fusion/docs/cli-reference.md

CLI Reference

← Docs index

Fusion’s command-line interface is exposed through the fn command.

Global Usage

fn <command> <subcommand> [options]

Global options

Option	Description
--project <name>, -P <name>	Target a specific registered project.
--help, -h	Show help output.

Project resolution order

When --project is not supplied, Fusion resolves project context in this order:

1. Explicit --project flag
2. Default project (set via fn project set-default <name>)
3. Current-directory auto-detection (.fusion/fusion.db lookup upward)

---

fn init

Initialize a new Fusion project in the current directory.

fn init
fn init --name my-project --path /absolute/path/to/project

During fresh initialization, Fusion also installs the bundled fusion skill into supported local agent homes when the target skill does not already exist:

• ~/.claude/skills/fusion
• ~/.codex/skills/fusion
• ~/.gemini/skills/fusion

fn init is non-destructive for these installs:

• Existing fusion skill directories are preserved (not overwritten).
• Per-target filesystem/permission failures are reported as warnings and do not fail project initialization.

---

fn update

Check for and install the latest @runfusion/fusion CLI release from npm.

fn update
fn update --check
fn update --global
fn update --json
fn upgrade

Option	Description
--check	Check only. Does not install. Exit code 1 when an update is available.
--global	Explicitly install globally (npm install -g @runfusion/fusion@latest). This is the default behavior.
--json	Output machine-readable status: currentVersion, latestVersion, updateAvailable, updated.

fn upgrade is an alias for fn update.

---

fn research

Manage persisted research runs from the CLI.

fn research create --query "Compare sqlite WAL vs rollback journal"
fn research create --query "Rust async runtime trade-offs" --wait --max-wait-ms 120000
fn research list --status failed --limit 20
fn research show RR-001
fn research export RR-001 --format json --output ./artifacts/research-RR-001.json
fn research cancel RR-001
fn research retry RR-001 --json

Subcommand	Description
fn research create --query <text> [--wait] [--max-wait-ms <ms>] [--json]	Create a run and optionally wait for completion.
fn research list | ls [--status <status>] [--limit <n>] [--json]	List recent runs (statuses: queued, running, cancelling, retry_waiting, completed, failed, cancelled, timed_out, retry_exhausted).
fn research show <run-id> [--json]	Show one run with timestamps, summary, and error details.
fn research export <run-id> [--format <json|markdown|pdf>] [--output <path>] [--json]	Export run results and persist an export record.
fn research cancel <run-id> [--json]	Request cancellation for an active run.
fn research retry <run-id> [--json]	Create a new retry run from a failed/timed_out run when lifecycle marks it retryable.

Disabled/setup behavior mirrors dashboard and agent surfaces:

• Feature disabled → FEATURE_DISABLED (enable project/global research settings)
• Missing credentials → MISSING_CREDENTIALS (configure provider auth)
• Provider unavailable or cooldown/rate limit → PROVIDER_UNAVAILABLE / RATE_LIMITED with retry metadata
• Invalid cancel/retry transitions are reported explicitly (INVALID_TRANSITION) with current status context
• Retry budget exhaustion and non-retryable failures are reported explicitly (RETRY_EXHAUSTED, NON_RETRYABLE_PROVIDER_ERROR)
• Non-retryable failures and invalid state transitions are surfaced as structured errors instead of generic failures

---

fn dashboard

Start the web dashboard (default port 4040, bound to 127.0.0.1).

fn dashboard
fn dashboard --port 5050
fn dashboard --host 0.0.0.0                # expose on LAN (use with care)
fn dashboard --token fn_yourStaticToken    # reuse a fixed token
fn dashboard --no-auth                     # disable bearer auth (local only)
fn dashboard --interactive
fn dashboard --paused
fn dashboard --dev

Option	Description
--port, -p	Dashboard HTTP port (default 4040).
--host	Host to bind (default 127.0.0.1, localhost only). Pass 0.0.0.0 to expose on all interfaces.
--token <token>	Bearer token to use. Default: $FUSION_DASHBOARD_TOKEN → $FUSION_DAEMON_TOKEN → auto-generated.
--no-auth	Disable bearer-token auth. Not recommended when binding to 0.0.0.0.
--paused	Start with the engine paused (automation disabled).
--interactive	Interactive port selection.
--dev	Start dashboard only (no AI engine, no planning/scheduler).

Interactive Terminal UI (TTY Mode)

When running in an interactive terminal (TTY), fn dashboard starts an interactive TUI with sectioned views for system status, logs, settings, and remote-access controls.

