This is the published mirror of FEATURES.md
in the repository — the authoritative inventory of every warden capability and
which surface can drive it.
warden exposes its features across five surfaces — the CLI (warden, aliased
wd), MCP (83 structured tools for an orchestrating agent), the /warden
skill, the web GUI, and the TUI cockpit.
Coverage legend: ✓ supported · — not applicable / not present on that surface ·
CLI-only features are marked and explained (host/process/interactive/secret
operations that are meaningless or unsafe over MCP/web).
Projects store complete agents[], pipelines[], terminals[], plans[], and
autopilots[] membership lists. Plan-run execution entities carry a plan_id
back-ref. Agents store parent_id, child_agents[], and child_pipelines[];
pipelines store parent_agent_id. Job agents belong to pipeline jobs, never
the owning agent’s child_agents[]. New projects open empty; reopening restores
hibernated members without auto-spawning an orchestrator. Agent UX states are
pending, busy, idle, need-input, done, orphaned, and rate_limited.
Spawn, inspect, message, and tear down per-task coding agents (Claude Code by
default; each in its own tmux session, most in a git worktree).
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Spawn an agent (ticket / prompt / typed) | start | spawn_agent | ✓ | ✓ | n | spawn-and-watch |
| Prompt-spawn (no repo, auto-typed) | start "<prompt>" | spawn_agent | ✓ | ✓ | n | spawn-and-watch |
| List agents | ls | list_agents | ✓ | ✓ | list | fleet-operations |
| Agent status / detail | status | get_agent | ✓ | ✓ | i | fleet-operations |
| Tail recent output | tail | get_agent_output | ✓ | ✓ | view | spawn-and-watch |
| Send text into an agent | send | send_to_agent | ✓ | ✓ | — | spawn-and-watch |
| Digest (catch-up summary) | digest | digest | ✓ | ✓ | i | rotation-digests |
| Attach to the live session | attach | CLI-only (interactive tmux) | ✓ | ✓ (terminal) | enter | spawn-and-watch |
| Adopt an existing Claude session | adopt | adopt_agent | ✓ | — | — | agents-lifecycle |
| Tear down (umbrella) | stop | stop_agent | ✓ | ✓ | x | fleet-operations |
| Finish cleanly (commit/push guard) | done (= stop --keep-worktree) | terminate_agent (force) | ✓ | ✓ | x | lifecycle-and-rails |
| Terminate | terminate (= stop --keep-record --keep-worktree) | terminate_agent | ✓ | ✓ | x | lifecycle-and-rails |
| Restore an orphaned agent | restore | restore_agent | ✓ | ✓ | r | agents-lifecycle |
Recover an archived orphaned agent with a live pane (tombstone-reaper safety net) | recover | recover_agents | ✓ | — | — | agents-lifecycle |
| Delete / hard-purge | delete (= stop --keep-worktree, record only) | delete_agent | ✓ | ✓ | D | fleet-operations |
Mandatory agent naming (auto-resolve when --name/name omitted; role conventions + Fast-Brain-engine slug + adjective-noun fallback; auto-disambiguate) | start (omit --name) | spawn_agent (omit name) | ✓ | ✓ | list name col | spawn-and-watch |
Live 3–5 word activity badge per agent (Fast-Brain; activity.interval, default 15s; keeps previous on failure) | automatic | — (REST activity field only) | — | — | agent row badge | tui-cockpit |
Fast-Brain PR title/body (KindPRSummary; per-field fallback; explicit title/body wins) | agent stop <AGENT> --pr | — | ✓ | — | — | lifecycle-and-rails |
Fast-Brain tier router (opt-in router.use_fast_brain, default off; lowest precedence, confidence ≥ 0.8) | start (omit --tier/--task/--role) | spawn_agent (omit pins) | ✓ | — | — | spawn-and-watch |
| Rename an agent | adopt --name / spawn name | spawn_agent (name) | ✓ | ✓ | info panel | fleet-operations |
| Tags (group / filter) | start --tag, ls --tag | spawn_agent (tags) | ✓ | ✓ | — | fleet-operations |
| Model selection | start --model / config | spawn_agent (model) | ✓ | ✓ | — | env-vars |
| Agent role (built-in persona + default flags) at spawn | start --role | spawn_agent (role) | ✓ | ✓ (Role select) | ctrl+r (new-agent) | agent-roles |
| Switch a running agent’s role (relaunch re-injects) | set-role | set_role | ✓ | — | — | agent-roles |
