sync: auto-stage dirty hub state (recovery)

2026-03-27 17:06:54 +01:00
parent 28b75b61fe
commit 2b165afc41
9 changed files with 2554 additions and 0 deletions
--- a/.claude/hooks/prompt-guard.py
+++ b/.claude/hooks/prompt-guard.py
@@ -0,0 +1,799 @@
+#!/usr/bin/env python3
+"""
+Crosslink behavioral hook for Claude Code.
+Injects best practice reminders on every prompt submission.
+Loads rules from .crosslink/rules/ markdown files.
+"""
+
+import json
+import sys
+import os
+import io
+import subprocess
+import hashlib
+from datetime import datetime
+
+# Fix Windows encoding issues with Unicode characters
+sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8')
+
+# Add hooks directory to path for shared module import
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from crosslink_config import (
+    find_crosslink_dir,
+    get_project_root,
+    is_agent_context,
+    load_config_merged,
+    load_guard_state,
+    load_tracking_mode,
+    save_guard_state,
+)
+
+
+def load_rule_file(rules_dir, filename, rules_local_dir=None):
+    """Load a rule file, preferring rules.local/ override if present."""
+    if not rules_dir:
+        return ""
+    # Check rules.local/ first for an override
+    if rules_local_dir:
+        local_path = os.path.join(rules_local_dir, filename)
+        try:
+            with open(local_path, 'r', encoding='utf-8') as f:
+                return f.read().strip()
+        except (OSError, IOError):
+            pass
+    # Fall back to rules/
+    path = os.path.join(rules_dir, filename)
+    try:
+        with open(path, 'r', encoding='utf-8') as f:
+            return f.read().strip()
+    except (OSError, IOError):
+        return ""
+
+
+def load_all_rules(crosslink_dir):
+    """Load all rule files from .crosslink/rules/, with .crosslink/rules.local/ overrides.
+
+    Auto-discovers all .md files in the rules directory. Files are categorized as:
+    - Well-known names: global.md, project.md, knowledge.md, quality.md
+    - Language files: matched by known language filename patterns
+    - Extra rules: any other .md file (loaded as additional general rules)
+
+    Files in rules.local/ override same-named files in rules/.
+    """
+    if not crosslink_dir:
+        return {}, "", "", "", ""
+
+    rules_dir = os.path.join(crosslink_dir, 'rules')
+    rules_local_dir = os.path.join(crosslink_dir, 'rules.local')
+    if not os.path.isdir(rules_dir) and not os.path.isdir(rules_local_dir):
+        return {}, "", "", "", ""
+
+    if not os.path.isdir(rules_local_dir):
+        rules_local_dir = None
+
+    # Well-known non-language files (loaded into specific return values)
+    WELL_KNOWN = {'global.md', 'project.md', 'knowledge.md', 'quality.md'}
+
+    # Internal/structural files (not injected as rules)
+    SKIP_FILES = {
+        'sanitize-patterns.txt',
+        'tracking-strict.md', 'tracking-normal.md', 'tracking-relaxed.md',
+    }
+
+    # Language filename -> display name mapping
+    LANGUAGE_MAP = {
+        'rust.md': 'Rust', 'python.md': 'Python',
+        'javascript.md': 'JavaScript', 'typescript.md': 'TypeScript',
+        'typescript-react.md': 'TypeScript/React',
+        'javascript-react.md': 'JavaScript/React',
+        'go.md': 'Go', 'java.md': 'Java', 'c.md': 'C', 'cpp.md': 'C++',
+        'csharp.md': 'C#', 'ruby.md': 'Ruby', 'php.md': 'PHP',
+        'swift.md': 'Swift', 'kotlin.md': 'Kotlin', 'scala.md': 'Scala',
+        'zig.md': 'Zig', 'odin.md': 'Odin',
+        'elixir.md': 'Elixir', 'elixir-phoenix.md': 'Elixir/Phoenix',
+        'web.md': 'Web',
+    }
+
+    # Load well-known files
+    global_rules = load_rule_file(rules_dir, 'global.md', rules_local_dir)
+    project_rules = load_rule_file(rules_dir, 'project.md', rules_local_dir)
+    knowledge_rules = load_rule_file(rules_dir, 'knowledge.md', rules_local_dir)
+    quality_rules = load_rule_file(rules_dir, 'quality.md', rules_local_dir)
+
+    # Auto-discover all files from both directories
