Changelog #

[Unreleased] #

• Fixed background (unattributed) kernel output missing from the expanded IPython cell view: it is now surfaced in the tool details and rendered under a "background output (unattributed)" label after stdout/stderr/result.
• Fixed a protocol interrupt during a REPL state restore leaving a mixed old/new namespace: names are now staged first and applied atomically with SIGINT parked across the apply, and an interrupt landing anywhere between a committed snapshot or restore and its request finishing is recovered instead of misreporting the completed operation as failed.
• Fixed the REPL snapshot writer leaving a new payload beside a truncated manifest on mid-write failures: payload and manifest now commit via unique same-directory temp files and atomic renames with guaranteed cleanup, and an interrupt during cleanup can no longer misreport a completed destructive snapshot as failed.
• Fixed the REPL runtime list_names request crashing the serve loop when the namespace held a non-string key; non-string keys are now skipped and every runtime request fails individually through the shared backstop instead of killing the loop.
• Addressed REPL host-swap review findings: reworded stale IPython-specific busy/restart messages for the default kernel and stopped restart() from resurrecting a concurrently killed REPL kernel.
• Fixed graceful REPL kernel shutdown() losing teardown ownership to its own child's exit handler, which made restart() misread the shutdown as superseded and never start the kernel again.
• Fixed REPL kernel start() waiting out the full 30s ready timeout when the kernel process fails to spawn; the spawn error now rejects startup immediately.
• Fixed a cell that rebound or ignored SIGINT (or a restored prior handler) permanently breaking protocol interrupts: the REPL runtime now re-asserts its SIGINT handler between cells.
• Fixed the Python REPL runtime surviving its owning process's death while a non-yielding cell runs: an owner-watchdog thread now hard-exits the runtime (killing live bash children first) when the owner process dies.
• Fixed an interrupt parked during a snapshot's prune window misreporting the completed destructive snapshot as failed; it is now consumed once the manifest is committed, and an interrupt landing just after a completed snapshot/restore request is consumed too instead of failing its done.
• Fixed a REPL runtime interrupt gap where an interrupt landing during a cell's trailing-expression repr or output drain was dropped; the request now stays interruptible until its done event is emitted, so a slow user repr can be cancelled.
• Fixed two REPL runtime request-lifecycle bugs: a cell closing sys.stdout/sys.stderr no longer kills the serve loop (done still arrives and later cells run), and an untargeted interrupt parked for a request that fails to compile is consumed with that request instead of spuriously cancelling the next cell.
• Fixed the REPL runtime leaking a finished cell's id onto late background-thread output: the current cell is now cleared right after the post-cell drain, so done stays the last event with that id and between-cell output carries a null id.
• Fixed bash() cells failing under strict-POSIX shells (dash) when the status pipe landed on a multi-digit fd.
• Fixed rlm.run outside a live kernel hanging forever instead of failing fast, which stalled CI shard 3 until timeout.
• Fixed two bash() spawn races: status-channel fds no longer leak when pipe creation fails mid-setup, and a status-socket gate keeps the command from starting until its pid is journaled (a kernel kill in that window now stops the child instead of orphaning it past the reaper).
• Fixed a compile-phase crash (e.g. RecursionError from a pathologically deep attribute chain) killing the REPL runtime instead of failing the one cell: any per-request failure now becomes error+done and the serve loop keeps running; rebinding sys.stdout/sys.stderr to flush-less objects no longer kills it either.
• Fixed the REPL runtime hanging before done when a cell closes fd 1/2 and a later open() reclaims the number: drain sync tokens now go through a private dup of the capture pipe, with a pump-liveness backstop so a dead pump can no longer wedge the serve loop.
• Fixed the Windows orphan reaper killing only the journaled bash() shell pid; it now uses taskkill /T so descendants die with the tree, matching the in-kernel bash() kill paths, and resolves taskkill via an absolute System32 path (with NoDefaultCurrentDirectoryInExePath) so a planted CWD taskkill.exe cannot hijack cleanup.
• Fixed a snapshot request with identical path and manifest_path silently clobbering the just-written state payload; the runtime now rejects it as a failed request.
• Fixed a snapshot request with a negative max_bytes/max_variable_bytes and prune_oversized writing an empty payload and then deleting every user variable; size caps must now be non-negative integers.
• Fixed an interrupt landing mid-snapshot leaving prune deletions half-applied: once the snapshot manifest is committed, SIGINT is deferred until every oversized name is removed, so the namespace always matches the on-disk snapshot.
• Hardened bash(): cancelling await bash(cmd) now kills the command's process group (background handles are unaffected), Windows helper binaries resolve via absolute System32 paths, kill() retries taskkill for already-reaped Windows trees, and orphan-journal enrollment fails closed when configured.
• Fixed cross-cell output misattribution in the REPL runtime: stream events are attributed at write time via context, and raw fd or user-thread output is emitted with a null id instead of being credited to whichever cell is running.
• REPL kernel: output from user threads, other cells' leftovers, and raw fd writes is no longer merged into the running cell's stdout; it is surfaced separately as unattributed background output.
• Hardened bash() further: the host now injects an absolute default shell into the kernel (no PATH lookup; /bin/bash else /bin/sh on POSIX), macOS start-id lookup uses /bin/ps, and Windows worker-teardown orphan kills go through hardened taskkill /T.
• Hardened Windows bash execution: the kernel shell is resolved only from trusted absolute paths (never PATH), and bash children are contained by kill-on-close job objects so a crashed kernel cannot leak process trees (taskkill remains only as a fallback when job creation fails).
• Hardened Windows bash() containment: children are now created directly inside the kill-on-close job (PROC_THREAD_ATTRIBUTE_JOB_LIST at CreateProcessW time), so no window exists in which a kernel kill can leak a suspended, never-run process; handle inheritance is restricted to exactly the child's stdio handles (PROC_THREAD_ATTRIBUTE_HANDLE_LIST), so concurrent spawns cannot leak each other's handles; the journal start-id query still runs only while the job-contained child is suspended, and bash() still raises instead of falling back to jobless taskkill when containment fails.
• Fixed a Windows bash() PID-reuse hazard: the child process handle is now retained through job cleanup and every taskkill-by-pid fallback (watch reap, kill(), cancel escalation, shutdown cleanup) and closed exactly once only after the handle is marked reaped, so a recycled pid can never be killed by the fallback.

[0.8.1] - 2026-08-26 #

• Fixed syntax highlighting in the expanded python tool-call view: triple-quoted strings spanning multiple lines now keep their string color instead of only the first line.
• Changed the default RLM maximum recursion depth for new sessions from 1 to 2.
• Changed ACP prompt requests to resolve only after all causally admitted subagent and parent work has settled.
• Changed the Cloudflare AI Gateway default model to claude-sonnet-4.5 after the catalog dropped the gateway's workers-ai mirror ids.
• Fixed ACP assistant chunks to identify message boundaries across autonomous turns.

[0.8.0] - 2026-08-21 #

• 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.

[0.7.4] - 2026-08-19 #

• Fixed model searches ranking stronger matches ahead of weaker signed-in matches while preferring signed-in providers for equivalent results (#539 by @eliebak).
• Fixed large IPython variables repeatedly slowing later turns by excluding them from persistent snapshots and removing them when context is compacted.
• Fixed daemon socket paths being used verbatim in identity derivations: on supported platforms, --daemon-socket spellings differing only by duplicate or trailing slashes now normalize to one canonical path, so worker-descriptor namespaces, daemon log files, and persisted descriptors agree.
• Added a thinking option to rlm.run for spawning subagents with an explicit reasoning level; invalid levels for the resolved child model fail spawn.
• Changed opening the agents view (full or scoped) with a draft prompt to auto-stash the draft instead of refusing; the draft is restored into the editor when the session is reopened.
• Fixed Shift+Enter no longer inserting a newline in terminals that send a literal \n (for example a Ghostty shift+enter=text:\n mapping): the byte decoded as ctrl+j and triggered the new edit-diff toggle instead of the editor newline.
• Removed a system prompt paragraph referring to an async bash() kernel helper and managed jobs that do not exist in the runtime.
• Changed RLM guidance to orchestrate independent workers in parallel, use available async shell helpers safely, end the turn instead of sleeping, polling, or blocking on long awaits, provide proactive outcome-focused progress updates from root agents, and use simplified technical English for user-facing prose.
• Fixed new top-level daemon sessions inheriting an RLM child depth from the supervisor process.
• Fixed active goals stalling after a mid-goal automatic compaction when the previous continuation prompt was already running: only undelivered continuations deduplicate, so a fresh continuation is queued instead of being suppressed.

