@levnikolaevich/hex-line-mcp 1.11.0 → 1.12.0

package/README.md

@@ -38,16 +38,16 @@ Advanced / occasional:
 | `verify` | Check if held checksums / revision are still current | Staleness check without full re-read |
 | `inspect_path` | Unified file-or-directory inspection | File metadata for files, tree or pattern search for directories |
 | `changes` | Compare file against git ref, shows added/removed/modified symbols | AST-level semantic diff |
-| `bulk_replace` | Search-and-replace across multiple files by glob | Compact summary (default) or capped diffs via `format`, dry_run, max_files |
+| `bulk_replace` | Search-and-replace across multiple files inside an explicit root path | Compact summary (default) or capped diffs via `format`, dry_run, max_files |
 
-### Hooks (PreToolUse + PostToolUse)
+### Hooks (SessionStart + PreToolUse + PostToolUse)
 
 | Event | Trigger | Action |
 |-------|---------|--------|
-| **PreToolUse** | Read/Edit/Write/Grep on text files | Size-aware redirect: cheap small operations may pass, expensive ones are redirected |
+| **PreToolUse** | Read/Edit/Write/Grep on text files | Redirect-first policy for text files; built-in tools stay reserved for binary/media and `.claude/settings*.json` exceptions |
 | **PreToolUse** | Bash with dangerous commands | Blocks `rm -rf /`, `git push --force`, etc. Agent must confirm with user |
 | **PostToolUse** | Bash with 50+ lines output | RTK: deduplicates, truncates, shows filtered summary to Claude as feedback |
-| **SessionStart** | Session begins | Injects a short no-discovery workflow for hex-line tools |
+| **SessionStart** | Session begins | Injects a short bootstrap hint; defers to the active output style when `hex-line` style is enabled |
 
 
 ### Bash Redirects
@@ -70,6 +70,8 @@ Hooks and output style are auto-synced on every MCP server startup. The server c
 
 Hooks are written to global `~/.claude/settings.json` with absolute path to `hook.mjs`. Output style is installed to `~/.claude/output-styles/hex-line.md` and activated if no other style is set. To activate manually: `/config` > Output style > hex-line.
 
+No extra manual setup is required after install. The startup sync uses the current Node runtime and a stable hook path under `~/.claude/hex-line`, so the hook command survives spaces in the home directory on Windows, macOS, and Linux.
+
 ## Validation
 
 Use the normal package checks:
@@ -94,6 +96,8 @@ Comparative built-in vs hex-line benchmarks are maintained outside this package.
 If a project already has `.hex-skills/codegraph/index.db`, `hex-line` automatically adds lightweight graph hints to `read_file`, `outline`, `grep_search`, `edit_file`, and `changes`.
 
 - Graph enrichment is optional. If `.hex-skills/codegraph/index.db` is missing, stale, or unreadable, `hex-line` falls back to standard behavior silently.
+- Graph enrichment is project-deterministic. `hex-line` only uses the graph database that belongs to the resolved current project scope.
+- Nested projects do not inherit graph hints from a parent repo index once a nested project boundary is detected.
 - `better-sqlite3` is optional. If it is unavailable, `hex-line` still works without graph hints.
 - `read_file`, `outline`, and `grep_search` stay compact: they only surface high-signal local facts such as `api`, framework entrypoints, callers, flow, and clone hints.
 - `edit_file` and `changes` surface the deeper review layer: external callers, downstream return/property flow, clone peers, public API risk, framework entrypoint risk, and same-name sibling warnings when present.
@@ -128,7 +132,7 @@ Use `replace_between` inside `edit_file` when you know stable start/end anchors
 
 ### Literal rename / refactor
 
-Use `bulk_replace` for text rename patterns across one or more files. Returns compact summary by default; pass `format: "full"` for capped diffs. Do not use it as a substitute for structured block rewrites.
+Use `bulk_replace` for text rename patterns across one or more files inside a known project root or directory scope. Pass `path` explicitly. In normal agent workflows that scope should be auto-filled from the current project root, not typed manually. Returns compact summary by default; pass `format: "full"` for capped diffs. Do not use it as a substitute for structured block rewrites.
 