| List the built-in role catalog | role list | list_roles | ✓ | ✓ (Role select) | list | agent-roles |
| Tiered model routing — derive the model tier from a task | start --task (also REST; no pipeline task:) | spawn_agent (task) | ✓ | — | — | agent-roles |
| Tiered model routing — pin the model tier for the quota-balanced resolver | start --tier (also REST and pipeline tier:) | spawn_agent (tier) | ✓ | — | ctrl+t (new-agent; live candidate table) | agent-roles |
| List the backend’s live model menu | models (--backend, --json) | CLI-only (agent-native; local worktree exec, no daemon round-trip) | ✓ | — | — | backend-superpowers |
| Backend selection (Claude / Aider / OpenCode / Codex / Crush / Goose / Cursor / Antigravity) | start --aicli | spawn_agent (backend) | ✓ | — | — (CLI/MCP pin; TUI is tier-first) | agent-backends — only claude is stable; codex and antigravity are β beta; aider, opencode, crush, goose, and cursor are 🧪 experimental |
Terminal session (kind=terminal) — managed $SHELL seat, no AI (prompt ignored); back-compat backend=terminal alias | start --kind terminal | spawn_agent (kind) | ✓ | ✓ (Terminals tab) | t (in the cursor’s project) | terminals |
Handoff — delegate (new / --to existing) or retire self (--retire) | handoff | handoff_agent | ✓ | — | — | rotation-digests |
Self-rotation (retire → successor) — alias for handoff --retire | rotate | rotate_agent | ✓ | — | — | rotation-digests |
| Fork an agent’s session into a new managed agent (Codex-only; branches the conversation, source keeps running; dirty-tree carry) | fork | fork_agent | ✓ | — | — | backend-superpowers |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
Task --type (development, analysis, spike, pr-review, code, docs, website, debug-ci, tests, other) | start --type | spawn_agent (type) | ✓ | ✓ | — | worktrees-task-types |
| Managed worktree per agent | automatic | automatic | ✓ | ✓ | — | worktrees-task-types |
| Scratch worktree (analysis/spike) | start --worktree | spawn_agent (worktree) | ✓ | ✓ | — | worktrees-task-types |
| In-repo opt-out (write-agent) | start --in-repo | spawn_agent (in_repo) | ✓ | ✓ | — | worktrees-task-types |
| List worktrees | worktree list (alias worktree ls; bare worktree) | list_worktrees | ✓ | — | — | worktrees-task-types |
| Prune orphaned worktrees | worktree prune (alias prune) | prune_worktrees | ✓ | — | — | fleet-operations |
| Clean merged branches + stale worktrees | workspace clean (alias clean) | — | ✓ | — | — | workspace-release-hygiene |
| Remove one agent’s worktree | remove-worktree (= stop --keep-record, worktree only) | remove_worktree | ✓ | ✓ | — | fleet-operations |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
Commit (auto-message from diff; [paths…], --amend/--force) | commit | commit (paths, amend, force) | ✓ | — | — | lifecycle-and-rails |
Push (protected-branch rails, --force-with-lease) | push | push | ✓ | — | — | lifecycle-and-rails |
Sync / rebase onto base (default: session/integration/default branch; --continue / --abort; non-zero on conflicts) | sync | sync (base, continue, abort) | ✓ | — | — | lifecycle-and-rails |
| Open / return the PR for the agent branch (agent keeps running) | git pr (--base, --title, --body/--body-file, --json) | create_pr | ✓ | — | — | lifecycle-and-rails |
| Run project checks (compact failures; works from subdirectories) | check | check | | | | |
| List configured checks without running them | check list | list_checks | ✓ | — | — | lifecycle-and-rails |
| SemVer release recommendation and categorized changelog; tag only on explicit confirmation | release (alias git release) | — | ✓ | — | — | workspace-release-hygiene |
| Agent-native diff review (backend’s own reviewer) | review (--base, --prompt, --backend) | CLI-only (agent-native; local worktree exec, no daemon round-trip) | ✓ | — | — | backend-superpowers |
| Machine-readable review findings (neutral JSON) | review --json | CLI-only | ✓ | — | — | backend-superpowers |
| git-guard hook (deny raw git mutations) | git-guard (hook) | enforced | ✓ | — | — | lifecycle-and-rails |