[0.7.3] - 2026-08-17 #

• Fixed assistant rendering when provider payloads contain null or sparse content blocks.
• Added authenticated host-request contracts with per-call request IDs, generation fencing, cancellation signals, and currentness checks.
• Fixed root daemon shutdown retaining cleanup ownership while kill events are in flight.
• Changed RLM family discovery to use a daemon-owned append-only spawn ledger with per-child display metadata instead of reconstructing topology from session files.
• Fixed long-running macOS supervisors losing ownership when system cleanup removed authority records from $TMPDIR.
• Fixed deleted RLM children leaking kernel snapshots while retaining their readable transcript tombstones.
• Changed Agents View subagent rows to show stable name · model/effort · summary metadata.
• Changed the default Cerebras model to the available gpt-oss-120b route and aligned cross-provider handoff fixtures with the generated catalog.
• Fixed the agent going silent after an automatic context compaction interrupted unfinished work: the tool loop now resumes when a threshold compaction fails or is skipped, and active goals keep continuing after a successful mid-goal threshold compaction.
• Changed the agents view splash hint from "type to start" to "type to search sessions".
• Added app.edits.expand (ctrl+j) to toggle edit diffs; diffs are now shown only by this toggle, and ctrl+o no longer affects them.
• Changed edit rendering so the ╰─ <path> +N -M summary line is always visible and ctrl+j toggles the diff inline beneath it, indented to the summary text.
• 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.

[0.7.2] - 2026-08-11 #

• Fixed Down Arrow focusing the Agents View entry before moving a nonempty prompt cursor to the end (ENG-5147).
• Added app.messages.expand (ctrl+p) to collapse or expand agent-to-agent messages separately from ctrl+o tool output.
• Added a ctrl+t expand hint to collapsed thinking blocks, matching the tool output hint.
• Changed expand/collapse hints to a consistent bracketed (Ctrl+O to expand) style across tool, message, summary, and error rows.
• Added a configurable copy action to login dialogs so raw sign-in URLs can be copied without selecting wrapped text (#643).
• Added privacy-safe pseudonymous product analytics for onboarding, command use, execution modes, run outcomes, TTFT, latency, usage, tools, retries, and compactions, with disclosure and opt-out controls (ENG-4682).
• Changed sent agent messages in the IPython cell UI to show only the message text with a ╰─ gutter when expanded, matching received messages, and hid the raw agent_message.send receipt dictionary.
• Fixed Homebrew installs attempting to self-update their versioned Cellar keg instead of directing users to brew upgrade prime-agent (#844)
• Fixed the agents view collapsing expanded subagent lists when returning from an opened agent (ENG-5105).
• Kept the subagent summary row visible and selectable while its list is expanded in the agents view, so pressing enter on it collapses the list again (ENG-5105).
• Added in-place editing of queued steering and follow-up messages: Alt+Up/Alt+Down browse the queue from the draft, Enter applies the edit as steering, Alt+Enter as a follow-up, and submitting an empty editor deletes the item; interrupts now preserve the queue (#838).
• Fixed workers with no live connection reporting as ready; stopping workers now report a stopping state, are hidden from live sessions, and no longer receive daemon-wide commands (#850).
• Fixed timed-out worker stops stranding dead-but-registered workers ("Session worker is not connected"); stops now finalize in the background once the process exits, and zombie processes are no longer counted as alive (#851).
• Fixed sessions becoming permanently unopenable after a stale worker registration was left behind; open/resume now self-heals by finishing the old cleanup and starting a fresh worker (#852).

[0.7.1] - 2026-08-07 #

• Fixed the bundled websearch skill description and missing-key guidance omitting the /login → MCP Connections step required to configure Serper.
• Fixed retry_worker cancelling its own recovery when a stopped session worker left a saved stop marker behind, leaving the session stuck at "Session worker is not connected".

[0.7.0] - 2026-08-05 #

Breaking Changes #

• Changed agent messages to always use steering delivery and removed delivery-mode options from the Python, CLI, RPC, and connection APIs. Code passing mode to agent_message.send, or a delivery mode over the CLI/RPC, must drop it.

Changed #

• Changed self-updates to report the previous and new Prime Agent versions.

Fixed #

• Fixed the subagent summary showing retained children as idle while they run follow-up work.

[0.6.1] - 2026-08-05 #

• Added reverse tab navigation to the /login configuration menu and moved the model scope shortcut to Alt+S.
• Fixed daemon startup crashes hiding their exit status and daemon log until the startup timeout.
• Documented the global idleEvictionMinutes daemon setting, including its default, valid values, and eviction/passivation behavior (#621).
• Fixed top-level --help omitting acp from the supported --mode values (#620).
• Fixed stop and rename becoming prompts when --daemon-socket precedes the command (#622).
• Fixed subagent terminal notices arriving as anonymous follow-up prompts instead of attributed agent messages, so a parent can now tell which child reported completion, failure, or cancellation, and a busy parent is steered at the next turn boundary rather than waiting to go idle (#617).
• Fixed ACP mode reporting a failed turn as a clean end_turn. A provider error, expired auth, or unusable model left session/prompt resolving with no updates at all, which reads to a client as a successful but empty turn; the turn now fails with the underlying error instead.
• Fixed ACP cwd mismatch metadata treating symlink aliases such as macOS /var and /private/var as different directories (#623).

[0.6.0] - 2026-08-04 #

Breaking Changes #

• Changed rlm(...) to return at task admission instead of waiting for the child to finish. It now yields a spawn handle (rlm_child_id, name, session_dir, model); RLMResult and its final answer, usage, and model-fallback warning are gone. A child reports back with agent_message.send(..., receiver_role="parent"), which arrives as an ordinary prompt and starts a parent turn. Code that read result.answer, or treated asyncio.gather(...) over rlm(...) as fan-in, must be updated.
• Changed agent_message.send to role-addressed delivery: pass receiver_role ("parent", "sibling", "child") plus receiver_name for siblings and children. The old positional send(target, message) form no longer works, and the separate roster() call is now agent_message.list_agents().
• Narrowed agent reach to the nuclear family: an agent may message or observe only its parent, siblings, and direct children. Top-level sessions are siblings of one another, so agent-to-agent between them still works; grandchildren and cousins must be reached by relaying through the intermediate child. Users are unaffected and still see every session.
• Requesting an unavailable subagent model now fails the spawn instead of silently falling back to the parent's model with a warning.
• Bumped the daemon schema revision to 13 for the parent-edge, depth, naming, and passivation wire changes; older clients and daemons are rejected cleanly at connect.

Added #

• Added --mode acp: Prime Agent now runs as an Agent Client Protocol agent over NDJSON on stdio, driving an AgentConnection in-process. IPython surfaces as an ACP execute tool call carrying its cell source, and capabilities ACP has no native concept for (subagents, autonomous gate state, rich IPython output, compaction, goals, heartbeats, continual-harness refinement) travel in a namespaced ai.primeintellect.prime-agent _meta envelope that vanilla ACP clients ignore. Documented in docs/acp.md.
• Added /rlm-max-depth to view or set the recursion cap for the current chat, with --global to change the default for new sessions.
• Added recursive navigation to the agents view: drill into any session's children and back out again, with each chat showing its own depth.
• Added a family roster via agent_message.list_agents(), listing parent, siblings, and children with name, id, depth, and status, including family members currently on disk.
• Added sibling-unique agent names, enforced at spawn and rename against loaded and unloaded sessions alike. The same name may be reused at different depths.
• Added an idleEvictionMinutes setting (default 90, off to disable) controlling idle eviction and passivation.

