Skip to content

A self-improving RLM agent for coding workflows and long-running autonomous tasks.

Rust
29.1 KB
Raw Blame

Changelog

Release notes for Prime Agent. Historical TypeScript release notes are recorded in the TypeScript changelog.

[0.10.2] - 2026-10-28

  • The static usage layer no longer tells agents to await commands before ending the turn: since #3580 daemonless sessions also wake on a queued [bash-done] turn, every session kind has completion notifications, and the no-polling rule (end the turn and collect the result on the done notification) is the one turn-end contract.
  • Agents view: leaving a session now returns the focus to that session's row, even after /switch, /new, /fork, or /resume in the chat, and when the row loads late.
  • tool_execution_end events on the RPC stream and in --mode json print output carry isError again, as the TypeScript version emitted them.
  • Azure OpenAI Responses requests no longer fail with 404 on the v1 API: they now post to <resource>/openai/v1/responses?api-version=v1 with the deployment as the request's model, instead of a /deployments/<name>/responses route Azure does not serve. A different AZURE_OPENAI_API_VERSION (for example preview, or a dated version against a custom /openai base URL) is sent as the api-version query instead, and an empty one falls back to v1.
  • A ! command or a model bash tool call whose shell dies from a signal (for example a crash, an OOM kill or kill) now reports 128 + the signal number as its exit code, so the model is told it failed, like the TypeScript version.
  • ACP: a failed session/prompt now carries kind, category and retryable in its error data next to details, so a host can tell an exhausted model connection (kind: "model_transport", the only retryable kind) from a context overflow (context_limit), an authentication or rate-limit error, or a harness failure.
  • The bundled websearch skill is now offered only when a Serper API key is set up (saved through /mcp, or SERPER_API_KEY), instead of being advertised on installs where every call fails; bundledSkills.websearch in settings still turns it on or off explicitly. A key saved in /mcp activates the skill in new sessions (/new).
  • A python_repl cell that runs longer than the new pythonCellTimeoutMs setting (default 60 seconds, 0 disables) now gets a note on stderr that points long shell work to bash() and long Python work to smaller cells. The cell is not stopped and keeps its result. Sessions without the daemon add no note.
  • When a running subagent is cancelled, deleted with rlm.delete_subagent, or stopped because its parent session is killed, the parent's semantic-edge ledger now records a child_cancelled line with the reason (user_cancel, parent_delete or close:killed) and no child_returned line beside it. A daemon restart or update records no such line.
  • A PageUp or wheel scroll that reaches the top of the loaded transcript while older messages are still loading is no longer lost: the view scrolls once the older page arrives.
  • @ file completion is now instant: the project's files are indexed in the background when the TUI starts, suggestions come straight from that index, and the menu opens in the same frame as the keystroke. Large file changes (such as moving a big directory) no longer block typing. Which files match, and their order, is unchanged.
  • Requests to Gemini models through the Google and Vertex AI providers no longer fail with 400 INVALID_ARGUMENT: the system prompt is now sent in the content format Google's API requires instead of as a plain string.
  • MCP OAuth login no longer fails when a server's advertised authorization server and its metadata issuer differ only by a trailing slash (for example TickTick).
  • Rows written around a compaction (a queued prompt, a rename or refinement notice) no longer show twice in the transcript. A steer or follow-up queued while you navigate the session tree or fork now waits until that finishes and runs on the branch you end up on, as the TypeScript version did.
  • Restoring a kernel snapshot no longer reopens saved file handles: a leftover handle opened for writing (even a closed one) used to empty its file on every restart. Restored file objects now come back closed, keeping their name and mode; open the file again to use it.
  • On openai-completions and Mistral providers, a turn whose tool calls were streamed without a function name no longer ends silently: the call gets a Tool unknown_tool not found error so the model can retry, and a nameless call next to named calls is dropped. Sessions that already stored such a call no longer fail every later request with 400 ... missing a function name.