 ### read_file
 
@@ -183,14 +187,29 @@ Edit operations (JSON array):
 ]
 ```
 
+Discipline:
+
+- Never invent `range_checksum`. Copy it from `read_file` or `grep_search(output:"content")`.
+- First mutation in a file: prefer `grep_search` for narrow targets, or `outline -> read_file(ranges)` for structural edits.
+- Prefer 1-2 hunks on the first pass. Once `edit_file` returns a fresh `revision`, continue from that state.
+
 Result footer includes:
 
 - `status: OK | AUTO_REBASED | CONFLICT`
+- `reason: ...` as the canonical machine-readable cause for the current status
 - `revision: ...`
 - `file: ...`
 - `changed_ranges: ...` when relevant
+- `recovery_ranges: ...` with the narrowest recommended `read_file` ranges for retry
+- `next_action: ...` as the canonical immediate choice: `apply_retry_edit`, `apply_retry_batch`, or `reread_then_retry`
 - `remapped_refs: ...` when stale anchors were uniquely relocated
-- `retry_checksum: ...` on local conflicts
+- `retry_checksum: ...` on local conflicts, narrowed to the exact target range when possible
+- `retry_edit: ...` when the server can synthesize a ready-to-retry edit skeleton from current local state
+- `retry_edits: ...` on conservative batch conflicts when every conflicted edit can be retried directly
+- `suggested_read_call: ...` when rereading is the safest next step
+- `retry_plan: ...` with a compact machine-readable next-call plan
+- `summary: ...` and `snippet: ...` instead of long prose blocks
+- `edit_conflicts: N` on conservative multi-edit preflight conflicts
 
 ### write_file
 
@@ -239,7 +258,7 @@ Not for `.json`, `.yaml`, `.txt` -- use `read_file` directly for those.
 
 ### verify
 
-Check if range checksums from prior read/search blocks are still valid, optionally relative to a prior `base_revision`. Returns a deterministic verification report with `status`, `summary`, and one line per checksum entry.
+Check if range checksums from prior read/search blocks are still valid, optionally relative to a prior `base_revision`. Returns a deterministic verification report with canonical `status`, `summary`, `next_action`, and compact entry lines.
 
 | Parameter | Type | Required | Description |
 |-----------|------|----------|-------------|
@@ -251,13 +270,16 @@ Example output:
 
 ```text
 status: STALE
+reason: checksums_stale
 revision: rev-17-deadbeef
 file: 1-120:abc123ef
 summary: valid=0 stale=1 invalid=0
+next_action: reread_ranges
 base_revision: rev-16-feedcafe
 changed_ranges: 10-12(replace)
+suggested_read_call: {"tool":"mcp__hex-line__read_file","arguments":{"path":"/repo/file.ts","ranges":["10-12"]}}
 
