Track what’s new in Fleet Prime Agent or browse all changelogs.
Week of November 16: Fleet Prime Agent 0.12.1 to 0.13.6
Fleet Prime Agent 0.12.1 and 0.12.2 shipped on November 12, 0.13.0 on November 15, 0.13.3 on November 16, and 0.13.6 on November 20.New features
Reorganized Settings (0.13.0) Settings now groups its sections into Fleet preferences, Prime Agent models, Prime Agent runtime, and Integrations, with a search box to find a setting. Each section has its own Save and Cancel, and drafts are kept while you move between sections. Choose whether Prime Agent settings save to the Current project (default) or to User defaults. Skills show their description, source, and loading status. See Settings.Attach files from the composer (0.12.2) A paperclip button beside + attaches files, and images show as chips in the composer. The button is disabled while a turn runs. The mode, model, and effort pickers share one keyboard-friendly popover: arrow keys move, Escape or a click outside closes it. See The composer.Refreshed interface (0.13.1) Fluid typography, a compact density, tighter command suggestions, and transcript scrolling behind the composer. The model picker shows model names only, truncated with the full name on hover. Spinners respect reduced motion. See Web workspace.Updates
- Breaking:
openUISyntaxVersion removed (0.12.1). Requests that still send it are rejected. OpenUI prompts are generated on the server (0.13.3). See the HTTP API reference.
- Save destination in the API (0.13.0).
PATCH /api/chat/settings accepts scope: "global" or "project". Without a scope, Fleet writes the project file only when the project already has settings. See the HTTP API reference.
- MCP servers (0.13.0). Argument and environment rows are validated when you add a server, replacing a server with the same name asks for confirmation, and sign-in guidance is shared with Connectors. See Settings.
- Stricter server checks (0.13.3). API errors use standardized codes. Session operations are rejected when their project is not registered. A request-aware origin check allows local origins only. See the HTTP API reference.
- Steadier streaming (0.13.3). The client validates stream frames but tolerates additional fields. Idle connections are reclaimed, replay payloads are capped, and a replay gap triggers recovery. Reopening a released session restores its messages and presentation. See Streaming.
- Thinking in the transcript (0.12.1). Thinking no longer shows as a nested tool trace. See Web workspace.
- Lazy Settings data (0.12.1). Provider, MCP, and full model catalog requests wait until you open Settings.
- Simplified workspace (0.13.6). The side panel unmounts after it slides closed, and a recovering session no longer breaks the project list.
- Toolchain and platform (0.12.1 and 0.13.2). The web app builds with React Compiler, and Oxlint with Oxfmt replaces Biome. Fleet requires Node.js 22.12.0 or later. See Install and Development.
- Fixes. A presentation saved while a turn finishes is persisted, a failed composer-intent save reports its HTTP error, cancelled intent responses are discarded, and keyboard command selection works reliably.
Week of November 9: Fleet Prime Agent 0.12.0
New features
Dashboards as attachments Ask for a dashboard and Prime Agent researches it with its normal tools. The board only lays out figures the reply already stated. The message shows an attachment that opens the board in the new Dashboards panel. A dashboard of the current chat reads the session’s own activity through the new session_outputs data tool. See Dashboards and artifacts.OpenUI Cloud builder Add a Thesys API key in Settings → Dashboards to let OpenUI Cloud build and edit boards. The key stays on the server. THESYS_API_KEY and FLEET_DASHBOARD_MODEL supply the key and model when Settings has not saved one. A board keeps the mode that created it. See Configure dashboards.Updates
- Render diagnostics. Setting a publishable OpenUI browser key (
VITE_OPENUI_CLIENT_KEY) sends render diagnostics to Thesys Reliability Monitoring. Events are sent without response text. See Configuration.