  • A subagent whose final model call fails with a provider error (after retries) now settles as error with the provider's message in rlm.collect, rlm.list_subagents and the factory, instead of done with no answer. If it had not replied yet, the parent gets a failure notice.
  • Requests through openai-completions, openai-responses and Mistral no longer send two Authorization headers when the model or the caller sets one (for example a custom provider with authHeader: true), so strict servers stop rejecting them.
  • A cost a provider reports in usage.cost is now used for any OpenAI-compatible provider, not only OpenRouter, so gateways in front of OpenRouter such as LiteLLM no longer record $0.
  • prime-agent mcp add --bearer-token-env-var now writes the bearerTokenEnvVar settings key the MCP engine reads, so HTTP MCP servers added this way send their bearer token instead of connecting without auth. Settings already written with bearer_token_env_var keep working.
  • Claude Haiku 5.x models on Anthropic and Bedrock now use adaptive thinking with an effort level instead of budget-based thinking, which the API rejects with a 400; on Bedrock, turning thinking off now disables it explicitly. Haiku 4.5 keeps budget-based thinking.
  • Prompt caching on openai-codex works again: every request in a session carries the same session id, which now stays stable across resumes and daemon restarts, so Codex WebSocket reuse and cached input work like the TypeScript version. Other providers also send the session id where the TypeScript version did (for example prompt_cache_key on OpenAI and Azure, and Mistral's affinity header), and side questions reuse the conversation's cache.
  • Bedrock works again: profiles that use credential_process now get credentials, requests are signed the way AWS expects instead of being rejected, and streamed replies are no longer empty. AWS_CONFIG_FILE and AWS_SHARED_CREDENTIALS_FILE are honored for credentials and region, and Claude 5 and newer models on Bedrock get prompt caching.
  • Fresh installs now ship the live model and MCP catalog, instead of a test snapshot that stayed listed until the first catalog refresh.
  • Gemini 3 and newer models on the Google and Vertex AI providers (for example gemini-4-argon and gemini-flash-latest) now send thinkingLevel instead of thinkingBudget, so the thinking levels actually differ, and thinking off sends the model's lowest level instead of a zero budget some models reject. Gemini 2.x keeps thinkingBudget.
  • On Anthropic-compatible endpoints that return thinking without a signature (such as MiMo, Kimi and DeepSeek), earlier thinking is now passed back as thinking instead of as visible text, so the model no longer echoes its reasoning more each turn. Claude models drop unsigned thinking, which they reject; the new requiresThinkingSignature compat flag overrides the per-model default.
  • A session with a large Python namespace no longer overflows its context after repeated compactions: each [python-state] notice lists at most 200 names or 4 KB (ending with +N more), and only the newest notice is kept in the model's context.
  • OpenCode and OpenCode Go requests carry the x-opencode-session header with the conversation id again, including compaction, refinement and branch-summary calls. Those three calls now send the session id to every provider for prompt caching, as the TypeScript version did, but always over SSE instead of the Codex WebSocket.
  • Prime Inference models dropped from the model catalog because the server sent no specs for them are now named in a log warning.
  • Kernel: rlm.factory.run(...), rlm.factory.help(), and every other rlm namespace call keep working after a cell runs import rlm, and dir(rlm.harness) now lists the harness methods.
  • Factory machines can loop back into an entry state: an entry state may now declare inputs as long as all of them are optional. Its first entry binds the null sentinel, and a later entry through a transition binds the real output, so an investigate/verify loop no longer needs a duplicate re-investigate state. A required input on an entry state is still rejected, with a message saying it must be optional.
  • A parent session's usage totals now include a grandchild's spend that lands after the parent already read the child's spawning turn. Before, that spend was dropped from the parent's totals, so long-running grandchildren were undercounted.
  • Reopening a saved session now publishes its passive subagents before the first Agents view snapshot, even when worker registration is delayed or the opening client disconnects.
  • A subagent spawned while its parent has no turn running (for example a factory run resumed between turns) now gets its task prompt at once instead of up to 60 s later. A parent turn aborted by a manual compaction or tree move no longer leaves later spawns waiting for that bound either.