Remote controls are available inside Interactive → Settings in the detail pane. Remote actions support:

• Switching active provider (tailscale / cloudflare) and explicit activation
• Manual tunnel lifecycle (start / stop)
• Persistent token regeneration (masked token display)
• Short-lived token generation with TTL input and expiry display
• URL + QR hand-off (always shows full authenticated URL)

⚠️ Remote URL/QR payloads include tokenized query data. Treat them like credentials and avoid sharing them in screenshots/chat/logs. Prefer short-lived links for ad-hoc phone login.

Remote action keys in Settings detail pane:

• C activate selected provider
• V start tunnel
• X stop tunnel
• P regenerate persistent token
• L enter TTL input mode and generate short-lived token
• U generate authenticated URL hand-off
• K request QR payload hand-off
• R refresh remote status/snapshot

Engine/runtime remote tunnel semantics used by dashboard + serve + TUI:

• Lifecycle states: stopped → starting → running → stopping (or terminal failed)
• Start/stop is process-supervised (spawn, SIGTERM, 5s default timeout, then SIGKILL)
• Provider switch is stop-first: the current provider is fully stopped before target startup is attempted
• Failed switch/start emits explicit failure status (switch_failed / invalid_config / start_failed) and never runs both providers concurrently
• Status/log subscribers receive redacted events (token-bearing args/env/log text masked)

QR hand-off behavior in TUI:

• format="text": renders the text payload directly
• format="image/svg": does not render raw SVG in terminal; shows the authenticated URL, expiry metadata, and a fallback instruction to open the URL on phone/browser

On startup, the TUI opens on the System section by default so you can immediately see host/port and access-token details.

Keyboard Navigation:

Key	Action
1-5	Switch to tab by number
n or →	Next tab
p or ←	Previous tab
r	Refresh stats (in Utilities tab)
c	Clear logs (in Utilities tab)
t	Toggle engine pause (in Utilities tab)
? or h	Toggle help overlay
q	Quit
Ctrl+C	Force quit

Logs Tab Navigation:

Key	Action
↑ or k	Move selection to older log entry
↓ or j	Move selection to newer log entry
Home	Jump to first log entry
End	Jump to last log entry
Enter, Space, or e	Toggle expanded view for selected entry
Esc	Close expanded view
w	Toggle wrap mode (long messages wrap vs. truncate)
f	Cycle severity filter (all → info → warn → error → all)

The Logs list uses a scrollable viewport across the full in-memory ring buffer. As long as an entry is still inside the buffer, you can reach it with ↑/↓ (or k/j) and jump to absolute bounds with Home/End.

Severity filtering is a view-only control for the current Logs pane. Cycling with f narrows the rendered list to the selected level (info, warn, or error), but all entries remain in the ring buffer and are shown again when you return to all.

In wrapped mode, long log messages are displayed with word wrapping. Long unbroken tokens (such as URLs or stack traces) are hard-wrapped at the available width. In expanded view, the full message is shown with complete wrapping for inspection.

During interactive TTY mode, streamed runtime text (including merge-session agent output) is routed into the Logs tab as stable line entries instead of being written directly into the alternate-screen terminal surface.

In non-TTY mode (CI, piped output, scripts), the dashboard falls back to plain console output to maintain compatibility with automated workflows.

Authentication

Unless --no-auth is passed, the dashboard API (including the terminal WebSocket) is protected by a bearer token. On first authenticated startup, Fusion resolves a token via the daemon-token manager and persists it in the existing global settings file (~/.fusion/settings.json, owner-only when supported). Later dashboard startups reuse the same stored token unless you explicitly override it.

On startup, Fusion prints both the resolved token and a click-to-open URL that embeds ?token=<token>:

fn dashboard
────────────────────────
→ http://localhost:4040
Auth:    bearer token required
Token:   fn_8f3a...
Open:    http://localhost:4040/?token=fn_8f3a...
         (the browser stores the token so you only need to click once)

On first visit the dashboard captures the token from the URL into localStorage (key fn.authToken) and strips it from the visible URL so the secret does not end up in browser history. Subsequent loads (including closing and reopening the tab) reuse the stored token.

