ln-012-mcp-configurator（LN 012 MCP configurator 搜索）

Paths: File paths (shared/, references/) are relative to skills repo root. Locate this SKILL.md directory and go up one level for repo root.

MCP Configurator

Type: L3 Worker Category: 0XX Shared

Invocation is agent-agnostic. Target surfaces are Claude-specific: this skill configures Claude Code MCP settings, hook/style sync, permissions, and Claude Code IDE extension permission mode.

MANDATORY READ

MANDATORY READ: Load shared/references/coordinator_summary_contract.md, shared/references/environment_worker_runtime_contract.md, and shared/references/worker_runtime_contract.md

Target MCP availability is not a precondition for this worker. If a server is not connected yet, use config files, package metadata, and CLI state to continue registration and cleanup work; report provider checks or live verification as skipped when they cannot run without MCP.

Input / Output

If summaryArtifactPath is provided, write the same summary JSON there. If not provided, return the summary inline and remain fully standalone. If runId is not provided, generate a standalone run_id before emitting the summary envelope.

Runtime

Runtime family: environment-worker-runtime

Phase profile:

1. PHASE_0_CONFIG
2. PHASE_1_CHECK_STATUS_AND_VERSION
3. PHASE_2_REGISTER_AND_CONFIGURE
4. PHASE_3_VERIFY_GRAPH_PROVIDER_DEPS
5. PHASE_4_HOOKS_AND_OUTPUT_STYLE
6. PHASE_5_GRAPH_INDEXING
7. PHASE_6_MIGRATE_ALLOWED_TOOLS
8. PHASE_7_PERMISSION_SURFACES
9. PHASE_8_WRITE_SUMMARY
10. PHASE_9_SELF_CHECK

Runtime rules:

• emit summary_kind=env-mcp-config
• standalone runs generate their own run_id and write the default worker-family artifact path
• managed runs require both runId and summaryArtifactPath and must write the summary to the exact provided path
• always write the validated summary artifact before terminal outcome

Output Contract

Always build a structured env-mcp-config summary envelope per:

• shared/references/coordinator_summary_contract.md
• shared/references/environment_worker_runtime_contract.md

Payload fields:

• servers
• permissions
• ide_extension
• hooks_synced
• output_style_synced
• allowed_tools_migrated
• graph_provider_status
• status

Server Registry

Two transport types: stdio (local process) and HTTP (cloud endpoint).

Workflow

Check Status & Version → Register & Configure → Verify Graph Provider Deps → Hooks → Permissions → IDE Extension Mode → Migrate → Report

Phase 1: Check Status & Version

Smart install: check MCP status AND package drift. npx -y caches aggressively — a connected server may still be backed by an older cached package.

Step 1a: Check MCP server status

Run claude mcp list -> parse each hex server:

**Step 1b: Version check for connected hex-* servers**

For each connected hex server, run in parallel:

npm view @levnikolaevich/${PKG} version 2>/dev/null

Then compare npm latest against the running local version:

1. npx cache probe: npm config get cache -> scan {cacheRoot}/_npx/**/node_modules/@levnikolaevich/${PKG}/package.json -> pick newest by semver/mtime

Use the npx cache version. If probe returns nothing, report running=unknown and treat the server as refresh-recommended rather than claiming it is current.

Note: hex packages run via npx -y, NOT global install. Never probe global npm paths.

Skip conditions:

Phase 2: Register & Configure

One pass: use Phase 1 state (do NOT re-run claude mcp list) -> remove unsupported -> register/update -> verify.

1. Reuse Phase 1 state — server map from Step 1a already has registration + connection status

- Fallback (standalone only): read ~/.claude.json + ~/.claude/settings.json

1. Remove unsupported servers:

1. Register missing OR update outdated servers:

- IF already configured AND connected AND cached local version matches -> SKIP - IF connected but outdated -> remove + re-add (forces npx to fetch latest) - IF dry_run: true -> show planned command - IF linear -> ask user: "Do you use Linear?" -> no -> SKIP

Registration commands (OS-dependent prefix):

1. Verify: claude mcp list -> check all registered show Connected. This is the only second claude mcp list call (post-mutation verify). Retry + report failures.

