hub
Prime Agent ships 58 audited features across architecture, agent loop, sessions, recursive subagents, long-running operation, customization, providers, and integrations. Each feature has a stable ID, user story, evidence links, and caveats. The explorer filters by category and maturity; eleven capabilities have dedicated deep pages.
Pinned repository file/line or test referenced in the audit.
All 58 audited features are visible without JavaScript; filters narrow the list as you type.
architecture · stable
arch-monorepo
As a contributor, I want separable UI, model, loop, host, and Python-runtime layers.
The workspace builds TUI, AI adapters, agent core, coding-agent host, and a Python runtime as distinct packages.
architecture · stable
arch-process
As a user, I want one broken session not to terminate every session.
A detached supervisor routes clients; each root tree has a worker; IPython runs separately; a catalog subprocess scans saved sessions.
architecture · stable
arch-connection
As a client author, I want the same runtime behind TUI and headless clients.
Clients render and send commands through AgentConnection; daemon-backed and in-process implementations expose coherent snapshots/events.
agent-loop · stable
agent-loop
As a user, I want streamed model output and iterative tool use until the task turn settles.
The loop emits lifecycle/message/tool events, streams provider deltas, executes tool calls, then processes steering, follow-ups, and host continuations.
agent-loop · stable
parallel-tools
As an extension/tool author, I want tools to declare safe execution ordering.
Tool batches execute in parallel by default, but a global sequential setting or any sequential tool forces ordered execution; results preserve source order.
agent-loop · stable
abort-retry
As a user, I want cancellation to stop active provider/tool work and transient failures to retry within policy.
Abort signals race provider/tool operations and produce aborted messages; the session has configurable exponential retry and retry events.
agent-loop · stable
ipython
As an agent, I want one programmable environment for files, shell commands, skills, and delegation.
The built-in tool lazily provisions a persistent IPython kernel, executes cells, captures stdout/results/errors/diffs/attachments, and retains namespace state.
agent-loop · stable
kernel-snapshot
As a long-running user, I want Python variables and imports to survive session revival.
Persisted sessions can write dill/JSON kernel snapshots and restore them after worker/kernel restart.
ui · stable
tui
As a terminal user, I want streaming chat, editor, overlays, menus, and status in one interface.
Interactive mode renders messages/tool calls, a multiline editor, selectors/overlays, image previews, notifications, and usage/context views.
ui · stable
editor
As a user, I want file mentions, completion, multiline input, pasted/dragged images, shell shortcuts, and an external editor.
The editor supports @ file search, tab path completion, multiline keys, clipboard/drag images, !/!! shell execution, and $VISUAL/$EDITOR.
commands · stable
slash-commands
As a user, I want discoverable in-session controls for auth, models, sessions, long-running work, exports, and configuration.
Autocomplete exposes canonical commands plus aliases; session-owned commands are parsed separately for daemon execution.
commands · stable
public-cli
As a shell user, I want commands to list, attach, stop, message, schedule, diagnose, update, and manage packages.
The public dispatcher validates and maps agents/list/attach/stop/rename/send/schedule/status/doctor/shutdown/package/update/model/session/config commands.
commands · stable
modes
As an integrator, I want human and machine interfaces over the same session runtime.
The CLI selects interactive TUI, one-shot print, JSON event stream, Prime-specific RPC, or standard ACP JSON-RPC.
commands · stable
typescript-sdk
As a TypeScript integrator, I want to create and control AgentSession directly without spawning a CLI process.
createAgentSession wires settings, authentication, model selection, tools, resources, extensions, persistence, MCP, and autonomous policy, then returns the session and extension load result.
sessions · stable
message-queue
As a user, I want to redirect active work or queue later work without waiting.
Enter/steer injects after the current assistant tool batch; Alt+Enter/follow-up waits for the agent to become idle; delivery can be all or one-at-a-time.
sessions · stable
session-jsonl
As a user, I want conversations recoverable and inspectable on disk.
Sessions append flat JSONL under the configured session directory, record cwd and tree entries, and support ephemeral no-session mode.
sessions · stable
session-tree
As a user, I want to revisit prior turns without destroying history.
Users can navigate the session tree, summarize abandoned branches, fork before a user message, clone at the current position, import JSONL, or resume by path/ID.
sessions · stable
session-export
As a user, I want portable or shareable session records.
/export writes styled HTML or JSONL; /share uploads a secret GitHub gist; export sanitizes untrusted content.
sessions · stable
side-questions
As a user, I want to ask a tangent without polluting the main session.
/btw and /side create inline side conversations whose messages are excluded from the persisted main conversation branch.
ui · stable
usage-context
As a user, I want to understand cost and context pressure across parent and child work.
/context and /usage report model context, token/cost totals, and own-versus-attributed child usage.
ui · stable
themes-keybindings
As a user, I want accessible appearance and remappable controls.