Changed #

• Changed finished subagents to stay on disk until something touches them, so memory scales with the active frontier rather than every subagent ever spawned. Lists show them without loading them, and attach, message, or transcript read wakes them on demand.
• Changed sessions to persist their parent edge and derived RLM depth, so tree position no longer has to be inferred from whatever happens to be in memory.
• Changed the supervisor to stop worker processes whose whole session tree has been idle past the threshold, and to passivate individually idle children inside still-busy workers.
• Replaced the child-agent inspector with a single subagent summary line under the prompt that opens the agents view scoped to that session's children.

Fixed #

• Fixed stop and rename rejecting custom daemon socket options.
• Fixed SIGINT in print mode leaving the session active until liveness reclaim.
• Fixed daemon startup failing permanently when an interrupted supervisor owner directory contained only stray files.
• Fixed agents-view fallback notices and scoped live sessions surviving transient refresh failures across chat returns.
• Fixed stopping completed subagents deleting their retained sessions.
• Fixed silent or cancelled RLM children leaving parents without a terminal status notice.
• Added missing argument hints to /name, /model, /export, and /import in autocomplete.

[0.5.1] - 2026-08-04 #

Fixed #

• Fixed /refine failing with an opaque JSON parse error when the refiner exceeded a fixed 4096-token output cap; output budgets now derive from the selected model, and a truncated reply reports the exhausted budget directly.

[0.5.0] - 2026-08-03 #

Breaking Changes #

• Reworked session input scheduling into a single session action lifecycle and store (daemon protocol 7, schema revision 8); older clients and daemons are rejected cleanly at connect.

Changed #

• Changed large daemon session loads to stream JSONL history and avoid retaining a second full-file copy in memory.
• Changed the agents view to render explicit session names in bold and the "(no messages)" placeholder in italics.
• Changed subagent guidance to retain reusable children and delete completed direct children once they are no longer needed.
• Changed top-level CLI help and documentation to expose autonomous mode, quality gates, and their limits.
• Changed daemon and RPC session state to report literal queued actions separately from active scheduler work.

Fixed #

• Fixed the blank line between the recap and the working hint so they render directly above each other.
• Fixed compaction retaining runtime resources after an explicitly deleted subagent had a transient cleanup failure.
• Fixed long-running thinking timers to display hours and days instead of unbounded minutes.
• Fixed overlapping daemon snapshot catch-ups closing healthy workers and preventing new sessions from starting.
• Fixed active scheduler work being reported as queued in session state.
• Fixed headless runs completing before queued follow-up work had finished.
• Fixed /compact consuming itself as its own successor action.
• Fixed daemon parse rejections dropping the command id, which left older clients waiting for a timeout instead of seeing the protocol error.
• Fixed --goal sessions never showing the objective to the model, which made seeded goals invisible to first turns and continuations.

[0.4.0] - 2026-08-01 #

Breaking Changes #

• Replaced the recursive daemon get_session_tree response with flat nodes linked by parentId (protocol 6); clients must support the new response shape.
• Removed /resume and bare --resume; browse sessions with left-arrow from a daemon chat, or use --resume <session-id|path> for a direct resume.

Added #

• Added ctrl+n to start a session from Agents View, and alt+enter to queue a reply as a follow-up while Enter steers a streaming session.
• Added session-owned /compact, /refine, /goal, and /autonomous commands with autocomplete to the Agents View reply composer, plus target-scoped /name and /kill commands.
• Added optional stable session names and initial prompts to /new.

Changed #

• Changed collapsed edit and IPython calls to show compact per-file line-change summaries while retaining full expanded diffs.
• Changed bare /effort to open a selector of the current model's supported reasoning levels, and removed token estimates from reasoning-effort displays.
• Improved session search ranking to prefer exact session-name and first-message matches before prefix, substring, and transcript fuzzy matches.

Fixed #

• Fixed deeply nested /tree sessions overflowing the daemon serializer by transferring and rebuilding the session tree iteratively.
• Fixed prime-agent agents opening a new chat for a process-local session.
• Fixed daemon startup after an interrupted supervisor leaves an empty ownership directory.
• Fixed /effort xhigh and /effort max being rejected before a model is active.
• Fixed IPython tracebacks emitting ANSI color codes.
• Fixed selected rows and selectors becoming nearly invisible on terminals whose background matches the selected theme color.
• Fixed startup waiting on private Prime Inference model authorization by caching authorization locally and refreshing stale entries in the background.

[0.3.3] - 2026-07-23 #

• Removed the bundled orchestration heartbeat skill from the model system prompt.
• Fixed feature hints crowding queued messages and side questions by placing them below the recap and hiding them while messages are queued (ENG-4741).
• Fixed /btw truncating long answers by rendering side questions in the scrollable transcript.
• Changed recognized slash commands to retain accent coloring after submission in live, replayed, and queued TUI surfaces while preserving Markdown arguments.
• Unified prompt, steering, follow-up, and session-command scheduling under session-owned admission with durable queue state and coordinated update/restart checkpoints.
• Unified Agents View and session resume into one searchable Running/Idle/Inactive session view with live heartbeat badges.
• Changed selection cursors from → to › across model selectors, scoped-models, and the theme default for consistency with tree and user-message selectors.
• Changed the queued follow-up hint connector from ↳ to ╰─ to match the tool-execution continuation connector.
• Changed /context tree connectors from ├ /└ to ├─ /└─ to match the tree selector and session picker.
• Changed the IPython cell queued marker from ▸ to ◇ to match the subagent and context-tree status icons.
• Changed slash-command autocomplete to separate argument hints and resource provenance, show only the selected command description, and summarize hidden results directionally.
• Fixed cancelled extension commands remaining alive when spawned processes ignored SIGTERM (#458 by @snimu).
• Fixed OAuth browser launch URLs being interpreted by the system shell.
• Added agent-callable refine skill so the model can schedule continual harness refinement from IPython via await refine.run() without blocking the current turn (#504 by @sethkarten).
• Changed long live session opens to render a bounded recent transcript tail while preserving full prompt history (#343 by @sethkarten).
• Changed /refine to run planning in the background so the conversation is not blocked during the LLM pass (#497 by @sethkarten).
• Added serialized headless refinement and --goal / --goal-token-budget for seeding durable session goals (#514 by @sethkarten).
• Added multi-turn /btw side conversations with transient in-pane bash commands (#512 by @ilijalichkovski).