Windows MSYS2 path validation (MANDATORY on win32): After registration, read ~/.claude.json -> verify each hex server's args[0] is "/c" not "C:/". If corrupted: fix via mcp__hex-line__edit_file (set_line the arg to "/c").

Error handling:

Phase 2b: Verify Graph Provider Dependencies

MANDATORY READ: Load skills-catalog/ln-012-mcp-configurator/references/hex_graph_provider_matrix.md and skills-catalog/ln-700-project-bootstrap/references/stack_detection.md.

After registration + connection, verify only the extra system binaries or packages needed for fuller MCP behavior. Assume the project already has its own runtimes and app dependencies. This phase does NOT install project dependencies, framework packages, or runtimes.

Step 1: Identify which servers are connected Reuse Phase 2 verification state. Only check deps for connected hex-* servers.

Step 2: Verify hex-line-mcp dependencies

Ripgrep install by platform:

If rg --version fails: ask user whether to auto-install (suggest platform-appropriate command). If user declines, WARN but continue — hex-line will degrade on grep.

Step 2a: EOL-risk advisory for hex-line

Inspect repo and user EOL policy, but do not rewrite it here:

• .gitattributes for eol= / text=auto
• .editorconfig for end_of_line
• git config --get core.autocrlf
• git config --get core.eol
• git config --get core.safecrlf

Report WARN when repo policy and user working-tree policy are likely to churn line endings during normal edits. This is advisory only — do not auto-edit repo policy files in ln-012.

Step 3: Detect current project language(s) for hex-graph

Detect from the current project root only:

1. docs/project/tech_stack.md if present
2. marker files in the project root
3. fallback source-extension scan

Use framework markers only to confirm language:

If no relevant language is detected for hex-graph, skip the optional provider checks and report SKIP.

Step 4: Verify hex-graph-mcp graph-specific providers and SCIP exporters

Use the MCP tool install_graph_providers from hex-graph-mcp as the source of truth:

• First call it with mode: "check"
• Reuse its instructions_for_agent output as the remediation text
• Only if the user agrees, rerun it with mode: "install"
• Treat the matrix below as reference policy, not as a second implementation source

For hex-graph, think only about graph-specific providers and optional SCIP exporters:

• DO install or recommend basedpyright, csharp-ls, phpactor, scip-python, scip-dotnet, and scip-php when their language is detected
• DO prefer the patched scip-python install on Windows until the upstream Windows fix is released
• DO mention HEX_GRAPH_SCIP_PYTHON_BINARY when a patched Python SCIP binary lives outside the default PATH
• DO prefer the isolated patched scip-php fork for PHP SCIP export when project-local Composer install is blocked or the upstream binary emits an empty artifact
• DO NOT install python, .NET, php, fastapi, django, laravel, nestjs, or any project dependency
• DO NOT install anything for framework overlays alone; Stage 4 framework support is parser/convention-based
• Ask the user before any provider install command

Step 5: Verify hex-ssh-mcp dependencies

No system dependencies (pure JS ssh2). Skip.

Step 6: Report dependency table

Print table with columns: Server | Dependency | Status | Action.

Error handling:

For hex-graph, also report the detected language(s) that drove the provider/exporter checks.

Phase 3: Hooks & Output Style [CRITICAL]

Do NOT call setup_hooks. hex-line now auto-syncs hooks and output style on MCP server startup.

