v2.6.0

Who is acting is now resolved one way everywhere: connected AI agents get identities of their own and act as themselves, login-free mode genuinely works, and lific init asks how you want to sign in instead of leaving that decision to a hand-edited config file. Alongside the identity work: two security fixes worth upgrading for on their own, comments you can edit and link to, and identifiers you can copy from wherever you see them.

Sign-in, identity, and agents (PR #23 by @zorro432)

• Login-free mode now works end to end. Running with required = false used to be half-broken: project reads passed while admin endpoints answered 403, because the "no auth here" signal never reached the handler-level gates. Identity is now resolved in one place for REST, MCP, and CLI alike, every request carries a real user, and a login-free instance can administer itself.
• lific init asks how you want to sign in. On a fresh install it offers Login-free or Passwords, states plainly what login-free means (anyone who can reach the server can administer it), binds login-free instances to loopback so that promise actually holds, and persists the choice to the config file. --auth-mode and --password cover scripted setups.
• Connected AI agents act as themselves. Approving a tool over OAuth mints a per-tool bot identity, stdio agents connected via lific connect --stdio carry their own token, and the audit log attributes their writes to the tool rather than to you. Disconnecting or deleting a bot revokes its OAuth tokens too, and reconnecting remembers which tool a client picked last time.
• lific connect grew a transport menu, with stdio preselected and remote and OAuth on offer. The flags (--stdio, --oauth, --url) behave exactly as before for scripted runs.
• MCP enforces the same authorization as REST. An agent can no longer do more through MCP than the same account could through the web UI.
• Hardening around the edges. Failed credentials now fail instead of quietly falling back to the first admin's identity; a server with web_auto_login enabled refuses to start on a non-loopback bind; connector configs that embed credentials are written with 0600 permissions; a new API key is bound to its user in the same insert that creates it; OAuth device-code redemption is atomic with the token mint.

Security

• An uploaded SVG can no longer run script on your instance. SVG is on the upload allowlist, and every image type was served with Content-Disposition: inline. Because an SVG is an XML document that may contain <script>, opening one directly (the "open image in new tab" path, not the rendered <img> in a comment) executed its script on Lific's own origin, where it could read the session token the web UI holds and act as whoever opened the file. Any account permitted to upload could use this to escalate against an admin who viewed the attachment. SVG uploads still work and still render in comments and pages; they now download rather than render when opened as a page of their own, and every attachment response carries Content-Security-Policy: default-src 'none'; sandbox as a second layer. Reported by @mjc.

• A malformed config file now stops the server instead of silently reverting to defaults. A config that failed to parse, including one broken by a single typo, produced a warning on stderr and then booted from the built-in defaults. Those defaults bind 0.0.0.0, allow self-service signup, and treat an empty cors_origins as "any origin", so an operator who had set host = "127.0.0.1" and allow_signup = false could be left running a materially more exposed instance with nothing but a log line to say so. A config file that exists but cannot be read or parsed is now a hard startup failure naming the file and the error. Unknown keys are rejected for the same reason: allow_signupp = false used to be ignored silently and now fails loudly. A missing config file is still perfectly fine and still starts on defaults. Reported by @mjc.

API

• "Authentication required" is now always a 403. Nineteen places in the REST API each wrote their own version of the "is anyone signed in?" check, and roughly half of them reported the failure as 400 Bad Request while the rest reported 403 Forbidden. The endpoints for your profile, password, sessions, API keys, connected tools, and comments were in the 400 group; they now answer 403 like everything else, with the message authentication required. If you have a client that treats a 400 from those endpoints as "you are signed out", it needs to look for 403 instead.

Web UI

• Comments have their own links, and #N becomes one. A comment can be linked directly and the browser scrolls to it, and a bare #12 in issue text resolves to that issue. (PR #24 by @unger1984)
• Edit and delete your own comments. Editing is inline with the full composer (markdown, mentions, quoting, attachments), deleting asks for confirmation in place, and an edited comment carries an edited label with the exact time in its tooltip. (PR #25 by @unger1984)
• Home shows a live activity rate. It seeds from the last 24 hours of activity you are allowed to see, then ticks along over the websocket instead of starting from a misleading zero. (PR #26 by @mjc)
• Copy an identifier from wherever you are looking at it: the detail-page breadcrumbs, a board card, or a list row. (PR #22 by @lardissone) The copy and peek buttons are reachable by keyboard, not hover alone, and pills name the identifier they copy so a screen reader announces something useful.
• Copy selected text from the selection toolbar, instead of reaching for the mouse to select and copy by hand.
• The desktop sidebar collapses, giving the board and the issue list the full width.
• The status chip in the issue topbar is a picker. Changing status no longer means opening the field below it.

Fixes

• Starter label presets no longer vanish after you add the first one.
• Editing a label no longer reloads the entire settings page.
• #N inside a numeric HTML entity is left alone rather than linkified into nonsense.
• Deeply nested plan steps stay readable instead of squeezing themselves into an ever-narrower column.
• Attachment links stay in sync on every MCP write path, not just some of them.
• A restore whose rollback fails now names the path the old database survives at, instead of leaving you to hunt for it.
• lific doctor checks the router production actually serves.
• Test databases are named from a counter rather than the clock, so a fast machine no longer collides two of them in the same millisecond.

Upgrading

• Two migrations run automatically on first start; they add the per-tool identity plumbing behind connected agents. No manual steps.
• Check that your lific.toml parses before rolling this out, because a config Lific previously ignored will now stop it from starting. lific doctor reports the config as a normal check and keeps running the rest, which makes it the right tool for this. The strictness is the fix, not a side effect: a config that fails to load is exactly the case that used to degrade quietly.
• If web_auto_login is enabled and [server] host is not loopback, the server now refuses to start rather than exposing a no-password login to the network. Bind 127.0.0.1 or turn auto-login off.
• If your config sets secure_cookies under [auth], remove it. It has never been read from the file (it is derived from whether server.public_url is https://), and unknown keys are now an error rather than being ignored.
• Audit entries written by connected tools are now attributed to the tool's own identity rather than to the operator who connected it.