-STALE 10-12 checksum: 10-12:oldc0de0 current=10-12:newc0de0
+entry: 1/1 | status: STALE | span: 10-12 | checksum: 10-12:oldc0de0 | current_checksum: 10-12:newc0de0 | next_action: reread_range | summary: content changed since checksum capture
 ```
 
 ### inspect_path
@@ -279,15 +301,15 @@ Inspect a file or directory path without guessing which low-level tool to call f
 
 ## Hook
 
-The unified hook (`hook.mjs`) handles four events:
+The unified hook (`hook.mjs`) handles three Claude hook events:
 
 ### PreToolUse: Tool Redirect
 
-Blocks built-in `Read`, `Edit`, `Write`, `Grep` on text files and redirects to hex-line equivalents. Binary files (images, PDFs, notebooks, archives, executables, fonts, media) are excluded.
+Applies redirect-first steering to built-in `Read`, `Edit`, `Write`, and `Grep` on text files. Binary/media files (images, PDFs, notebooks, archives, executables, fonts, media) stay on built-in tools. `.claude/settings.json` and `.claude/settings.local.json` at project root or home are also allowed on built-in tools.
 
 ### PreToolUse: Bash Redirect + Dangerous Blocker
 
-Intercepts simple Bash commands (`cat`, `head`, `tail`, `ls`, `find`, `grep`, `sed -i`, etc.) and redirects to hex-line tools. Blocks dangerous commands (`rm -rf /`, `git push --force`, `git reset --hard`, `DROP TABLE`, `chmod 777`, `mkfs`, `dd`).
+Intercepts simple Bash commands (`cat`, `head`, `tail`, `tree`, `find`, `stat`, `wc -l`, `grep`, `rg`, `sed -i`, etc.) and redirects covered cases to hex-line tools. `ls`/`dir` are redirected only for recursive listing. Dangerous commands (`rm -rf /`, `git push --force`, `git reset --hard`, `DROP TABLE`, `chmod 777`, `mkfs`, `dd`) are blocked.
 
 ### PostToolUse: RTK Output Filter
 
@@ -298,7 +320,11 @@ Triggers on `Bash` tool output exceeding 50 lines. Pipeline:
 3. **Deduplicate** -- collapses identical normalized lines with `(xN)` counts
 4. **Truncate** -- keeps first 15 + last 15 lines, omits the middle
 
-Configuration constants in `hook.mjs`:
+### SessionStart: Bootstrap Hint
+
+Injects a compact startup reminder. If the `hex-line` output style is active, the hook emits only a minimal bootstrap hint plus `ToolSearch('+hex-line read edit')` fallback. Otherwise it injects the short preferred read/edit workflow directly, including the scope rule: use file paths for file tools and the current project root for repo-wide tools such as `bulk_replace`.
+
+Hook policy constants in `lib/hook-policy.mjs`:
 
 | Constant | Default | Purpose |
 |----------|---------|--------|
@@ -308,7 +334,7 @@ Configuration constants in `hook.mjs`:
 
 ### SessionStart: Tool Preferences
 
-Injects a short operational workflow into agent context at session start: no `ToolSearch`, prefer `outline -> read_file -> edit_file -> verify`, and use targeted reads over full-file reads.
+Injects a short operational workflow into agent context at session start. If schemas are not loaded yet, it includes the `ToolSearch('+hex-line read edit')` fallback. Primary flow stays `outline -> read_file -> edit_file -> verify`, with targeted reads over full-file reads.
 
 ## Architecture
 
@@ -318,6 +344,8 @@ hex-line-mcp/
   hook.mjs            Unified hook (PreToolUse + PostToolUse + SessionStart)
   package.json
   lib/
+    hook-policy.mjs   Shared hook policy: redirects, thresholds, danger patterns
+    setup.mjs         Startup autosync for hook + output style
     read.mjs          File reading with hash annotation
     edit.mjs          Anchor-based edits, diff output
     search.mjs        ripgrep wrapper with hash-annotated results
@@ -397,7 +425,7 @@ The PostToolUse hook normalizes Bash output (replaces UUIDs, timestamps, IPs wit
 <details>
 <summary><b>Can I disable the built-in tool blocking?</b></summary>
 
-Yes. Remove the PreToolUse hook from `.claude/settings.local.json`. The MCP tools will still work, but agents will be free to use built-in Read/Edit/Write/Grep alongside hex-line tools.