Week of November 2: Fleet Prime Agent 0.11.3
Fleet Prime Agent 0.10.3 through 0.11.3 shipped between November 3 and November 6.New features
Generative UI reads Fleet data (0.11.0) Generated OpenUI programs can call Query(...) through read-only tools: workspace_tree, list_sessions, list_resources, and session_stats. New components include Tabs, Accordion, TextArea, Slider, Steps, KeyValueList, Empty, Skeleton, and Image, plus the @Duration, @Bytes, @Tokens, @Percent, and @ShortPath functions. Form state survives reloads and session switches. See the HTTP API reference.Prime Agent runtime 0.10.2 (0.10.3) Fleet now installs Prime Agent 0.10.2, with upstream fixes for Google, Azure OpenAI, Bedrock, and Anthropic providers, MCP OAuth issuers with a trailing slash, and stuck Python cells. The protocol is unchanged, so existing sessions keep working. The first launch after updating downloads the new runtime once. See Upgrading the runtime.Updates
- Large custom provider catalogs (0.11.1). Catalogs with many models, such as Command Code with 95+ models, no longer fail validation and disappear from Settings → Providers. You edit custom provider models in a searchable list with Load more paging. See Providers and models.
- Thought rows (0.10.4 and 0.11.3). Each stretch of thinking gets its own Thought row, including thinking after a tool call. Long Thoughts split into paragraphs. Kernel cells, the REPL, and tool traces say CPython instead of IPython. See Web workspace.
- Tool status (0.10.3 and 0.10.5). Failed tool calls, including Python cells that raise an exception, stay marked as failed after the turn finishes. A tool you stop mid-run shows Cancelled. See Web workspace.
- Chat and sidebar polish (0.10.5). Enter submits the rename dialogs, and untitled sessions show New chat in the sidebar. On phones, the Main and Activity tabs show only their icons. See Web workspace.
- Runtime download check (0.11.2). Fleet refuses runtime downloads that are not on the pinned Fleet release host. See Security.
Week of October 26 — Fleet Prime Agent 0.10.2
New features
Native Prime Agent runtime On first launch, Fleet provisions the pinned Rust Prime Agent runtime (0.10.1) and verifies the archive and executable checksums (0.8.0 and 0.9.0). Fleet talks to the daemon directly and refuses incompatible daemons, including a TypeScript daemon. See Upgrading the runtime.Edit settings and provider credentials in Fleet Fleet now writes settings and provider credentials itself, with locked atomic writes and owner-only credential files (0.9.0). Providers that have only a stored credential now appear. See Providers and models.Activity traces and Activity tab The new Activity tab shows activity traces with credentials redacted (0.8.0). See Web workspace.Updates
- Steadier sessions and MCP status (0.9.0). The MCP list shows whether a credential actually works. The selected model is kept across reloads. Chat sessions recover after an invalid stream event. See Web workspace.
- Model catalog from the daemon (0.10.0). The daemon now serves the model catalog. Providers configured only through an environment variable, such as
ANTHROPIC_API_KEY, count as configured. A custom models.json provider without apiKey gets a diagnostic. Sign-in guidance names /mcp login <name> and /login. Saved settings show in the model picker without a restart. See Providers and models.
- Installer fixes (0.10.0 and 0.10.1). The installer keeps
prime-agent-runtime beside the binary, which fixes the “runtime bundle … missing” error, and includes fleet-prime. Binary-only installs under ~/.prime/fleet/runtimes repair themselves on the next start. See Install.
- Tab polish (0.10.2). Tabs share one tab strip, and right-panel tabs keep their labels. See Web workspace.
Week of October 19 — Fleet Prime Agent 0.7.4
Updates
@qredence/fleet 0.7.4. Released October 23. It ships the SSE replay, Session tree and model picker fixes from the October 12 entry. It also removes presentation sidecars when you delete a session. See Web workspace.
- Health check drops
kernel. On main after 0.7.4: Fleet no longer prewarms its own IPython kernel, because the Prime Agent runtime provisions one. GET /api/health no longer reports kernel. See the HTTP API reference.