[0.3.2] - 2026-07-20 #

• Fixed invalid --resume session IDs being submitted as prompts, with nearest-session guidance instead (ENG-4722).
• Changed /model to show all public models with authenticated providers first and open provider authentication when an unavailable model is selected (ENG-4575).
• Changed the shared configuration menu to cycle tabs with Tab, use Shift+Tab for model scope, show an Escape close hint, preserve arrow-key search editing, and remove the model selector's provider shortcut.
• Fixed searchable selectors retaining their previous scroll position after the query changed.
• Changed interactive, print, JSON, RPC, piped-stdin, and no-session clients to use the same daemon-owned runtime while preserving their existing commands, output protocols, and lifecycle behavior (ENG-4685).
• Added RPC controls for schedules, heartbeats, agent messaging, and live session observation (ENG-4685).
• Fixed daemon-backed headless startup, rollback routing, RPC wire compatibility, and duplicate client runtime preparation (ENG-4685).
• Fixed heartbeat-owning subagents appearing completed, showing completion checkmarks below the prompt, being omitted from active subagent counts, or remaining visible after deletion.
• Fixed the heartbeat tray and manager showing heartbeats from unrelated sessions.
• Fixed daemon backpressure triggering redundant catch-up snapshots for events already queued by the socket.
• Added dedicated stable and beta installers, with stable advancing on version bumps and beta advancing on every commit to main.
• Fixed incompatible daemon builds crashing startup or respawning after shutdown, with capability negotiation, verified provenance, and convergent force shutdown (ENG-4687).
• Changed tool-result and announcement images to show compact metadata instead of terminal graphics (#437 by @snimu).
• Changed top-level CLI help to show concise common options and commands without loading runtime resources (ENG-4688).
• Fixed completed subagents cancelling their RLM heartbeats before the first run (ENG-4652).
• Changed the fullscreen follow shortcut from Alt+Down to Ctrl+Shift+Down for more reliable terminal input (ENG-4684).
• Added user-requested model selection for subagents with bounded account-authorized discovery and explicit parent-model fallback warnings (ENG-4649).
• Added subtle feature hints to longer-running agent turns (ENG-4521).
• Fixed active heartbeats not resuming after Prime Agent updates (ENG-4657).
• Fixed the Agents View reordering sessions whenever prompts or heartbeats updated their activity timestamps (ENG-4650).
• Added parent-scoped subagent lifecycle APIs: create children with readable default or orchestrator-chosen names, recover running or completed children through rlm.list_subagents(), continue them through agent messaging, and close/remove them with rlm.delete_subagent().
• Changed shell commands to use discoverable agent, schedule, package, model, session, update, doctor, and full-shutdown verbs without exposing the background daemon hierarchy (ENG-4538).
• Fixed unsupported Node versions crashing before startup by requiring Node 22.8.0 or newer and showing upgrade guidance before loading the CLI (ENG-4260).
• Added @ file-path autocomplete to new-agent and reply prompts in the Agents View.
• Fixed slow daemon clients becoming stuck when newer session snapshots arrived during catch-up.
• Fixed queued messages getting stranded when an agent turn ended (ENG-4653).
• Changed /traces upload-all to pace requests within the platform rate limit, honor bounded Retry-After responses, and support interruption.
• Fixed resuming a daemon-resident session to attach the requesting client to its existing worker without disturbing other clients (ENG-4656).
• Fixed daemon-owned updates terminating their updater before the daemon restart and session restore completed (ENG-4606).
• Fixed first-launch Prime login and kept onboarding visible between team and model selection (ENG-4658).
• Fixed active heartbeat sessions appearing under Needs Input or Completed instead of a dedicated Heartbeats section (ENG-4654).
• Fixed stashed prompts being lost when leaving and reopening a session from the Agents View (ENG-4659).
• Added a combined heartbeat indicator and manager for pausing, resuming, or stopping user and agent heartbeats (ENG-4536).

[0.3.1] - 2026-07-15 #

• Added /fast for OpenAI Fast mode on supported ChatGPT models (ENG-4620).
• Changed wrapped diff rows to use a blank hanging gutter.
• Fixed team-gated Prime Inference routes being missing from model selectors by merging the authenticated team catalog during model refresh (ENG-4645).
• Added confirmation when fullscreen text selection copies to the clipboard (ENG-4644).
• Added an agent-run edit total above the recap.
• Changed edit tool calls to always show full diffs while keeping IPython source collapsed until Ctrl+O expands it.
• Changed tool expansion hints to appear only on the latest tool row instead of every tool call (ENG-4583).
• Changed IPython kernels to set NO_COLOR=1, preventing ANSI color escapes from inflating %%bash output.
• Fixed update restarts starting concurrent daemon supervisors or unlinking a replacement supervisor's socket (ENG-4600).
• Fixed worker recovery races and made daemon shutdown-all converge across hidden supervisors (ENG-4603).
• Changed provider, model, and MCP setup to use one tabbed configuration menu (ENG-4539).
• Changed the shared configuration menu to show prominent, responsive tabs with configurable navigation shortcuts (ENG-4534).
• Fixed IPython edit diffs replacing syntax highlighting with a single foreground color (ENG-4616).
• Fixed Prime Inference login leaving new sessions without a persisted model selection (ENG-4573).
• Fixed empty prompt placeholders hiding the input caret.
• Fixed automatic model selection preferring other configured providers over Prime Inference's GLM 5.2 default.
• Fixed missing ripgrep blocking subagents and added actionable installation guidance for the optional search helper (ENG-4572).
• Removed the shared worker snapshot spill cache to prevent concurrent workers from deleting each other's snapshot chunks (ENG-4601).
• Fixed narrow slash-command descriptions ending abruptly or clearing the prompt background, and added a content-sized popup above the input with the same distinct surface as /btw (ENG-4542).
• Fixed snapshot transfers terminating resident workers, stranding partial readers, or rejecting identical retries (ENG-4602).
• Fixed the resume picker opening on an older session instead of the newest session (ENG-4630).
• Fixed tool-only responses rendering directly against the preceding user prompt.

[0.3.0] - 2026-07-13 #

• Changed daemon and headless execution to isolate each root session tree in a recoverable worker process, with protocol-v2 chunked snapshots, compact streaming, attachment-local backpressure, session leases, and unchanged print, JSON, and RPC interfaces.
• Added autonomous mode with host-side continuations, configurable limits, and quality gates for evaluator-controlled runs (#278 by @sethkarten).
• Added /traces preview and /traces upload-all for inspecting the current payload and backfilling saved parent and subagent traces.
• Changed /traces upload and /traces upload-all to be explicit one-shot uploads that do not enable automatic sharing.
• Changed trace uploads to retry transient network and HTTP failures with bounded exponential backoff and jitter.
• Fixed Prime Inference credential and team-header precedence to prefer PRIME_API_KEY, then the Prime CLI config, then auth.json.
• Fixed aborted autonomous gates leaving detached process trees and supervisor recovery retaining intentionally stopped workers after stale scheduler locks.
• Fixed supervisor replacement surfacing fatal socket errors or recovering roots that were intentionally stopped (ENG-4526).
• Fixed daemon catch-up snapshots being disposed mid-transfer or triggering resets that cleared drafts, local queues, dialogs, active UI state, or in-flight reasoning traces.
• Fixed compact daemon streams occasionally duplicating the first token of an assistant response.
• Fixed subagent prompts and usage counters flickering or disappearing during daemon resyncs and large parallel runs, and added compact fixed-width recap rows.
• Fixed stale heartbeat jobs reopening sessions after they were archived, deleted, explicitly shut down, concurrently terminated, or lost resident worker ownership (ENG-4519).
• Fixed heartbeat starvation by moving durable schedules into per-session artifacts and running them concurrently in their owning resident workers, independent of supervisor replacement (ENG-4527).

[0.2.9] - 2026-07-13 #

• Changed tool call groups to use one blank row above and below without blank rows between consecutive calls.
• Changed the session tree to show only user messages by default.
• Changed agent-to-agent messages to render as directional rows, with received messages expandable in chat and sent messages shown below their Python cell (ENG-4531).
• Fixed IPython state restore notices rendering as full user messages when prompts were queued or restored (ENG-4530).
• Changed bare /mcp to open the Services menu while preserving explicit list, login, and logout subcommands (ENG-4535).
• Added /btw and /side for one-turn inline side questions that use the current context without changing the main session (ENG-4509).
• Changed scheduled heartbeat prompts to steer (interrupt the current turn) by default, with a steer/follow_up delivery mode selectable via /heartbeat --steer|--follow-up and the rlm_heartbeat skill's delivery_mode argument.
• Changed the new-chat splash to show only version, model, and cwd metadata and rotate among five example prompts.
• Fixed self-updates losing restored daemon sessions to a socket cleanup race and leaving open session or agents-view windows disconnected.
• Changed daemon connection errors to report the failed operation, session identity, recovery steps, socket, and diagnostic log instead of raw protocol reasons.
• Changed the Agents View and new-chat splashes to keep one blank row above the butterfly.
• Fixed Agents View retrying after an intentional daemon shutdown instead of stopping with restart guidance.
• Fixed stale heartbeat jobs reopening archived, deleted, or concurrently terminated sessions (ENG-4519).
• Fixed onboarding blocking normal TUI use by reopening login or model selection after startup (ENG-4537).
• Fixed IPython Bash cells with leading blank lines being labeled and previewed as Python (ENG-4529).
• Fixed recap layout shifts by keeping the previous recap visible until its replacement arrives (ENG-4533).
• Changed the new-chat tray to hide shortcut guidance while typing and keep the agents link visible.

