Changelog

Chats interrupted by a restart now recover automatically

If Boss was restarted mid-conversation, the affected chat would silently stall, leaving you with no response and no clear way forward. Boss now detects those interrupted chats on startup and resumes them automatically, so a restart in the background is much less likely to leave work stranded.

Related to this, Boss no longer reports a conversation as still working when it quietly died partway through a background step. The status you see now reflects what actually happened.

Improvements

• When Boss shuts down, it now finishes handling any in-flight requests before closing out its internal connections. This reduces the chance of a dropped response at shutdown time.
• Detaching from a chat you have manually renamed no longer overwrites that name with an auto-generated title.
• Skill configuration that involves lenses or drafters is now discovered and connected correctly. Previously, certain role bindings could be silently dropped during setup, meaning those capabilities were never used.
• Boss is better at confirming its own tools are actually available before attempting to use them, and gives clearer information when they are not.
• When a readiness check times out, Boss now includes more detail about what went wrong, making it easier to understand why a session did not start as expected.
• The proof command now accepts --help and -h on every subcommand, and error messages name the accepted commands and flags rather than giving a generic failure.

Fixes

• Fixed an issue where a chat shown as “Working” could be stuck in that state after the underlying task had already failed.
• Fixed an issue where detaching from a chat could silently discard a name you had set manually.
• Fixed an issue where certain skill extensions were dropped during startup, meaning configured tools were never loaded or dispatched.
• Fixed an issue where Boss might attempt to use its own command-line tool without first confirming the tool was present and executable, leading to unclear errors.
• Fixed an issue where an interrupted background step left the chat status permanently showing as active.

Idle session cleanup

Boss now automatically cleans up terminal sessions that have been sitting idle, freeing up resources without any manual intervention. This happens in the background by default, and you can adjust or disable the behavior in your configuration if you prefer to manage sessions yourself.

Improvements

• When you attach to a session from the web terminal, Boss is better at resuming chats that had lost track of their terminal session, rather than leaving them stuck.
• Boss is now smarter about distinguishing a session that was cleaned up intentionally versus one where something went wrong, so it makes better decisions about which chats need attention.

Fixes

• Fixed an issue where adding a new account over SSH would swallow keystrokes, making it impossible to type your account label.
• Fixed an issue where pasting an image while connected over plain SSH would show a raw file path instead of the image. Boss now shows a clear message explaining that image paste requires a direct connection, so you know what to do rather than seeing a confusing path string.
• Fixed an issue where cleaning up an idle terminal session could incorrectly mark the associated chat as having exited, causing unnecessary errors or unexpected behavior.

Attended chats can now spin up sub-agents automatically

When you chat with Boss directly (an attended session), Boss can now bring in sub-agents to help with the work without you having to configure anything extra. Previously this required explicit setup; now it happens by default, with an opt-out available in settings if you prefer to keep sessions single-agent.

Improvements

• Node.js toolchains are now visible to the Boss service. If you installed Node via nodenv, nvm, or a similar tool, Boss’s background service previously could not see that installation and would fail silently or use the wrong version. You can now also add custom paths through a daemon_path_extra setting, and the Boss doctor view shows which paths the service is actually using.
• Usage limits are checked correctly across models. An account with headroom in its overall weekly budget could still get blocked because per-model limits were being applied too aggressively. Boss now checks the right budget window, so sessions are less likely to be stopped when there is capacity available.
• Gates that fail to start are now reported as failures. If a gate command crashed or timed out before returning a result, Boss would treat it the same as a gate that deliberately blocked work, making it hard to tell whether a job was skipped on purpose or because something went wrong. Failures now show up as a distinct failed row so you can see what actually happened.
• Verify-only checklist items are now checked properly. Acceptance criteria marked as verify-only were not being seen by the gate that reviews changes, so they could pass without ever being confirmed. Boss now correctly flags those items and requires them to be cleared before work ships.

Fixes

• Fixed an issue where the status message shown while Boss is waiting would include a redundant label prefix, making the line read awkwardly.

Paste support in mobile chat

You can now paste text into a chat session on mobile. A “Paste” button appears in the chat toolbar on touch devices; tapping it reads from your clipboard and drops the text into the conversation. If the clipboard content is too large or the chat is not ready to receive input, Boss holds the sheet open so you know the paste did not go through silently.