  • /refine and automatic refinement work again with Prime Inference team billing: they sent no team header and failed with "Refiner did not return a JSON object".

[0.10.1] - 2026-10-21

  • A corrupted line in the RLM spawn ledger is now skipped and logged instead of failing every ledger read, so session listings and new subagent spawns keep working.
  • Idle passivation now defers while a command to the session is still in flight (for example adding a heartbeat or cron job), so such a job no longer stays dormant until the session is resumed.
  • On macOS the daemon no longer spawns a ps process for every client connection and every process-liveness check; start times and zombie states are read in-process from the kernel's process record, cutting idle battery drain.
  • The chat title and the agents view now include the spend of deleted subagents, and both show the same number.
  • Large sessions reopen faster: the usage scan is saved when a session closes and reused on the next open.
  • Removed the -t/--tools, -nt/--no-tools and -nbt/--no-builtin-tools flags. They were silently ignored; they now fail as unknown options.
  • /update in the TUI no longer hangs on an installer prompt or swallows keystrokes: the installer now runs detached from the terminal.
  • The daemon no longer wakes up every second for each session it keeps running across a restart or update: it now waits on the operating system's process-exit notification (kqueue on macOS, pidfd on Linux), which cuts idle battery drain. On macOS this also stops a ps process being spawned every second for each such session.
  • ACP sessions and TUI --no-session sessions are now owned by the client that created them: they are never idle-passivated, and their agents-view row is removed when they stop instead of lingering as a passive row that cannot be resumed.
  • Messaging a subagent whose worker was idle-passivated now wakes it and delivers the message, instead of failing with "Unknown active session".
  • Nightly updates download the nightly installer directly instead of going through the stable installer first.
  • A malformed escape sequence containing a non-ASCII character (for example, Option-key input mixed with Cyrillic or accented text) is now passed through as typed text instead of crashing the terminal UI.
  • A session file whose header id is not a plain file name (for example ../x or an absolute path) is now treated as an invalid session, so its artifacts can no longer be written outside the session-artifacts folder.
  • prime-agent list and other daemon commands no longer hang until they time out when a busy daemon sends a large response slowly.
  • A subagent that fails because its task cannot be delivered, or whose worker becomes unreachable, now sends a failure notice with the error to its parent, and a goal waiting on it resumes.
  • Cancelling a running prompt by its admission id now reports it as owned instead of cancelled, and with cancelOwned the cancel now stops the turn.
  • On Linux the daemon no longer spawns a ps process each time it checks the identity of a process that has already exited (stale session leases, session reuse, update checks).
  • Agent runs no longer stall the session while recording git state: the git checks at the start and end of each run now run in parallel in the background instead of blocking other work on the session.
  • Long lines inside fenced code blocks now wrap onto the next row instead of being cut off at the edge of the terminal.
  • An ipython cell that mostly runs shell commands now shows as bash in its collapsed card, with its first bash() command and how many more it ran, including commands run through helper functions.
  • A prompt stashed with ctrl+s now returns to the empty prompt box after you send another prompt, as it did in the TypeScript version.
  • ACP sessions run one prompt turn at a time: a prompt sent while a turn runs, or while a cancel is stopping it, is refused instead of overlapping it; a cancel now stops only its own turn, and a failed turn no longer reports itself as finished.
  • The installer's --force no longer stops the daemon the install runs under, waits for the normal drain before signalling a stuck daemon, and no longer leaves a temp directory behind when it exits early.
  • After a daemon restart or update, a parent session lists its subagents again, and messaging one wakes it as the parent's subagent, instead of rlm.list_subagents() coming back empty.
  • New prime-agent factory list | import | export commands manage the machine library of shareable MACHINE.md factory templates; the bundled seeds (builder, pr-manager, review-sweep) ship inside the runtime so every install sees them, personal machines live under the agent dir, and an invalid machine never persists.
  • await rlm.factory.run('<name>') runs a library machine directly without creating a harness entry; a machine that exists but is broken names its exact errors, and an existing export target is refused instead of silently overwritten.