Week of October 12 — Fleet Prime Agent
These fixes are on main after @qredence/fleet 0.7.3 and ship with the next release.Updates
- Gateway-compatible defaults for custom providers. When you save a custom
openai-completions provider whose base URL is not api.openai.com, Fleet sets compat.supportsStore to false. Models whose id starts with gpt-6 or gpt-5.6 default to reasoning: true. Values you already set are kept. See OpenAI-compatible endpoints.
- Model picker lists working models only. The composer’s model selector now shows only configured models you added in Settings. See Pick a model.
- Steadier SSE replay. The session event stream now tracks overflow and reset correctly, so a reconnect resumes from the right point. See Out-of-turn pushes.
- Session tree stays aligned. Session Tree entries now map to the right transcript messages across compactions, custom messages, and tool results. See Session tree.
Week of October 5 — Fleet Prime Agent 0.7.1 to 0.7.3
Updates
- Prime Agent runtime 0.9.8.
@qredence/fleet 0.7.3 pins the stock Prime Agent runtime at 0.9.8 (from 0.9.3), with the matching prime-agent-core and prime-agent-ai tarballs. The daemon protocol identity is unchanged. See The pin.
- Subagent progress notes. Subagent rows and the subagent thread header show the child’s latest progress note. See Tool cards.
/autonomous budget flags. The command hint lists --max-continuations, --max-turns, --max-tokens, --timeout-ms, and --gate. /scoped-models names the Alt+M cycling key. See Slash commands.
- Cleaner composer suggestions. Ghost-text completion no longer suggests runtime-injected continuation messages as if you had typed them. See The composer.
- Dependency patches. 0.7.1 and 0.7.2 update bundled dependencies with no user-facing behavior change.
Week of September 21 — Fleet Prime Agent 0.7.0
Fleet Prime Agent had no user-facing changes from September 28 to October 4.New features
Redesigned controls and sidebar @qredence/fleet 0.7.0 redesigns the buttons, scroll areas, and sidebar of the web workspace. Drag to resize the sidebar, throw it to collapse it, hover for a peek overlay, or use the [ and ] keyboard shortcuts. Rename, delete, and fork dialogs share one dialog, and resource panels can be resized from the keyboard. See Web workspace.
Week of September 14 — Fleet Prime Agent 0.6.2 to 0.6.6
New features
Session Tree panel The Session Tree right panel (0.6.3) shows the branches of the current session. Browse earlier entries, highlight them in the transcript, and rewind from history. See Session tree.Composer autocomplete The composer shows a dimmed completion drawn from your earlier prompts (0.6.4). Tab accepts it and Escape dismisses it. In 0.6.5 the suggestion becomes session-aware: prompts from the current session win, and the suggestion only extends as far as matching prompts agree. A suggested slash command only fills the composer, so nothing runs until you press Enter. See The composer.Composer command routing With a TYPESAFE_API_KEY set and the Settings → Chat toggle on, a message that describes a built-in command is recognized. Read-only commands run on submit, and state-changing ones are offered for confirmation. Anything else goes to the agent as before. See Environment variables.Updates
- Display settings. Density, reduced-motion, and transcript-follow settings now take effect, and overlapping turn-progress UI is collapsed (0.6.2). See Settings.
- Composer controls. The Agent and Model selectors share one trigger style, and the plus button opens slash commands without inserting
/ (0.6.2). See The composer.
- One web app package. The design workspace is merged into the web app (0.6.6). Nothing changes in the interface, and no action is needed when you upgrade. See Project structure.
Week of September 7 — Fleet Prime Agent 0.5.8 to 0.6.1
New features
Message a subagent from its tab Subagent tabs now have their own composer, so you can message a child directly while the parent session keeps running (0.5.8). POST /api/chat accepts an optional childId alongside the parent sessionId. See Subagent tabs and the HTTP API.Edit queued messages Edit steering and follow-up messages in place while they wait in the queue (0.6.0). Saving an empty edit deletes the message. See The composer.MCP connections in Settings Settings → MCP lists your MCP server connections (0.6.0). Add a server, sign in to OAuth servers, sign out, or remove a connection. See Settings.Updates
- Prime Agent runtime 0.9.3. 0.5.8 pins the stock runtime at 0.9.3 with daemon and Codex model discovery fixes. See Upgrading the runtime.