| check-guard hook (redirect broad runs) | check-guard (hook) | enforced | ✓ | — | — | lifecycle-and-rails |
| root-guard / boundary enforcement | guard (hook) | enforced | ✓ | — | — | lifecycle-and-rails |
| Branch + CI status (per agent) | branches | get_branch_status | ✓ | ✓ | — | observability |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
Create from YAML spec / template ([spec.yaml] or -f; --start, --json) | pipeline create | create_pipeline | ✓ | ✓ | — | pipelines |
List pipelines (default: current project; --all / --project / --status / --json) | pipeline list | list_pipelines | ✓ | ✓ | view | pipelines |
Show one pipeline’s jobs/output (--prompts / --json / --watch / --all-jobs) | pipeline show | show_pipeline | ✓ | ✓ | i | pipelines |
| Start | pipeline start | start_pipeline | ✓ | ✓ | — | pipelines |
Cancel (confirms when jobs are live; --yes; cannot restart) | pipeline cancel | cancel_pipeline | ✓ | ✓ | — | pipelines |
| Pause | pipeline pause | pause_pipeline | ✓ | ✓ | — | pipelines |
| Resume | pipeline resume | resume_pipeline | ✓ | ✓ | — | pipelines |
| Show one job | pipeline job show | show_pipeline (job in payload) | ✓ | ✓ | — | pipelines |
| Retry a failed job | pipeline job retry | retry_pipeline_job | ✓ | ✓ | — | pipelines |
| Edit a pending job | pipeline job edit | edit_pipeline_job | ✓ | — | — | pipelines |
| Emit a job’s handoff output | pipeline emit | emit_pipeline_output | ✓ | — | — | pipelines |
Delete a pipeline record (always confirms; --yes) | pipeline delete | delete_pipeline | ✓ | — | — | pipelines |
| Validate a spec (no daemon) | pipeline validate | validate_pipeline | ✓ | — | — | pipelines |
| List built-in templates | pipeline template list | list_pipeline_templates | ✓ | — | — | pipelines |
| Plan-owned pipeline control (pipeline pause/resume/cancel/delete refuse) | plan pause|resume|stop | control_plan | ✓ | — | — | pipelines |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Shared-context set | ctx set | ctx_set | ✓ | ✓ | — | shared-context-messages |
| Compare-and-set (claim a task) | ctx cas | ctx_cas | ✓ | — | — | shared-context-messages |
| Append | ctx append | ctx_append | ✓ | — | — | shared-context-messages |
| Get / list / delete | ctx get/list/del | ctx_get / ctx_list | ✓ | ✓ | — | shared-context-messages |
| Directed message to an inbox | msg send | send_message | ✓ | ✓ | — | shared-context-messages |
| Read inbox | msg inbox | read_inbox | ✓ | ✓ | — | shared-context-messages |
| Wait for a message (park/wake) | msg wait | wait_for_message | ✓ | — | — | shared-context-messages |
| File-conflict detection | collab conflicts | get_collaboration_status | ✓ | ✓ | — | shared-context-messages |
| Who is editing a file | collab who-is-editing | who_is_editing_file | ✓ | ✓ | — | shared-context-messages |
Project memory — show/edit .warden/memory.md | memory (--raw, --path, --edit) | CLI-only (local, no daemon round-trip) | ✓ | — | — | project-memory |
Project memory — projected into every spawn (memory.inject) | config (memory.inject, default on) | automatic (all backends but aider) | ✓ | — | — | project-memory |
Project memory — auto-curation from digests (memory.curate) | config (memory.curate, default off) | automatic on completion (proposes unverified entries to the working tree; never commits) | ✓ | — | — | project-memory |
Project memory — local grounding in the REPL (memory.ground) | repl → /memory <q> (/mem, /ask) + project_memory tool | REPL-only (Fast-Brain, no full agent turn) | ✓ | — | — | project-memory |
| Projects — first-class daemon projects (open local/remote, close, new, list) | projects (list, open, open-local, open-remote, new, close) | — | — | ✓ | ✓ | project-groups |
Projects — zero-touch auto-registration — launching an agent/pipeline/terminal/plan in a directory auto-registers (or reopens) its project; projects open* is optional | automatic on start / agent start / pipeline create / plan create / plan run / terminal launch (--project overrides) | automatic on spawn_agent | — | ✓ | ✓ | project-groups |