  • Each model request now carries a request id header, and each session writes a semantic-edge ledger (request, subagent and compaction events) matching the TS agent.
  • ACP mode over the daemon now rejects an initialize request with invalid params (for example a string protocolVersion), as the TypeScript version does.
  • --mode acp no longer falls back to a separate in-process engine when the background service cannot start: it now requires the daemon, and a startup failure prints Error: and exits 1, like the TypeScript version. A session/cancel sent as a request instead of a notification now answers "method not found", like the TypeScript version.
  • ACP mode now reports heartbeat schedule changes and shell commands that another client runs in the session (as tool calls), and leaves out empty tool output, like the TypeScript version.
  • ACP session/new now clears the connection's MCP servers when admitting them fails or times out, so servers the worker applied before a lost acknowledgement no longer stay admitted, matching the TypeScript behavior.
  • ACP prompts now finish only once the subagents they started have finished: the completion update reports how many were still running, the final update reports zero, and cancelling or closing the session cancels them. Closing stdin no longer waits for running work to finish.
  • ACP mode over the daemon now honors the session flags: without --no-session the conversation is saved like any other session and --resume, --continue, --fork and --session-dir open or place it; only --no-session sessions belong to the editor's connection and end when it disconnects, and session/close stops the running work without ending the session. Startup failures (an unknown or busy session) now exit with an error instead of starting, and DO_NOT_TRACK / the telemetry setting now also disable the daemon session's telemetry.
  • ACP prompts and other long daemon requests (waiting for a turn, idle or headless completion, compaction, tree navigation, session start) now wait up to 24 hours like the TypeScript daemon instead of failing with "Session worker timed out" after 10 minutes, and a session worker that dies mid-request now fails it right away instead of after its restart.
  • The daemon now always delivers a turn's events before the reply that ends the turn, and ACP mode no longer waits up to 30 seconds after some turns before answering.
  • Added the rlm.rename(new_name, session_id=...) kernel function: renames the calling session (no session_id) or a direct child selected by spawn handle, list_subagents() row, or session id — never by child name — reusing spawn-name validation and daemon-owned sibling uniqueness; the renamed session gets a Session renamed old->new`` transcript notice (suffixed by parent when a parent renamed it), and the kernel readiness check now requires a runtime with rlm.rename.
  • Agents view: the rename affordance (ctrl+r / /name) now also applies to subagent rows, which carry the same renameable session identity as top-level agent rows.
  • A request that a crashed session worker never answered now fails with "Daemon worker socket closed", the same error the TypeScript daemon reports.
  • Client-owned sessions (--no-session ACP and TUI sessions) now stop 30 seconds after their client disconnects without closing them, or 30 seconds after a daemon restart if the client does not reconnect, instead of running on hidden until the daemon stops.
  • Clicking a compaction summary, a harness refinement notice, a skill invocation, or an injected prompt (heartbeat, goal, subagent status) in the transcript expands or collapses it again.
  • openai-completions streams no longer hang forever when a server keeps the response body open past the data: [DONE] marker (a persistent SSE channel with no content length, chunked terminator, or half-close): the stream now ends at the marker like the TypeScript SDK, so the turn runs its tool call instead of waiting silently for an EOF that never comes.
  • A maxTokens configured in models.json (or a per-model override) now reaches the provider unchanged instead of being silently clamped to the 32000 default output ceiling; catalog models keep the default ceiling (#755).
  • The Windows one-liner is real again: the release pipeline publishes install.ps1 (the stable-channel render, the channel's own base + fallback to beta when stable carries no Windows build) on every beta cut instead of waiting for a first Rust stable that has not been cut, and the README's PowerShell command points at the copy that actually serves (the merged repo file, the same route the Windows battery installs from) instead of the domain prefix that 404s.
  • Session resume no longer wedges when the saved state held a bash() handle: snapshots skip live process handles, whose restore reopened a stale fd number and killed the restored kernel's event loop. Restoring also refuses to reopen any pickled raw fd number (legacy snapshots, user pipe objects), failing just that saved name instead of the whole kernel. A kernel that drops a protocol frame now logs it to kernel-stderr.log instead of failing silently. A timed-out request whose code was still being sent to a stalled kernel no longer leaves the kernel permanently busy; the kernel restarts instead.