- Stale default models fall back. When a saved default model id no longer exists in a known provider’s catalog, Fleet uses that provider’s first available model instead of failing the turn (0.5.8). See Pick a model.
- Slash command arguments. Commands that take arguments stay in the composer until you submit them, and command-only sessions get a real title instead of “(no messages)” (0.6.0). See Slash commands.
- Reorganized Settings. The Settings dialog groups its sections into Workspace, Models, and Advanced, and the sidebar footer shows the installed Fleet version (0.5.8). See Settings.
- Reliability fixes. Staged attachments survive failed sends, answered questions close correctly, and projects can no longer target GitHub CLI credential directories (0.5.11).
Week of August 29 — Fleet Prime Agent
New features
Session insights panel with live activity timeline The web UI adds a session insights side panel with a live activity timeline, so you can follow what the agent is doing as events arrive instead of scrolling the transcript. See Web interface.Inline OpenUI artifacts with canvas rendering Chat responses can now stream OpenUI artifacts that render inline in the thread, including HTML artifacts on a canvas. Artifact browsing improves with a new file tree component and a deeper workspace tree. See Web interface.Persistent session presentation The web UI now streams agent activity and artifacts through typed presentation contracts and persists them with the session. Reopening a session restores rendered activity and artifacts rather than replaying a raw transcript. See Sessions and branching.Updates
- Stock Prime Agent runtime. The web surface now runs the stock Prime Agent runtime at a pinned version instead of a vendored copy, so runtime behavior tracks upstream releases directly. See Deployment and installer.
- Unified tool envelopes and structured events. Tool calls and results now stream as typed envelopes with correlation IDs and a discriminated error shape. The thread UI renders structured tool events, streaming compaction artifacts, and auto-retry error envelopes. See Streaming protocol and Tool cards.
- Streams survive session resets. Active chat streams are preserved across session resets, so a reset no longer drops an in-flight response. See Streaming chat.
Bug fixes
- Provider models and OpenUI actions preserved. Editing a provider keeps registry content and model metadata intact. OpenAI-compatible model discovery results are normalized and persisted, and unsupported API families are rejected. Settled reasoning blocks and reopened OpenUI actions keep working. See Providers and models.
- Installer respects existing installs. The installer ships the
fleet-prime shim without overwriting an existing prime-agent installation. See Deployment and installer.
Week of August 17 — Fleet Prime Agent
New features
Fleet Prime Agent — self-improving RLM agent with persistent sessions and a standalone web UI Fleet Prime Agent is a new product in the Qredence suite. Prime Agent runs as a daemon-backed agent with persistent IPython-backed sessions, subagent spawning via rlm, and a standalone Qredence web UI. Sessions support branching, tool cards render every tool call inline, and streaming chat flows through the typed AgentSessionEvent / ChatStreamEvent / AgentMessage contracts. See Introduction to Fleet Prime Agent and Getting started.Bring-your-own provider and model selection Prime Agent’s AI layer abstracts providers so you can pick your provider and model per session, including OAuth sign-in for hosted providers and MCP servers for extending the tool surface. See Providers and models.Coding agent with CLI, interactive mode, skills, and extensions The bundled coding agent ships a CLI, an interactive mode, a skills system, extensions, and a refinement loop, all backed by a session runtime that persists conversation and IPython state across runs. Trigger refinement or sign-in with slash commands (/refine, /login). See Slash commands, Sessions and branching, and the coding agent overview.Daemon protocol and web API Prime Agent exposes a daemon protocol for local integrations and a web API for the standalone UI, both documented with request/response shapes and streaming semantics. See Daemon protocol and Web API.Installer and release pipeline A one-shot installer sets up the runtime, and the release pipeline publishes signed builds for the daemon, CLI, TUI, and web surface. See Installer and Release pipeline. Last modified on November 20, 2026