work.json

A fullscreen-scaling layout fills a 4K wall TV or reflows to a single-column phone — each surface keeps its own per-device layout.

Goal-tracked KPIs with RAG status, leaderboards, an agent-of-the-period crown, a team-goal thermometer — and confetti when the target falls.

Tiered refresh (20s Talk · 60s tickets · 5-min aggregates), SLA breach countdowns, a war-room takeover, sound alerts and drill-downs.

First-run wizard, OS-native encrypted secrets, admin tokens, client-scoped access links with server-enforced isolation and SSRF-guarded fetches.

Classifies and assigns by skill, load and a fairness rotation — with holds, auto-reassign on no-ack, escalation and an SLA clock that respects support hours.

Zendesk, Freshdesk, Freshservice, Jira SM, HaloPSA, SuperOps, NinjaOne and ConnectWise PSA. Signature-verified webhooks in; assignment written straight back per ticket's source.

An LLM (OpenAI / Anthropic / custom) adds sentiment, a summary and a suggested area — advisory by default, with a deterministic fallback.

In-app setup UI, API keys encrypted in a secret vault, admin role, hardened sessions and exactly-once intake with crash recovery.

Walks the reverse import graph from the changed files — honouring tsconfig path aliases and dynamic imports — down to the minimal set of affected tests.

A lockfile, tsconfig or runner-config change runs the whole suite rather than guess. It never silently skips a test it can't reason about.

Tracks per-test pass/fail history and ranks tests by a flake score, so a pared run's failures are ones you can trust — flaky and broken told apart.

Off by default and never load-bearing — when enabled it only adds heuristic suggestions for dynamic links static analysis can't see, always flagged.

Declarative allow/deny matched on tool names, paths, hosts and arguments — enforced at the boundary, under the Anthropic SDK, the Vercel AI SDK or LangChain.

Per-run budgets for call count, token spend and wall-clock, plus loop detection that halts a runaway the moment it starts repeating itself — terminally.

Flag sensitive calls as ask and pause for a human; approve or reject with a reason that flows straight into the trace.

Every call — allowed, blocked or halted — appended to a JSONL trace with secrets redacted and flushed per event, then rendered with leash view.

Pause before any step and land in the runner's shell with the exact environment, secrets and working directory it would have — the thing act doesn't do.

Edit a file or an env var in the paused container and re-run just that step — iterate in seconds instead of push-wait-read-logs.

A gitignored .secrets file is injected into the container, so your breakpoint shell sees the same values your step does — never committed, never logged.

Reads your existing .github/workflows/*.yml — there's no new pipeline DSL to learn and nothing to rewrite.

Bloomberg-style ticker + function codes (FA, GP, DES, ERN) and full <GO> syntax, with autocomplete and a Launchpad home screen.

Multi-year financial statements and analyst targets, price charts, company descriptions and earnings — sourced from unofficial free APIs, no keys required.

World indices and movers with sparklines, a GICS sector monitor, commodities, treasury yields, FX and crypto, plus a live news wire.

No build step — plain JS/CSS and canvas charts. Backend runs identically on Node/Express or a Cloudflare Worker on the free tier.