+Yes. To downgrade redirects to advice, set `.hex-skills/environment_state.json` to `{ "hooks": { "mode": "advisory" } }`. To remove the hook entirely, delete the `hex-line` hook entries from `~/.claude/settings.json`. To disable the MCP server for one project, add `hex-line` to `~/.claude.json -> projects.{cwd}.disabledMcpServers`.
 
 </details>
 

package/dist/hook.mjs

@@ -54,10 +54,12 @@ function normalizeOutput(text, opts = {}) {
 }
 
 // hook.mjs
-import { readFileSync, statSync, writeSync } from "node:fs";
-import { resolve } from "node:path";
+import { readFileSync, writeSync } from "node:fs";
+import { resolve as resolve2 } from "node:path";
 import { homedir } from "node:os";
-import { fileURLToPath } from "node:url";
+
+// lib/hook-policy.mjs
+import { resolve } from "node:path";
 var BINARY_EXT = /* @__PURE__ */ new Set([
   ".png",
   ".jpg",
@@ -98,7 +100,9 @@ var OUTLINEABLE_EXT = /* @__PURE__ */ new Set([
   ".tsx",
   ".py",
   ".cs",
-  ".php"
+  ".php",
+  ".md",
+  ".mdx"
 ]);
 var REVERSE_TOOL_HINTS = {
   "mcp__hex-line__read_file": "Read (file_path, offset, limit)",
@@ -109,7 +113,7 @@ var REVERSE_TOOL_HINTS = {
   "mcp__hex-line__outline": "Read with offset/limit",
   "mcp__hex-line__verify": "Read (re-read file to check freshness)",
   "mcp__hex-line__changes": "Bash(git diff)",
-  "mcp__hex-line__bulk_replace": "Edit (text rename/refactor across files)"
+  "mcp__hex-line__bulk_replace": "Edit (text rename/refactor across files inside an explicit root path)"
 };
 var TOOL_HINTS = {
   Read: "mcp__hex-line__read_file (not Read). For writing: write_file (no prior Read needed)",
@@ -122,12 +126,12 @@ var TOOL_HINTS = {
   ls: "mcp__hex-line__inspect_path for tree or pattern search (not ls/find/tree). E.g. pattern='*-mcp' type='dir'",
   stat: "mcp__hex-line__inspect_path for compact file metadata (not stat/wc/file)",
   grep: "mcp__hex-line__grep_search (not grep/rg). Params: output, literal, context_before, context_after, multiline",
-  sed: "mcp__hex-line__edit_file for hash edits, or mcp__hex-line__bulk_replace for text rename (not sed -i)",
+  sed: "mcp__hex-line__edit_file for hash edits, or mcp__hex-line__bulk_replace with path=<project root> for text rename (not sed -i)",
   diff: "mcp__hex-line__changes (not diff). Git diff with change symbols",
   outline: "mcp__hex-line__outline (before reading large code files)",
   verify: "mcp__hex-line__verify (staleness / revision check without re-read)",
   changes: "mcp__hex-line__changes (git diff with change symbols)",
-  bulk: "mcp__hex-line__bulk_replace (multi-file search-replace)"
+  bulk: "mcp__hex-line__bulk_replace with path=<project root> (multi-file search-replace)"
 };
 var DEFERRED_HINT = "If schemas not loaded: ToolSearch('+hex-line read edit')";
 var BASH_REDIRECTS = [
@@ -136,9 +140,7 @@ var BASH_REDIRECTS = [
   { regex: /^tail\s+(?!-[fF])/, key: "tail" },
   { regex: /^(less|more)\s+/, key: "cat" },
   { regex: /^ls\s+-\S*R(\s|$)/, key: "ls" },
-  // ls -R, ls -laR (recursive only)
   { regex: /^dir\s+\/[sS](\s|$)/, key: "ls" },
-  // dir /s, dir /S (recursive only)
   { regex: /^tree\s+/, key: "ls" },
   { regex: /^find\s+/, key: "ls" },
   { regex: /^(stat|wc)\s+/, key: "stat" },
@@ -152,34 +154,13 @@ var TOOL_REDIRECT_MAP = {
   Grep: "Grep"
 };
 var DANGEROUS_PATTERNS = [
-  {
-    regex: /rm\s+(-[rf]+\s+)*[/~]/,
-    reason: "rm -rf on root/home directory"
-  },
-  {
-    regex: /git\s+push\s+(-f|--force)/,
-    reason: "force push can overwrite remote history"
-  },
-  {
-    regex: /git\s+reset\s+--hard/,
-    reason: "hard reset discards uncommitted changes"
-  },
-  {
-    regex: /DROP\s+(TABLE|DATABASE)/i,
-    reason: "DROP destroys data permanently"
-  },
-  {
-    regex: /chmod\s+777/,
-    reason: "chmod 777 removes all access restrictions"
-  },
-  {
-    regex: /mkfs/,
-    reason: "filesystem format destroys all data"
-  },
-  {
-    regex: /dd\s+if=\/dev\/zero/,
-    reason: "direct disk write destroys data"
-  }
+  { regex: /rm\s+(-[rf]+\s+)*[/~]/, reason: "rm -rf on root/home directory" },
+  { regex: /git\s+push\s+(-f|--force)/, reason: "force push can overwrite remote history" },