Project groups — named collections of projects shown in the TUI tree; incremental /members add/remove and bulk replace | project-groups (list, show, create, update, delete, members add|remove) | — | — | ✓ (read) | ✓ | project-groups |
| Open restores members — opening a project restores the members that were hibernated when it was last closed; it does not auto-spawn an orchestrator, so a project with no restorable members opens empty | automatic on project open | — | — | — | ✓ | project-groups |
| Peer awareness — grouped orchestrators learn their Project Group name and sibling orchestrator names via context injection at every (re)launch | automatic (daemon-wired via PeerContextFn) | — | — | — | — | project-groups |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| List pending approvals | approvals | list_approvals | ✓ | ✓ | cockpit | approvals-supervised |
| Answer an approval | approve | approve | ✓ | ✓ | a | approvals-supervised |
| Auto-approve toggle | auto-approve <id> on|off | set_auto_approve | ✓ | ✓ | — | approvals-supervised |
| Auto-approve rule policy (tool/glob/regex/paths, per-agent, identical-prompt circuit breaker) | auto-approve rules|allow|deny|clear|enable|disable | set_auto_approve_policy | ✓ | ✓ | — | approvals-supervised |
Workspace-trust prompt auto-answer (Claude/Codex/Antigravity in-pane; Cursor via --trust; on by default) | config (trust_workspace) | — | — | — | — | approvals-supervised |
| Model-assisted prompt recognition (a reworded/unknown menu is read by Fast-Brain, verified against the pane, then follows the normal policy; on by default) | config (recognize_prompts) | — | — | — | — | approvals-supervised |
Learned prompt shapes — a model-read prompt is learned on its first verified success so the next one needs no model call; list/forget (REST GET/DELETE /api/v1/known-prompts[/{id}]; config known_prompts_max, known_prompts_prune_days) | approval known list|forget | list_known_prompts, forget_known_prompt | ✓ | — | — | approvals-supervised |
| Fast-Brain approval arbiter (optional; confidence threshold, destructive guard, circuit breaker, and fail-open escalation remain authoritative) | config (auto_approve.use_fast_brain) | automatic while arbitrating | ✓ | — | — | fast-brain |
| Set permission mode (running agent) | set-permission-mode | set_permission_mode | ✓ | ✓ | — | approvals-supervised |
| Supervised spawn (gate every prompt) | start --supervised | spawn_agent (supervised) | ✓ | ✓ | — | approvals-supervised |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
Create schedule (cron / at / now; agent or pipeline; same options as start) | schedule create | create_schedule | ✓ | — | — | scheduling |
| List schedules (table with state: enabled / disabled / done / failed) | schedule list | list_schedules | ✓ | — | — | scheduling |
| Show one schedule (full fire payload + last run) | schedule show | get_schedule | ✓ | — | — | scheduling |
| Test a schedule (fire once, next run unchanged) | schedule run | run_schedule | ✓ | — | — | scheduling |
| Edit a schedule (only the flags passed) | schedule edit | update_schedule | ✓ | — | — | scheduling |
| Enable / disable schedule | schedule enable / disable | enable_schedule / disable_schedule | ✓ | — | — | scheduling |
Delete schedule (confirms; --yes to skip) | schedule delete | delete_schedule | ✓ | — | — | scheduling |
Scheduled-run session linkage (schedule_id on sessions; scheduled-agents capability) | — | — | ✓ | — | — | scheduling |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Snapshot worktree + transcript | snapshot create | snapshot_create | ✓ | — | — | snapshots |
| List snapshots | snapshot list | snapshot_list | ✓ | — | — | snapshots |
| Restore a snapshot (rails) | snapshot restore | snapshot_restore | ✓ | — | — | snapshots |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Live resource metrics | stats | get_metrics | ✓ | ✓ | — | observability |
| Metrics history (time-series) | stats --history | get_metrics (history) | ✓ | ✓ | — | observability |
| Memory-pressure gate / headroom | doctor / spawn gate | get_pressure | ✓ | ✓ | — | observability |