Built-in dark/light/Prime themes and project/global theme resources pair with configurable editor/app keybinding maps.
rlm · stable
rlm
As an agent, I want to delegate bounded context to independent child agents.
await rlm(prompt, name, model) returns an admission handle immediately; the host starts a child AgentSession and answers arrive through agent messages or files.
rlm · stable
rlm-registry
As a parent agent, I want to list, message, and delete retained children after compaction or restart.
The parent-scoped registry persists child identity/status, rehydrates completed daemon children, supports exact selector deletion, and leaves transcript artifacts on disk.
rlm · stable
a2a
As an agent or user, I want active family members to coordinate directly.
Messages target parent/child/sibling/family members, support auto/steer/follow_up delivery, and enforce size/rate/pending-queue safety limits.
rlm · stable
observe
As an operator, I want to inspect another active root or child session without attaching as its owner.
Agent observation returns snapshots and streams wrapped observed-session events; closing the target emits a close signal.
long-running · stable
daemon-continuity
As a user, I want work to continue after closing the terminal and recover after supervisor/worker failure.
Resident workers outlive the TUI, supervisors adopt live workers, per-root crashes retry on a bounded schedule, and snapshots/replay resynchronize clients.
long-running · stable
leases-journal
As an operator, I want concurrent clients and reconnects not to duplicate writes or side effects.
Canonical-path leases prevent concurrent transcript writers; clientId+commandId journals return completed results but mark uncertain mutations instead of replaying them.
long-running · stable
compaction
As a long-running user, I want old context summarized before the model window overflows.
Manual, threshold, and overflow compaction summarize older history, preserve recent context and kernel state, emit lifecycle events, and can retry overflowed prompts.
long-running · stable
goals
As a user, I want an objective to remain active across turns with budget accounting.
/goal creates/edits/pauses/resumes/clears a durable objective; only goal.complete() marks success; tokens/time/continuations and optional budget are tracked.
long-running · stable
autonomous
As a user, I want unattended work to continue only within budgets and objective checks.
Autonomous mode injects continuations until all shell gates pass or continuation/turn/token/time limits stop it; unchanged failed workspaces are not rerun blindly.
long-running · stable
heartbeats
As a user or agent, I want periodic prompts to re-enter an existing session.
/heartbeat manages one user heartbeat; rlm_heartbeat manages multiple agent-owned jobs with independent labels/delivery modes; both persist per session.
long-running · stable
schedules
As an automation user, I want prompts delivered at a future time or cron cadence.
prime-agent schedule adds/lists/cancels jobs targeted at agents; due ticks are claimed before delivery and missed ticks coalesce.
long-running · stable
refine
As a user, I want evidence-backed lessons persisted without rewriting the base prompt.
/refine reviews a trajectory and applies small updates to supplemental prompts, memories, skill descriptions, or subagent specs; snapshots support rollback and optional global scope.
configuration · stable
settings
As a user, I want defaults globally and narrower overrides per project.
~/.prime/agent/settings.json is deep-merged with .prime/agent/settings.json; CLI/env settings have documented precedence for selected fields.
configuration · stable
context-prompts
As a project owner, I want repository conventions injected predictably.
Prime Agent walks global/ancestor/cwd AGENTS.md or CLAUDE.md files; SYSTEM.md replaces and APPEND_SYSTEM.md extends the default prompt; CLI can disable or override.
customization · stable
skills
As a user, I want reusable workflows loaded progressively and executable from IPython.
Agent Skills SKILL.md metadata is progressively disclosed; pyproject-backed skills install editable packages and expose importable/callable Python modules; /skill:name forces invocation.
customization · stable
builtin-skills
As a new user, I want key workflows available without installing separate packages.
Built-in skills ship at lowest precedence and can be disabled/overridden; Python-backed skills bridge host features and services.
customization · stable
extensions
As a developer, I want custom tools, commands, UI, policies, providers, and lifecycle hooks.
jiti-loaded TypeScript factories register tools/commands/shortcuts/flags/providers, intercept lifecycle/model/tool/input events, render custom TUI, and persist entries.
customization · stable
packages
As a user, I want installable bundles of extensions, skills, prompts, and themes.
package install/remove/list/update resolves npm or git sources, persists global/project configuration, installs production dependencies, and loads declared resources.
customization · stable
prompt-themes-resources
As a user, I want project-scoped reusable prompts and presentation assets.
ResourceLoader discovers package/global/project/explicit paths with precedence and exclusions; /reload refreshes keybindings, extensions, skills, prompts, themes, and context.
integrations · stable
mcp
As a user, I want Linear, Notion, and custom remote MCP services callable from IPython.