+  { regex: /git\s+reset\s+--hard/, reason: "hard reset discards uncommitted changes" },
+  { regex: /DROP\s+(TABLE|DATABASE)/i, reason: "DROP destroys data permanently" },
+  { regex: /chmod\s+777/, reason: "chmod 777 removes all access restrictions" },
+  { regex: /mkfs/, reason: "filesystem format destroys all data" },
+  { regex: /dd\s+if=\/dev\/zero/, reason: "direct disk write destroys data" }
 ];
 var COMPOUND_OPERATORS = /[|]|>>?|&&|\|\||;/;
 var CMD_PATTERNS = [
@@ -189,11 +170,24 @@ var CMD_PATTERNS = [
   [/pip install/i, "pip-install"],
   [/git (log|diff|status)/i, "git"]
 ];
-var LINE_THRESHOLD = 50;
-var HEAD_LINES = 15;
-var TAIL_LINES = 15;
-var LARGE_FILE_BYTES = 5 * 1024;
-var LARGE_EDIT_CHARS = 1200;
+var HOOK_OUTPUT_POLICY = {
+  lineThreshold: 50,
+  headLines: 15,
+  tailLines: 15
+};
+function buildAllowedClaudeSettingsPaths(cwd, home) {
+  const cwdNorm = cwd.replace(/\\/g, "/");
+  const homeNorm = home.replace(/\\/g, "/");
+  return [
+    resolve(cwdNorm, ".claude/settings.json"),
+    resolve(cwdNorm, ".claude/settings.local.json"),
+    resolve(homeNorm, ".claude/settings.json"),
+    resolve(homeNorm, ".claude/settings.local.json")
+  ].map((entry) => entry.replace(/\\/g, "/").toLowerCase());
+}
+
+// hook.mjs
+import { fileURLToPath } from "node:url";
 function extOf(filePath) {
   const dot = filePath.lastIndexOf(".");
   return dot !== -1 ? filePath.slice(dot).toLowerCase() : "";
@@ -203,16 +197,8 @@ function getFilePath(toolInput) {
 }
 function resolveToolPath(filePath) {
   if (!filePath) return "";
-  if (filePath.startsWith("~/")) return resolve(homedir(), filePath.slice(2));
-  return resolve(process.cwd(), filePath);
-}
-function getFileSize(filePath) {
-  if (!filePath) return null;
-  try {
-    return statSync(resolveToolPath(filePath)).size;
-  } catch {
-    return null;
-  }
+  if (filePath.startsWith("~/")) return resolve2(homedir(), filePath.slice(2));
+  return resolve2(process.cwd(), filePath);
 }
 function isPartialRead(toolInput) {
   return [toolInput.offset, toolInput.limit, toolInput.start_line, toolInput.end_line, toolInput.ranges].some((value) => value !== void 0 && value !== null && value !== "");
@@ -238,7 +224,7 @@ function isHexLineDisabled(configPath) {
   if (_hexLineDisabled !== null) return _hexLineDisabled;
   _hexLineDisabled = false;
   try {
-    const p = configPath || resolve(homedir(), ".claude.json");
+    const p = configPath || resolve2(homedir(), ".claude.json");
     const claudeJson = JSON.parse(readFileSync(p, "utf-8"));
     const projects = claudeJson.projects;
     if (!projects || typeof projects !== "object") return _hexLineDisabled;
@@ -264,7 +250,7 @@ function getHookMode() {
   if (_hookMode !== void 0) return _hookMode;
   _hookMode = "blocking";
   try {
-    const stateFile = resolve(process.cwd(), ".hex-skills/environment_state.json");
+    const stateFile = resolve2(process.cwd(), ".hex-skills/environment_state.json");
     const data = JSON.parse(readFileSync(stateFile, "utf-8"));
     if (data.hooks?.mode === "advisory") _hookMode = "advisory";
   } catch {
@@ -317,47 +303,26 @@ function handlePreToolUse(data) {
   const hintKey = TOOL_REDIRECT_MAP[toolName];
   if (hintKey) {
     const filePath = getFilePath(toolInput);
-    const fileSize = getFileSize(filePath);
     if (BINARY_EXT.has(extOf(filePath))) {
       process.exit(0);
     }
     const resolvedNorm = resolveToolPath(filePath).replace(/\\/g, "/");
-    const cwdNorm = process.cwd().replace(/\\/g, "/");
-    const homeNorm = homedir().replace(/\\/g, "/");
-    const claudeAllow = [
-      cwdNorm + "/.claude/settings.json",
-      cwdNorm + "/.claude/settings.local.json",
-      homeNorm + "/.claude/settings.json",
-      homeNorm + "/.claude/settings.local.json"
-    ];
-    if (claudeAllow.some((p) => resolvedNorm.toLowerCase() === p.toLowerCase())) {
+    const claudeAllow = buildAllowedClaudeSettingsPaths(process.cwd(), homedir());
+    if (claudeAllow.includes(resolvedNorm.toLowerCase())) {
