documentation

You run cargo clippy and get 87 warnings across 13 packages. Most AI tools fix them one at a time. They see text, not data. They can't filter by severity, group by package, or spawn concurrent agents.

How Superglide handles it

Superglide runs on Nushell, where commands return typed data structures instead of text. Modern build tools like Cargo already support JSON output. Superglide makes that structured data actionable.

# Get structured data from Cargo
let warnings = cargo clippy --message-format=json | from json -o

# Analyze which packages need the most work
$warnings | where message?.level in [warning error]
  | get package_id | uniq -c | sort-by count

Output: 18 warnings in ai-gateway-impl, 10 in tui-app, 9 in terminal-emulator...

Better: build dependency graph of errored packages, fix from leaves up:

# Get package metadata and dependency graph
let packages = cargo metadata --format-version=1
  | from json | get packages
  | select id dependencies

# Get packages with errors
let errored = $warnings | get package_id | uniq

# TODO: build dependency subgraph of only errored packages
# TODO: topological sort to get layers (leaves first)
# TODO: new Nushell primitive for dag-par-each?
#       executes each layer in parallel, waits for layer to complete

# For each layer (pseudo-code):
$warnings | where package_id in $current_layer
  | group-by package_id
  | par-each { |pkg| subagent "fix package" }

The agent builds a dependency subgraph containing only packages with errors, then fixes them topologically. Leaf packages (with no errored dependencies) fix first in parallel. Next layer waits for dependencies to complete. Changes propagate correctly up the tree.

Works for any bulk operation

The pattern is the same for test failures, dependency updates, refactoring tasks, or documentation generation. If the tool outputs JSON, Superglide can parallelize it.

# Fix all failing tests in parallel
cargo test -- --format json
  | from json -o
  | where type == "test" and event == "failed"
  | par-each { |test| subagent "diagnose and fix" }

The model receives queryable records instead of text it has to parse. Each agent works independently. The operations compose.

Terminal-native architecture

Superglide is designed for modern terminal emulators like Ghostty, WezTerm, and Kitty. It implements a multiplexed TUI similar to tmux or Zellij: one daemon runs service backends, and the TUI attaches as a client.

Sessions persist when you detach. Your terminal workflow continues uninterrupted. No context switching between terminal and IDE. The AI becomes another composable tool in your existing environment.

Getting started

Superglide ships as a single binary with all dependencies embedded. Authenticate through the dashboard, then run superglide in any directory. Logs write to ~/.superglide/logs/ as JSONL. State persists in ~/.superglide/chats.