OAuth or bearer-authenticated HTTP MCP servers expose discovered tools as async Python methods through McpIntegration; built-ins enable by login.
integrations · stable
trace-sharing
As a researcher, I want to inspect and explicitly share current/all traces.
/traces controls opt-in sharing, previews sanitized trace payloads, and uploads with a Prime API key scoped for agent_traces.
models · stable
providers
As a user, I want one agent over subscription, API-key, cloud, and OpenAI-compatible providers.
The AI package normalizes streaming text/thinking/tool calls/usage across built-in provider adapters and a generated model catalog.
models · stable
auth
As a user, I want login for subscriptions and secure local API-key fallback.
/login supports ChatGPT/Codex, Claude, Copilot, providers, and MCP; auth.json is mode 0600; key resolution is CLI, auth file, env, then models.json.
models · stable
models-config
As a self-hosting user, I want Ollama, LM Studio, vLLM, or custom endpoints.
models.json can define supported API-compatible providers/models; extensions can register custom implementations and OAuth flows; model resolution supports provider/id and thinking suffixes.
models · stable
thinking-model-scope
As a user, I want model-specific reasoning control and a curated cycle list.
/effort selects supported levels, /fast toggles OpenAI service tier, /scoped-models filters cycling, and unsupported levels are clamped.
models · stable
images
As a user, I want to attach screenshots and send them only to compatible models.
File args, paste, and drag produce image content; images can resize/block; adapters route image tool results according to provider limits.
environments-evals · integration
prime-env
As a researcher, I want to discover, scaffold, install, and publish verifiers environments.
The bundled prime-intellect skill documents prime env list/info/install/init/push and points to the official prime CLI.
environments-evals · integration
prime-evals
As a researcher, I want small local smoke evals before hosted scale.
The bundled skill documents prime eval run with model/sample controls and --hosted --follow; ACP also supports evaluation harnesses.
environments-evals · integration
prime-training
As a researcher, I want to launch hosted post-training or use self-managed RL.
The bundled skill documents prime lab setup, prime train model/capacity/config/run workflows, and identifies prime-rl as the power-user self-managed path.
environments-evals · integration
prime-sandbox-compute
As a user, I want optional external isolation and infrastructure around agent work.
The bundled skill routes to prime sandbox, inference, availability/pods, storage, and tunnel products.
environments-evals · stable
acp-eval
As an evaluator, I want standards-based prompting, streaming, cancellation, stop reasons, and Prime metadata.
ACP exposes initialize/session new/prompt/cancel/close, maps text/reasoning/tools, and places subagents/gates/goals/heartbeats/refinement metadata under a namespaced _meta key.
security · stable
trust
As an operator, I want clear boundaries before allowing autonomous code execution.
README/docs warn that model Python, shell commands, skills, and extensions run with user permissions; workers/kernels isolate lifecycle, not security.
privacy · new-in-inspected-main
telemetry
As a privacy-conscious user, I want to know what is collected and disable it.
Default aggregate events include version/OS/mode/outcomes/performance/counts but exclude prompts, code, paths, commands, credentials, raw errors, and identity/hardware; settings/env/offline can disable.
distribution · stable
installer-update
As a user, I want reproducible installation and updates without manually managing npm.
The published installer resolves channel/version, downloads release tarball and SHA256SUMS, verifies checksums, installs, and can prepare the kernel; update supports coordinated resident-worker checkpoints.
quality · stable
ci
As a maintainer, I want formatting/type/build checks plus package, process, and kernel tests.
CI builds/checks on Node 22 and runs agent, AI, TUI, sharded coding-agent, daemon process, and kernel jobs; nightly process stress is separate.
distribution · active
release
As a maintainer, I want all packages versioned together with stable and rolling beta artifacts.
Release scripts version packages in lockstep; workflow builds/checks/packs, uploads immutable tarballs/checksums to R2, advances manifests, and creates GitHub releases.
quality · developer-tooling
benchmarks
As a maintainer, I want performance regressions measurable.
Scripts/tests benchmark daemon fanout/attach, startup, TUI rendering, tool statistics, and high-worker process stress.
Categories are fixed to the audited inventory taxonomy: architecture, agent-loop, ui, commands, sessions, rlm, long-running, configuration, customization, integrations, models, environments-evals, security, privacy, distribution, and quality. Filter by either category or maturity, then jump to the deep page or coverage anchor.
Eleven features have dedicated deep pages. The remaining cards still expose stable anchors, pinned evidence, and caveats so the explorer can serve as a complete coverage backstop.
Each card points to the repository line or documentation paragraph that supports it. Deep-page cards still list the source chips even when the coverage URL resolves elsewhere.
The complete audited inventory contains 58 stable feature IDs. Use the explorer above to search and filter every record; the coverage index below provides direct, no-JavaScript anchor links.
Deep pages are linked on their cards. Stable anchors still make the explorer complete, and every card remains searchable by feature ID.