Add bundled interface fonts, text size and letter spacing to CompozyOS Appearance

Terminals that agents open now stay in the Terminal list until you open one, instead of opening a desktop window each time. Before, every agent terminal created a window that outlived the process, so desktops filled with windows that restored as "This terminal isn't here" and the Terminal dock item could show a terminal as open when none was running.

When an agent's terminal exits, or any terminal is cleaned up, its windows close on every desktop, and Undo or Reopen cannot bring them back. Terminal windows left over from before a daemon restart are removed at startup. Terminal output history and recordings are kept.

Migration notes: no action needed. Open an agent's terminal from the Terminal list or from the link in the transcript.

Restarting CompozyOS with active always-on agents no longer stalls startup. Before, each agent's main session resumed before the daemon was listening, so its CompozyOS tools waited out a connection deadline and failed before a retry, adding that delay per agent before the desktop app was ready. Always-on agents now resume once the daemon is accepting connections, and connect on the first attempt.

compozy app status also no longer rejects the desktop app's state while it is still starting.

Migration notes: no action needed.

Several problems that could leave an always-on agent's check-ins stuck are fixed:

• A check-in reply that wraps its proposal in a sentence or two of commentary is accepted. Before, any text around the JSON rejected the reply, and repeated rejections could pause the agent. A reply with more than one candidate is still rejected, and the failure report now names the reason.
• After you fix the triage model or runtime configuration that made a check-in fail, the agent schedules its next check within its normal cadence instead of waiting up to a day. Its status explains the earlier failure and that you can run a check now.
• A manual heartbeat wake uses the HEARTBEAT.md currently on disk, the same policy status shows. Before, a file edited in the repository could look active in status while a wake returned heartbeat_no_policy.
• A heartbeat status request that asks for session health without a session ID is rejected as invalid input instead of returning session_not_found.

Migration notes: no action needed.

The Atlassian extension 1.0.1 connects to https://mcp.atlassian.com/v2/mcp, the endpoint Atlassian recommends. On some sites, Jira, Confluence and Rovo Search calls through the old endpoint failed with 401 Unauthorized; scope does not match even after a successful sign-in.

Migration notes: update the Atlassian extension to 1.0.1 from the Marketplace and sign in again when prompted.

A parent session that watches a child with a session monitor now gets one wake when the child stops early, from that monitor. Before, it could also get a second early-stop wake that assumed the child should have opened a pull request and could arrive after the child had already moved on. The early-stop wake now asks the parent to review the child's answer, and is dropped if the child has started a newer turn.

Compact session records from native tools now include pending reply watches, so a parent can see what it is waiting on without requesting the full record.

A Claude subagent's answer is no longer lost or marked failed when the provider reuses its call ID in a later turn, and a background subagent's late summary is kept after its turn finishes.

Migration notes: no action needed.

A parent session busy with a long turn now receives replies and needs-attention wakes from its child sessions during that turn. Before, monitor wakes were delivered while child replies and attention wakes waited until the turn ended. Each reply keeps its own identity, and a needs-attention wake is dropped when the child has moved on or a session monitor already covers that state. When a pending permission or question changes, the parent gets a fresh wake, so the current state is never lost.

A child that answers a follow-up from the session that created it is no longer flagged as stopped before delivering just because it opened no pull request.

Migration notes: no action needed.

A new Claude session started with a full model ID such as claude-opus-5-5 could fail at startup when Claude's model list offered only older IDs or aliases, while the same model worked on another attempt. New sessions, including input-only sessions, now register the selected model the same way resumed sessions already did, so they start on that exact model. If Claude refuses or ignores the model, the session fails once with an error that names the provider and the model. It never switches to an alias or an older model.

Migration notes: no action needed.

A Claude session pinned to a full model ID such as claude-opus-5-5 now resumes on that model after a sign-in recovery. Before, a reloaded session could see only a stale or offline model list that left out the pinned model, so recovery failed with a model error even after you signed back in and Usage showed the account as healthy.

CompozyOS now checks Claude's local sign-in state before it selects a model, keeps the session's exact model through the reload, and reports the real cause when recovery still fails: a missing sign-in, a model the provider refuses, or the last replacement error. A parent watching a child that is waiting for you to sign in is now told so, instead of being told the child is waiting for permission.

Migration notes: no action needed.