Precedence when resolving the token:

1. --no-auth (disables auth middleware entirely)
2. --token <token> flag
3. FUSION_DASHBOARD_TOKEN environment variable
4. FUSION_DAEMON_TOKEN environment variable (back-compat with fn daemon)
5. Stored token in ~/.fusion/settings.json
6. New generated token persisted to ~/.fusion/settings.json (first authenticated run)

To override defaults without changing stored settings, export one of the env vars:

export FUSION_DASHBOARD_TOKEN=fn_my_override_token
fn dashboard

To revoke/reset access, choose the behavior you want:

• Temporary override: set --token / env var for the current run.
• Persistent reset: clear daemonToken from ~/.fusion/settings.json (or rotate it via fn daemon --token-only/token rotation workflow), then restart dashboard.
• Client logout: clear fn.authToken in browser localStorage so clients must re-authenticate with the current server token.

Optional provider: Factory AI via Droid CLI

When the published CLI bundle includes the vendored @fusion/droid-cli extension, users can enable Factory AI — via Droid CLI in Settings → Authentication.

Requirements:

• a working Droid CLI binary (droid on PATH by default, or a custom plugin droidBinaryPath)
• successful local login (droid auth login)
• Fusion restart after toggling the provider on (to reload extensions)

Authentication status checks in Settings → Authentication use the same effective Droid binary path as the Droid runtime plugin, so custom binary-path installs are detected correctly.

---

fn serve

Start Fusion as a headless node (API server + AI engine, no frontend UI).

fn serve [--port <port>] [--host <host>] [--paused] [--daemon]
fn serve --interactive

Option	Description
--port, -p	Port for the API server (default 4040).
--host	Host to bind (default 127.0.0.1, localhost only). Pass 0.0.0.0 to expose on all interfaces.
--paused	Start with engine paused (automation disabled).
--interactive	Interactive port selection.
--daemon	Enable bearer token authentication for CLI client connections.

fn serve uses the same project-scoped Remote Access manager as fn dashboard. When remote access is enabled/configured, the headless server exposes /api/remote/* control/status endpoints and applies the same hybrid token validation rules for remote routes (persistent token + optional short-lived token registry).

Headless operators should use the same lifecycle/API flow as dashboard mode:

• POST /api/remote/provider/activate
• POST /api/remote/tunnel/start
• POST /api/remote/tunnel/stop
• GET /api/remote/status
• POST /api/remote-access/auth/login-url

GET /remote-login?rt=<token> is intentionally public for phone-link handoff, but token validity is still enforced server-side.

For end-to-end setup, risk guidance, and troubleshooting, see docs/remote-access.md.

For programmatic consumers, these endpoints map to the engine tunnel manager contract:

• getStatus() for current snapshot
• start(provider, config) / stop() / switchProvider(...)
• subscription hooks for live status and log updates (used by stream/poll clients)

---

fn daemon

Start Fusion daemon (API server + AI engine, always requires bearer token authentication).

fn daemon [--port <port>] [--host <host>] [--token <token>] [--paused] [--interactive] [--token-only]

Option	Description
--port, -p	Port for the daemon server (default: auto-assigned).
--host	Host to bind (default 127.0.0.1, localhost only). Pass 0.0.0.0 to expose on all interfaces.
--token	Set a specific daemon token. If not provided, a random token is generated and printed.
--paused	Start with engine paused (automation disabled).
--token-only	Only generate/show the token without starting the server.
--interactive	Interactive port selection.

---

fn desktop

Launch the Fusion desktop app (Electron).

fn desktop
fn desktop --dev
fn desktop --paused
fn desktop --interactive

Option	Description
--dev	Launch with hot-reload (connects to Vite dev server).
--paused	Launch with automation paused.
--interactive	Interactive port selection.

---

fn task

Task lifecycle and task operations.

Creation and planning

fn task create "Fix login race condition"
fn task create "Fix bug" --attach screenshot.png --depends FN-010
fn task create "Investigate flaky runner" --node edge-runner
fn task plan "Design a new authentication flow"

For AI-guided task specification, see Planning mode.

Planning mode

Use planning mode to turn a rough idea into a triage task through an interactive AI-guided Q&A flow.

fn task plan [description]

description is optional. If you omit it, the CLI prompts for an initial idea (Describe your idea:) before creating the planning session.

Planning questions are interactive and use these types:

• text (multi-line; finish with DONE on its own line)
• single_select (pick one option)
• multi_select (pick one or more comma-separated options)
• confirm ([Y/n], Enter defaults to yes)

Planning flow:

1. Create planning session from your description/idea.
2. Answer the current question.
3. Receive either a follow-up question or completion summary.
4. Review summary (title, description, suggested size, dependencies, key deliverables).
5. Confirm creation (or skip confirmation with --yes).
6. Task is created in triage when confirmed.

