Changelog

Releases

Published releases, newest first, generated from the version history.

v1.6.0
latest

your phone runs skep: one conversation, HIGH-risk approvals, /update

Everything merged since 1.5.0 (#116–#119):

  • v129 (#116): a Telegram-bound chat is one conversation. Replies to turns started in the web UI reach Telegram (a chat that has read private files sends only a line saying the reply stayed in the web UI), the Queen's follow-on card after a button press arrives on the phone, and a card that blocks the chat is sent again when you write from the phone; a card the phone can confirm shows what it acts on.
  • docs (#117): a guide to running skep always-on on an Ubuntu 24.04 server (docs/guide-ubuntu-server.md).
  • self-update (#118): serve checks PyPI every 12 hours and announces a newer release once; /update on Telegram (which needs channel_can_confirm on) installs it with the tool that installed skep (pipx or uv tool), checks that skep --version reports the new version, and restarts.
  • v130 (#119): a person you name as a channel's admin can approve HIGH-risk cards from Telegram, Discord or Slack (a card folding more items than a messenger can show in full still needs the web UI); the chat shows live thinking/working status; the channel deck gains /stop, /btw, /approve, /deny, /compact, /usage and /model, with command menus generated from it; Telegram photos reach the model.
v1.5.0

the daily driver: alarms, a policy tree, workers that land

Everything merged since 1.4.2 (#102, #104–#114):

  • graphite-nav (#102): the sidebar's bottom links become one Settings gear, with every page inside it.
  • v121 (#104): when a scheduled prompt tries an off-menu tool call, an alarm goes to the chat's bound channel; Telegram link previews are off so they can't leak notes; script and chained schedule output is marked private and untrusted when a chat reads it.
  • v122 (#105): research, YouTube transcripts and a 16:30 mail summary (python -m skep.mcp_servers.mail --digest 5) work from your phone.
  • v123 (#106): a schedule can only narrow what its workers may do.
  • v124 (#107): Policies gains a browsable tree that explains the shipped rules, and "Test an action" can check a role or a role-bound schedule.
  • v126 (#108): unedited seed skills refresh, template grant changes are audited, and worker-written memory stays out of the Queen's automatic context.
  • v127 (#109, #112): workers that can do the work they are given; re-verify installs Node dependencies from a supported lockfile for a recognized verify command. --auto-approve is retired: it still parses and has no effect; a patch lands without a click only when the project turns on auto_apply_verified_patch and pins a real check.
  • docs-site (#110, #111): the website is rebuilt on Astro, every claim checked against the code; the hand-written site is gone.
  • v128 (#113, #114): fixes from a browser field test of the Queen: it calls tools in parallel on the ChatGPT plan again, /workon moves the chat onto its new project, and object results in commands and cards show labeled fields instead of a raw JSON dump.
v1.4.2

a thread-first web UI, the Queen on a ChatGPT plan

Everything merged since 1.4.1 (#93, #94, #101):

  • v117: a chat keeps the project you pick and starts in workspace; the Queen can create projects.
  • v118: the Queen runs on a ChatGPT plan, with the Codex login as a provider.
  • v119: the main key stays on its own endpoint.
  • v120-F1: zsh here-documents work inside the macOS sandbox (TMPPREFIX is set beside TMPDIR; /tmp stays closed).
  • H1, H1-G, Graphite A: the web UI is thread-first (sidebar, Inbox, context panel) in the Graphite look; Policies opens on what agents may do, and Settings and Setup are one place each.
v1.4.1

the release gate runs a real sandbox

Published 2026-11-16: the first release since 1.2.0, carrying the 1.2.1, 1.3.0 and 1.4.0 work above.

The v1.4.0 tag was pushed, but its release gate failed and nothing was published, so 1.4.1 is the published form of 1.4.0, with one fix to the gate. scripts/linux-sandbox-docker-smoke.sh runs bwrap as root inside a container. Root's bwrap creates namespaces directly, which needs CAP_SYS_ADMIN and CAP_NET_ADMIN; Docker drops both and its default seccomp and AppArmor profiles refuse the mounts. Before ES-F1 skep then ran the worker unconfined and the smoke passed without testing a sandbox; since ES-F1 it fails closed, which is what failed the 1.4.0 gate. The smoke now gives its one throwaway container those two capabilities with both profiles lifted.

Getting that far exposed a real bug: the Docker image had no ps, which the monitor and run re-adoption call, so a sandboxed run in the published image crashed. The image now installs procps.

This release also carries work that was versioned but never published:

not published v1.4.0 the edge sweep

Tagged 2026-11-15, but its release gate failed and nothing was published: this work shipped in v1.4.1 on 2026-11-16.

