Changelog

Fixes

• Fixed an issue where some database connections could fail or behave unreliably due to a misconfigured security layer between Boss and its connection pooler.

Fixes

• Fixed an issue where database connections used during startup could interfere with normal connections, which could cause intermittent failures when Boss was initializing or updating.

Boss now surfaces pending approval prompts

When Claude needs your input during a session, Boss detects the approval prompt and marks the session with a question indicator. You can see at a glance that a session is waiting on you, without having to open it and scroll through the output to figure out why nothing is happening.

Improvements

• Smarter verification during build, review, and repair. Boss now runs a tighter set of checks at each stage, so you spend less time waiting on tests that do not apply to the work at hand.
• Stronger acceptance criteria. Boss is better at catching acceptance criteria or evidence requirements that can never be satisfied or that do not actually confirm anything useful. These are flagged earlier rather than silently passing.
• Visiting www.bossanova.dev now redirects to the main site correctly.

Fixes

• Fixed an issue where a session doing background work after a turn ended could appear idle or stuck instead of showing that it was still active.

Epic orchestration is here

Boss can now plan and run multi-part work as a coordinated epic: a top-level goal broken into child tasks, each with its own budget, scheduled and resumed automatically. When child tasks finish, Boss evaluates whether the overall goal is met before closing out the epic. This means longer, more complex jobs no longer need manual hand-holding between steps.

Improvements

• Per-repo test selection. You can now declare which tests Boss should run for a given repo in a .boss-skills.json file. This gives you control over test scope without changing anything in your CI configuration.
• Area detection is more accurate. Boss parses the “Key changes” section of a plan to understand which parts of the codebase are affected. Some tokens were being misread as file paths, causing Boss to either look for files that don’t exist or ignore areas it should have caught. This is fixed.

Fixes

• Fixed an issue where the sign-up prompt in the terminal could disappear unexpectedly when the footer was refreshed.
• Fixed a website issue where missing pages returned the homepage instead of a proper 404, and the sitemap is now available at /sitemap.xml.

Smarter repair rounds and more reliable agent work

Boss’s internal repair and review cycles got a meaningful overhaul this release. When a sub-task finishes, Boss now checks whether any results are worth salvaging before discarding them, so recoverable work is no longer thrown away on failure. Boss also re-verifies claims made by sub-agents before treating them as settled findings, rather than accepting them at face value.

The “finalize” step of a session is also more informative: the CLI now reports the actual post-merge state, surfaces the session’s effective model (and flags when a default was overridden), and suggests the nearest valid flag when you mistype one.

Improvements

• Boss now checks whether an artifact from a failed run still exists before deciding to discard it, keeping usable work rather than starting over unnecessarily.
• Sub-agent findings are now re-verified against the actual repository before being accepted as evidence, so a stale or incorrect claim is less likely to influence the final result.
• The boss CLI now suggests the closest matching flag when an unrecognized flag is passed, and gives a clearer description of why a flag was rejected.
• The boss CLI now shows the session’s active model, and notes when the model in use differs from the account default.
• boss restart --json is now available, giving scripts a stable, machine-readable outcome from a daemon restart.
• The finalize command now carries its own rejection reason out to the caller instead of silently dropping it.
• Boss can now report the resolved chat ID in the response envelope for chat wait operations.
• Repair rounds now handle skipped derivations more explicitly, routing them to an appropriate next step rather than leaving them unresolved.

Fixes

• Fixed an issue where restarting Boss on macOS could be throttled by the system, causing slow restarts or connection errors. Boss now correctly identifies itself to the system process manager so this no longer happens.
• Fixed an issue where a failed sub-task could discard side effects and recoverable work that should have been preserved.
• Fixed an issue where several helper functions returned a misleading success result instead of an error when called incorrectly, making problems harder to diagnose.
• Fixed an issue where proof capture could silently skip steps due to missing recipes or budgets that did not fit the available constraints.
• Fixed an issue where the merge status reported after a session completed could reflect a stale pre-merge state rather than the actual outcome.
• Fixed an issue where a plan could accumulate mutable fields (estimate, priority) that should not have been written into the stored plan body.
• Fixed an issue where a concurrent finalize run could delete a newer plan that a parallel process had already written.
• Fixed an issue where certain internal timing checks could produce unreliable results under load, causing intermittent test failures that masked real problems.
• Fixed an issue where a file-path claim from a sub-agent could be accepted as verified without being checked against the actual repository layout.