| Fleet insights (parallelizable pairs, etc.) | insights | insights | ✓ | — | — | insights |
| Cost umbrella (spend + savings in one) | cost (cost spend / cost savings) | spend + savings | ✓ | — | — | savings |
| Token-savings ledger | savings (alias cost savings) | savings | ✓ | — | — | savings |
| Cost governance ($ spend rollup) | spend (alias cost spend) | spend | ✓ | ✓ | $ in ls | savings |
| Budget gate (soft $ cap on spawn) | config (tokens.budget_gate) / spawn gate | automatic | ✓ | — | — | observability |
| Full-text search | search | search | ✓ | ✓ | — | cli |
| Browse archived agents | history | history | ✓ | ✓ (archive) | — | cli |
| Action audit trail | audit log | audit_log | ✓ | — | — | observability |
| Event stream (SSE) | — | — | — | ✓ | — | observability |
PTY event journal & stream (runtime.pty: shadow|on; redacted, replayable, cursor-resumable WS at /api/v1/sessions/{id}/pty) | config (runtime.pty) | — | — | API | — | pty-runtime |
Plain-terminal survival & native attach: independent PTY holder (daemon restart/update/crash keeps the shell; authenticated re-adoption, never a silent relaunch), attach with controller lease + fenced input, cockpit terminal pane (runtime.pty: on) | attach <id> (env WARDEN_ATTACH=observe|tmux; WARDEN_PTY_HOLDER=off reverts) | — | — | API (/pty, /attach) | cockpit terminal pane | pty-runtime |
PTY prompt engine: adapter-normalized prompts, TOCTOU-revalidated auto-answers, learned shapes, native posture modes (runtime.pty: on; agy first) | config (runtime.pty) | — | — | — | — | pty-runtime |
| Desktop / webhook / Slack notifications | config (notify.enabled, notify.webhook_*) | automatic | ✓ | ✓ | — | observability |
Context/token guard (gauge, alert, auto-/compact) | config (tokens.guard) | automatic | ✓ | ✓ | gauge | observability |
Force-compact a busy critical agent (interrupt→/compact→resume) | force-compact (config tokens.force_compact) | set_force_compact | ✓ | — | — | observability |
| Crash / anomaly detection (OOM, loop, pre-crash) | config | automatic | ✓ | ✓ | — | observability |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Export session metadata | export | export_sessions | ✓ | — | — | cli |
| Import session metadata | import | import_sessions | ✓ | — | — | cli |
| Save a spawn preset | preset save / library save-preset | CLI-only (local config authoring) | ✓ | — | — | cli |
| List presets | preset list | CLI-only | ✓ | — | — | cli |
| Save a prompt template | prompt-template save / library save-prompt | CLI-only (local config authoring) | ✓ | — | — | cli |
| List prompt templates | prompt-template list | CLI-only | ✓ | — | — | cli |
| Fill a prompt template into a spawn | start --prompt-template <name> --set VAR=value | CLI-only | ✓ | — | — | cli |
| Browse presets + prompt templates + pipeline templates (one umbrella) | library list | library_list | ✓ | — | — | cli |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| List registered plugins | plugin list | list_plugins | ✓ | — | — | plugins |
| Custom task types via plugins | config | enforced at spawn | ✓ | — | — | plugins |
| Lifecycle hook events (pre/post spawn, commit, check, teardown) | config | enforced | ✓ | — | — | plugins |
The browser GUI (served by the daemon) is a URL-routed shell (/cockpit
home · /tui · /pipelines · /metrics · /terminals · /archive · /others
· /agent/<id> — deep-linkable, back/forward, shareable). It provides: the
Cockpit home (Fleet header + agent grid), a full-screen TUI launcher
(top-bar ▢ TUI button) that streams the literal warden tui into the browser, a
Pipelines tab with a live DAG, a Metrics
tab (per-agent and fleet-total CPU/memory, per-agent context, fleet size,
tokens saved — two-column on desktop, single-column on mobile), a Terminals
tab (PTY viewport + New terminal; terminals are excluded from the agent grid/counts),
an Archive tab, the Others catch-all (attention queue, conflicts, activity;
sits last), in-browser attach terminals, a header-button Context & Messages
overlay, spawn modal, bulk actions, keyboard shortcuts, and theming.
