Back to Skills

lsp-safe-edit

blackwell-systems
Updated 6 days ago
53
2
53
View on GitHub
Designdesign

About

This skill safely applies code edits by first previewing changes and comparing LSP diagnostics before writing to disk. It only commits edits if they don't introduce new errors, and surfaces quick fixes when issues arise. Developers should use it for risk-free refactoring and multi-file modifications with built-in validation.

Quick Install

Claude Code

Recommended
Primary
npx skills add blackwell-systems/agent-lsp -a claude-code
Plugin CommandAlternative
/plugin add https://github.com/blackwell-systems/agent-lsp
Git CloneAlternative
git clone https://github.com/blackwell-systems/agent-lsp.git ~/.claude/skills/lsp-safe-edit

Copy and paste this command in Claude Code to install this skill

Documentation

Requires the agent-lsp MCP server.

lsp-safe-edit

Wrap any code edit with a before/after diagnostic comparison. Speculatively previews the change in-memory before touching disk, then diffs errors introduced vs. resolved after applying. If errors appear, surfaces code actions to fix them.

Prerequisites

LSP must be running for the target workspace. If not yet initialized, call mcp__lsp__start_lsp with the workspace root before proceeding.

Auto-init note: agent-lsp supports workspace auto-inference from file paths. Explicit start_lsp is only needed when switching workspace roots.

Input

  • target file(s): One or more files to be edited (absolute paths).
  • description of change: What you intend to edit and why.

Workflow

Step 1 — Open target file(s)

Call mcp__lsp__open_document for each file that will be edited:

mcp__lsp__open_document(file_path: "/abs/path/to/file.go", language_id: "go")

Step 2 — Capture baseline diagnostics (BEFORE)

Call mcp__lsp__get_diagnostics for each target file. Store as BEFORE. For multi-file edits, collect diagnostics for all files involved.

BEFORE = mcp__lsp__get_diagnostics(file_path: "/abs/path/to/file.go")

If the server returns an empty list immediately after open, wait briefly and retry — LSP analysis is async.

Step 3 — Speculative preview (preview_edit)

Before touching disk, call mcp__lsp__preview_edit to preview the error delta of the intended change:

mcp__lsp__preview_edit(
  file_path: "/abs/path/to/file.go",
  start_line: <N>,
  start_column: <col>,
  end_line: <N>,
  end_column: <col>,
  new_text: "<replacement text>"
)

Returns net_delta (new errors introduced minus errors resolved) without writing to disk.

Decision:

net_deltaAction
≤ 0Proceed — edit improves or does not worsen error state
> 0Pause. Report introduced errors to user and ask: "Proceed anyway? [y/n]"

If net_delta > 0 and user says "n", stop. Do not apply the edit.

Multi-file edits: preview_edit covers one file at a time. For edits spanning multiple files, run it for each file independently and sum the deltas. If any file shows net_delta > 0, pause before continuing.

When to skip Step 3: If the intended change is a new file (Write), there is no existing file to simulate against. Skip to Step 4.

Step 3b — Refactor preview with simulate_chain (renames and signature changes)

Use this step instead of or after Step 3 when the change is a rename, signature change, or any edit with dependent follow-on edits (e.g., updating all call sites after adding a parameter).

simulate_chain applies a sequence of speculative edits in-memory and reports whether the cumulative change is safe — without touching disk:

mcp__lsp__simulate_chain({
  "workspace_root": "/abs/path/to/workspace",
  "language": "go",
  "edits": [
    {
      "file_path": "/abs/path/to/file.go",
      "start_line": <N>, "start_column": <col>,
      "end_line": <N>,   "end_column": <col>,
      "new_text": "<replacement>"
    },
    // additional dependent edits (e.g. call site updates) ...
  ]
})

Returns:

  • cumulative_delta — net error change across all steps
  • safe_to_apply_through_step — how many steps are safe to apply in sequence

Decision:

cumulative_deltasafe_to_apply_through_stepAction
≤ 0= total stepsAll steps safe. Proceed to Step 4.
≤ 0< total stepsSafe up to that step. Review remaining steps.
> 0anyNet regression. Report to user before proceeding.

When to use Step 3b:

  • Renaming an exported symbol and updating its call sites
  • Adding/removing a parameter and updating all callers
  • Any multi-file refactor where edits are order-dependent

When to skip Step 3b:

  • Simple in-place edits with no dependent follow-on edits (Step 3 is sufficient)
  • New file creation (no existing text to simulate against)

Step 4 — Apply the edit to disk

Apply the change using the Edit or Write tool:

  • Use Edit for targeted replacements in an existing file.
  • Use Write only when creating a new file or doing a full rewrite.
Edit(file_path: "/abs/path/to/file.go", old_string: "...", new_string: "...")