+    language_rules = {}
+    all_files = set()
+
+    try:
+        if os.path.isdir(rules_dir):
+            for entry in os.listdir(rules_dir):
+                if entry.endswith('.md') or entry.endswith('.txt'):
+                    all_files.add(entry)
+    except OSError:
+        pass
+
+    if rules_local_dir:
+        try:
+            for entry in os.listdir(rules_local_dir):
+                if entry.endswith('.md') or entry.endswith('.txt'):
+                    all_files.add(entry)
+        except OSError:
+            pass
+
+    for filename in sorted(all_files):
+        if filename in WELL_KNOWN or filename in SKIP_FILES:
+            continue
+        if filename in LANGUAGE_MAP:
+            content = load_rule_file(rules_dir, filename, rules_local_dir)
+            if content:
+                language_rules[LANGUAGE_MAP[filename]] = content
+        elif filename.endswith('.md'):
+            content = load_rule_file(rules_dir, filename, rules_local_dir)
+            if content:
+                lang_name = os.path.splitext(filename)[0].replace('-', '/').title()
+                language_rules[lang_name] = content
+
+    return language_rules, global_rules, project_rules, knowledge_rules, quality_rules
+
+
+# Detect language from common file extensions in the working directory
+def detect_languages():
+    """Scan for common source files to determine active languages."""
+    extensions = {
+        '.rs': 'Rust',
+        '.py': 'Python',
+        '.js': 'JavaScript',
+        '.ts': 'TypeScript',
+        '.tsx': 'TypeScript/React',
+        '.jsx': 'JavaScript/React',
+        '.go': 'Go',
+        '.java': 'Java',
+        '.c': 'C',
+        '.cpp': 'C++',
+        '.cs': 'C#',
+        '.rb': 'Ruby',
+        '.php': 'PHP',
+        '.swift': 'Swift',
+        '.kt': 'Kotlin',
+        '.scala': 'Scala',
+        '.zig': 'Zig',
+        '.odin': 'Odin',
+        '.ex': 'Elixir',
+        '.exs': 'Elixir',
+        '.heex': 'Elixir/Phoenix',
+    }
+
+    found = set()
+    cwd = get_project_root()
+
+    # Check for project config files first (more reliable than scanning)
+    config_indicators = {
+        'Cargo.toml': 'Rust',
+        'package.json': 'JavaScript',
+        'tsconfig.json': 'TypeScript',
+        'pyproject.toml': 'Python',
+        'requirements.txt': 'Python',
+        'go.mod': 'Go',
+        'pom.xml': 'Java',
+        'build.gradle': 'Java',
+        'Gemfile': 'Ruby',
+        'composer.json': 'PHP',
+        'Package.swift': 'Swift',
+        'mix.exs': 'Elixir',
+    }
+
+    # Check cwd and immediate subdirs for config files
+    check_dirs = [cwd]
+    try:
+        for entry in os.listdir(cwd):
+            subdir = os.path.join(cwd, entry)
+            if os.path.isdir(subdir) and not entry.startswith('.'):
+                check_dirs.append(subdir)
+    except (PermissionError, OSError):
+        pass
+
+    for check_dir in check_dirs:
+        for config_file, lang in config_indicators.items():
+            if os.path.exists(os.path.join(check_dir, config_file)):
+                found.add(lang)
+
+    # Also scan for source files in src/ directories
+    scan_dirs = [cwd]
+    src_dir = os.path.join(cwd, 'src')
+    if os.path.isdir(src_dir):
+        scan_dirs.append(src_dir)
+    # Check nested project src dirs too
+    for check_dir in check_dirs:
+        nested_src = os.path.join(check_dir, 'src')
+        if os.path.isdir(nested_src):
+            scan_dirs.append(nested_src)
+
+    for scan_dir in scan_dirs:
+        try:
+            for entry in os.listdir(scan_dir):
+                ext = os.path.splitext(entry)[1].lower()
+                if ext in extensions:
+                    found.add(extensions[ext])
+        except (PermissionError, OSError):
+            pass
+
+    return list(found) if found else ['the project']
+
+
+def get_language_section(languages, language_rules):
+    """Build language-specific best practices section from loaded rules."""
