⚡ Claude Terminal v1.3.7
Your tickets next to your sessions, a Git tab in every chat that follows its pull requests, and automations that know which account they spend.
The headline
🎫 Tickets, starting with LinearConnect your issue tracker with a personal API key and a new Tickets screen lists every ticket of your workspace, as a list or a board you can drag cards across. Start a Claude session straight from a ticket, on the tracker's branch if you want, and each chat gets a Tickets tab with the tickets it is working on (#238, #239). |
🔀 A Git tab in every chatBeside Changes, each session now has a Git tab: the pull requests it opened or worked on, grouped by what needs you (a red check, a conflict, a review to address), with their CI as a dot on the tab button. Fix it, Resolve, Update and Address review draft the right prompt in the composer (#237). |
🤖 Automations that know which account they spendEvery automation used to run on the machine-wide login, whatever project it targeted. Each one now picks its Claude account (the default, the project's binding, or a named account), can be skipped when that account is above a usage threshold, and the Cost tab gains an Automations view with runs, cost, cost per run and accounts used. |
🔐 Connect an account the way the CLI does"Connect an account" now runs the CLI's own |
🎫 Tickets
| What it does | |
|---|---|
| Connecting | Settings > Tickets takes a personal API key, checks it with the tracker before saving anything, and stores it in your OS credential store. The settings file keeps the workspace and the user, never the key |
| The Tickets screen | Every ticket of the workspace, with "mine" chips, search, Status / Assignee / Priority / Labels filters, grouping and sorting. A List / Board toggle, where dropping a card on a column changes its state |
| Sub-issues | Shown under their parent, Linear style, with their progress (1/3) on the list, the board and the detail
|
| Detail pane | Description, sub-issues and comments. State, assignee and priority change from there, and a refusal from the tracker puts the old value back |
| Start a session | Opens a chat on the project you pick with the ticket linked and its content in the composer, unsent. Optionally checks out the branch the tracker suggests and moves the ticket to "started" |
| In the chat | A Tickets tab per session, @tickets to put a ticket's current status, description and comments in your prompt, and a recap of the session drafted as a comment, published only when you click
|
| Detection | Tickets a session creates or names are suggested in the conversation, and linked only if you say yes. Terminal sessions get the same through a button in the session bar |
Linear is the first tracker. The layer underneath is provider-neutral, and every tracker has to pass the same contract tests.
🛡️ Security
- Agents can no longer approve another tab's permission prompts. The MCP tool that answers a tab's prompt can now only deny. Letting one agent grant another tab's permissions is back as an opt-in under Settings > Claude > Agent orchestration, with a warning while it is on. Reported and fixed by @bleleve
- Agents can no longer type into a terminal tab that is showing the Claude Code approval dialog, unless that same opt-in is on
- Always Allow now allows only the tool you clicked, instead of switching the whole conversation to bypass permissions. If you preferred the old behaviour, it is an opt-in setting under the permission modes, with a warning while it is on
Important
Both opt-ins are off after the update and cannot be turned on by an agent through the settings tools. If you run one agent that drives the others from start to finish, turn on Agent orchestration yourself.
Everything else
- Folders of repositories: a project that is a plain folder holding several repositories now works with them. The Git tab, the toolbar actions, the Changes panel and the dashboard list the repositories one or two folders down and act on the one you pick, remembered per project
- The account an automation used is shown on task cards and in run history
- The usage bar froze on the last figure while an account was at its limit, because the usage endpoint answers 429 exactly then. The figures now come from the chat stream as well (#234, #235)
- An account that hit its limit kept reading 100% after the reset (#233)
- macOS asked for the login keychain password again and again, after every update and every few minutes. Claude's Keychain items are now read through
/usr/bin/security, the only app they trust (#231) - macOS reported downloaded builds as damaged. The bundle is now signed ad hoc, so it opens through Privacy & Security > Open Anyway, with no terminal involved (#241)
- An account switch on macOS could show credentials to other local users while they were written. They no longer go through the command line
- The context gauge read over 300% on models that serve a 1M window under a plain id, such as Opus 5.5. It now takes the window Claude reports
- One Escape closed two modals when one was opened from another, which cancelled a sign-in started from the account switch offer
- Deleting an account an automation uses is refused before the confirmation, not after it
- Automation sessions were kept for project paths containing a dot or a space, and older agent steps ignored the xhigh effort level
- Ticket trackers are one adapter file each, checked by a contract suite run against a fake tracker kept deliberately unlike Linear, so the core cannot quietly assume Linear. The design is in
design/issue-trackers.md - Automation spend is recorded per step in an append-only log, because a step's transcript is deleted as soon as it ends and the Cost tab could never see it
🙏 Thanks
| @bleleve | Most of this release: Tickets and the Linear adapter, the chat Git tab, connecting accounts through claude auth login, the usage fixes, the macOS Keychain and signing fixes, and the private report and fix of the agent permission issue
|
📜 Full changelog, 24 commits
test(remote) wait for the background last-connection write instead of racing it
feat(automations) let each automation pick the Claude account it runs as
fix(workflows) delete automation sessions where the CLI writes them
feat(automations) add per-automation cost, usage gate and account visibility
fix(workflows) honour xhigh effort on legacy agent steps
fix(accounts) refuse an automation's account before confirming deletion
fix(accounts) read the CLI's Keychain items through /usr/bin/security (#231)
fix(accounts) keep Keychain writes off argv and harden the security pipe
fix(chat) read the context window the CLI reports instead of guessing it
fix(chat) scope Always Allow to the tool instead of bypassing the session
feat(chat) add a Git tab to each session: its pull requests by what they need, and their CI (#237)
fix(usage) stop an account that hit its limit reading 100% after the reset (#233)
feat(tickets) add a provider-neutral Tickets layer (#238)
feat(tickets) connect Linear with a personal API key (#239)
fix(usage) read the session figures off the chat stream when the endpoint answers 429 (#235)
fix(build) ad-hoc sign the macOS bundle (#241)
feat(accounts) connect an account through claude auth login (#240)
fix(modal) let only the top modal answer Escape and Tab
feat(accounts) identify saved accounts by their OAuth profile
fix(mcp) let chat_respond_permission deny, never grant
feat(settings) opt in to agents granting another tab's permissions
feat(permissions) opt in to Always Allow bypassing the conversation
feat(git) detect repositories in the sub-folders of a project folder
chore(release) bump version to 1.3.7
Windows Claude-Terminal-Setup-1.3.7.exe · macOS .dmg (Apple silicon & Intel) · Linux .AppImage