For multi-file edits, apply each file's changes before collecting post-edit diagnostics (Step 5). If any individual edit fails, stop and report before applying remaining files.

Step 5 — Capture post-edit diagnostics (AFTER)

Call mcp__lsp__get_diagnostics again for each edited file. Store as AFTER.

AFTER = mcp__lsp__get_diagnostics(file_path: "/abs/path/to/file.go")

For multi-file edits, collect diagnostics for all files and merge the results.

Step 6 — Compute the diagnostic diff

Compare BEFORE and AFTER:

  • Introduced = diagnostics in AFTER not in BEFORE (new problems).
  • Resolved = diagnostics in BEFORE not in AFTER (fixed problems).

Match by (file, line, message) tuple to handle line-number shifts. Treat error and warning severity separately.

Step 7 — Surface code actions if errors were introduced

If any new error-severity diagnostics appear, call mcp__lsp__suggest_fixes at each error location to surface quick fixes:

mcp__lsp__suggest_fixes(
  file_path: "<file>",
  start_line: <error line>,
  start_column: 1,
  end_line: <error line>,
  end_column: 999
)

Report available code actions to the user:

Errors introduced (2):
  file.go:34 — undefined: MyType
    → Code action: Import "mypackage" (quickfix)
  file.go:51 — cannot use int as string
    → No code actions available

Apply code actions? [y/n/select]

If the user accepts, apply the code action's WorkspaceEdit via mcp__lsp__apply_edit, then re-collect diagnostics and re-diff.

Step 8 — Format (optional)

If the diagnostic diff is clean (net change ≤ 0), offer to format the edited file via the language server:

mcp__lsp__format_document({ "file_path": "/abs/path/to/file" })

Returns TextEdit[]. If non-empty, apply immediately:

mcp__lsp__apply_edit({ "workspace_edit": <TextEdit[]> })

Skip if the user did not ask for formatting, or if there are unresolved errors (fix errors before formatting).

Step 9 — Report using DiagnosticDiffFormat

Output the final summary:

## Edit Summary

Files changed: N
Errors introduced: A  →  Errors resolved: B  (net: A-B)
Warnings introduced: C  →  Warnings resolved: D

### Introduced errors
- file.go:34 — undefined: MyType

### Resolved errors
- file.go:12 — unused variable: x

Decision Guide

Net changeAction
0Safe. No new errors.
NegativeNet improvement — errors resolved. Safe.
Positive (after code actions)Do NOT commit. Offer to revert.

When net change > 0 after code actions:

  1. Show the full list of remaining introduced errors.
  2. Offer to revert using the original old_string in a follow-up Edit call.
  3. Wait for user decision before proceeding.

Do not commit or stage files when net change > 0.


Multi-file workflow

For edits spanning multiple files (e.g., changing a function signature and all its call sites):

  1. Open all files in Step 1.
  2. Collect BEFORE diagnostics for all files.
  3. Simulate each file independently in Step 3 — sum net_delta values.
  4. Apply edits file by file in Step 4 — stop on first failure.
  5. Collect AFTER diagnostics for all files and merge.
  6. Check code actions on any file showing new errors.

Report the combined diagnostic diff across all files in the final summary.

GitHub Repository

blackwell-systems/agent-lsp
Path: skills/lsp-safe-edit
0
agentskillsai-agentsai-toolingclaudeclaude-codecode-intelligence

Related Skills

executing-plans

Design

Use the executing-plans skill when you have a complete implementation plan to execute in controlled batches with review checkpoints. It loads and critically reviews the plan, then executes tasks in small batches (default 3 tasks) while reporting progress between each batch for architect review. This ensures systematic implementation with built-in quality control checkpoints.

View skill

requesting-code-review

Design

This skill dispatches a code-reviewer subagent to analyze code changes against requirements before proceeding. It should be used after completing tasks, implementing major features, or before merging to main. The review helps catch issues early by comparing the current implementation with the original plan.

View skill

connect-mcp-server

Design

This skill provides a comprehensive guide for developers to connect MCP servers to Claude Code using HTTP, stdio, or SSE transports. It covers installation, configuration, authentication, and security for integrating external services like GitHub, Notion, and custom APIs. Use it when setting up MCP integrations, configuring external tools, or working with Claude's Model Context Protocol.

View skill

web-cli-teleport

Design

This skill helps developers choose between Claude Code Web and CLI interfaces based on task analysis, then enables seamless session teleportation between these environments. It optimizes workflow by managing session state and context when switching between web, CLI, or mobile. Use it for complex projects requiring different tools at various stages.

View skill