Session reordering

You can now manually reorder sessions on the home screen and sessions list. Press Alt+Up or Alt+Down to move the selected session, and Boss keeps that order in place across refreshes. The order is also respected when you view sessions through the web.

Features

• Daemon display names are now more useful. Boss previously reported the raw DHCP hostname as a daemon’s display name, which on macOS often produces something meaningless. Boss now derives a recognizable name from the machine. You can also set a custom name with boss settings --daemon-name without going through interactive menus, and boss daemon doctor now shows you the name your daemon presents in the cloud.

• Daemon filters work correctly when multiple daemons share a name. If two daemons in your fleet happen to have the same label, the sessions and repositories filters now qualify them so you can tell which is which. Previously they looked identical and were impossible to distinguish.

• The daemon picker in org fleets is no longer missing entries. When you belong to an organization, daemons that already had visible sessions were being left out of the daemon filter list. All daemons now appear.

• Colleague sessions no longer show “Daemon offline”. When an org member shares a session, Boss routes it correctly. Previously, a transport-layer gap meant those sessions always appeared offline.

• Boss plan check commands are validated more carefully. If a check step named a make target or test file that does not exist, Boss would accept it and then fail to run it later. Boss now resolves the full command at setup time and catches the problem earlier.

Improvements

• Setup script progress bars now render cleanly in place instead of producing garbled output.
• iTerm tab integration is now correctly limited to macOS, avoiding unexpected behavior on other platforms.

Fixes

• Fixed an issue where completed boss-plan runs could be marked as failed due to minor cosmetic state drift unrelated to the actual result.
• Fixed an issue where manually set session order was ignored when viewing or moving sessions through the cloud.

Fixes

• Fixed an issue where stopping Boss while it was shutting down could signal the wrong process. Boss now confirms a process is actually running before treating it as the active one, so stop and restart commands reach whatever is genuinely serving.
• Fixed an issue where Boss could hang longer than necessary during shutdown. If active requests stopped draining, Boss now exits the wait early rather than sitting out the full timeout.
• Fixed an issue where plugin processes could keep running after Boss itself crashed. Plugins now watch for their parent process and exit on their own if Boss goes away unexpectedly.
• Fixed an issue where the diagnostics tool incorrectly blamed console ownership for a problem it had not actually confirmed. The doctor now only reports causes it has established, and names the correct steps to take.
• Fixed an issue where plans containing markdown tables were always flagged as drifted during write-back verification, even when nothing had changed. Boss now correctly recognizes table formatting that the upstream service normalizes on its own.

More reliable PR check reading, smarter repair rounds, and better budget handling

Boss reads the status of PR checks to decide whether work is done. In this release, it correctly handles checks that are pending, skipped, or not yet reported, rather than misreading them as failures or false successes. This means Boss is less likely to spin on a PR that is actually fine, or to move on from one that is not.

Repair rounds now escalate more intelligently. When a problem keeps coming back across multiple fix attempts, Boss works through a defined ladder of strategies rather than repeating the same approach. It also checks each piece of evidence behind a review finding individually before deciding to apply, decline, or publish a fix, so it is less likely to act on a premise that does not hold.

Features

• Boss now tracks how often each internal quality check fires across a run, giving a clearer picture of where work is being spent.
• When a repair budget runs short, Boss now explains why a fix round was not funded rather than silently dropping into a read-only review state.

Improvements

• Skills (the scripts Boss runs to do its work) are now written in a shell-portable style, so they behave the same whether the underlying shell is bash, zsh, or fish. Previously, some constructs worked in one shell but silently misbehaved in others.
• When Boss writes a description back to your issue tracker, it now reads the content from a file that has already been validated, rather than passing it through an argument. This avoids limits on argument length and keeps the content consistent with what was checked.
• Boss now reports whether it is running under the system-level watchdog or a user-level one, and surfaces that in the status and doctor output. If the watchdog is not loaded, it points you to the install command rather than suggesting a restart.