A profile with [providers.<id>] enabled = false for a provider that has no built-in or configured definition no longer stops the daemon from starting. Disabled providers are no longer validated until you enable them, and provider usage skips a provider it cannot resolve and logs a warning instead of failing for every provider. An enabled provider, or the default provider, must still be fully defined.

Migration notes: no action needed.

A pull_request monitor can now watch a pull request opened from a fork. Pass the branch as owner:branch, for example alice:fix-login. Before, the monitor registered but never saw the pull request, because only branches in the base repository were matched. A plain branch name still selects a branch in the base repository, and a malformed owner:branch value is rejected when you create the monitor.

A canceled monitor also no longer reports its cancellation time as fired_at.

Migration notes: no action needed.

Loop and Goal status now shows the correct source for each applied Agent Client Protocol (ACP) option, the settings CompozyOS sends to the agent when it starts: options added by the agent definition show agent, and options the provider could not apply no longer list a source. Before, the report could leave an agent-added option without a source or keep one for an option that was skipped.

A Goal that rotates to a fresh session after its context fills now asks for the reasoning level you requested, including the provider default. Before, it could ask for the level the provider had reported, which failed the Goal's runtime check and stopped recovery.

Migration notes: no action needed.

compozy automation jobs trigger <id> now returns as soon as the job's agent accepts the prompt, with the run id and running status, and exits 0. The run keeps going in the daemon, which still owns retries, the final result and session cleanup. Before, the command waited for the whole agent turn, so a turn longer than 30 seconds made the CLI report a request timeout while the run kept going, and you never got its run id.

The HTTP/UDS trigger route and the compozy__automation_jobs_trigger native tool return the running run the same way.

v0.6.0 was the last release that migrated or silently ignored the retired declarations below (#1518). v0.6.1 removes those compatibility shims (#1593). Most old declarations now fail the normal validation for their file instead of being migrated or ignored, and CompozyOS no longer rewrites them. The table notes the exceptions. Current declarations behave as before.

v0.6.1 removes the one-release compatibility translation for Loop tool pins created before native tool result views shipped in v0.6.0 (#1432). Tool actions compare their pinned input and selected output contract directly with the live tool. The automatic full-result fallback, pre-view input digest translation, and table translating superseded output-schema digests have been removed.

When your computer sleeps through several runs of a recurring job, the job now follows its catch_up_policy when your computer wakes up. Before, a daemon that kept running through the sleep replayed every missed run regardless of policy. coalesce and run_once_on_catchup run once for the latest missed time, replay runs each missed time in order, and skip_missed checks the earliest missed run against the grace period: outside it, the backlog is recorded as one skipped run; within it, the latest missed time runs once. The job keeps its original schedule afterwards.

Migration notes: no action needed.

compozy__skill_view now returns a skill document once, in the result's structured content, with a short summary as display text. Before, it returned the full document twice, so a large reference could exceed the result size budget and come back as an artifact to page through instead of inline text. Clients that read only text still receive the complete document.

compozy__tool_artifact_read now returns 32 KiB pages by default so each page fits comfortably in one response. You can still request up to 64 KiB per page.

Migration notes: no action needed.

In a long session, the transcript no longer flickers or stutters when the first tool call of a turn appears. Before, once a session held more than a page of history, each new entry made the app treat the whole loaded transcript as changed and redraw every message. Messages that did not change now keep their place and are not redrawn.

Migration notes: no action needed.

A Terminal window in the Web UI could stay on "Connecting…" forever, with no prompt and no error, most often with the daemon in WSL2 and the browser on Windows. Each connection attempt now has a 10-second deadline. A stalled attempt is retried automatically and the window shows "Reconnecting…". After three stalled attempts in a row, the window says it couldn't connect to the terminal and offers Reconnect.

Migration notes: no action needed.

compozy__terminal_exec now accepts a managed worktree by ID or name, the same way opening a terminal does. Use the native tool's worktree input, the HTTP/UDS request field, or compozy terminal exec --worktree <id-or-name>. Before, a session not bound to a worktree could open a terminal there but could not run a command in it: exec rejected the checkout as invalid_cwd.

Relative working directories resolve inside the selected worktree, and the worktree cannot be removed while the command is still running.

Migration notes: no action needed.

If the app hit an error while syncing the window layout, it now reconnects on its own. Before, the menubar kept showing "Can't save window layout — retrying" with no detail, every dock item stayed disabled, and window commands did nothing until you reloaded the app. The message now shows the underlying error, the app reconnects with increasing delays, and the dock works again as soon as the layout syncs.

Migration notes: no action needed.