Untether 0.36.0 — live Claude sessions, steerable follow-ups and safer unattended runs

[github-actions[bot]]

TL;DR

Claude Code sessions now stay live after an answer. Background agents report back in the chat, and you can steer a run while it's working. Unattended crons and webhooks no longer wait on a button nobody can tap, and Codex safe mode is now properly read-only. This release has six breaking changes, so please read Heads up before you upgrade.

✨ What's new for you

• Live Claude sessions. When Claude keeps working after its reply (background agents, monitors, scheduled wake-ups), those turns now arrive as their own messages, with working approval buttons. Follow-ups go straight into the running session instead of waiting for it to exit (#776).
• Background task panel. The progress message lists background agents with elapsed time, tokens and current step. After the answer, a single status message keeps you up to date.
• Steer or queue follow-ups. Choose per chat or topic whether a message sent mid-run waits for the turn to end (queue) or folds into the running turn (steer) (#775). Use /steer or /queue for a one-off, or set the default on the new /config Follow-up page.
• Context use at a glance. Claude's status line now ends with how full the context window is, e.g. 62% ctx, and compaction shows as a 🗜️ Context compacted row instead of looking like a stall. API retries and safeguard stops get their own rows too.
• Stop at limit actually stops. With Stop at limit on, reaching your daily budget pauses new runs until midnight, with a one-off Run anyway button. Today's spend now survives restarts.
• Replies carry context. Reply to a message (or a selected quote) and the agent sees the text you replied to. Thanks to @Dvredin for this one.
• Per-cron model and effort. A cron can set its own model and reasoning. Claude Code's own auto permission mode is now available alongside Untether's plan-auto.
• Better exports. /export sends the full transcript as a .md or .json file, covering every turn of a live session.
• For shared chats. The opt-in approval_originator_only setting lets only the person who started a run answer its approvals and questions.
• A new home. Untether now has its own website at untether.cc. The help centre stays at littlebearapps.com/help/untether.

🛠️ Fixes you'll notice

• Approval buttons only work in the chat they were posted in, and a second tap says "Already answered".
• /cancel and /new in a forum topic only affect that topic, and /cancel stops a loop for good.
• The "waiting for you" reminder says what it's waiting for and disappears once you answer.
• Tidier Telegram formatting: tables, numbered lists and one-item-per-line replies keep their shape, and CLAUDE.md no longer becomes a link.
• Voice notes no longer hear "Clawde" or "Clawed Code".
• Codex, OpenCode and Pi keep your saved session when a resumed run fails for an unrelated reason, and Codex and OpenCode chats get proper /usage totals.
• OpenCode 2.x is refused with a clear message (Untether supports the 1.x CLI).
• Plus well over 100 smaller fixes to progress messages, background tasks, costs, triggers and file transfer.

🧹 Under the hood

• Security hardening: file-transfer deny globs match at any depth and through symlinks, /browse needs a project-bound chat, and logs mask credentials in URLs.
• Config is parsed once per edit rather than on every read, and hot-reload no longer misses quick edits.
• Dependency floors are raised past recent anyio and aiohttp advisories.
• Release testing now focuses on the four supported engines: Claude Code, Codex, OpenCode and Pi.

⚠️ Heads up

• Gemini CLI and AMP are deprecated. Both still ship and load, and no removal is scheduled. They're just no longer supported: Google ended Gemini CLI for individual accounts, and AMP refuses headless use. /config marks them with ⚠️. Antigravity CLI support is planned for 0.36.1.
• auto is now plan-auto. Untether's old auto (plan, then auto-approve) is renamed plan-auto, and auto now means Claude Code's own auto mode. Chat settings migrate automatically; if you set permission_mode = "auto" in untether.toml and want the old behaviour, change it to plan-auto.
• Prompting modes now prompt. default, manual and acceptEdits (what /planmode off sets) now ask before shell commands, web fetches and MCP tools.
• Unattended runs deny approvals. Cron and webhook runs no longer wait on buttons; approvals and questions are denied and the run carries on or reports back. Give every Claude cron that should act on its own an explicit permission_mode such as plan-auto, auto or dontAsk. The startup message lists affected crons.
• Codex safe mode is read-only. Edits, builds and tests fail in Safe. Switch chats or crons that need to change files to full auto.
• Bypass flags in extra_args are refused. Flags such as --dangerously-skip-permissions or --yolo stop that engine loading. Remove them; the security guide has the full list.
• With Loop mode off, Claude can't schedule its own recurring tasks. It's pointed to Loop mode or /at instead. [loop] own_schedule = false restores the old behaviour.

⬆️ Upgrade

Read Upgrading to v0.36.0 first; it also lists the config keys to remove if you ever roll back to 0.35.4. Then upgrade with the tool you installed with:

uv tool upgrade untether
# or
pipx upgrade untether

Restart Untether (systemctl --user restart untether, or send /restart from Telegram) and run untether doctor for a quick preflight check.

---

Links: Full changelog · GitHub release · PyPI · Website · Help centre

💬 Questions? Ask in Q&A. Got an idea? Share it in Ideas. Built something with Untether? Post it in Show and tell. Found a bug? Open an issue.