Fixes

• Fixed an issue where PR checks in a pending or skipped state were misread as failures, causing Boss to retry work that did not need retrying.
• Fixed an issue where a repair run with a very small budget could silently choose a terminal state and stop attempting fixes without explanation.
• Fixed an issue where the doctor command could report a problem on a healthy machine that uses unattended (system-level) supervision and has no user-level agent installed.
• Fixed an issue where a failing spawn was not correctly attributed to the watchdog, making the cause harder to identify.

Unattended background operation on macOS

Boss can now run as a true background service on multi-user macOS machines, staying alive and picking up work without anyone logged in to the desktop. Previously, Boss relied on a supervision method tied to the GUI session, which meant it could not reliably restart itself unattended on a shared or headless Mac. The new unattended mode uses a lower-level mechanism that survives user switches and logouts.

You can see which supervision mode Boss is using by running boss status or boss doctor. If something goes wrong with the background service, Boss now reports its actual state rather than guessing.

Features

• Boss now has a .boss-skills.json configuration reference on the docs site, including a minimum viable config and a boss init step in the quick-start guide. If you have been unsure how to configure Boss for your repo, this is the place to start.
• When Boss reads a plan before starting work, it now re-checks that any facts, file references, and coordinates in the plan still match the actual code. Stale assumptions are flagged before work begins rather than discovered mid-build.
• Boss can now recover review comments and resolve threads even when it has exhausted its GitHub API quota, by falling back to an alternative read path. Before this, quota exhaustion could leave review threads stuck and unresolved.

Improvements

• Gate runs now write a clear, durable verdict so Boss always knows whether a check passed, failed, or hit an environment problem. Previously, ambiguous exit conditions could leave a run’s outcome unknown.
• Temporary files created during planning now have predictable names and are cleaned up safely even when multiple runs overlap.
• When Boss commits work to a pull request, it handles a broader set of commit shapes correctly, including bootstrap commits that are intentionally empty. Previously, some legitimate commits were rejected or misreported.
• When a task description is synced from your issue tracker, Boss now verifies the result matches what was written, accounting for formatting differences the tracker may introduce. This prevents a mismatch from blocking work unnecessarily.
• Fixed an issue where saving configuration could silently ignore a failure to sync a local directory, making it look like the save succeeded when it had not.
• Fixed an issue where creating an organization and then switching context could lose the newly created organization if the switch step failed.
• Fixed an issue where Boss plan gates could reject valid plan shapes, file citations, and check commands, causing legitimate work to be blocked.
• Fixed an issue where certain background operations could be routed to the wrong account context in a multi-organization setup.

Daemon restart and supervision fixes

Running boss daemon restart could previously kill a running daemon and then fail to bring a new one back up, leaving you with no service at all. That is fixed: Boss now detects which mode the daemon is actually running in before restarting, so it restarts in the right mode and leaves you with a working daemon afterward.

On macOS, boss daemon doctor could also miss cases where the daemon was registered with launchd but would never actually start, then silently fall back to a less reliable unsupervised mode without telling you. Doctor now checks the launch history and surfaces that situation clearly instead of hiding it.

Improvements

• Sessions in the dashboard now lay out more evenly: the repository column takes only as much space as it needs, and the session name column gets the rest.
• Boss can now notify you when an AI session (Codex) is waiting for a reply choice, not just when it is asking a direct question. If a session was paused waiting for you to pick an option, you would not have been told. Now you will be.
• Fixed an issue where sessions could disappear from the list if the daemon serving them reported a different location than expected. Sessions now track liveness against the daemon that owns them, so they stay visible.

Fixes

• Fixed an issue where restarting the daemon could destroy a healthy instance without successfully replacing it.
• Fixed an issue where boss daemon doctor would silently downgrade to unsupervised mode when it could not verify launchd supervision, instead of warning you.