[0.2.8] - 2026-07-09 #

• Added built-in Herdr integration that reports agent lifecycle state to Herdr panes automatically, without requiring herdr integration install pi.
• Changed Escape to interrupt active work with a visible abort notice, double Escape to open the session tree from an empty prompt or clear an idle draft, and ? to show shortcuts (ENG-4489).
• Changed new-chat guidance to show concise shell, command, file, and shortcut hints, with Agents View first and ? for shortcuts after the model and effort (ENG-4489).
• Changed ? shortcut help to appear as a temporary compact panel below the transcript, while /hotkeys shows the full reference without Ctrl+Z (ENG-4489).
• Fixed Escape repeats around autocomplete, queued draft restoration, whitespace-only drafts, and active background work (ENG-4489).
• Fixed the agents-view splash shifting when opening an agent session (ENG-4517).
• Changed /model to sort featured flagship models above a provider's long tail (with a numeric-aware alphabetical tiebreak), so the full Prime Inference catalog doesn't flood the picker.
• Fixed selector prompts and choices filling their background through the terminal's right edge.
• Changed automatic harness refinement to be enabled by default while keeping autoRefine.enabled: false as the opt-out.
• Fixed non-numeric autoRefine.turnInterval and autoRefine.cooldownMs settings falling back to defaults instead of silently enabling a noisy auto-refine loop.
• Fixed all session-resume entry points to share a searchable full-screen picker, stream results while loading, and support renaming (ENG-4513).

[0.2.7] - 2026-07-08 #

• Changed subagent and refinement guidance to favor non-blocking subagent tasks by default, use disk-backed tracking for long-running fan-out, inspect or message live subagents when agent observation/messaging skills are available, and capture reusable delegation roles, procedures, facts, preferences, and prompt addendums with /refine.
• Changed attach_image to resize and compress large inline image attachments before storing them for rendering and replay (#340 by @sethkarten).
• Fixed heartbeat and goal continuation prompts rendering like ordinary user messages (ENG-4482).
• Fixed /heartbeat guidance to show stop and the every <duration> <instruction> interval syntax (ENG-4484).
• Fixed Ctrl+C canceling the active turn, bash command, and IPython kernel execution deterministically, with a compact recovery prompt and model-visible reset notice when an interrupted IPython cell keeps running (ENG-4490).
• Fixed login dialogs in fullscreen so sign-in URLs can be selected natively (ENG-4480).
• Fixed /model opening and selection staying blocked on live model refreshes (ENG-4505).
• Fixed provider auth failures leaving stale credentials shown as connected in /login (ENG-4491).
• Fixed typing into the prompt after highlighting an inline subagent (ENG-4494).
• Fixed session-targeted heartbeat jobs staying scheduled after sessions are killed or saved sessions are deleted (#332).
• Fixed self-updates interrupting and automatically resuming daemon sessions instead of waiting for long-running work to finish.
• Fixed provider errors being surfaced instead of retried within the retry budget (ENG-4503).
• Fixed Agents View returning from fullscreen sessions without flashing primary scrollback (ENG-4508).

[0.2.6] - 2026-07-06 #

• Fixed the installer splash flickering during animation and resize by stabilizing full-screen redraws and removing misleading synthetic percentages (ENG-4481).
• Fixed Prime Inference auth syncing with Prime CLI login and team selection.
• Fixed provider auth failures showing provider-specific /login commands instead of the /login selector.
• Removed the legacy pi-mono bash and edit built-in tools; use IPython %%bash cells and the Python edit skill instead.

[0.2.5] - 2026-07-06 #

• Added daemon-backed user orchestration with agent-to-agent messaging and read-only observation of active sessions (#207 by @sethkarten).
• Added an orchestration heartbeat skill for compact multi-session progress, blocker, and action summaries (#207 by @sethkarten).
• Added an opt-in auto-refine review hook that can ask whether /refine should run after turn intervals or compaction checkpoints (#201 by @sethkarten).
• Added opt-in fullscreen mode with a scrollable transcript, pinned prompt bar, mouse selection, and /fullscreen controls (#316).
• Added prompt stashing so a draft can be temporarily saved, a separate prompt or command can run, and the draft is restored afterward (#321).
• Added resume support to the agents view so stored sessions can be attached and managed without leaving the view (#318).
• Added subagent delegation guidance to encourage parallel and background rlm calls when recursion is available (#306 by @alexzhang13).
• Changed fullscreen TUI rendering to be enabled by default (#325).
• Changed --resume to accept an optional session path or ID (#319).
• Changed the installer onboarding splash to show ordered setup phases with a percentage instead of cycling detail text (#327, ENG-4376).
• Changed provider stream failures to show classified diagnostics and request IDs, with structured agent logs for debugging (#313).
• Fixed daemon-hosted extensions sharing the wrong Herdr pane environment across concurrent sessions (#303).
• Fixed parallel subagent guidance failing on first use by pre-importing asyncio in the IPython kernel bootstrap (#315).

[0.2.4] - 2026-07-01 #

• Changed the agents view to list only sessions the daemon is actively holding, and stopped the daemon from auto-restoring on-disk sessions on startup, so a restarted daemon no longer surfaces a wall of weeks-old sessions; sessions come back via /resume or --resume (#295).
• Changed the kernel install progress line to name the current step and show a percentage instead of a static message (#293).
• Changed the CLI to honor a -- end-of-options separator, so arguments after it are passed through instead of parsed as flags (#296).
• Changed provider stream failures to retry transient errors (content filter trips and prose 5xx responses) instead of failing the turn (#297).
• Fixed IPython and bash tool calls failing for the rest of a run after a session was rebuilt, by rebinding built-in tools to the live runtime at call time (#299).
• Fixed the kernel venv not rebuilding when the bundled runtime source changed, by tracking a content hash of the runtime (including its pyproject.toml) in the staleness check (#291).
• Fixed a large subagent fan-out spawning every IPython kernel at once and starving the machine, by bounding concurrent kernel boots (default min(16, 2*cores), override with PRIME_AGENT_MAX_CONCURRENT_KERNEL_BOOTS) (#294).
• Added a Python forkserver (on by default on Linux, opt out with PRIME_AGENT_KERNEL_FORKSERVER=0) that forks subagent kernels from one pre-imported template process instead of a full cold boot each time, with automatic fallback to direct spawn on any failure (#298, #300).
• Fixed empty tool results on OpenAI-style providers being sent as a literal "(see attached image)" placeholder, which made models hallucinate a nonexistent image (#290).