• With --yes, final confirmation is skipped and the task is created immediately.
• Without --yes, the CLI asks Create this task? [Y/n]:; answering no cancels creation.

Option	Description
--yes	Skip final confirmation before creating the planned task.
--project <name>, -P <name>	Run planning mode against a specific registered project.

Planning session limit: maximum 1000 planning sessions per hour.

Query and logs

fn task list
fn task show FN-001
fn task logs FN-001 --follow --limit 50 --type tool

fn task logs now exposes full agent-log content for each entry type. In particular, thinking, tool_result, and tool_error entries preserve full multiline output (including stderr/stack details) so you can inspect raw tool responses directly from the CLI stream.

fn task show <id> includes routing and provenance context when available:

• task node override
• project default node fallback
• unavailable-node policy value
• source provenance line (Source: <origin>), including parent task / GitHub issue URL context when present

Execution and status

fn task move FN-001 in-progress
fn task update FN-001 2 done
fn task log FN-001 "Updated API contract"
fn task retry FN-001
fn task pause FN-001
fn task unpause FN-001

Node routing controls

fn task set-node FN-001 edge-runner
fn task clear-node FN-001

Notes:

• set-node resolves either node name or node ID.
• set-node and clear-node are blocked while the task is in progress.
• Use fn node list / fn node show <name> to discover node IDs and status.

Collaboration and guidance

fn task comment FN-001 "Needs stricter validation"
fn task comment FN-001 "Reviewed with QA" --author "alex"
fn task comments FN-001
fn task steer FN-001 "Reuse existing auth middleware"

Completion, maintenance, and history

fn task attach FN-001 ./trace.log
fn task merge FN-001
fn task duplicate FN-001
fn task refine FN-001 --feedback "Add rollback handling"
fn task archive FN-001
fn task unarchive FN-001
fn task delete FN-001 --force

GitHub integration

fn task pr-create FN-001 --title "Fix login race" --base main
fn task import owner/repo --labels bug --limit 10
fn task import owner/repo --interactive

---

fn project

Manage registered projects in multi-project mode.

fn project list --json
fn project add my-app /path/to/app --isolation child-process
fn project show my-app
fn project info my-app
fn project set-default my-app
fn project detect
fn project remove my-app --force

Subcommands: list|ls, add, remove|rm, show, info, set-default|default, detect.

---

fn node

Manage external execution nodes.

fn node list --json
fn node connect edge-runner --url https://node.example.com --api-key $NODE_API_KEY --max-concurrent 4
fn node disconnect edge-runner --force
fn node show edge-runner
fn node health edge-runner

Subcommands: list|ls, connect, disconnect, show|info, health.

---

fn mesh

Mesh network status.

fn mesh status [--json]

Subcommands: status.

---

fn mission

Mission hierarchy operations.

fn mission create "Platform hardening" "Security and reliability initiative"
fn mission list
fn mission show mission_123
fn mission delete mission_123 --force
fn mission activate-slice slice_456

Subcommands: create, list|ls, show|info, delete, activate-slice.

---

fn agent

Agent runtime operations.

fn agent stop AGENT-001
fn agent start AGENT-001
fn agent mailbox AGENT-001
fn agent import <source> [--dry-run] [--skip-existing]
fn agent export ./output-dir --company-name "My Company" --company-slug my-company

Subcommands: stop, start, mailbox, import, export.

fn agent import

Import agents from companies.sh packages. Supports single manifest files, team packages, and archives.

Source formats:

• Single AGENTS.md manifest file
• Companies.sh package directory with COMPANY.md, TEAM.md, and AGENTS.md
• Archive files (.tar.gz, .tgz, .zip)

Options:

Option	Description
--dry-run	Preview import without creating agents or skill files
--skip-existing	Skip agents with names that already exist in Fusion

Team hierarchy: When importing a companies.sh package with team structure, the importer preserves manager/report relationships for both fresh and partial imports. Manifest-style manager references such as ceo, ../ceo/AGENTS.md, and already-valid Fusion agent IDs are resolved to actual Fusion reportsTo agent IDs before agents are created, and --skip-existing reuses matching existing managers when available instead of flattening the org tree.

Skill imports: When importing from a package directory or archive, the importer also imports any package skill manifests (skills/*/SKILL.md). Skills are written to {project}/skills/imported/{company-slug}/{skill-slug}/SKILL.md. Existing skill files at the target path are skipped (not overwritten). Single AGENTS.md file imports do not include package skills.