       process.exit(0);
     }
     if (resolvedNorm.includes("/.claude/")) {
       redirect("Protected .claude/ path. Use built-in tools for .claude/ config files.");
     }
     if (toolName === "Read") {
-      if (isPartialRead(toolInput)) {
-        process.exit(0);
-      }
-      if (fileSize !== null && fileSize <= LARGE_FILE_BYTES) {
-        const ext2 = filePath ? extOf(filePath) : "";
-        const hint = filePath && OUTLINEABLE_EXT.has(ext2) ? `Use mcp__hex-line__outline(path="${filePath}") for structure, then mcp__hex-line__read_file(path="${filePath}") with ranges.` : filePath ? `Use mcp__hex-line__read_file(path="${filePath}"). Built-in Read wastes edit context.` : "Use mcp__hex-line__read_file. Built-in Read wastes edit context.";
-        advise(hint, DEFERRED_HINT);
-      }
       const ext = filePath ? extOf(filePath) : "";
-      const outlineHint = filePath && OUTLINEABLE_EXT.has(ext) ? `Use mcp__hex-line__outline(path="${filePath}") for structure, then mcp__hex-line__read_file(path="${filePath}") with ranges to read only what you need.` : filePath ? `Use mcp__hex-line__read_file(path="${filePath}") with ranges or offset/limit` : "Use mcp__hex-line__inspect_path or mcp__hex-line__read_file";
-      redirect(outlineHint, "Do not use built-in Read for full reads of large files.\n" + DEFERRED_HINT);
+      const rangeHint = isPartialRead(toolInput) ? " Preserve the same offset/limit or ranges in read_file." : "";
+      const outlineHint = filePath && OUTLINEABLE_EXT.has(ext) ? `Use mcp__hex-line__outline(path="${filePath}") for structure, then mcp__hex-line__read_file(path="${filePath}") with ranges to read only what you need.${rangeHint}` : filePath ? `Use mcp__hex-line__read_file(path="${filePath}") with ranges or offset/limit.${rangeHint}` : "Use mcp__hex-line__inspect_path or mcp__hex-line__read_file";
+      redirect(outlineHint, "Use hex-line for text-file reads to keep hashes, revision metadata, and graph hints in one flow.\n" + DEFERRED_HINT);
     }
     if (toolName === "Edit") {
-      const oldText = String(toolInput.old_string || "");
-      const isLargeEdit = Boolean(toolInput.replace_all) || oldText.length > LARGE_EDIT_CHARS || fileSize !== null && fileSize > LARGE_FILE_BYTES;
-      if (!isLargeEdit) {
-        const editHint = filePath ? `Prefer mcp__hex-line__edit_file(path="${filePath}") for hash-verified edits.` : "Prefer mcp__hex-line__edit_file for hash-verified edits.";
-        advise(editHint);
-      }
       const target = filePath ? `Use mcp__hex-line__grep_search or mcp__hex-line__read_file, then mcp__hex-line__edit_file with path="${filePath}"` : "Use mcp__hex-line__grep_search or mcp__hex-line__read_file, then mcp__hex-line__edit_file";
-      redirect(target, "For large or repeated edits: locate anchors/checksums first, then call edit_file once with batched edits.\n" + DEFERRED_HINT);
+      redirect(target, "Use hash-verified edits for text files. Locate anchors/checksums first, then call edit_file once with batched edits.\n" + DEFERRED_HINT);
     }
     if (toolName === "Write") {
       const pathNote = filePath ? ` with path="${filePath}"` : "";
@@ -433,11 +398,14 @@ function handlePostToolUse(data) {
   }
   const lines = rawText.split("\n");
   const originalCount = lines.length;
-  if (originalCount < LINE_THRESHOLD) {
+  if (originalCount < HOOK_OUTPUT_POLICY.lineThreshold) {
     process.exit(0);
   }
   const type = detectCommandType(command);
-  const filtered = normalizeOutput(lines.join("\n"), { headLines: HEAD_LINES, tailLines: TAIL_LINES });
+  const filtered = normalizeOutput(lines.join("\n"), {
+    headLines: HOOK_OUTPUT_POLICY.headLines,
+    tailLines: HOOK_OUTPUT_POLICY.tailLines
+  });
   const filteredCount = filtered.split("\n").length;
   const header = `RTK FILTERED: ${type} (${originalCount} lines -> ${filteredCount} lines)`;
   const output = [
@@ -455,9 +423,9 @@ function handlePostToolUse(data) {
 }
 function handleSessionStart() {
   const settingsFiles = [
