its own branch, following the repository's AGENTS.md. Commits carry Graft trailers (Graft-Agent, Graft-Model, Graft-Task).
  3. Review: Graft writes an intent card, checks it against the diff and applies graft.yaml: trivial and small low-risk changes can be approved without a human; security paths, core contracts and large changes require the maintainer.
  4. Done: merged with CI green, acceptance criteria checked, docs updated, estimate actuals recorded on the issue.