| Feature | Where | Docs |
|---|
| URL routing (deep links, back/forward) | all routes | web-mission-control |
| Agent grid + Fleet header / busy-idle badges | Cockpit (/cockpit, home) | web-mission-control |
| Attention queue / approvals | Others (/others) | web-mission-control |
| Pipelines tab + live DAG | Pipelines (/pipelines) | web-mission-control |
| Metrics: per-agent + fleet-total CPU/mem, per-agent context, fleet size, tokens saved (2-col responsive) | Metrics (/metrics) | observability |
| Terminals tab (PTY viewport + New terminal; excluded from the agent grid/counts) | Terminals (/terminals) | web-mission-control |
| Archive (history) | Archive (/archive) | web-mission-control |
| In-browser attach terminal | Agent (/agent/<id>) | web-mission-control |
Full-screen TUI launcher — the three-pane cockpit streamed edge-to-edge into the browser (same panes, shortcuts, real shells & live agent sessions; Ctrl+Q exits); self-healing (survives daemon restarts; a wedged session is auto-rebuilt, or force one with warden tui --rebuild-web-cockpit) | top-bar ▢ TUI button (/tui) | web-mission-control |
| Per-agent backend logo on cards + group-by Agent (claude/aider/…; empty ⇒ claude) | Cockpit (/cockpit) | agent-backends |
| Context & messages | header 🗒 overlay | web-mission-control |
Spawn modal (+ New agent) — incl. a Role dropdown (built-in role picker, defaults general) | header button | agent-roles |
| Bulk actions | Bulk action bar | web-mission-control |
| Conflicts panel / activity feed | Others (/others) | web-mission-control |
| Keyboard shortcuts + help | global | web-mission-control |
REST API + OpenAPI (/api/docs) | daemon | api-openapi |
A terminal mission-control. Keys: n spawn · t new terminal in the cursor’s project · enter attach ·
i info/inspector · a approve · x terminate · D delete · d digest · r refresh ·
f filter · g/G top/bottom · o/p panes · s sort · c context ·
M-t/M-a/M-p rotate the terminal pane over terminals / the agent pane over agents / the agent pane over pipeline agents (add Shift — M-T/M-A/M-P — to reverse; each grabs focus on the pane it drives; or the config-free Ctrl-b prefix fallback — Ctrl-b then t/a/p — for terminals that don’t send Alt/Option as Meta, e.g. macOS Terminal.app / iTerm2) ·
? help · q quit. A ◆ and a bold name badge mark the agent/terminal currently shown in a pane (tracks both enter-open and the M-t/M-a/M-p rotation).
Includes a pipeline view and per-job info.
| Feature | Where | Docs |
|---|
| Agent list + live status (incl. per-agent backend token; empty ⇒ claude) | control pane → Agents section | tui-cockpit |
| Control-pane navigator tree — grouped by project; a project’s agents, pipelines, plans and terminals nest under it | control pane | tui-cockpit |
Terminals nest under their project (created on demand, never auto-spawned); t creates a terminal in the project under the cursor | control pane → project tree / terminal pane | tui-cockpit |
Inspector (i) — agent & pipeline detail | inspector | tui-cockpit |
Approvals cockpit (a) | cockpit | tui-cockpit |
Digest (d) | inspector | tui-cockpit |
| Pipeline view | pipeline pane | tui-cockpit |
Agent sub-trees (spawned agents nest under parent; h/l collapse; tombstone on parent delete) | main pane | tui-cockpit |
| Spawn / attach / terminate / delete | keybindings | tui-cockpit |
Role picker in the new-agent form (ctrl+r; built-in catalog, defaults general) | new-agent form | agent-roles |
Tier picker + live candidate table (ctrl+t; auto / tier-1/2/3; headroom bars; resolver picks backend+model) | new-agent form | agent-roles |
| Shift-to-select (native copy under tmux mouse mode) | help hint | tui-cockpit |
Native tmux window when launched inside tmux (no nesting; auto via $TMUX) | tui --tmux-native | tui-cockpit |
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| MCP server (stdio) | mcp | n/a (is the server) | ✓ | — | — | mcp-and-skill |
/warden Claude skill | shipped in skills/ | drives MCP/CLI | ✓ | — | — | mcp-and-skill |
| Interactive REPL | repl (aliases interactive, i) | CLI-only (interactive REPL) | — | — | — | repl |
↳ deterministic / commands (no model) | /agents, /spawn, /tell, … /help | — | — | — | — | repl |
↳ line editor: history, reverse-search, live / menu, Tab completion, colour | readline-backed | — | — | — | — | repl |
| ↳ guided argument forms (pick-lists + free text, Fast-Brain pre-fill) | bare /spawn, /spawn+ <prompt> | — | — | — | — | repl |
warden detects the coding-agent CLIs installed on this machine (claude, codex,
aider, …) and persists
each in an embedded ScrivaDB store (~/.warden/backends) with a billing tier,
an enabled flag, and at most one default. The store is warden’s single source of truth for
which backends exist and how they’re tiered: autopilot’s cost-tier ladder reads from it.
warden’s own internal thinking runs on Fast-Brain instead (thinking-mode and the reserved local row are retired). Detection is a fact
(installed / binary path / detected-at) a rescan reconciles; tiering is a
preference (tier / default / enabled) a rescan preserves.