  • Internal: daemon log, panic, and test assertion messages no longer print session ids or fixture API keys (CodeQL rust/cleartext-logging cleanup).
  • ACP mode now resolves a prompt turn whose final model response stopped at the provider's output-token limit with stopReason: "max_tokens" instead of end_turn, so a truncated answer is no longer indistinguishable from a finished one; autonomous-run limits keep their existing stop reasons (the TypeScript version does not map the truncation case).
  • A spawned subagent's kickoff task now renders as an agent message from the parent (labeled [task from parent]) instead of an unlabeled user message, and spawn prompts are no longer expanded as slash commands or prompt templates, matching the TypeScript version.
  • prime-agent update on a Homebrew-managed install (Cellar formula or Caskroom cask) no longer attempts an installer self-update: it now refuses and points at brew upgrade prime-agent (or brew upgrade --cask prime-agent).
  • Model output streamed over HTTP no longer shows replacement characters in place of CJK characters or emoji when a provider's chunk boundary lands inside one: the stream holds the split bytes and decodes the character once it is complete.
  • The / and @ autocomplete dropdowns now open, and drag auto-scroll keeps scrolling, while a reply streams, instead of waiting for the stream to pause.
  • With many agents running, session workers send at most one roster update at a time and no longer look up model details while holding the session lock, which cuts background CPU during subagent fan-outs and keeps state queries responsive.
  • Listing agents (CLI list, resume preflight, the agent roster) and spawning a named session no longer stall behind one unresponsive session worker: resident summaries are read concurrently under a short per-worker bound that degrades an unreachable worker to its recovering row, and the saved-session and scheduled-jobs roster scans moved off the supervisor's runtime threads.
  • Session workers whose registration reaches an interrupted stop now retire and release their session lease instead of repeatedly registering and re-running the stop after a daemon restart.
  • Print and JSON runs now release their session lease after the runtime stops, including error exits and panic unwinding, so detached session writers finish or are cancelled before another run can reopen the session.
  • Daemon housekeeping now reclaims dead-owner session leases and obsolete cache files, retains recent socket logs, and removes retired worker recovery journals only after their writer is proven gone. Journals with unverifiable ownership and flat TypeScript update status records are preserved. Premature explicit lease releases on RPC disposal and orphan teardown have been removed.
  • Print-mode goals consume eligible empty-turn backoff wakeups and report the no-progress continuation cap rather than silently losing further progress.
  • Opening a session no longer scans the whole RLM spawn ledger and the whole session-artifacts tree: only the opened parent's own subagents are re-listed, and a create that makes a brand-new session skips the scan, so session opens stay fast as the ledger grows.
  • Keep a submitted prompt draft consumed when the daemon connection closes before its reply, preventing an accidental duplicate submission after reconnect. Inspired by Pi Durable and Pi Pocket's §05 “Exactly once” explanation.
  • Replacement session workers stay supervised after re-registration, and failed worker relaunches after an abandoned update resume crash recovery instead of leaving the session unsupervised.
  • Agents view: the selected agent's row now expands to show its latest rlm.progress_note (live child progress) while it runs.
  • Progress notes fit narrow panes and clear before an aborted run is published as idle.
  • HTTPS clients and Codex WebSockets now trust operating system certificate authorities alongside bundled roots, and honor custom CA bundles supplied through NODE_EXTRA_CA_CERTS, including the default proxy streaming path.
  • Daemon recovery journals now preserve intact records and safely remove an incomplete trailing record after a crash, including a cut in Unicode text, so subsequent records survive reopening. Healing syncs worker recovery data before replacement and stops on journal read failures instead of rewriting unreadable data.
  • A parent session's usage and cost totals include its subagents' usage again: child usage attribution rows are written to the parent's transcript file, so live totals count child cost and a resumed parent keeps it, matching the TypeScript version.
