Roadmap
devtwin is an early local tool, not yet a packaged release. The order below favors daily use: protect private state, prove it on real projects, connect sessions, then bring over an existing setup.
Available today
Section titled “Available today”- Scoped project onboarding with preview and fingerprint approval, read-only toward the repository.
- Immutable context records: decisions, facts, constraints, unknowns and more, with supersession and history.
- Bounded context briefs with freshness labels.
- Agent proposals that you review and publish.
- Private, selected-record handoff packets.
- Bundled skills for Claude Code and Codex, with journaled install, uninstall, recovery and rollback check.
- Live mode: a private store under
~/.devtwin, a pinned kit clone and your real skill folders. - Personal profile,
devtwin-setupskill and skill templates. - Manual stack watch with approved hosts and dated digests.
- Private drafts.
- Skill builder and third-party skill sources, installing into disposable pilot homes.
- A read-only
doctor. - Lessons: agents propose, you accept, and the agent hook delivers matching lessons while the agent works.
- The agent hook for Claude Code and Codex, which also blocks owner-only commands from agent shells.
- A guided
setupthat walks through each setup step. - An organization hub for sharing accepted lessons and “working on” notes with colleagues.
- Preferences captured from your corrections and delivered through a managed block in your own agent instruction files, and output templates for an RCA, a report and a decision note.
Lessons, the hook, setup and the hub are tested against local fixtures. They have not yet run against a real hub host.
| Priority | Work | Status |
|---|---|---|
| 1 | Backup and restore of private state, with recovery guidance and upgrade compatibility. | Planned. Deferred so live activation could be proven first. |
| 2 | Live activation on a real machine: fresh Claude Code and Codex sessions finding the skills, with rollback verified. | Implemented. First activation on a real machine passed on 2026-10-25, including fresh non-interactive Claude Code and Codex sessions. Interactive sessions are still to be checked. |
| After 2 | Identity routing: bind each scope to its own GitHub and Slack accounts; read and draft first, sending later. | Designed, sequenced after live activation. |
| 3 | Session start and end hooks: resume and handoff adapters with a reviewable worker lifecycle. | Partly done. The agent hook delivers lessons and blocks owner-only commands; resume and handoff adapters are planned. |
| 4 | Selective migration from an existing personal setup: dry-run mapping, approved records only. | Planned. |
| 5 | More connectors beyond GitHub and Slack, with account and workspace verification. | Planned. |
| 6 | Company knowledge hub adapter with reviewed, sanitized promotion. | Implemented as the organization hub. |
| Next | Preferences and templates: preferences captured from corrections you confirm, a managed block in ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md, and output templates. | Implemented; see Preferences and templates. |
| Next | Release radar: a weekly digest of vendor releases for your stack, with schedules and notifications. | Planned. |
| Next | Setup review: a monthly reminder to review your setup, and real-agent acceptance checks. | Planned. |
| 7 | Skill export: explicit file selection with privacy and rights review. | Planned. |
| 8 | Weekly skill updater and scheduler, disabled until you enable it. | Planned. |
| 9 | Open-source release: versioned releases and a reviewed release process. | Planned. |
Priorities 5 to 8 don’t block offline daily use. Nothing on this list connects an account, publishes, installs or migrates anything on its own.
Frozen until the daily-use work ships
Section titled “Frozen until the daily-use work ships”Skill-builder and skill-source features are frozen until priorities 1 to 4 ship. Bug and safety fixes continue. The create-from-template path (skill template-list, template-seed, template-verify and the templates) is the one exception.
Known gaps being tracked
Section titled “Known gaps being tracked”- CLI records are attributed to
humanwith a fixed run ID, whoever runs the command. record publish --applydoes not check who the controller is.- The default onboarding request ID comes from your flags, not file contents; pass a new
--request-idafter the repository changes. - Bundled skills are checked for command routes, not every flag.
See Limitations for the full list, and the repository for current work.