[wan9chi]

Motivation

vp run loads every package's vite.config.* before it plans anything, even when the run touches one package. To load configs only when a run needs them (#944), every reader of the task graph has to go through methods that can load what they read. Today the planner, the task selector and the session read petgraph nodes and edges directly, and call load_task_graph() first so that everything is loaded.

This PR moves every reader onto such methods without changing what loads: everything still loads up front.

Changes

• vt_graph::lazy_graph::LazyGraph: a generic graph that loads a partition's nodes the first time any of them is read, and resolves a node's edges the first time they're read. Resolving edges can read nodes, loading their partitions, but not other nodes' edges, so a dependsOn cycle across packages can't make two loads wait on each other. Nodes live in an append-only elsa::FrozenMap, so reads return plain references.
• TaskLoader holds the per-package code that was in IndexedTaskGraph::load: config loading, its checks, and dependsOn resolution. The root config loads with the package list, since its settings apply to the whole workspace.
• IndexedTaskGraph exposes only reads: task, tasks, depends_on, query_tasks, get_script_hook and list_tasks. Tasks are keyed by TaskId (package index and task name) instead of petgraph node indices, and a query returns a BTreeMap of the tasks it selected.
• The planner, the task selector and the session use those reads. TaskGraphLoader, Session::ensure_task_graph_loaded, TaskNodeIndex and the petgraph TaskGraph type are removed, and Session::task_graph() returns the graph. vite-plus's bench/benches/workspace_load.rs calls ensure_task_graph_loaded, which becomes task_graph().

IndexedTaskGraph::load still reads every config and resolves every dependsOn, through the same reads, so output and errors don't change and no snapshot changes. One difference: when several packages have config errors, the first one reported can differ. Before, every config was read and checked before any task was created; now each package's config and tasks load together, the root's first, then the others in package order. The error that lists tasks setting cache fields at the top level still covers every package; plan fixture top_level_cache_fields_workspace_error checks that with the root and a package.

LazyGraph has unit tests for what each read loads, a cycle across partitions, a retry after a failed load, and duplicate edges.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches core task-graph loading, querying, and planning with widespread async API changes, though behavior is intended to stay the same aside from multi-error ordering.

Overview
Refactors the workspace task graph so all reads go through async APIs built on a new LazyGraph, instead of a fully built petgraph and TaskGraphLoader.

LazyGraph loads graph partitions on first access (nodes per package, edges on first dependsOn read), with stable node refs via elsa::FrozenMap. TaskLoader holds the per-package config/script loading and dependency resolution that used to live in IndexedTaskGraph::load.

IndexedTaskGraph is now keyed by TaskId (not TaskNodeIndex). Callers use task, tasks, depends_on, query_tasks, list_tasks, and get_script_hook. Query results are a BTreeMap of QueriedTask instead of a petgraph of indices.

Session caches the graph in tokio::sync::OnceCell (**task_graph()** replaces ensure_task_graph_loaded). vt_plan takes a loaded IndexedTaskGraph directly and plans from TaskNode references.

IndexedTaskGraph::load still runs load_all so every package and dependsOn is validated up front (same outcomes; which config error surfaces first can change when several packages fail). TopLevelCacheFieldsCollector::absorb aggregates misplaced cache fields across packages; a new plan snapshot fixture covers root + package errors together.

^{Reviewed by Cursor Bugbot for commit 8584d82. Bugbot is set up for automated code reviews on this repo. Configure here.}

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	cargo/​elsa@​1.11.2

View full report

[github-actions]

fspy benchmark

linux

dynamic/launch             change  -0.33%  [ -3.62% ..  +3.71%]  overhead   +57.02%
dynamic/access             change  +0.86%  [ -3.03% ..  +6.65%]  overhead   +19.91%
dynamic/access-relative    change  -0.58%  [ -3.42% ..  +1.80%]  overhead   +71.63%
dynamic/access-contended   change  +1.54%  [ -6.70% .. +10.25%]  overhead    +9.45%
static/launch              change  +0.86%  [ -4.28% ..  +5.18%]  overhead  +202.22%
static/access              change  +0.41%  [ -0.97% ..  +2.04%]  overhead +1479.74%
static/access-relative     change  +0.34%  [ -1.68% ..  +2.85%]  overhead +1971.84%
static/access-contended    change  -0.22%  [ -1.53% ..  +1.52%]  overhead +1832.25%

macos

dynamic/launch             change  -0.81%  [ -3.77% ..  +3.22%]  overhead  +244.39%
dynamic/access             change  +0.34%  [-10.69% ..  +3.42%]  overhead    +5.94%
dynamic/access-relative    change  -0.34%  [ -8.23% ..  +5.88%]  overhead  +258.59%
dynamic/access-contended   change  -2.91%  [-49.31% ..  +3.27%]  overhead    -0.73%

windows

dynamic/launch             change  +0.37%  [ -6.25% ..  +7.97%]  overhead   +25.21%
dynamic/access             change  +0.00%  [ -2.00% ..  +2.55%]  overhead    +1.69%
dynamic/access-relative    change  +0.34%  [ -2.99% ..  +5.48%]  overhead    +2.12%
dynamic/access-contended   change  -0.68%  [-13.21% .. +11.36%]  overhead    -0.52%

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 424f37e. Configure here.}