[0.2.3] - 2026-06-30 #

• Added built-in Linear and Notion integrations that the agent drives from Python in the kernel (no new agent tools); each is a bundled skill that talks to the service's official MCP server and auto-discovers its tools. They ship disabled and turn on after you sign in via the Services tab in /login or /mcp login, with credentials stored in the existing auth.json (#280).
• Added an attach-image skill that loads an on-disk image (PNG, JPEG, GIF, WebP) into the model's context as a viewable attachment so a vision-capable model can directly see screenshots, diagrams, charts, or scanned pages (#274).
• Changed subagents to be first-class sessions: opening a subagent now attaches to its own session and renders through the same rich chat UI as the main conversation instead of a laggy parent-rebuilt transcript, finished subagents stay viewable in the list and sort below running ones, and the detail view shows the subagent's own recap and animated working status (#282).
• Changed session lifecycle handling so the agents view now lists every live session (not only daemon-resident ones), fixing reports of sessions going missing; abandoned new chats that were never sent a message are discarded instead of lingering (#269).
• Changed the IPython kernel to stay alive across compaction: variables, imports, and helpers the agent defined are no longer wiped, and the model is instead told which names remain defined (#267).
• Changed local slash commands like /context, /system-prompt, /logs, /changelog, and /hotkeys to echo the typed command into the chat so their output is anchored to a visible command instead of floating (#270).
• Changed session recaps to use a non-reasoning model (Qwen3-30B instruct), which reliably closes the recap tag instead of occasionally surfacing a dangling "..." (#284).
• Changed the heartbeat scheduler to defer /heartbeat and internal heartbeat cron jobs while the target session is already working, rescheduling the next interval instead of piling a prompt onto a busy agent (#265).
• Changed Ctrl+O on IPython and bash cells to keep the same summary line in place and just attach the full code and output beneath it (aligned under the code gutter), instead of restructuring the block on expand (#288).
• Removed the "call at most one built-in tool per turn" instruction from the system prompt, allowing the agent to invoke multiple built-in tools in a single turn (#210).
• Fixed historical session replay re-emitting inline terminal image escape payloads; history now shows lightweight image fallback labels while live tool results still render images inline (#281).
• Fixed pressing back from a subagent opened directly from the agents view dropping you into the parent's chat; it now returns to the agents view, with a "back to agents" hint (#271).
• Fixed the agents view resetting the highlight to the first row when returning to it; selection now sticks to the session you had open across reorders and reattaches (#268).
• Fixed freshly created chats being titled by their session ID until their file flushed; they are now titled by their first prompt immediately (#264).
• Fixed opening a session from the agents view failing when its original working directory no longer exists; it now opens in a fallback directory with a notice instead of breaking (#287).

[0.2.2] - 2026-06-25 #

• Added a bundled websearch skill (Google search via the Serper API) that loads by default. Add a Serper key via /login ("Serper (web search)"); it is stored with your other credentials and supplied to the skill automatically. The skill can be disabled with bundledSkills.websearch: false and overridden by a same-named skill in any user, project, package, or --skill location (#86).
• Added image input support for vision-capable Prime Inference models (Claude, GPT-5.x, Grok, Kimi K2.7 Code, Qwen3-VL), which previously dropped attached images as unsupported (#261).
• Added a live subagent tree above the working loader showing each in-flight subagent with a prompt excerpt, tool-use and token counts, and its recap once generated; finished subagents drop out of the tree (#254).
• Changed the prompt bar to show the active model and thinking level on the left and always show context token count and percentage used on the right, instead of only surfacing context usage past the halfway point (#252).
• Changed the /model picker to rank results by most-recently-used, so models you actually pick float to the top and break ties among equally-good fuzzy matches (#251).
• Changed the collapsed bash and IPython tool previews to pick the most informative line via a shared heuristic, skipping low-signal setup lines and redacting long blobs and secret-looking values (#248).
• Changed subagents to render as an inline, scrollable list below the prompt with arrow-key navigation and prompts that elide shared prefixes, replacing the full-screen subagent viewer; running subagents and in-progress markers now animate so the agent never looks crashed (#247).
• Fixed context overflow appearing at ~50% remaining for Prime Inference Claude models by correcting their context window to 200k and counting prompt tokens only (excluding output) for the context indicator and compaction trigger (#246).

[0.2.1] - 2026-06-23 #

Fixed #

• Fixed daemon session recaps disappearing while a new turn regenerated them (#239).
• Fixed bundled built-in skills missing from the packaged release layouts (#240).

[0.2.0] - 2026-06-23 #

Added #

• Added /effort (alias /thinking) to set the reasoning level, with argument autocomplete that lists the levels the current model supports.
• Added a /system-prompt command that shows the exact prompt last sent to the model, labelling it honestly when no turn has run yet.
• Added a /rename alias for /name and a Ctrl+R shortcut in the Agents View to rename the selected session inline.
• Added support for feeding pasted images into model context: pasted images become atomic editor markers, are validated and resized, held in a bounded registry, and dropped when the active model lacks vision.
• Added edit diffs to the collapsed IPython view, rendering file edits as a wrapped, full-width relative-path diff prefixed with the cell status marker.

Changed #

• Replaced the Shift+Tab thinking-level cycle with the /effort command, exposing a max thinking level on Claude models that support it.
• Changed the /goal and /effort commands to stay highlighted in the editor while their argument is being typed.
• Changed queued follow-up messages to render below the execution indicator.
• Changed the RLM system prompt to align its shared sections exactly with rlm-harness, including the environment block and pre-installed package hints.
• Changed trace uploads to be observable and resilient: failures surface the underlying cause, outcomes are logged to agent-traces.log, /traces shows the resolved endpoint, and transient failures retry once.

Fixed #

• Fixed Prime Agent formatting breaking when resizing to a small screen, where tool-output colors bled into the padding at narrow widths.
• Fixed onboarding showing no models after entering a provider key by refreshing the scoped model list after login.
• Fixed silent daemon replacement reading as random crashes by logging shutdown/replacement decisions, and offering to stop a stale-version daemon at startup instead of erroring out.

Performance #

• Improved session load, context building, and listing to scale linearly: file loads decode per line over a raw buffer, and branch/context building uses push+reverse instead of per-entry unshift.
• Improved daemon responsiveness under large session loads by parsing session files off the event loop, so loading one big session no longer freezes the other sessions the daemon hosts.

[0.1.9] - 2026-06-22 #

Added #

• Added the Prime brand splash to the new-chat view.

Changed #

• Changed daemon attach to send slimmer snapshots and to avoid saved-session disk scans in the Agents View, speeding up switching between sessions.

Fixed #

• Fixed daemon out-of-memory crashes when listing saved sessions by streaming the listing, preserving large session row metadata, and ignoring oversized tool rows when computing session activity.

[0.1.8] - 2026-06-21 #

Added #

• Added daemon shutdown --all to stop every Prime Agent daemon on the machine, hardened against recycled PIDs and able to force-kill wedged daemons.
• Added git context to session traces: each trace records the repository URL, branch ref, and HEAD commit, captured at end of turn and carried over when a session is forked.

Changed #

• Changed prime-agent to open a new chat by default at launch instead of the previous session, with the daemon session created lazily on the first message and empty chats discarded on quit.
• Changed sending a message from the Agents View to open the chat for that session.
• Changed model resolution to persist the selected model across updates and default Prime Inference to Claude Opus 4.8 when no model has been chosen.

Fixed #

• Fixed prime-agent attaching to a stale daemon left running by a previous version after self-update: update now stops the old daemon and starts the new version (confirming first when busy sessions would lose work), and a stale daemon that cannot be replaced fails loudly instead of a silent broken attach. Both shutdown paths now poll the socket until it stops listening, so a transient hiccup cannot spawn a duplicate daemon.

[0.1.7] - 2026-06-18 #

Added #

• Added session and RLM heartbeats: a persistent, user-controlled heartbeat re-prompts a long-running session on a schedule via daemon-backed cron jobs, exposed through the heartbeat slash command and a bundled rlm-heartbeat Python skill, plus a cron CLI command to list jobs.

Changed #

• Changed collapsed IPython tool calls in the TUI to render as a single-line summary instead of a multi-line block.

[0.1.6] - 2026-06-17 #

Added #

• Added a daemon ps CLI command that lists every Prime Agent daemon running on the machine, with confirmation before shutdown and guards against killing a shared or still-reachable daemon.
• Added opt-in trace uploads: /traces enables background upload of persisted session JSONL files to the Prime Inference trace endpoint.
• Added agent summaries and live status to Agents View, generated daemon-side per session and refreshed on sweep.
• Added crash-stack capture for the daemon: output routes to a rotating per-socket log file under <agentDir>/logs/, client-side crashes write to client-errors.log, and a /logs command shows the log directory.

Changed #

• Changed startup notices (app-update, extension-update, and tmux warnings) to surface on the Agents View instead of being appended to every chat session.

Fixed #

• Fixed IPython kernel state being lost across session resume: kernel variables are now snapshotted under session-artifacts, restored on resume, deleted with the session, and dropped on compaction.
• Fixed the viewport jumping when toggling tool-output expansion in the TUI; the viewport now stays anchored across expand/collapse.
• Fixed non-persisted (e.g. /tmp) sessions creating an RLM working directory they did not need.