-    resolve(process.cwd(), ".claude/settings.local.json"),
-    resolve(process.cwd(), ".claude/settings.json"),
-    resolve(homedir(), ".claude/settings.json")
+    resolve2(process.cwd(), ".claude/settings.local.json"),
+    resolve2(process.cwd(), ".claude/settings.json"),
+    resolve2(homedir(), ".claude/settings.json")
   ];
   let styleActive = false;
   for (const f of settingsFiles) {
@@ -471,8 +439,7 @@ function handleSessionStart() {
     } catch {
     }
   }
-  const prefix = styleActive ? "Hex-line MCP available. Output style active.\n" : "Hex-line MCP available.\n";
-  const msg = prefix + "<hex-line_instructions>\n  <deferred_loading>If hex-line schemas not loaded, run: ToolSearch('+hex-line read edit')</deferred_loading>\n  <exploration>\n    <rule>Use outline for structure (code + markdown), not Read. ~10-20 lines vs hundreds.</rule>\n    <rule>Use read_file with offset/limit or ranges for targeted reads.</rule>\n    <rule>Use grep_search before editing to get hash anchors.</rule>\n  </exploration>\n  <editing>\n    <path name='surgical'>grep_search \u2192 edit_file (fastest: hash-verified, no full read needed)</path>\n    <path name='exploratory'>outline \u2192 read_file (ranges) \u2192 edit_file with base_revision</path>\n    <path name='multi-file'>bulk_replace for text rename/refactor across files</path>\n  </editing>\n  <tips>\n    <tip>Carry revision from read_file into base_revision on edit_file.</tip>\n    <tip>If edit returns CONFLICT, call verify \u2014 only reread when STALE.</tip>\n    <tip>Batch multiple edits to same file in one edit_file call.</tip>\n    <tip>Use write_file for new files (no prior Read needed).</tip>\n  </tips>\n  <exceptions>Built-in Read OK for: images, PDFs, notebooks, Glob (always), .claude/settings.json</exceptions>\n</hex-line_instructions>";
+  const msg = styleActive ? "Hex-line MCP available. Output style active.\n<hex-line_instructions>\n  <deferred_loading>If hex-line schemas not loaded, run: ToolSearch('+hex-line read edit')</deferred_loading>\n  <note>Follow the active hex-line output style for primary tool choices.</note>\n  <exceptions>Built-in tools stay OK for images, PDFs, notebooks, Glob, .claude/settings.json, and .claude/settings.local.json.</exceptions>\n</hex-line_instructions>" : "Hex-line MCP available.\n<hex-line_instructions>\n  <deferred_loading>If hex-line schemas not loaded, run: ToolSearch('+hex-line read edit')</deferred_loading>\n  <exploration>\n    <rule>Use outline for structure (code + markdown), not Read. ~10-20 lines vs hundreds.</rule>\n    <rule>Use read_file with offset/limit or ranges for targeted reads.</rule>\n    <rule>Use grep_search before editing to get hash anchors.</rule>\n  </exploration>\n  <editing>\n    <path name='surgical'>grep_search \u2192 edit_file (fastest: hash-verified, no full read needed)</path>\n    <path name='exploratory'>outline \u2192 read_file (ranges) \u2192 edit_file with base_revision</path>\n    <path name='multi-file'>bulk_replace(path=&quot;&lt;project root&gt;&quot;) for text rename/refactor across files</path>\n  </editing>\n  <tips>\n    <tip>Auto-fill path from the active file or project root. Do not leave repo scope implicit.</tip>\n    <tip>Never invent range_checksum. Copy it from fresh read_file or grep_search blocks.</tip>\n    <tip>Prefer set_line or insert_after for small local changes and replace_between for larger bounded rewrites.</tip>\n    <tip>Carry revision from read_file into base_revision on edit_file.</tip>\n    <tip>If edit returns CONFLICT, call verify \u2014 only reread when STALE.</tip>\n    <tip>Avoid large first-pass edit batches. Start with 1-2 hunks, then continue from the returned revision.</tip>\n    <tip>Use write_file for new files (no prior Read needed).</tip>\n  </tips>\n  <exceptions>Built-in tools stay OK for images, PDFs, notebooks, Glob, .claude/settings.json, and .claude/settings.local.json.</exceptions>\n</hex-line_instructions>";
   safeExit(1, JSON.stringify({ systemMessage: msg }), 0);
 }
 var _norm = (p) => p.replace(/\\/g, "/");