Improvements

• Adding an account while connected to a remote host now works correctly. Previously, Boss tried to resolve the setup locally instead of routing it through the remote connection, which caused the flow to fail. The account registration prompt also fits properly in the window now instead of being clipped.
• Boss now handles the case where Claude’s usage-check service is temporarily rate-limited. Instead of silently falling back to a full billable API call, it backs off and retries. Retry delays stated by the server are respected, with a 30-minute ceiling.
• Incomplete or placeholder plans are now caught before they reach the session tracker, so you are less likely to see malformed or stub content appear in your plan view.

Fixes

• Fixed an issue where a temporary authentication hiccup overnight could sign you out of Boss entirely. Boss now waits out short refresh stalls and recovers on its own.

Git operations are faster and more reliable

Boss does a lot of work behind the scenes with git: fetching your code, checking branches, opening draft pull requests. In this release, several improvements land together that make that work noticeably more stable.

SSH connection sharing. Boss now reuses a single SSH connection per host rather than opening a new one for every git operation. This cuts the per-operation overhead from repeated handshakes, so sequences of git actions (fetch, then push, then open a PR) complete faster.

Automatic retry on transient errors. Temporary network hiccups, brief server-side refusals, and similar one-off failures no longer stop a session in its tracks. Boss now detects these as recoverable, waits a moment, and tries again automatically.

Draft PR creation retries. If opening a draft pull request fails transiently (for example, a brief outage on the git host), Boss will try again on its own rather than leaving the session in a broken state.

Improvements

• When a draft PR is temporarily failing to create, Boss now shows a “retrying” status instead of a generic error, so you can tell the difference between something it is actively working on and something that needs your attention.
• Usage limits are now evaluated more accurately when deciding whether to start a full review, so Boss is less likely to either block work it could handle or underestimate the cost of a large job.

Fixes

• Fixed a connection timeout that could occur when Boss was talking to Codex.
• Fixed an issue where a draft PR retry could fire against a branch with no commits, which would cause a confusing failure.

PARTIAL: a new outcome when a build finishes with deferred work

When Boss completes a build but some tasks could not fully close out, it can now report a PARTIAL outcome rather than forcing a binary pass or block. You will see this in the PR it opens: the result is clearly labeled, the completed work is recorded, and the deferred items are noted for follow-up. Previously there was no clean way to express “most of this is done” and the build would have to pick one of the harder outcomes instead.

This also unblocks a merge-gate condition that had no producer before, so builds that land in this state can now continue flowing through the pipeline correctly.

Improvements

• When Boss decides to block a build rather than merge it, commits now receive their tag before the block is published. Previously, those commits were left untagged, which caused downstream CI checks to fail on every blocked build.
• Boss is better at distinguishing between decisions it should always stop and report on versus decisions it can make autonomously and record in the PR body. Those recorded decisions now consistently appear in the pull request for review.
• Two operations (image extraction and label creation) are now treated as optional when the connected tracker does not support them. Boss will no longer fail a build just because a tracker integration does not implement those calls.
• Fixed an issue where plan attachments were silently dropped on certain Node versions. Boss now verifies that each attachment was received correctly after uploading, so a failed upload does not go unnoticed.

Fixes

• Fixed an issue where the BLOCKED route could leave commits without a tag, causing commit linting CI to turn red on builds that were intentionally blocked rather than merged.
• Fixed an issue where plan attachment uploads would silently succeed without actually sending any files on older Node versions.

boss init sets up a new repo in one step

Running boss init in a repository now writes a .boss-skills.json file for you. Boss inspects the repo, detects your build commands and project structure, and produces a starting config you can commit and adjust. If something cannot be detected automatically, Boss tells you what it skipped and why, so there are no silent gaps.

This replaces the previous approach of writing that config file by hand or copying one from another project.