1.3.0 was bumped but never tagged, so this is the first published release since 1.2.0 and carries both: the roles-and-delegation work above, and the edge sweep, a live-model campaign against the policy engine, the sandbox and the landing path (#56–#90). What an operator sees:

  • One decision point. REST, CLI, hooks, ACP, the Queen's file and shell lanes, read_url, search_web, MCP and browser calls all decide through one policy_v3.decide(). A deny binds every spelling of a command, reaches through sh -c, and a policy write is checked before it lands.
  • Taint asks before acting. A chat that has read private data or taken in outside content asks before it sends or acts; /trust clears a chat's taint. High-risk cards need your desk, repeated asks fold into one card, and a dispatch card lists what the run will get.
  • A stricter sandbox. It fails closed; it hides rc files, login stores, unused LLM keys, agent sockets and ssh-agent; a git hook can't escape it, and on Linux Landlock confines the worker. An engine's own sandbox is switched off under skep's (codex/claude_code). An opt-in provider-key relay (provider_key_relay, default off, ADR 0062) keeps the provider key out of the worker, and codex can run on a ChatGPT plan login.
  • No unverified landing. A shell worker needs a pinned verify command, landing can't half-fail beside worktree cleanup, and approvals claim a card atomically before they act.
  • Honest output. Logs and cards mask secrets and channel tokens, skep doctor and the decision log report what actually happened, follow-ups survive a restart, and skep's own log lines reach serve.log.
  • Tests. The web UI's JavaScript runs under node --test, and CI fails a PR with changed lines no test runs (diff coverage ≥ 90%).

The four pre-1.2.1 template names are still accepted as aliases.

not published v1.3.0 first-run: roles and delegation

Never tagged, released or uploaded: this work first shipped in v1.4.1 on 2026-11-16.

The first-run O4, R1, R2 and R3 work above, as a minor release. What an operator sees:

  • A chat runs in a role (builder, operator, assistant, or a custom one): the picker on a new chat, /role, skep role use. The role is a policy layer that only tightens the approved baseline, and skep policy why --role names it.
  • The Queen dispatches under its role, and a worker it dispatches holds no more than it does: exceeds_delegator rows, why --delegated-by.
  • Workers take an explicit role (skep run --role, dispatch_run role=) and a team run — a manager in a role fanning out to up to three workers — is skep team / batch_dispatch manager= / POST /api/teams, one ledger thread. Run cards show the chain.
  • New chats show a one-line why on every verdict card and a turn tally until dismissed (skep onboard --quiet).

The four pre-1.2.1 template names are still accepted as aliases in 1.3.0 (ADR 0056 open question 3 proposed retiring them here; not done in R3).

not published v1.2.1 first-run: onboard onto a locked baseline

Never tagged, released or uploaded: this work first shipped in v1.4.1 on 2026-11-16.

The first-run O1, O2, O3 and O5 work above, as a patch release. What an operator sees:

  • skep onboard (or skep onboard --template builder --yes) takes a fresh install from nothing to a locked policy baseline in five resumable steps; skep onboard --check is skep doctor plus the onboard rings and the egress truth.
  • Three templates — builder, operator, assistant — replace the four old documents (aliases until 1.3.0); skep policy explain <name> prints each as an allow / ask / deny table.
  • skep policy approve --template NAME locks the baseline (<home>/policy/baseline.json + baseline.lock + the approvals ledger); an edit fails closed with baseline_tampered until skep policy approve --reapprove.
  • A fresh install asks for everything until a baseline is approved (no_baseline); an existing home keeps deciding as before and is told how to approve one.
  • skep acp install-hint names the baseline the governor answers from.

Air smoke (docs/acp-governor.md): pipx install skep && skep onboard --template builder --yes, then skep acp install-hint.

v1.2.0

the ACP governor

Published 2026-11-09; the release commit is dated 2026-11-06.

The acp-governor B0–B2 work above: skep acp govern|doctor|install-hint and skep policy decisions. Operator notes: a fresh home starts with shell and filesystem on engine v2; an existing home keeps its switch. Install with pipx install skep && skep acp doctor -- npx -y @agentclientprotocol/claude-agent-acp. ADR draft 0054 moves to docs/adr/ on acceptance.

v1.1.0

the policy engine

The first release since 1.0.2 (2026-08-03): v107–v116 and the policy- engine rewrite P0–P3 above. Operator note: the engine switch ships OFF — flip shell first, after the golden parity suite is green; install the Claude Code hook from docs/policy-hook.md. Linux/macOS CI gates stay the same, with job timeouts so an unreachable apt mirror fails in minutes.

v1.0.2

the v106 audit round

The published form of the v106 round above (its release notes count the fixes as F1–F13), plus parallel CI. Published to PyPI and as a GitHub release on 2026-08-03.

v1.0.1

the public launch

The first public release: the repository opens here (see the top of this file), published to PyPI and as a GitHub release on 2026-07-30. Everything above this line is the internal history that led to it.

Earlier releases and the round-by-round record are in the full version history.