paseo-companion

Paseo Companion

Paseo is a daemon-managed CLI for launching, observing, and steering AI coding agents. Every agent gets a stable ID that serves as the handle for all subsequent operations — follow-ups, logs, waiting, archiving.

This skill is a runtime adapter. It documents how to drive the Paseo CLI; it does not decide whether a task should run through Paseo at all, on the local daemon or a remote one, or with which model. Those choices belong to the orchestration layer (team-lead-orchestration) and any local routing profile that applies (e.g. a personalized routing skill).

paseo <command> [options]

Typical workflows

Run a single task and wait for completion

paseo run "implement the new auth flow"

Blocks until the agent finishes. Add --wait-timeout 30m to cap the wait.

Send a follow-up to the same agent

Reuse before relaunch — if an agent already exists for related work, continue it:

paseo send <id> "now add tests for the new endpoint"

<id> accepts a unique prefix or the agent name, not just the full UUID.

Run multiple agents in parallel

Detach agents, do other work, then join:

api_id=$(paseo run -d --json --title api-lane "implement the API" | jq -r .id)
ui_id=$(paseo run -d --json --title ui-lane "implement the UI"  | jq -r .id)

# ... do other work ...

paseo wait "$api_id"
paseo wait "$ui_id"

If wait times out, the agent is still running. Inspect state or wait again; do not relaunch the same work.

Target a remote daemon

When a remote daemon is exposed through a tunnel or Daytona preview, pass --host to every command that should address that daemon:

paseo provider ls --host "$REMOTE_PASEO_HOST"
paseo run --host "$REMOTE_PASEO_HOST" -d --provider opencode --mode orchestrator --cwd /workspace --prompt-file ./task.md
paseo wait --host "$REMOTE_PASEO_HOST" --timeout 1800 <id>
paseo logs --host "$REMOTE_PASEO_HOST" <id> --tail 20

Use the host:port form, not a full http://... preview URL.

Isolate work in a git worktree

paseo run --worktree feature-x "implement feature X"
paseo run --worktree experiment-y --base develop "try approach Y"

Iterate until tests pass

paseo loop run "fix the failing tests" \
  --verify-check "npm test" \
  --max-iterations 10

Use --verify "<prompt>" for an agent-based verifier instead of a shell command.

Monitor a running agent

paseo logs <id> -f            # stream live output
paseo attach <id>             # interactive stream (Ctrl+C detaches, doesn't stop)
paseo inspect <id>            # detailed metadata snapshot

List and manage agents

paseo ls                      # active agents
paseo ls -a                   # include archived
paseo stop <id>               # interrupt a running agent
paseo archive <id>            # soft-delete
paseo delete <id>             # hard-delete

Common options

--provider <provider/model>   # pick provider/model (decided by routing profile, not this skill)
--cwd /path/to/repo            # set working directory
--host 10.0.0.8:6767           # target a remote daemon
--json                         # machine-readable output
-d                             # detach (return immediately, print agent ID)
--title auth-lane              # stable human name for tracking
--label area=backend           # metadata for filtering
--prompt-file ./task.md        # read prompt from file (for long/complex prompts)
--image screenshot.png         # attach an image to the prompt

Command map

Goal	Command
Launch and wait	paseo run "PROMPT"
Launch detached	paseo run -d "PROMPT"
Continue an agent	paseo send <id> "PROMPT"
Wait for completion	paseo wait <id>
Stream logs	paseo logs <id> -f
Attach interactively	paseo attach <id>
List agents	paseo ls [-a]
Stop/archive/delete	`paseo stop
Target remote daemon	add --host <ip>:<port>

When to read references

You need to...	Read
Inspect, stop, archive, delete agents; update metadata or labels	references/agent-management.md
Discover providers/models, select a provider, switch modes, enable thinking	references/providers-and-modes.md
Use git worktrees for isolation, manage Paseo-created worktrees	references/worktree-and-cwd.md
Set up verification loops or recurring scheduled tasks	references/loop-and-schedule.md
Create persistent terminals, send keystrokes, capture output	references/terminal.md
Set up inter-agent chat rooms or handle permission requests	references/chat-and-permit.md
First-time setup, daemon start/stop/restart, connect to remote daemon	references/daemon-and-onboarding.md
Drive agents on a remote daemon through preview/tunnel host routing	references/remote-host-orchestration.md
Script/automate Paseo output, JSON/YAML formats, schema validation	references/output-formats.md

Non-negotiables

• Reuse before relaunch. If an agent already exists for related work, paseo send to it — don't spin up a new one.
• Wait, don't poll. Never loop on paseo ls / paseo inspect. Use paseo wait <id> (blocks efficiently) or paseo logs <id> -f (streams).
• Timeout doesn't stop. If wait times out, the agent is still running. Use paseo stop <id> to actually interrupt.
• Carry --host through. A remote agent ID is only useful when later wait, logs, send, and inspect target the same daemon host.
• Quote prompts. For multi-line or escape-heavy prompts, use --prompt-file.