Features

• Rename sessions from the home screen. Press r on any session in the list to rename it in place, without opening the session first.
• boss session mcp shows what MCP tools are actually available. Running this command reports the MCP servers and tools resolved for a given session, and where each one came from. Useful when you want to confirm which tools Boss can see before or during a run.
• boss tail now includes agent logs. Log output from the underlying coding agent is now readable through boss tail, alongside the existing service logs. Both log formats are handled automatically.
• Smarter default review. When cross-model review or Compound Engineering is available on your account, Boss now uses them during the review phase automatically, without any extra configuration.
• Knowledge extension role in builds. Boss builds now include a knowledge-extension phase that runs after review, giving extensions a dedicated place to contribute structured knowledge to a build.
• Build time cap raised to four hours. The maximum wall-clock time for a boss-build run is now four hours, up from three.

Improvements

• Better config detection for new and unfamiliar repos. Boss now derives default build commands and targets from what the repo actually declares, rather than assuming a fixed project shape. This affects both boss init and how Boss configures itself when no .boss-skills.json is present.
• Session rows no longer show contradictory status. When a session is actively running, any leftover status hint from a previous failure is now shown dimly rather than at the same weight as the live activity. A past failure on a live row no longer looks like a current one.
• Degraded review tier now fixes what it finds. Previously, when the full review tier was unavailable, the degraded fallback would detect problems and stop. It now attempts to fix them within a bounded repair pass.
• Tracker integration in boss-epic is more accurate. Boss now correctly distinguishes between a tracker tool that is configured but unreachable and one that was never configured, so you see the right status instead of a generic “unreachable” message.
• Capability checks run against the right context. A fix ensures that capability checks run against your actual working project, so tracker tools and other integrations no longer incorrectly report as missing.
• Stale background service is now detected and flagged. If Boss upgraded via Homebrew but the background service is still running the previous version, Boss now warns you on startup and explains how to restart it. Previously this could cause confusing behavior with no obvious cause.
• Liveness signal on running sessions is more precise. Boss now distinguishes between a session that is genuinely making progress and one that is showing a spinner with no new output, so status displays are more meaningful.
• Zero-change sessions are documented. The built-in skill and MCP tool descriptions now explain how to run a session that reads or analyzes a codebase without making any changes.

Fixes

• Fixed an issue where skipped tests could be counted as passing when Boss verified a test run.
• Fixed an issue where tracker tools could silently fail to resolve because the MCP server key did not match what Boss expected.
• Fixed an issue where boss init could write through a symlink when refusing to overwrite an existing config file.
• Fixed an issue where running boss doctor after a Homebrew upgrade could show a false connection error before the service restarted cleanly.

Rename your chats

You can now rename any chat session directly from the chat list. Press r on any chat to give it a new name, which makes it easier to find conversations later when your list starts to grow. There is also a new boss chat rename command if you prefer working from the terminal.

Features

• Orphaned session cleanup. Boss now automatically detects and cleans up terminal panes that were left running after a session ended unexpectedly. These ghost panes would previously just accumulate in the background.
• Smarter no-op runs. Sessions that finish without producing any output no longer get stuck showing as blocked. Boss now recognizes a clean zero-output finish for what it is.
• Non-GitHub repos handled gracefully. If your project is not hosted on GitHub, Boss no longer flags the end of a session as a failure requiring attention. It simply finishes quietly.

Improvements

• Boss loads existing session history faster at startup. It now skips over files that clearly have not changed before doing a deeper check, which reduces the amount of work done up front.
• Starting a new chat is less likely to leave behind a partially created session if something goes wrong mid-creation. Boss now rolls back the terminal pane if the session record cannot be saved.
• Background history repair work no longer competes with the time budget Boss needs to start up responsively. The two tasks now run independently.

CLI gets a full --json mode and new commands for scripting

Boss’s command-line tool has grown significantly in this release, making it much more useful for scripting, automation, and working with Boss remotely. Nearly every read command, boss ls, boss show, boss chats, and session checks, now accepts a --json flag that returns structured, stable output you can pipe into other tools. Error responses follow the same envelope, so scripts can reliably tell what went wrong.

Several new commands also landed: boss merge lets you trigger a merge from the terminal directly. boss chat new opens a fresh chat, and boss chat send gains a --wake-if-asleep option so you can send a message to a session that has gone quiet without manually reviving it first. boss agents lists the agent runners that are currently loaded. boss new picks up --json, tracker fields, and a --tmux-unattended flag for headless use.