Tiers: free · subscription · pay_per_use · unclassified .
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| List the registry (installed, tier, default, enabled, limited) | backends list (alias ls) | list_backends | ✓ | ✓ (🧩 backends panel) | b | backend-registry |
| Rescan installed CLIs (reconcile detection, preserve prefs) | backends rescan | rescan_backends | ✓ | ✓ (⟳ Rescan) | r | backend-registry |
| Set a backend’s billing tier (free|subscription|pay_per_use|unclassified) | backends tier <id> <tier> | set_backend_tier | ✓ | ✓ (Tier dropdown) | t (cycle) | backend-registry |
| Set the single default backend | backends default <id> | set_default_backend | ✓ | ✓ (Default radio) | d/enter | backend-registry |
| Enable / disable a backend | backends enable|disable <id> | — (REST PATCH /backends/{id}; no MCP tool) | ✓ (via CLI) | ✓ (Enabled checkbox) | e/space | backend-registry |
| Fast-Brain internal thinking (classification, naming, commit messages, digest narration, memory curation, REPL planning, crash/stall/CI diagnosis, and priority admission — latency-bounded, fail-open) | automatic | automatic | ✓ | — | — | fast-brain |
The registry supersedes the deprecated autopilot.brain.backends ladder and
autopilot.brain.allow_pay_per_use gate — those keys are imported once on the
first boot after upgrade, then ignored (edit tiers via the surfaces above).
| Feature | Status | Docs |
|---|
| Node device authorization | Shipped: warden login provisions a local certificate identity for a compatible Hub | warden login device flow |
| Node-side relay wire contract and accept-side scope enforcement | Shipped building blocks; Hub service and dial-out connector are planned | relay contract |
| Versioned daemon API shared by web and future clients | Shipped | node API client contract |
| Android app and Hub-routed mobile connection | Planned; use the mobile-responsive web dashboard today | Android connectivity status |
Plans are ScrivaDB-canonical records. Autopilot is the long-running execution
mode that gives a Plan a manager, isolated workers, a daemon-owned landing loop,
and one owner-merged final PR. See Plans, Using
plans, and Autopilot
for the model and safety boundaries.
| Feature | CLI | MCP | Skill | Web | TUI | Docs |
|---|
| Create, inspect, and list canonical Plans | plan create|show|list | create_plan, get_plan, list_plans | ✓ | ✓ | project tree | using-plans |
| Modify a pending Plan definition or task DAG (revision-checked; non-pending plans reject changes) | plan update|edit|task add|edit|rm | update_plan | ✓ | ✓ | — | Plans |
| Run a Plan in autopilot, pipeline, orchestrator-worker, or manual mode | plan run <id> --mode … | run_plan | ✓ | ✓ | r | using-plans |
| Track task status and executor state | plan task status, plan show --watch | update_task_status | ✓ | ✓ | project tree | using-plans |
| Pause, resume, stop, or destructively restart an executor | plan pause|resume|stop|restart | control_plan, restart_plan | ✓ | ✓ | — | Autopilot guide |
| Archive / unarchive a Plan (preserve unmerged branches and task progress) | plan archive|unarchive | archive_plan, unarchive_plan | ✓ | ✓ | a | using-plans |
| Autopilot manager / worker / resolver topology, guardian, overwatch, and usage-limit recovery | automatic; autopilot status | autopilot_status, brain_consult | ✓ | ✓ | project tree | Autopilot concepts |
| Daemon-owned gate/CI fix loop and worker landing to the per-plan integration branch | automatic; autopilot land is an escape hatch | land | ✓ | ✓ | — | Autopilot guide |
| One final integration → default-branch PR; daemon gates and polls, owner merges | automatic; autopilot status | autopilot_complete (manager-only) | ✓ | ✓ | — | Autopilot concepts |
These operate on the host, the daemon process, the local shell, or the bearer
secret — they are intentionally not exposed over MCP or web, because doing so
would either be meaningless (process/host control) or a security smell (handing
out / rotating the very token that guards the MCP and HTTP channels).