After all Phase 2 registrations complete:

1. Trigger one harmless hex-line call to start the server, for example mcp__hex-line__inspect_path({path: "{project_path}"})
2. Verify the startup sync results below

Hooks (in ~/.claude/settings.json):

1. PreToolUse hook — redirects built-in Read/Edit/Write/Grep/Glob to hex-line equivalents (Glob -> inspect_path(pattern=...))
2. PostToolUse hook — compresses verbose tool output (RTK filter)
3. SessionStart hook — injects MCP Tool Preferences reminder
4. disableAllHooks: After verifying hooks are synced, read ~/.claude/settings.json, set disableAllHooks: false (merge, preserve all other fields), write back. This is appropriate here because the user explicitly invoked setup and wants hooks active. autoSync does NOT manage this flag — it only syncs hook file content and entries.

Output Style: 5. Copies output-style.md to ~/.claude/output-styles/hex-line.md 6. Sets outputStyle: "hex-line" if no style is active (preserves existing style)

Verification: Confirm all of the following after the first hex-line tool call:

• ~/.claude/settings.json contains the 3 hex-line hook entries with current command path
• disableAllHooks: false (set explicitly by this skill, not by autoSync)
• ~/.claude/output-styles/hex-line.md exists and matches package content
• outputStyle: "hex-line" is set only when no other style was already active

Hex-line workflow reminder: the synced output style must teach all of the following:

• carry revision into same-file follow-up edits as base_revision
• run verify before delayed or mixed-tool follow-up edits on the same file
• reuse retry_edit, retry_edits, retry_checksum, and retry_plan directly
• treat line-ending policy as preserved file state, not something edit_file should silently normalize

Note: autosync is idempotent. Any later hex-line startup re-checks installed hook and style content and updates only when they drift.

Phase 4: Graph Indexing

After hex-graph registration + connected status:

1. mcp__hex-graph__index_project({path: "{project_path}"}) — build initial code knowledge graph
2. If graph-backed reads or edits later look stale, rerun mcp__hex-graph__index_project({path: "{project_path}"}) — the indexing path is idempotent and refreshes changed files only

Skip if hex-graph not registered or not connected.

Phase 5: Migrate allowed-tools [CRITICAL]

Scan project commands/skills to replace built-in tools with hex-line equivalents in allowed-tools frontmatter.

Tool mapping:

Steps:

1. Discover .claude/commands/*.md via mcp__hex-line__inspect_path(path=".claude/commands", pattern="*.md", max_entries=0) and .claude/skills/*/SKILL.md via mcp__hex-line__inspect_path(path=".claude/skills", pattern="SKILL.md", max_entries=0)
2. For each file: parse YAML frontmatter, extract allowed-tools
3. For each mapping entry: a. If built-in present AND hex equivalent absent -> add hex equivalent, remove built-in (except Read and Bash) b. If built-in present AND hex equivalent already present -> remove built-in (except Read and Bash) c. Preserve ALL existing mcp__* tools not in the replacement table
4. Write back updated frontmatter (preserve quoting style)

Skip conditions:

Phase 6: Grant Permissions

Instruction files and MCP Tool Preferences content are owned by ln-014-agent-instructions-manager. This skill must not create or rewrite CLAUDE.md or AGENTS.md.

Ensure built-in tools and MCP server prefixes are in ~/.claude/settings.json -> permissions.allow[].

Built-in tools (ensure present):

MCP servers (per configured):

1. Read ~/.claude/settings.json (create if missing: {"permissions":{"allow":[]}})
2. Consolidate existing entries: For each universal tool (Bash, Read, Write, Edit, Grep, WebSearch, WebFetch, and each mcp__{name}):

- If specific variants exist (e.g. Bash(gh repo:*), Bash(node...), WebFetch(domain:docs.github.com)), remove them — the universal entry covers all - Pattern: entry.startsWith("Tool(") → remove if universal Tool will be added

1. Add universal entries that are not already present
2. Write back (2-space indent JSON)

Idempotent: existing universal entries skipped. Only specific sub-entries are consolidated into universal ones.

