v0.8.0

• Fixed an OAuth login that finishes after its server was retargeted arming the old-endpoint token against the new URL: credentials are endpoint-bound at issuance, and the host and kernel only use a token bound to the configured endpoint. Breaking: generic MCP OAuth credentials stored before this release lack the binding and require one /mcp login <server>.
• Fixed mcp add keeping a stored mcp:<name> credential when the entry was new: any add now drops the name's credential, so tokens for authored non-catalog skills (e.g. slack) cannot replay to a user-configured URL.
• Fixed kernel MCP shutdown budgets exceeding the host's kill deadline; graceful close now finishes inside it, and a kernel that exits without a shutdown_reply no longer stalls shutdown for the full deadline.
• Fixed a shutdown race that could leave an MCP server process running after its generation was dropped from the registry.
• Fixed the kernel MCP regression test and the Python runtime tests not running in CI.
• Fixed first IPython calls after an upgrade failing with a raw "Operation was not possible or timed out": kernel startup now tolerates cold venv boots (30s budget; crashes still fail fast via the exit handler), and zmq socket-teardown rejections surface as actionable retriable kernel errors.
• Fixed headless completion reporting a clean finish when a post-compaction continuation failed to start: ACP and print-mode idle waiters now see the failure, while interactive idle behavior is unchanged.
• Added a pre-imported generic MCP API and shell/TUI commands to manage persistent Streamable HTTP and stdio servers in user settings.
• Breaking: removed the documented catalog-name override — an mcpServers entry named after a built-in integration (e.g. linear) no longer repoints the built-in at a custom url/bearerTokenEnvVar; it now disables the built-in skill and is not served by the generic runtime. Rename the entry (e.g. linear-proxy) to keep using a custom endpoint via the generic API. This closes a credential-replay surface where name-keyed tokens could be sent to an override URL.
• Fixed agents overlooking enabled generic MCP connections by advertising their names and pre-imported mcp API usage in the system prompt.
• Fixed /mcp management feedback disappearing during resource reload and limited server details in TUI output to names and transports.
• Fixed credentials configured as env var names resolving to the literal variable name when the variable is set but empty; an empty env var now reports a missing credential (#1468).
• Fixed ACP rejecting an immediate follow-up prompt when injected work restarted the session; follow-ups now queue behind in-flight work, and cancellation drops queued follow-ups before they start.
• Added correlated ACP terminal-quiescence metadata, resident session settlement, and fail-closed daemon input fencing; prevented recovery state from persisting runtime credentials or model configuration.
• Fixed explicit RLM child deletion leaving hidden unsettled work after runtime teardown, including reporting cleanup failures and notifying the parent when deletion completes.
• Added changelog fragments (packages/<pkg>/.changes/*.md) with a CI check and release-time aggregation, eliminating [Unreleased] merge conflicts.
• Fixed the queued-message browse controls (Option+Up) rendering in the same style as typed prompt text inside the input box; the header is now dimmed like other hints so it cannot be mistaken for part of the prompt.
• Fixed IPython kernels and forkserver processes outliving their owner after a hard crash: kernels now arm ipykernel's parent-death poller via JPY_PARENT_PID, the forkserver watches its parent pid, and both pids are registered in the orphan process journal for supervisor recovery.
• Fixed a pid-reuse race for forked IPython kernels: signaling and liveness now go through the forkserver (the kernels' parent) instead of raw pid operations from Node, and the orphan journal's inactive record is only written on a confirmed kill outcome.
• Added session-scoped ACP MCP servers through the kernel MCP program API (#1378 by @hallerite).
• Changed the subagents summary under the prompt into a bordered agents tile with color-coded running/idle/inactive counts and a right-aligned open hint.
• Enabled /fast with OpenAI API-key authentication for GPT-5.4/GPT-5.5/GPT-5.6 and updated the unavailable message (#1595).
• Fixed /goal re-prompting a parent that had correctly delegated to subagents and ended its turn: the continuation now waits until descendant work settles, then resumes automatically.
• Changed post-compaction continuation error classification to typed AgentContinueError codes instead of matching error message text.
• Fixed the working-status elapsed timer (e.g. "Waiting · 5s") restarting at 0s after leaving and re-entering a session or re-attaching to it; the timer is now anchored to the in-flight turn's user message and keeps counting.
• Added a session_before_refine extension hook: extensions can replace /refine and auto-refine planning with their own proposal (for example using a cheaper model — see examples/extensions/custom-refinement.ts) or skip a refinement round; rollbacks bypass the hook and extension edits go through the normal apply-time validation. Also documents refine_complete.
• Added a durable [refinement] transcript message after each refinement showing the applied harness edits (expandable to exact before/after diffs via the shared tool-output toggle), and a live loader while a user-issued /refine runs.
• Fixed the Agents View heartbeat refresh failing entirely ("Cannot list heartbeats while session worker is failed") when any resident worker was terminally failed: failed workers are now excluded from the global catalog while recovering and disconnected workers still fail closed.
• Refreshed MCP providers immediately after server changes so OAuth connections can be started without restarting Prime Agent.

Beta (v0.8.0-beta.7.1.c0c1310)

Automated beta build from main (c0c1310ff39e16ad5f3e762c45546f5fbb02e440).

v0.7.5

• Fixed self-update falling back to an unpublished npm package when the custom release domain is unavailable.

v0.7.4

• Added fullscreen mouse controls, persistent selection, and Linux primary-selection copy support (#8).
• Added GitHub Releases fallback for automated artifact publishing and installer distribution in forks without Cloudflare R2 credentials (#5).
• Changed Ctrl+C to clear the active draft while interrupting work or arming the exit shortcut.
• Fixed Backspace failing to remove a leading shell-mode ! prefix.

v0.7.3

• Changed product analytics telemetry to default to disabled (opt-in instead of opt-out).
• Fixed fullscreen wheel scrolling in Ghostty while retaining application link clicks; set terminal.fullscreenMouse to false to use native Cmd-click instead.
• Changed the agents view to sort idle and inactive sessions by last message time, newest first, while keeping running agents in stable creation order.
• Fixed openai-codex models being invisible to rlm subagents and find_models because model discovery reported Prime Agent's own version as the Codex client version (#1375 by @bilelrais).
• Added a working hint that recommends sharing traces with Prime Intellect to help train open-source LLMs.
• Restored bare prime-agent --resume opening the agents view and the /resume [id|path] slash command; bare commands open the agents view and an argument resumes that session in place.
• Fixed URLs not opening on click in fullscreen mode on terminals such as Ghostty; clicking a link in the transcript, dock, or overlays now opens it in the browser.
• Fixed ctrl+p ("Toggle agent message expansion") only toggling received agent messages; it now expands and collapses sent agent messages together with received ones.