The Rust Development Home

LSP, quality, and expert agents in one bundle

microsoft/amplifier-bundle-rust-dev

The Frame

One bundle for the full Rust development experience

microsoft/amplifier-bundle-rust-dev composes LSP code intelligence, quality tools, and expert agents so Rust gets written, verified, and understood without ever leaving the agent.

The verified proof: it started by absorbing navigation that already worked. →

LSP code intelligencerust-analyzer for understanding
Quality toolscargo fmt · clippy · cargo check · stub detection
Expert agentsrust-dev + code-intel apply Rust judgment

Proof · It already navigated

The first commit absorbed the LSP that already worked

Commit 6ddede7 — "absorb LSP content from lsp-rust bundle" — wired rust-analyzer as the server, with diagnostics, rename, codeAction, inlayHints, goToImplementation, and customRequest all enabled.

Real, pre-existing code intelligence. But navigation only gets you halfway. →

# first feature commit (2026-02-14)
6ddede7 feat: absorb LSP content
        from lsp-rust bundle

# rust-lsp.yaml
server.command: [rust-analyzer]
checkOnSave.command: clippy
capabilities: 6 enabled

The Tension

Navigation stops at understanding — it never verifies

rust-analyzer helps you read the code, but there was no in-agent cargo fmt, no clippy, no cargo check, and no stub detection. An edit could compile-fail without surfacing back into context.

So the bundle adds the missing verification layer. →

Understanding was coveredhover, definitions, references, implementations
Verification was notno in-agent cargo fmt / clippy / cargo check
Completeness was notno stub / TODO detection on edits

The Mechanism

The rust_check tool runs four checks in one pass

One tool, one pass: formatting (cargo fmt), linting (clippy), types (cargo check), and stub detection for todo!()/unimplemented!()/unreachable!() plus // TODO and // FIXME — with test files and legitimate patterns exempted.

Both concerns need one engine and expert judgment behind them. →

1

format

cargo fmt

2

lint

clippy

3

types

cargo check

4

stubs

todo!/unimplemented!/unreachable! · TODO/FIXME

The Setup · One home, not bolted-on parts

One shared checker library, two named expert agents

A single 431-line checker library backs both the tool and the hook. Two agents ship with the bundle: rust-dev, the quality expert, and code-intel, the LSP/rust-analyzer navigation specialist.

Now the payoff can lean on both. →

431

lines in the shared checker library

2

expert agents: rust-dev + code-intel

The Payoff

The hook fires on every edit and injects results into context

The hooks-rust-check module triggers on *.rs write/edit events, runs the quality checks (lint, types, format), and injects the feedback straight into the agent's context — verification becomes automatic, not a step you remember.

Writing, verifying, and understanding Rust now all happen inside the agent. →

# rust-quality.yaml — hooks-rust-check
file_patterns: ['*.rs']
checks: [lint, types, format]
report_level: warning
auto_inject: true  # → into agent context

The Turn · It pays its own way

v0.2.0 made the Rust home free when you're not writing Rust

v0.2.0 (commit c971e3a) moved always-on context into the code-intel agent via a context-sink pattern — the author's estimate is ~3,371 tokens saved off always-on for non-Rust sessions.

Compose the whole thing, pay only when you use it. →

~3,371

tokens saved off always-on (author's estimate)

v0.2.0

context-sink refactor, git tag present

The Takeaway

Compose understanding, verification, and judgment into one includable bundle

That's the pattern rust-dev generalizes: one "Full Rust development experience" that composes LSP + quality + expert agents. And you can still include just the half you need — LSP-only or quality-only.

One bundle, three concerns, composable down to a single behavior.

LSP only

rust-dev:behaviors/rust-lsp — navigation without quality hooks

Quality only

rust-dev:behaviors/rust-quality — checks without LSP

The full home

behavior-rust-dev composes both plus the expert agents

Sources

Research Methodology

Repository: microsoft/amplifier-bundle-rust-dev (public, default branch main; created 2026-02-14, pushed 2026-05-18)

Feature status: Active — bundle.yaml/pyproject declare 0.1.0; git tag v0.2.0 present (sub-behaviors rust-lsp/rust-quality at 0.2.0)

Research performed against a fresh independent clone + gh API:

gh repo view microsoft/amplifier-bundle-rust-dev --json createdAt,pushedAt,isPrivate,defaultBranchRef
git show 6ddede7 --stat — first commit absorbing LSP from lsp-rust
cat behaviors/rust-dev.yaml behaviors/rust-lsp.yaml behaviors/rust-quality.yaml
wc -l src/amplifier_bundle_rust_dev/checker.py — 431 lines
git show c971e3a --stat --format='%an|%ai' — v0.2.0 context-sink refactor
git rev-list --count HEAD (12) · git shortlog -sne HEAD

Gaps / estimates: Token figures ~2,558 / ~813 / ~3,371 are the author's own in-commit estimates, not independently re-measured. "Edits fail silently" is narrative framing, not a repo quote. No adoption/usage numbers were found or claimed.

Primary contributors (HEAD/main): Brian Krabach (9 commits — bundle, core library, agents, v0.2.0); Salil Das (3 commits — Microsoft OSS governance files).