Phase 6b: Verify IDE Extension Permission Mode

The Claude Code IDE extension (Cursor / VSCode) has its own claudeCode.initialPermissionMode setting that overrides permissions.defaultMode from ~/.claude/settings.json when Claude runs through the IDE. Default value is "default" — this silently ignores any project-level bypassPermissions from .claude/settings.local.json. To actually enable bypass when running through the IDE, two extension settings must be set together.

This phase is detection-first. It does NOT modify IDE settings without explicit user consent because permission mode is a security decision and IDE settings are vendor-specific.

Step 1: Detect installed Claude Code IDE extensions

For each IDE: glob the extensions directory for anthropic.claude-code-*. If at least one match exists, the extension is installed.

**Step 2: Read package.json from extension to enumerate claudeCode.* settings**

The relevant keys (verified against extension package.json contributes.configuration.properties):

Step 3: Read user settings JSON for each detected IDE and report current state

Always include in the report: a one-line explanation of why .claude/settings.local.json defaultMode is silently ignored when running via IDE extension.

Step 4: Cross-check against .claude/settings.json and .claude/settings.local.json defaultMode

If user's project settings declare permissions.defaultMode = "bypassPermissions" but the IDE extension is in default-prompt or bypass-blocked state → emit a WARN in the summary explaining the override:

WARN: Project .claude/settings.local.json declares defaultMode: bypassPermissions, but the Claude Code IDE extension for {IDE} forces --permission-mode {detected} at session start. The project setting is ignored when running through the IDE. To align them, run this skill with apply_ide_override=true and confirm at the prompt, or manually set claudeCode.initialPermissionMode and claudeCode.allowDangerouslySkipPermissions in {settings_path}.

Step 5: Optional remediation (only with explicit user consent)

This phase only writes IDE settings when the caller passes apply_ide_override=true (a separate flag distinct from the main dry_run). Even with the flag, the skill MUST prompt the user with a single confirmation question listing exactly which keys will change in which file.

When applying changes:

1. Backup first: copy settings.json to settings.json.bak.{timestamp} before any write
2. Read → deep-merge → write: preserve every other claudeCode.* key (preferredLocation, useTerminal, respectGitIgnore, etc.) and every non-claudeCode.* key
3. Write only the two keys dictated by the desired mode:

- For bypassPermissions: set claudeCode.allowDangerouslySkipPermissions: true AND claudeCode.initialPermissionMode: "bypassPermissions" - For acceptEdits: set claudeCode.initialPermissionMode: "acceptEdits" (gate not required) - For default: remove both keys (or set to false / "default")

1. Instruct full restart: print "IDE restart required (full Quit + reopen, NOT Reload Window) — extension reads settings only at activation"
2. Record in summary: include before/after values, backup path, and the explicit user consent timestamp

Skip conditions:

Why this phase is detection-first by default:

• IDE settings are per-IDE-vendor and per-user — auto-modifying them across machines is intrusive
• bypassPermissions is a security-sensitive choice; Anthropic explicitly guarded it with allowDangerouslySkipPermissions for IDE contexts (intended for sandboxes without internet)
• Users may legitimately want different modes per IDE (e.g. acceptEdits in Cursor for daily work, bypassPermissions in a separate VSCode profile for sandbox automation)
• Aligns with the no-agent-config-modification rule from the host project memory: detection always, writes only on explicit consent

Phase 7: Report

Status table:

MCP Configuration:
| Server    | Transport | Version | Status        | Permission | Detail                  |
|-----------|-----------|---------|---------------|------------|-------------------------|
| hex-line  | stdio     | 1.5.0   | configured    | granted    | global npm (hex-line-mcp) |
| hex-ssh   | stdio     | 1.2.0   | updated       | granted    | was 1.1.6, now 1.2.0     |
| context7  | HTTP      | —       | configured    | granted    | mcp.context7.com        |
| Ref       | HTTP      | —       | configured    | granted    | api.ref.tools (key set) |
| linear    | HTTP      | —       | skipped       | skipped    | user declined           |

