blackwell-systems/agent-lsp

Breaking

• get_change_impact renamed to blast_radius. Same handler, same parameters, new name.

Added

Inspector evolution (8 improvements to /lsp-inspect):

• Batch mode (--top N): directory-level inspection with ranked output
• Comparison mode (--diff): branch-only issue detection
• Fix suggestions: exact fix text for every finding
• Confidence tiers: verified/suspected/advisory
• Blast-radius severity calibration using caller counts
• Results persisted to .agent-lsp/last-inspection.json
• MCP resource inspect://last for programmatic access
• Inspector now has 12 check types (was 8)

Concurrency analysis (language-agnostic, 25 languages):

• 4 new inspector checks: unrecovered_concurrent_entry, unchecked_shared_state, channel_never_closed, shared_field_without_sync
• blast_radius: sync_guarded metadata on mutex-protected types
• find_callers: cross_concurrent flag traces through goroutine/thread boundaries
• /lsp-concurrency-audit skill (24th skill): field-level safety report

Agent DX improvements:

• explore_symbol: type info + source + callers + refs in one call
• safe_apply_edit: preview + auto-apply when net_delta == 0
• Auto-diagnostics: errors_after/warnings_after in symbol edit responses
• Indexed indicator: indexed: true/false in blast_radius/find_references/find_symbol
• Proactive diagnostic regression notifications via DiagChangeTracker
• Intent aliases: callers, explore, safe_edit

Other:

• blast_radius: scope: "all" for unexported dead code detection
• blast_radius: filter: "untested" for coverage gap queries

Summary

24 skills. 65 tools. 12 inspector check types. 30 CI-verified languages.

v0.9.0: Symbol Editing, Proactive Notifications, Intent-Based Naming

60 tools. 30 languages. 22 skills. This release reshapes how agents interact with agent-lsp: tools are named for what you want to do, not what LSP calls it; edits happen at the symbol level, not the line level; and the server pushes state changes instead of waiting to be asked.

Breaking: 7 tools renamed

Tool names now match agent intent instead of LSP protocol jargon:

| Old name | New name |
|----------|----------|
| get_info_on_location | inspect_symbol |
| get_document_symbols | list_symbols |
| get_workspace_symbols | find_symbol |
| get_code_actions | suggest_fixes |
| get_references | find_references |
| call_hierarchy | find_callers |
| simulate_edit_atomic | preview_edit |

Go function signatures unchanged. Only the MCP tool name strings are affected. Update your mcp-assert fixtures and any hardcoded tool references.

Symbol-level editing (4 new tools)

Edit code by symbol name instead of line numbers:

• replace_symbol_body: "replace function X with this implementation" without knowing coordinates
• insert_after_symbol / insert_before_symbol: add code adjacent to a named symbol
• safe_delete_symbol: delete only if zero references (checks find_references first, refuses if callers exist)

All four use the shared ResolveSymbolByNamePath resolver that supports dot-notation (MyStruct.Method) and overload indexing (Method[0]). Skills updated: /lsp-edit-symbol now uses replace_symbol_body as the primary path; /lsp-dead-code offers cleanup via safe_delete_symbol.

Proactive server notifications

Four channels push state changes to agents without requiring a tool call:

1. Diagnostic changes (2s debounce): "file X now has 3 errors"
2. Workspace ready (one-shot): "language server indexing complete"
3. Process health (crash/recovery): "language server exited unexpectedly"
4. Stale references (3s debounce): "source files changed, cached references may be outdated"

All channels wired automatically on start_lsp. Best-effort delivery; clients that ignore notifications are unaffected.

Passive mode

Connect to an already-running language server via TCP instead of spawning a new process:

{ "tool": "start_lsp", "args": { "root_dir": "/project", "connect": "localhost:9999" } }

Reuses the IDE's warm index with zero duplicate memory. Supported by gopls (gopls -listen=:9999), clangd, and other servers with TCP listen mode.

Provider-agnostic skill awareness

Four-layer reinforcement architecture ensures agents know about the 22 skills regardless of provider:

1. Instructions field: every MCP client receives a task-to-tool mapping on connect
2. Per-response hints: every tool response suggests the logical next action
3. On-demand workflows: prompts/get("lsp-refactor") loads full instructions
4. Phase enforcement: errors include recovery guidance when agents skip steps

agent-lsp init now writes provider-specific rules files alongside the MCP config: CLAUDE.md for Claude Code (with "use this, not that" table), .cursor/rules/ for Cursor, .clinerules for Cline, .windsurfrules for Windsurf, GEMINI.md for Gemini CLI.

Also

• group_by=symbol on get_diagnostics: group errors by owning function/type
• Cross-referencing in tool descriptions: tools suggest related tools
• Confidence assertions: "net_delta 0 means safe to apply without verification"
• Go test path auto-fix: bare internal/notify now works (auto-prefixes ./)
• CleanupStaleDaemons wired into daemon startup (found by /lsp-dead-code)
• Nil sender crash fixed in notification channels (found by /lsp-inspect)
• interface{} to any across 85 files
• 814 lines of new tests