+    sections = []
+    for lang in languages:
+        if lang in language_rules:
+            content = language_rules[lang]
+            # If the file doesn't start with a header, add one
+            if not content.startswith('#'):
+                sections.append(f"### {lang} Best Practices\n{content}")
+            else:
+                sections.append(content)
+
+    if not sections:
+        return ""
+
+    return "\n\n".join(sections)
+
+
+# Directories to skip when building project tree
+SKIP_DIRS = {
+    '.git', 'node_modules', 'target', 'venv', '.venv', 'env', '.env',
+    '__pycache__', '.crosslink', '.claude', 'dist', 'build', '.next',
+    '.nuxt', 'vendor', '.idea', '.vscode', 'coverage', '.pytest_cache',
+    '.mypy_cache', '.tox', 'eggs', '*.egg-info', '.sass-cache',
+    '_build', 'deps', '.elixir_ls', '.fetch'
+}
+
+
+def get_project_tree(max_depth=3, max_entries=50):
+    """Generate a compact project tree to prevent path hallucinations."""
+    cwd = get_project_root()
+    entries = []
+
+    def should_skip(name):
+        if name.startswith('.') and name not in ('.github', '.claude'):
+            return True
+        return name in SKIP_DIRS or name.endswith('.egg-info')
+
+    def walk_dir(path, prefix="", depth=0):
+        if depth > max_depth or len(entries) >= max_entries:
+            return
+
+        try:
+            items = sorted(os.listdir(path))
+        except (PermissionError, OSError):
+            return
+
+        # Separate dirs and files
+        dirs = [i for i in items if os.path.isdir(os.path.join(path, i)) and not should_skip(i)]
+        files = [i for i in items if os.path.isfile(os.path.join(path, i)) and not i.startswith('.')]
+
+        # Add files first (limit per directory)
+        for f in files[:10]:  # Max 10 files per dir shown
+            if len(entries) >= max_entries:
+                return
+            entries.append(f"{prefix}{f}")
+
+        if len(files) > 10:
+            entries.append(f"{prefix}... ({len(files) - 10} more files)")
+
+        # Then recurse into directories
+        for d in dirs:
+            if len(entries) >= max_entries:
+                return
+            entries.append(f"{prefix}{d}/")
+            walk_dir(os.path.join(path, d), prefix + "  ", depth + 1)
+
+    walk_dir(cwd)
+
+    if not entries:
+        return ""
+
+    if len(entries) >= max_entries:
+        entries.append(f"... (tree truncated at {max_entries} entries)")
+
+    return "\n".join(entries)
+
+
+
+def get_lock_file_hash(lock_path):
+    """Get a hash of the lock file for cache invalidation."""
+    try:
+        mtime = os.path.getmtime(lock_path)
+        return hashlib.md5(f"{lock_path}:{mtime}".encode()).hexdigest()[:12]
+    except OSError:
+        return None
+
+
+def run_command(cmd, timeout=5):
+    """Run a command and return output, or None on failure."""
+    try:
+        result = subprocess.run(
+            cmd,
+            capture_output=True,
+            text=True,
+            timeout=timeout,
+            shell=True
+        )
+        if result.returncode == 0:
+            return result.stdout.strip()
+    except (subprocess.TimeoutExpired, OSError, Exception):
+        pass
+    return None
+
+
+def get_dependencies(max_deps=30):
+    """Get installed dependencies with versions. Uses caching based on lock file mtime."""
