Console

The Console is the dashboard at http://localhost:3001 when you run the stack locally. It has two spaces, switched at the top of the sidebar: PAL, your personal agent, and PAL Runtime, where you build, ship and watch agents. The switch remembers your last space. Every number on these pages is a real API read; a source that fails shows — or an error with a retry, never a made-up value.

Keyboard: ⌘K asks PAL; g o Overview, g a Agents, g r Runs, g v Sandboxes, g t Today, g h Holds, g s Settings; ? shows the cheat sheet.

PAL Runtime

Overview

The Runtime home reads like an on-call handover: one sentence on whether things are fine, the four numbers that matter (runs, success, spend, sandboxes), what ran when (48 half-hour slots), and what needs attention. Each recent failure has an inline Why? that asks the server for an explanation grounded in that run’s real events. Cross-tenant ids are a 404.

Agents and the lifecycle tabs

An agent page lays the work out as a sequence. The numbered rail is the lifecycle; the unnumbered tabs are configuration. The URL carries the tab (?tab=), and old links such as ?tab=runs still resolve.

Tab	Answers	Related
Design	Prompt, model, tools	Agents
Test	Talk to it	Testing
Evals	Suites and scores	Evaluations
Gate	Ready to ship?	Promotion
Promote	Sandbox, staging, prod	Promotion
Observe	Runs and cost	Observability
Debug	Why it failed	Debugging and support
Schedule	Cron schedules	Scheduling
Identity & policy	Who the agent is and what it may do	Agent identity and policy
Settings	Everything else

Runs

The run list and a run’s detail page, with the event stream and the step waterfall. A run’s trace is the redacted tree of steps, model calls, tool calls and child runs, built from the journal (GET /v1/runs/{id}/trace). Explain on a run gives a short, grounded reading of what happened.

Reliability and SLOs

Each agent has a Reliability page (/agents/<name>/reliability): SLO targets with error-budget burn, a run-activity chart, and opt-in sampled production evals with per-version drift. All verdicts, such as burn alerts and drift, come from the server; the page lays them out.

Incidents

/incidents lists anomalies and SLO burn found by the observability sweep: cost, latency, error rate, tool use or quality moving well outside an agent’s own baseline. Detection is deterministic. A model may write the summary, but only if it cites the incident’s own runs. Acknowledging and resolving are recorded in an audit table.

Debug

The Debug tab points the debug agent at a run, an incident or a failing eval case. It reads the evidence the server collects (run trace and journal, the failing version’s snapshot against the last good one, promotion audit, failed eval cases, budget, open incidents) and returns a root cause that cites it, plus fix proposals. Nothing is applied until a person confirms. The CLI equivalent is pal debug (CLI). Replaying a past failure against a candidate version is not a server action yet; pal test --rehearse does it from the terminal.

Promotions

/promotions is tenant-wide: what is waiting on a human, and the append-only audit of who shipped what to sandbox, staging and prod. It uses the same endpoints and approval flow as an agent’s Promote tab. Approving needs an owner or admin session.

Sandboxes

/runtime lists your microVMs: state, placement and lifecycle (pause, resume, fork, terminate), polled every 5 seconds. If a poll fails it keeps the last good rows and says they are stale. See Sandbox lifecycle.

The rest of the Runtime sidebar: Deployments, Evals, Experiments, Analytics, Budgets, Connectors, Channels and Marketplace.

PAL

The PAL space is your personal agent’s home. It opens on Today, a ranked “what needs you” list: held replies you can approve, edit or discard in one click, suggested life events, open commitments and threads waiting on you. The ranking is a deterministic order plus one model pass that may only reorder and annotate.

Page	What is there
Today	The Brief
Holds	Replies PAL is waiting on your yes for (Holds and trust)
Automations	Life-event feed and per-kind trust (auto, ask, off)
People & memory	Contacts, VIPs and groups
Activity	Recent activity
Documents	Your personal documents
Voice	How PAL learns to sound like you (Teach it your voice)
Channels	iMessage and WhatsApp (Set up PAL)
Assistant	The assistant’s controls

First sign-in lands on Welcome, where you pick PAL or PAL Runtime. Both paths open a conversation with the onboarding agent; a step-by-step wizard is the labelled fallback.

Ask PAL (⌘K)

The command palette takes plain language and turns it into one validated action: navigate, answer, explain a run, create an agent, or pause, resume or terminate a sandbox. Routes are allowlisted, ids must come from the context the model was given, and any change stops at a confirm card first. The browser never calls a model provider; it goes through the control plane.

Settings

Settings has its own navigation: General, Model providers, API keys, Team, Devices, Billing, MCP, Data protection and Audit log. Some entries show only to roles that can use them.

• Devices (/settings?tab=devices): pair a Mac or phone to your workspace with a one-time code, list paired devices, revoke one.
• Data protection (/settings/dlp): the egress DLP rules the sandbox broker applies to outbound requests. Anyone can read them; changes need settings:write. Matched values are never logged, only the decision. See Egress.
• Audit log (/settings/audit): owner and admin only. There is no single tenant-wide audit endpoint, so this page merges the two that exist, promotions and sandboxes, and says so. Per-agent policy audit is on the agent’s Identity & policy tab.

Theme

Dark is the default. The ☀/☾ toggle in the top bar switches to light and the choice persists in your browser.