| Feature | CLI | Why CLI-only | Docs |
|---|
| Run / manage the daemon | daemon | process control on the host | install |
| Bearer token generate / show / rotate (+ read-only token) | token | the secret that protects every other surface; show --readonly for the view-only WARDEN_READONLY_TOKEN | remote-access |
| Local-LLM model picker (memory-ranked) | llm suggest | reads host hardware to size the orchestrator model | repl |
| Configuration view / init / path | config | local file authoring | env-vars |
Data schema ledger + boot guard (<data>/schema.json: integer schema_version separate from semver; daemon refuses a data dir it must not open — newer, needing migration, or mid-migration — before touching a store; pre-ledger installs are stamped once at the schema-1 baseline; fresh installs start on schema 2 (single warden-db); schema-1 data is cut over once by the journaled, verified migrate --apply that update runs, with full rollback; schema_version on /healthz, checked by update) | automatic (daemon; version prints the binary’s schema); migrate --plan/--apply/--resume/--restore, update | a property of the on-disk data dir and the daemon process | troubleshooting |
| Update and data-safety operator runbook (read-only update and migration planning; rollback, offline repair, scoped factory reset, and backup-gated legacy-store cleanup) | update --plan, rollback, migrate, repair, factory-reset | host data lifecycle and recovery | updates-and-data-safety |
Planned upgrade & waypoint path planner (warden update --plan / wd update --plan: evaluates release manifest, computes minimal waypoint hops, reports breaking API changes, preflights whole-data stores, and estimates downtime/disk without mutating state) | update --plan (alias wd update --plan) | host binary and data upgrade planning | install |
Migration registry runner (warden migrate / wd migrate: --check, --apply, --resume, --restore; idempotent, journaled in schema ledger, verified before advancing ledger; daemon never runs destructive migrations at boot) | migrate (alias wd migrate) | offline data format upgrade execution | install |
Update transaction snapshot & rollback (warden rollback / wd rollback: data snapshot taken in <data>/backups/pre-<ver>-<ts>/ before update with hardlink/copy and checksums, journaled in schema ledger; rollback performs plain binary swap when no schema change, or restores snapshot with loss-of-changes confirmation; 2 updates / 14 days retention) | rollback (alias wd rollback) | transactional upgrade safety | install |
Whole-store repair (warden repair all --resolve-history=live-wins: offline check and repair for all ScrivaDB stores, quarantine backup, resolving duplicate live keys and revision-regression conflicts) | repair all | offline whole-data repair; daemon must be stopped | install |
Data directory init & install parity (warden init: stamps schema ledger, ensures all store directories, runs baseline migrations; shared code path with scripts/install.sh) | init | idempotent host data setup | install |
| Health / environment doctor | doctor (--reconcile-membership offline membership repair) | host diagnostics | troubleshooting |
| Install missing dependencies | setup | installs missing host dependencies such as tmux, git, gh (via the OS package manager + official installers) | install |
| First-run tutorial | tutorial | interactive walkthrough | quickstart |
| Shell completion | completion | shell integration | install |
| Hook entry points (guards) | hook / *-guard | invoked by Claude Code hooks | lifecycle-and-rails |
Scoped factory-reset (--scope runtime|data|full; --backup, --keep-config, --keep-backends, --prune-worktrees; requires --yes) | factory-reset | destructive host wipe; daemon must be stopped for the data phase | cli |
| Version | version | — | — |
Every fleet/data feature is reachable over MCP (including the
umbrella stop_agent). The only
CLI-exclusive features are the host/process/interactive/secret commands in
§17 (plus interactive attach/repl, the local-config preset /
prompt-template authoring commands, and the agent-native, worktree-local
review / models verbs — they exec in the agent’s worktree with no daemon
round-trip, so they have no MCP twin by design), which are
CLI-only by design. New parity tools added for full coverage: digest,
get_metrics, savings, spend, search, history, audit_log, list_worktrees,
list_plugins, get_pressure, set_auto_approve, set_auto_approve_policy,
set_permission_mode,
prune_worktrees, export_sessions, import_sessions, rotate_agent,
handoff_agent, pause_pipeline, resume_pipeline, retry_pipeline_job,
edit_pipeline_job, emit_pipeline_output, delete_pipeline,
validate_pipeline, list_pipeline_templates, library_list,
create_schedule, get_schedule, update_schedule, run_schedule, enable_schedule, disable_schedule, delete_schedule, fork_agent, set_role, list_roles,
list_backends, rescan_backends, set_backend_tier, set_default_backend. (set_thinking_mode is a retired no-op.) (Enable/disable a backend is CLI/web/TUI + REST PATCH /backends/{id} only — intentionally not an MCP tool.)