IDE Extension Permission Mode:
| IDE     | Extension Version | initialPermissionMode | allowDangerouslySkipPermissions | Effective State | Conflict with project defaultMode |
|---------|-------------------|-----------------------|---------------------------------|-----------------|-----------------------------------|
| Cursor  | 2.1.92            | bypassPermissions     | true                            | bypass-active   | aligned                           |
| VSCode  | (not installed)   | —                     | —                               | no-ide          | n/a                               |

Critical Rules

1. Write only via sanctioned paths. Register servers via claude mcp add. Write to ~/.claude/settings.json ONLY for permissions (permissions.allow[]) or when explicit hook/style verification requires correction after autosync fails
2. Verify after add. Always run claude mcp list after registration to confirm connection
3. Ask before optional servers. Linear requires explicit user consent
4. npx -y for all hex MCP. Never npm i -g — npx provides process isolation and avoids EBUSY on Windows. On Windows, wrap with cmd /c npx (see Phase 2 OS prefix table)
5. Remove unsupported servers. Clean up servers no longer in the registry
6. Grant permissions. After registration, add mcp__{server} to user settings
7. Minimize claude mcp list calls. Phase 1 runs it once (discovery). Phase 2 reuses that data. Only Phase 2 Step 4 runs it again (post-mutation verify). Max 2 calls total
8. Always check npm drift. Connected!= up to date. Compare npm latest against the newest locally cached npx package version before skipping
9. MSYS2 path safety. On Windows with Git Bash/MSYS2, always prefix claude mcp add with MSYS_NO_PATHCONV=1. After registration, verify args[0] in .claude.json is "/c" not "C:/". Fix inline if corrupted.
10. Verify graph-specific deps after install. After Phase 2 registration, check system binaries for hex-line, graph-specific optional providers, and optional SCIP exporters for detected project languages. Auto-install only with user consent. Never install project runtimes or framework packages here.
11. Report EOL churn risk, do not hide it. If .gitattributes, .editorconfig, or Git config suggest working-tree line-ending rewrites, warn explicitly instead of silently changing repo policy here.
12. Non-destructive config writes. Always read → merge → edit. Never overwrite config files from scratch. Preserve all keys/sections not owned by this skill.
13. IDE extension settings are detection-first. Phase 6b reads claudeCode.* keys from Cursor / VSCode user settings and reports the effective permission mode, but never writes IDE settings without apply_ide_override=true AND an explicit user prompt confirmation. IDE settings are vendor-specific and changing bypassPermissions is a security decision the user must own.

Anti-Patterns

Definition of Done

• MCP packages installed and versions verified against npm registry (Phase 1)
• Missing servers registered and verified connected (Phase 2)
• Outdated servers re-registered with latest version (Phase 2)
• Graph-specific dependencies verified: ripgrep available, detected hex-graph providers and SCIP exporters reported (Phase 2b)
• Hooks auto-synced after first hex-line startup (PreToolUse, PostToolUse, SessionStart, ConfigChange, PermissionDenied) and disableAllHooks: false (Phase 3). MCP config notes: _meta["anthropic/maxResultSizeChars"] up to 500K for large tool results (Claude Code 2.1.91); MCP_CONNECTION_NONBLOCKING env var for -p mode (2.1.89); disableSkillShellExecution setting prevents skill shell execution (2.1.91)
• Output style installed (Phase 3)
• Permissions granted for all configured servers (Phase 6)
• IDE extension permission mode detected for installed Cursor / VSCode and reported with WARN if it overrides project defaultMode (Phase 6b). Writes only with apply_ide_override=true AND user consent.
• Project allowed-tools migrated (Phase 5)
• Status table with version column displayed (Phase 7) including IDE extension permission mode row
• Structured summary returned
• Summary artifact written to the managed or standalone runtime path

Version: 1.7.0 Last Updated: 2026-04-07