[0.1.5] - 2026-06-16 #

Added #

• Added rich syntax-aware diff rendering for IPython file edits in the TUI: the edit Python skill emits structured edit results that the interactive view renders as a colored, full-width unified diff inside the cell.
• Added a subagent spawn-program panel to Agents View: expand a subagent group and press Ctrl+O to toggle a panel showing the IPython cell that called rlm.run to spawn them.
• Added slash-command alias resolution so command aliases resolve to their canonical command consistently across interactive mode and Agents View, including in autocomplete.

Changed #

• Moved goals out of the harness tool surface into a bundled goal Python skill (goal.get / goal.create / goal.complete) backed by session state; the only built-in tool is now ipython, and the rlm.run comm channel is generalized into a typed host bridge.
• Changed the RLM system prompt to prefer Python for reading and searching files, porting the IPython guidance from rlm-harness.
• Spaced out the Agents View shortcut hints for readability.

Fixed #

• Fixed slow opening of long agent sessions: the JSONL socket reader is now O(n) instead of O(n^2) on large records, the session tree is fetched lazily instead of embedded in the attach snapshot, SessionManager.open() no longer parses the session file twice, and context building avoids copying every entry on the hot path.
• Fixed open() to stay consistent with the full loader when a session file begins with a blank line.
• Fixed goal-completion usage accounting that could overcount tokens.

[0.1.4] - 2026-06-15 #

Added #

• Added a /refine command and a session-backed rlm.harness continual-learning state (prompt notes, memory, reusable skills, and subagent specs) that persists globally across sessions, with explicit CRUD methods, a refinement log, and global rollback. The compact harness overview is injected into the system prompt, and /refine re-reads state before applying so concurrent writes are not clobbered.
• Added an edit built-in Python RLM skill for targeted single-occurrence string replacement in existing files, callable from the kernel or as a shell command.

Changed #

• Changed the IPython control prompt to require %%bash as the first line of a shell cell to match the rlm-harness.

[0.1.3] - 2026-06-12 #

Added #

• Added a /context command showing a tree overview of the main agent and all sub-agents with per-agent tokens, cost, and context-window usage, plus session totals and a token/cost breakdown.
• Added /clear as an alias for /new.

Changed #

• Changed /usage to be an alias for the new /context command.

Fixed #

• Fixed the stale "no models available" warning appearing for sessions that already have a working model.
• Fixed the ! and !! bash shortcuts in interactive mode by running bash through the agent connection, restoring streaming output, history, and Ctrl+C abort for both in-process and daemon-attached clients.

[0.1.2] - 2026-06-12 #

Fixed #

• Fixed the model selector showing no models after logging in with Prime Inference during onboarding by reloading auth storage from disk when the model registry refreshes (#151).

[0.1.1] - 2026-06-11 #

Fixed #

• Fixed first launch to run onboarding before opening the Agents View (#147).
• Fixed multiline status errors in Agents View to render as a single flattened line so they cannot overlap the input (#146).
• Fixed slash commands in the main Agents View (#149).

[0.1.0] - 2026-06-11 #

Breaking Changes #

• Changed InteractiveMode construction to require an AgentConnection and explicit UI services or local session host.

Added #

• Added a two-step Ctrl+X stop/delete interaction for selected agents in Agents View.
• Added a daemon-backed Agents View as the default local interactive entrypoint.
• Added versioned daemon protocol metadata, sequenced session events, attach snapshots, replay status, and artifact references for future Swarm gateway wrapping.
• Added an AgentConnection client boundary with in-process and daemon adapters for interactive-mode decoupling.
• Added daemon mode and CLI controls for starting on demand, creating, listing, attaching, detaching, killing, renaming, and prompting live sessions.
• Added rich TUI attach for already-active daemon sessions via --session <selector> and live daemon <selector> shorthand.
• Added a built-in skill-creator skill that teaches the agent to create new skills: markdown layout, frontmatter rules, placement and precedence, and the Python-backed skill contract (package layout, run() convention, optional CLI, kernel venv behavior) with a test-verified working template.
• Added built-in skills shipped with prime-agent, starting with prime-intellect: ecosystem knowledge and prime CLI workflows for verifiers environments, evaluations, Hosted Training, sandboxes, inference, and compute. Built-in skills have the lowest precedence (user, project, and package skills with the same name win) and can be disabled with the enableBuiltinSkills setting or --no-skills.
• Added a session-backed rlm.harness state helper for reset-free prompt notes, memory, skills, subagent specs, and refinement events.
• Added /refine to update editable harness state with Create/Update/Delete edits and rollback support based on refinement history.

Changed #

• Changed Agents View Ctrl+C handling to mirror the interactive chat view: the first press shows a bottom hint and the second exits Prime Agent.
• Changed keybinding hints to render arrow keys as ↑, ↓, ←, and →.
• Changed Agents View to keep transient status and reply text out of the agent list area.
• Changed Agents View Ctrl+X so the first press only stops sessions that are actively running.
• Changed daemon-owned chat sessions opened from Agents View to show a ← agents tray hint when the input is empty.
• Changed active session creation to use per-session runtime config so active sessions can use different cwd, model, auth, and tool settings.
• Changed interactive Ctrl+C to interrupt the current operation first and exit only on a second press while the exit hint is visible; Escape now clears the input bar without interrupting the agent.
• Changed the IPython system prompt section to use the upstream rlm-harness IPYTHON_CONTROL_PROMPT: IPython is framed as a persistent control environment, not the target project's runtime. Shell commands should use %%bash cells instead of !cmd escapes. The agent should not install dependencies into the IPython kernel but use the project's own environment instead.
• Removed the .venv interpreter hint from the system prompt (no longer needed with the control-environment framing).

Fixed #

• Fixed confusing transcript formatting around thinking blocks and tool calls: ipython cells and default-shell tools (bash and extension tools) now share one panel style with a subtle neutral background instead of a status-colored box or a left rail, and tool status headers name the tool (python · done · 7ms, bash · running) so they no longer read as floating labels for the preceding thinking block. Themes gain a required toolPanelBg color for the panel background.
• Fixed prime-agent to detect a daemon left running by a previous version after self-update: the daemon now reports its app version on connect, and idle stale daemons are restarted automatically (daemons with active sessions are left running with a warning).
• Fixed Agents View listing daemon-owned subagents as top-level selectable agents instead of nested child rows.
• Fixed Agents View opening saved or stale sessions by creating a daemon runtime from the saved session file before attaching.
• Fixed Agents View delete confirmation so the red stopped confirmation expires after two seconds without removing the stopped session row.
• Fixed Agents View selected-row highlighting so it spans the full terminal width after prompt wrapping changes the layout.
• Fixed Agents View prompt bar to show a placeholder for creating a new session.
• Fixed Agents View opening sessions with the dashboard cwd's model registry, which could incorrectly show the model selector for daemon-owned sessions from another cwd.
• Stopped showing changelog entries automatically on install, first launch, and update startup.
• Fixed multi-line IPython, assistant, and child-agent errors to collapse internal tracebacks by default while preserving full details on expand.
• Fixed child-agent navigation to show contextual keybinding hints and a visible focused tray marker.
• Fixed the release installer to ask before bootstrapping the IPython kernel runtime during install, avoiding default first-run uv prompts inside the TUI.
• Fixed browser sign-in links to show plain URLs when terminal hyperlinks are unsupported.
• Fixed the release installer splash to keep its logo geometry stable across terminal resizes.

Removed #

• Removed the interactive ! / !! bash shortcuts; use IPython for shell commands.

[0.0.10] - 2026-06-08 #

Added #

• Added an inline input prompt indicator to the interactive editor.
• Added contextual keybinding hints and a visible focused tray marker for child-agent navigation.
• Added OS-specific shortcut labels in keybinding hints, rendering Cmd/Option on macOS and capitalized key names elsewhere.