The import summary reports:

• Skills imported (new skills written)
• Skills skipped (already exist at target path)
• Skill errors (invalid manifests or write failures)

Examples:

# Import a single agent manifest
fn agent import ./ceo/AGENTS.md

# Import a full companies.sh package (includes agents and skills)
fn agent import ./my-company/

# Import from archive
fn agent import ./package.tar.gz

# Preview without creating
fn agent import ./package/ --dry-run

# Skip existing agents
fn agent import ./package/ --skip-existing

---

fn message

User mailbox operations for sending and managing direct messages with agents.

fn message inbox
fn message outbox
fn message send AGENT-001 "Please prioritize FN-222"
fn message read MSG-123
fn message delete MSG-123

Subcommand	Description
fn message inbox	List your inbox messages (newest first, up to 20).
fn message outbox	List messages you sent (newest first, up to 20).
fn message send <agent-id> <content>	Send a user→agent message and print the created message ID.
fn message read <id>	Show one full message by ID and auto-mark it as read if unread.
fn message delete <id>	Permanently delete one message by ID.

Mailbox behavior

• inbox header shows unread totals as Inbox (<count> unread).
• Unread inbox rows are prefixed with ●; read rows have no dot.
• Inbox sender labels use Agent <id> for agent senders and raw user IDs for user senders.
• Outbox recipient labels use Agent <id> for agent recipients.
• Inbox/outbox previews are truncated to 80 characters with a trailing ellipsis (…).
• send success output includes ✓ Message sent: <message-id> plus the destination agent.
• read prints full metadata (Message, Type, From, To, Time) and the complete message body.
• read exits with code 1 when the message ID is not found.
• delete removes the message immediately and prints ✓ Message <id> deleted.

Options

Option	Description
--project <name>	Route mailbox operations to a specific registered project (resolved via project context). Supported by all fn message subcommands.

Related command

fn agent mailbox <agent-id> is separate from fn message: it inspects an agent-owned mailbox (agent inbox view), while fn message ... manages the CLI user mailbox.

---

fn settings

Show and manage settings.

fn settings
fn settings set maxConcurrent 4
fn settings set defaultNodeId node_abc123
fn settings set unavailableNodePolicy fallback-local
fn settings export [--scope global|project|both] [--output <file>]
fn settings import <file> [--scope global|project|both] [--merge] [--yes]

Option	Description
--scope	Scope selector for settings export and settings import: global, project, or both (default: both).
--output	Custom output file path for settings export.
--merge	Merge imported values with existing settings (used by settings import).
--yes	Skip confirmation prompt during settings import.

---

fn git

Project git operations.

fn git status
fn git fetch
fn git fetch upstream
fn git pull --yes
fn git push --yes

---

fn backup

Database backup lifecycle.

fn backup --create
fn backup --list
fn backup --restore .fusion/backups/fusion-2026-04-08.db
fn backup --cleanup

---

fn plugin

Plugin lifecycle management.

fn plugin list
fn plugin install <path>
fn plugin uninstall <id> --force
fn plugin enable <id>
fn plugin disable <id>
fn plugin create <name>

Subcommands: list|ls, install, uninstall, enable, disable, create.

---

fn skills

Browse and install agent skills from skills.sh.

fn skills search <query> [--limit <n>]
fn skills install <owner/repo> [--skill <name>]

Subcommands: search, install.

Option	Description
--limit	Max search results (default: 10, max: 50). Used by search.
--skill	Install a specific skill by name. Used by install.

---

Useful option flags by context

Option	Used by
--project, -P	Most project-scoped commands (for example: fn task ..., fn message ..., fn agent mailbox, fn settings, fn research, fn mission, fn node, fn plugin, fn skills)
--port, -p	fn dashboard, fn serve, fn daemon
--host	fn serve, fn daemon
--interactive	fn dashboard, fn serve, fn daemon, fn desktop, fn task import, fn project add
--paused	fn dashboard, fn serve, fn daemon, fn desktop
--dev	fn dashboard, fn desktop
--attach	fn task create
--depends	fn task create
--node	fn task create
--feedback	fn task refine
--yes	confirmation-skipping flows (task plan, settings import, git pull/push, etc.)
--limit, -l	fn task import (default: 30, max: 100), fn skills search (default: 10, max: 50)
--labels, -L	fn task import
--skill	fn skills install

For configuration details used by these commands, see Settings Reference.