Features

• boss merge is now a CLI command. You can merge a session from the terminal, and it returns structured output when used with --json, including stable error codes.
• boss chat new and --wake-if-asleep. Start a new chat session from the terminal, or send a message that automatically wakes a sleeping session.
• boss agents. Lists the agent runners Boss has loaded, useful when you want to see what is active without opening the UI.
• boss ls, boss show, boss chats, and session checks all support --json. Consistent structured output across the read surface makes it straightforward to use Boss in scripts or external tooling.
• Chat status and last output now appear in boss chats. The list view shows each chat’s current status alongside a timestamp for its most recent output, giving you a quick health check from the terminal.
• boss new gains tracker fields, --tmux-unattended, and --json. You can now attach tracker metadata when creating a session and start it in a headless tmux mode, with structured output available for both.
• Remote chat status now works with --remote. Previously, chat status could not be read when connecting to a remote Boss instance. That limitation is gone.
• boss merge now includes the merge strategy in its output, both in the terminal and via the MCP merge_session tool, so you can see exactly which strategy was applied.
• Scroll a chat pane with Shift+PgUp / Shift+PgDn (or Shift+Up / Shift+Down) without leaving the keyboard.
• Account rotation now prefers the account with the most headroom before its quota resets, so weekly allowances are less likely to go to waste when you have multiple accounts configured.

Improvements

• Boss skills can now operate over the CLI transport as a fallback when MCP is not available, so certain operations no longer stall silently in environments without MCP.
• The Boss MCP server is no longer wired into every agent spawn by default, reducing unnecessary overhead for agent runs that do not need it.
• When syncing plans with Linear, image URLs with expiring credentials are now re-signed correctly, markdown is normalized consistently, and estimate values are clamped to valid ranges. This prevents round-trip mismatches that could cause plan comparison to fail unexpectedly.
• Support for two legacy third-party skill integrations has been removed. If you were not using those, nothing changes.

Fixes

• Fixed an issue where scrolling in a chat pane could target the wrong tmux client, causing the scroll to have no effect.
• Fixed an issue where mouse reporting could be lost after re-attaching to a reused pane.
• Fixed an intermittent startup issue where reconnecting to Boss too quickly after a restart could leave the connection in a bad state.

Review quality and speed

The automated review process Boss runs on pull requests got a significant overhaul in this release. Reviews now run their checking rounds in parallel instead of one after another, which means the overall review time drops noticeably on larger PRs. Boss also now picks a review depth based on how much time is available: when there is a healthy budget, it runs the full stack; when time is shorter, it switches to a lighter mode that still catches real problems without stalling the build.

The other big change is convergence. Previously, the review loop could get stuck in a pattern where it flagged a new finding, a fix was applied, and the next round surfaced something related but unverified, repeating indefinitely. Boss now tracks whether each finding was actually confirmed by a follow-up check before it is counted. Findings that cannot be independently verified are no longer treated as blockers, and the loop has a clear exit condition so reviews reliably complete.

Improvements

• Boss now runs review rounds in parallel rather than serially, reducing the time spent waiting for review results.
• The review process has a defined convergence threshold: once findings have been verified across rounds, Boss stops re-reviewing and moves on.
• The PR description now shows which review tier was used, so you can see at a glance whether the full or lighter review ran.

Fixes

• Fixed an issue where several internal tools reported success even when they had actually failed, which could cause gates to pass incorrectly. Security checks, linting wrappers, and a handful of other steps now fail closed when something goes wrong instead of silently reporting a clean result.
• Fixed an issue where GitHub review reply comments could lose their markdown formatting or fail silently due to how the reply body was passed. Replies are now written through a file, which handles special characters and longer content reliably.
• Fixed an issue where Boss could crash or lose output when a linting tool produced a large amount of text.
• Fixed an issue where certain file paths containing non-ASCII characters were skipped by security checks.
• Fixed an issue where sensitive details could appear in delivery error messages. That information is now redacted before it is reported.
• Fixed an issue where remote sessions on machines using the fish shell could fail because shell-specific command wrapping was missing.