  • Attribution retries are idempotent: each row keeps the stable identity of its frozen report, and a re-delivered report appends nothing once that identity is already in the transcript, so a partial save or a lost cursor cannot double-count child usage.
  • Wrap complete sign-in URLs and login actions in narrow terminals, giving the onboarding login panel more space when needed.
  • Upgrading from the npm package no longer deletes the native prime-agent command when npm and the installer share a bin directory, even if the npm uninstall fails or is interrupted. The installer now fails when the installed command cannot answer --version, instead of reporting success with a broken command.
  • On macOS, socket paths are now checked against the native 103-byte limit, and a long TMPDIR (for example under Nix) no longer breaks the daemon: its sockets move to /tmp/prime-agent-<uid> when the default location would be too long.
  • Settings changes preserve unrelated and concurrently updated fields, and no longer overwrite a settings file that failed to parse. Adding the first MCP server also recovers from the null server map written by older settings saves.
  • mcp.config host requests now resolve integrations live: an MCP server connected or configured mid-session (e.g. through /mcp) is dispatchable from the kernel without restarting the session. Added a configurable auto-compaction percentage with five-point settings steps, retention-aware triggering, and preserved model output headroom.
  • Add an enabled-by-default, global-only Harness refinements toggle in /settings → Agents. Project settings cannot override it. Turning it off prevents refinement generation and loading saved harness rows into future model requests; existing saved state is preserved.
  • Disabled refine.run() requests report a refusal without scheduling work. Plans completed while disabled are discarded. Undelivered digests are deferred until a turn admitted with refinements enabled; toggling does not re-arm a consumed digest.
  • Earlier replies and summary prose may already reflect refinements, and sent requests cannot be recalled. Start a new disabled session for a conversation that has never used saved refinements.
  • Factory execution also requires harness access for stored specs and subagent references: factory.run() refuses while refinements are off, including library-machine lookup. It does not start children or change saved state.
  • On macOS, a long TMPDIR no longer breaks the daemon socket for installer and npm installs: when the default socket path would exceed the 103-byte limit, it moves to /tmp/prime-agent-rust-<uid>.
  • Factory resident states now reject retries, and the factory help plus the pr-manager seed's fixing prompt spell out the recovery posture a re-run needs when an earlier attempt already changed the working tree.
  • The kernel's model.info and compact.status context window now follow daemon model switches (set_model, cycle_model, and provider failover) instead of reporting the session-creation model. Slack is available in the MCP connector list with browser login, token refresh, and support for its registered desktop callback port.
  • Terminal worker stops on macOS (and Linux kernels without pidfd) now send SIGTERM and SIGKILL via kill(2), gated by the start-id identity check before each signal, instead of silently reporting the worker as survived.
  • Saved-session catalog rows no longer include allMessagesText; full session transcripts remain unchanged. Current Rust and Desktop clients do not use this field, but older TypeScript clients and older typed consumers that require it are incompatible with the new daemon.
  • Kernel startup no longer fails on a raw EACCES when the kernel venv's directory is not writable by the current user: a venv that still runs the runtime yields its python with the missing Python skills disabled behind the unavailable-skill notice, and a venv that cannot run the runtime errors naming the venv and the PRIME_AGENT_KERNEL_VENV / PRIME_AGENT_KERNEL_PYTHON overrides.
  • Agents view and /resume: the newest 200 sessions load first, scrolling loads more, and a search loads the rest. A footer shows how many older sessions are not loaded yet.
  • Remove automatic session archiving and the sessionArchiveMaxAgeDays and sessionArchiveMaxSessions settings. The agents view and /resume page through the live session store instead.
  • Keep lazy recovery of files archived by older Rust versions: a saved-session wake by ID, ID prefix or exact name restores a matching archived file when no live session matches. Existing live files are never overwritten; other archived files stay in place until addressed.
  • The agent prompt now sanctions ending a turn with a text-only reply and forbids placeholder no-op tool calls (such as a pass-only cell) used to acknowledge a message or signal idleness.