Changed #

• Changed the IPython system prompt to the upstream rlm-harness IPYTHON_CONTROL_PROMPT: IPython is framed as a persistent control environment rather than the target project's runtime, shell commands use %%bash cells instead of !cmd, and project imports, tests, and dependency checks run through the project's own environment. Removed the .venv interpreter hint.
• Changed interactive Ctrl+C to interrupt the current operation first and exit only on a second press while the exit hint is visible; Escape now clears the input bar without interrupting the agent.
• Changed the Prime theme to tone down the flashy neon purple and lime green in favor of a calmer dusty lavender and sage green.

Fixed #

• Fixed missing ripgrep to surface a clean inline warning at startup and fail sub-agent runs with a clear message, while routing kernel diagnostics into captured stderr.
• Fixed IPython kernel startup to avoid blocking the session, cancelling child RLM runs on session abort and reporting bootstrap progress through a start-options handler.
• Fixed the subagent tool-expansion keybinding so it toggles expanded tool output inside the child-agent detail view.
• Fixed browser sign-in links to show plain URLs when the terminal does not support hyperlinks.
• Fixed the auth selector to preserve the selected provider's login type.
• Stopped showing changelog entries automatically on install, first launch, and update startup.

[0.0.9] - 2026-06-04 #

[0.0.8] - 2026-06-04 #

Added #

• Added an onboardingCompleted setting and a dedicated Prime Inference onboarding splash that prompts users authenticated only via the Prime CLI to choose a model before their first turn.

Changed #

• Changed the system prompt to frame the agent as a general-purpose agent that uses code to solve tasks rather than a pure coding agent, with guidance that shell state does not persist across !cmd/%%bash cells while Python kernel state does.

Fixed #

• Fixed the onboarding flow so model selection, manual API-key entry, cancellation, and the "model already ready" path all resolve correctly and mark onboarding complete instead of re-prompting on every launch.
• Fixed the release installer to ask before bootstrapping the IPython kernel runtime during install (defaulting to bootstrap when no terminal is detected) and to avoid stalling on an interactive uv prompt.

[0.0.7] - 2026-06-01 #

Added #

• Added Prime team selection during Prime Inference login so team inference costs use the selected Prime CLI context.
• Added Python-backed skills that install into the persistent IPython kernel and are exposed alongside markdown skills.

Changed #

• Changed the Prime Agent install script to use a bounded animated Prime Lab splash with centered progress and confirmation prompts.
• Changed startup onboarding to guide unauthenticated users through login and model selection before the first agent turn.
• Changed installer npm and Node.js setup progress to keep command output hidden behind the splash and rotate detail text.

Fixed #

• Fixed update notifications and package docs to point at prime-agent update and use compact one-line alerts.
• Fixed Prime CLI credentials from prime login to make Prime Inference models available on startup.
• Fixed first-run search helper downloads to run quietly instead of printing over onboarding.
• Fixed stale no-model and tmux/update startup notices from appearing during successful onboarding.

Removed #

• Removed the unused small Prime logo export.

[0.0.6] - 2026-05-27 #

Changed #

• Changed installer startup so npm and Node.js setup output stays hidden behind the bounded Prime splash with rotating detail text.
• Changed postinstall to optionally bootstrap the fd and rg search helpers (gated by an env flag) alongside the kernel, and made search-helper downloads default to silent.

Fixed #

• Fixed Prime Inference auth so credentials from prime login are read from the Prime CLI config and surfaced as a prime_cli auth source, making Prime Inference models available on startup without a separate login.
• Fixed initial model selection to skip a saved default model that no longer has configured auth.
• Fixed first-run search-helper downloads to run quietly instead of printing over onboarding.

[0.0.5] - 2026-05-26 #

Added #

• Added a centered-overlay menu system for onboarding and a redesigned Prime onboarding splash and Prime Inference login dialog with browser sign-in plus a manual API-key fallback.
• Added theme support for adapting interactive surfaces to the detected terminal foreground/background colors.

Changed #

• Changed startup onboarding to guide unauthenticated users through login and model selection before the first agent turn.
• Changed the model selector and OAuth/provider selectors to render as centered surface menus rather than inline CLI lists.
• Changed update and package-update notifications to compact one-line alerts pointing at prime-agent update.

[0.0.4] - 2026-05-21 #

Added #

• Added system prompt note listing pre-installed Python packages (requests, httpx, pyyaml, tomli, python-dotenv, pandas, numpy, scipy, beautifulsoup4, lxml, pydantic).
• Added DEFAULT_RLM_EXTRA_UV_ARGS constant and kernel bootstrap installation of those packages; updated prompt to reference the constant instead of a hardcoded list.

Fixed #

• Fixed the RLM kernel package prompt to show importable module names and reject PRIME_AGENT_KERNEL_PYTHON overrides missing default kernel packages.

[0.0.2] - 2026-05-20 #

Added #

• Added a persistent ipython tool backed by a Jupyter kernel so Python variables and imports survive across tool calls.
• Added the RLM harness system prompt and prime-agent-runtime bridge so IPython code can call rlm.run to spawn recursive child agent sessions.
• Added automatic IPython runtime bootstrap with uv-managed Python, ipykernel, and prime-agent-runtime.
• Added subagent UI surfaces for recursive runs, including compact tray status, full-width detail views, and structured child transcripts rendered like the main chat.
• Added /goal for long-running objectives that continue after normal follow-ups drain until the model marks the goal complete.
• Added a pi-style installer script and R2-backed private npm tarball release pipeline for Prime Agent.
• Added Prime Inference as a selectable built-in OpenAI-compatible provider with PRIME_API_KEY authentication and openai/gpt-5.5 as the default model.
• Added a first-class /login Prime Inference browser auth flow that imports usable Prime CLI credentials or obtains a new key through the Prime challenge flow.
• Added /usage to show token, cost, and context usage on demand.

Changed #

• Changed the default active built-in tool set to ipython.
• Changed compaction to restart the active IPython kernel so summarized sessions release in-memory Python state.
• Changed recursive background work to use normal Python async tasks with rlm.run instead of a separate RLM background API.
• Changed completed IPython cell rendering to use width/version-aware caching, reducing TUI redraw lag in long sessions.
• Changed collapsed IPython cells to show compact input and output previews with a single expansion hint.
• Changed auto-compaction checks to use the current context estimate and stop between long tool-loop turns before resuming after compaction.
• Changed the goal status UI to use a compact lower-tray indicator instead of repeating the full objective in chat.
• Changed IPython prompt guidance to prefer !cmd and %%bash for shell commands.
• Changed kernel bootstrap to prompt before installing uv and skip postinstall bootstrap unless explicitly enabled.
• Changed the app update check and self-update flow to read the Prime Agent release manifest and install manifest tarballs directly.

Fixed #

• Fixed tarball self-updates to install the tarball without first uninstalling the same logical package.
• Fixed IPython kernel startup to let ipykernel bind OS-assigned ports instead of randomly selecting fixed ports.
• Fixed RLM child usage aggregation so parent session totals include recursive child runs after session reloads.
• Fixed the RLM child-agent detail viewer to render messages, thinking, and tool output with the main chat presentation, open at the latest transcript output, and use terminal scrollback for native scrolling.
• Fixed rlm.run comm handlers to log failures and drain in-flight child runs during kernel disposal.
• Fixed raw tab rendering in TUI-backed transcript views so painted backgrounds survive indentation.
• Fixed auto-compaction threshold checks during trailing context and tool-result growth.

Removed #

• Removed install/update telemetry pings to pi.dev and the related setting and environment override.
• Removed the RLM background API; recursive agents now use rlm()/rlm.run() with normal Python async tasks for background work.
• Removed the legacy read, write, grep, find, and ls built-in tools.
• Removed the local TPS extension that posted token/cache stats after each agent response.

[0.0.1] - 2026-05-18 #

Added #

• Initial Prime Agent release, forked from pi-mono: a persistent ipython tool backed by a Jupyter kernel as the default tool set, recursive RLM subagents via rlm.run, /goal for long-running objectives, an auto-bootstrapped uv-managed kernel runtime, Prime-branded TUI, and an R2-backed tarball release pipeline with a pi-style installer.