Roadmap
What works, what is in the open, what is not planned
This page lists what has shipped and what is tracked in public. It carries no dates and no promises: a feature is on it when the code is released.
Works today
In the current release
Each capability is stated with its limit.
Current release: v1.6.0. The full record is in the changelog.
Runs in a sandbox
Each task gets a disposable git worktree. Sandbox-mode runs are confined by Seatbelt on macOS, bubblewrap with Landlock on Linux, or podman if you opt in, and are refused when no backend applies.
Limit. The built-in worker can also run in workspace mode, which has no OS wall.
Re-verification
skep re-runs the project’s pinned verify command on a clean checkout with the patch applied, and records one of four results.
Limit. A project with no pin falls back to the worker’s own command, and the record says so.
Landing by approval
A patch waits for your approval and lands as a commit on skep/<task_id> or a branch you name, never the default branch. A pull request is opened only when you ask.
Limit. Auto-landing is opt-in: it takes only a re-verified patch and writes only a skep/* branch.
One policy engine
Layered allow, ask and deny rules with skep policy why, explain, rsop and decisions; a locked baseline built from three templates; skep onboard for the first run.
Limit. The decision log is a local file. It is rotated, never deleted, and not tamper-evident.
Five engines
Claude Code, Codex, Aider and pi as external agents, plus skep’s own worker on the model provider you configure. Codex can use a ChatGPT plan login, per project.
Limit. External agents need a pinned verify command; Claude Code and Codex need their auth environment variable.
Roles and delegation
A role is a prompt, a voice and a policy layer that only narrows. skep team runs a manager and up to three tasks under one record.
Limit. Still one operator per home: a team is one person’s agents.
Agents you run elsewhere
A Claude Code PreToolUse hook and an ACP governor answer an outside agent’s permission questions from your policy and record them.
Limit. Neither is a sandbox. The hook covers four tool kinds; an ACP agent in a mode that never asks is only observed.
The chat assistant
The same conversation in the web UI, the terminal, Telegram, Slack and Discord. Changes are proposed as cards you confirm.
Limit. A messenger can confirm four low-risk action classes. Shell commands, policy changes and landings stay in the web UI.
Schedules and tools
Recurring runs through skep schedule and skep tick, MCP servers behind the same policy, skill shelves, memory and notes.
Limit. Scheduled runs use the built-in worker, not a project’s pinned external agent.
Releases
A version tag runs the release gate, publishes to PyPI and cuts the GitHub release. CI runs on Linux and macOS.
Limit. The macOS leg runs lint, types, unit tests and smoke; coverage and the sandbox smoke are Linux-only.
Shipped
Release by release
Generated from the version history.
- v1.6.0 · your phone runs skep: one conversation, HIGH-risk approvals, /update 2026-11-20
- v1.5.0 · the daily driver: alarms, a policy tree, workers that land 2026-11-19
- v1.4.2 · a thread-first web UI, the Queen on a ChatGPT plan 2026-11-18
- v1.4.1 · the release gate runs a real sandbox 2026-11-16 , also carrying: the edge sweep; first-run: roles and delegation; first-run: onboard onto a locked baseline
- v1.2.0 · the ACP governor 2026-11-09
- v1.1.0 · the policy engine 2026-10-22
- v1.0.2 · the v106 audit round 2026-08-03
- v1.0.1 · the public launch 2026-07-30
Earlier releases are in the version history.
In the open
Being worked on in public
Planned work is whatever is in the issue tracker and the open pull requests. Nothing else is promised here.
Checked on 2026-11-18: the issue tracker has no open issues, and there is one open pull request, which its author has marked as parked (a role that pins its own model provider). Nothing else is scheduled in public.
If something you need is missing, open an issue. A report of what broke on your machine, or of a claim on this site that the code does not keep, is the most useful kind.
Not planned
Deliberately not planned
These are not in skep and not on the list above. Each would change what skep is, so each waits for evidence that people who use it need it.