Grep Can't Read Code

Giving AI agents IDE-grade code intelligence

The Frame

An agent can grep for text — but text can't understand code

Agents already have grep. The Amplifier LSP bundles hand them the same language-server intelligence an IDE gives a human, so they navigate by meaning instead of string match.

Meaning is not a slogan here — it's a shipped engine. Here's the evidence.

Text matching can't tella trait impl from a comment
Text matching can'ttrace a call graph
Text matching can't verifywhether an edit still compiles

Proof

tool-lsp ships 14 real LSP operations, backed by tests

Exactly 14 operations: 7 always-on base ops plus 7 language-gated extended ops, no overlap. Grouped as Navigation, Call hierarchy, Verification, Refactoring, Inspection, Extensions.

A substantial, shipped engine. So why isn't grep enough on its own?

14

LSP operations exposed by tool-lsp

7 + 7

Base ops + extended ops, zero overlap

17

Test files in the tool-lsp module

519

Lines in tool.py

The Tension

Grep returns text matches — so an agent navigating by string is navigating blind

Grep gives text matches that may include comments, strings, and similar names. LSP gives semantic results — actual code relationships. For code navigation, semantic wins.

The fix is to match by meaning. Start with finding a function's real callers.

GrepMay match comments, strings, similar names
LSPSemantic results — actual code relationships
Rule of thumbUse grep for text searching; use LSP for semantic understanding

Mechanism

findReferences returns the actual calls, not text that looks like a call

To find all calls to a function, findReferences is semantic — it finds the actual calls. Grep may match comments, strings, or similar names that merely resemble the name.

Finding one function's callers is the start. Now trace the whole graph.

Task: find all calls to a function
findReferencesSemantic — finds actual calls
grepMay match comments, strings, similar names

Mechanism

incomingCalls and outgoingCalls trace the complete call graph

These call-hierarchy operations map the complete call graph — something grep cannot do reliably. Text search has no way to follow who calls whom across a codebase.

Navigation by meaning is powerful. But the real payoff is verifying the agent's own edits.

incomingCalls / outgoingCallsTrace call graphs
Maps the complete call graphBase operations, always on
grepCannot do this reliably

The Payoff

After an edit, the agent reads compiler errors and fixes them itself

The diagnostics → codeAction → apply cycle is the AI equivalent of a developer watching compiler output and applying suggested fixes. Checking for errors after editing is simply not possible with grep.

This verify loop is production-tuned per language. For Rust, that means killing cold-start delay.

1diagnosticsCheck for errors introduced by the edit
2codeActionGet suggested fixes for those errors
3applyThe AI equivalent of watching compiler output

Mechanism · Rust

For Rust, rust-analyzer runs as a persistent background service

A 5-minute (300-second) idle timeout keeps rust-analyzer warm to eliminate cold-start indexing, then auto-shuts down after 5 minutes with no connected clients. customRequest unlocks rust-analyzer extensions like macro expansion.

Real intelligence, production-tuned. So when do you still reach for grep?

300s

Idle timeout — 5 minutes, keeps rust-analyzer warm

30–90s

First-session cold start on large projects, without persistence

Takeaway

Use LSP for meaning; use grep for text — and be honest about the limits

The durable rule: LSP for semantic understanding, diagnostics, and refactoring; grep for text searching. The bundle earns trust by documenting rust-analyzer's real limits, with fallbacks.

Credibility comes from honest limits, not just capabilities.

typeHierarchy unsupportedWorkaround: goToImplementation
Async call-hierarchy gapsWorkaround: findReferences as fallback
Cold start 30–90s on large projectsSolved by the persistent server

Sources

Research Methodology

SHIPPED & CONSOLIDATED

Data as of: repositories last pushed 2026-05-18 (release tag v0.2.0 on both engines).

Repositories (all microsoft/*): the general engine microsoft/amplifier-bundle-lsp provides tool-lsp (14 operations, 17 test files) and the code-navigator agent; Rust intelligence lives in microsoft/amplifier-bundle-rust-dev (code-intel agent, rust-analyzer config, rust-lsp.md), which absorbed the deprecated microsoft/amplifier-bundle-lsp-rust forwarding stub on 2026-02-15.

Research performed:

Operation count & split: gh api repos/microsoft/amplifier-bundle-lsp/contents/modules/tool-lsp/amplifier_module_tool_lsp/tool.py --jq .content | base64 -d then python3 len(OPERATIONS) (14; BASE=7, EXTENDED=7, overlap 0)
Grep vs LSP & verify loop: gh api repos/microsoft/amplifier-bundle-lsp/contents/context/lsp-general.md --jq .content | base64 -d
Call graph & code-navigator: gh api .../agents/code-navigator.md
Test files: gh api repos/microsoft/amplifier-bundle-lsp/git/trees/HEAD?recursive=1 --jq '[.tree[].path | select(startswith("modules/tool-lsp/tests/test_"))] | length' (17); wc -l tool.py (519)
Idle timeout 300s: gh api repos/microsoft/amplifier-bundle-lsp-rust/commits/7a8c587 --jq .commit.message; gh api repos/microsoft/amplifier-bundle-rust-dev/contents/behaviors/rust-lsp.yaml --jq .content | base64 -d
rust-analyzer limits & customRequest: gh api repos/microsoft/amplifier-bundle-rust-dev/contents/context/rust-lsp.md --jq .content | base64 -d
Consolidation & deprecation: gh api repos/microsoft/amplifier-bundle-lsp-rust/contents/bundle.yaml --jq .content | base64 -d; gh api repos/microsoft/amplifier-bundle-lsp-rust/git/trees/HEAD?recursive=1
Contributors: gh api 'repos/microsoft/amplifier-bundle-lsp-rust/commits?per_page=100' --jq '[.[] | select(.commit.author.name=="Brian Krabach")] | length'

Primary contributors: Brian Krabach (8 of 9 lsp-rust commits; drove absorption into rust-dev); Salil Das (1 of 9 lsp-rust commits plus MIT LICENSE / OSS governance PRs).

Gaps: v0.2.0 is a git tag on both engines; the in-repo bundle.yaml manifests still read version 0.1.0. The lsp-rust stub bundle.yaml reads version 2.0.0.