+    cwd = get_project_root()
+    deps = []
+
+    # Check for Rust (Cargo.toml)
+    cargo_toml = os.path.join(cwd, 'Cargo.toml')
+    if os.path.exists(cargo_toml):
+        # Parse Cargo.toml for direct dependencies (faster than cargo tree)
+        try:
+            with open(cargo_toml, 'r') as f:
+                content = f.read()
+                in_deps = False
+                for line in content.split('\n'):
+                    if line.strip().startswith('[dependencies]'):
+                        in_deps = True
+                        continue
+                    if line.strip().startswith('[') and in_deps:
+                        break
+                    if in_deps and '=' in line and not line.strip().startswith('#'):
+                        parts = line.split('=', 1)
+                        name = parts[0].strip()
+                        rest = parts[1].strip() if len(parts) > 1 else ''
+                        if rest.startswith('{'):
+                            # Handle { version = "x.y", features = [...] } format
+                            import re
+                            match = re.search(r'version\s*=\s*"([^"]+)"', rest)
+                            if match:
+                                deps.append(f"  {name} = \"{match.group(1)}\"")
+                        elif rest.startswith('"') or rest.startswith("'"):
+                            version = rest.strip('"').strip("'")
+                            deps.append(f"  {name} = \"{version}\"")
+                        if len(deps) >= max_deps:
+                            break
+        except (OSError, Exception):
+            pass
+        if deps:
+            return "Rust (Cargo.toml):\n" + "\n".join(deps[:max_deps])
+
+    # Check for Node.js (package.json)
+    package_json = os.path.join(cwd, 'package.json')
+    if os.path.exists(package_json):
+        try:
+            with open(package_json, 'r') as f:
+                pkg = json.load(f)
+                for dep_type in ['dependencies', 'devDependencies']:
+                    if dep_type in pkg:
+                        for name, version in list(pkg[dep_type].items())[:max_deps]:
+                            deps.append(f"  {name}: {version}")
+                            if len(deps) >= max_deps:
+                                break
+        except (OSError, json.JSONDecodeError, Exception):
+            pass
+        if deps:
+            return "Node.js (package.json):\n" + "\n".join(deps[:max_deps])
+
+    # Check for Python (requirements.txt or pyproject.toml)
+    requirements = os.path.join(cwd, 'requirements.txt')
+    if os.path.exists(requirements):
+        try:
+            with open(requirements, 'r') as f:
+                for line in f:
+                    line = line.strip()
+                    if line and not line.startswith('#') and not line.startswith('-'):
+                        deps.append(f"  {line}")
+                        if len(deps) >= max_deps:
+                            break
+        except (OSError, Exception):
+            pass
+        if deps:
+            return "Python (requirements.txt):\n" + "\n".join(deps[:max_deps])
+
+    # Check for Elixir (mix.exs)
+    mix_exs = os.path.join(cwd, 'mix.exs')
+    if os.path.exists(mix_exs):
+        try:
+            import re
+            with open(mix_exs, 'r') as f:
+                content = f.read()
+                # Match {:dep_name, "~> x.y"} or {:dep_name, ">= x.y"} patterns
+                for match in re.finditer(r'\{:(\w+),\s*"([^"]+)"', content):
+                    deps.append(f"  {match.group(1)}: {match.group(2)}")
+                    if len(deps) >= max_deps:
+                        break
+        except (OSError, Exception):
+            pass
+        if deps:
+            return "Elixir (mix.exs):\n" + "\n".join(deps[:max_deps])
+
+    # Check for Go (go.mod)
+    go_mod = os.path.join(cwd, 'go.mod')
+    if os.path.exists(go_mod):
+        try:
+            with open(go_mod, 'r') as f:
+                in_require = False
+                for line in f:
+                    line = line.strip()
+                    if line.startswith('require ('):
+                        in_require = True
+                        continue
+                    if line == ')' and in_require:
+                        break
+                    if in_require and line:
+                        deps.append(f"  {line}")
+                        if len(deps) >= max_deps:
+                            break
+        except (OSError, Exception):
+            pass
+        if deps:
+            return "Go (go.mod):\n" + "\n".join(deps[:max_deps])
+
+    return ""
+
+
+def build_reminder(languages, project_tree, dependencies, language_rules, global_rules, project_rules, tracking_mode="strict", crosslink_dir=None, knowledge_rules="", quality_rules=""):
+    """Build the full reminder context."""
+    lang_section = get_language_section(languages, language_rules)
+    lang_list = ", ".join(languages) if languages else "this project"
+    current_year = datetime.now().year
+
+    # Build tree section if available
+    tree_section = ""
+    if project_tree:
+        tree_section = f"""
+### Project Structure (use these exact paths)
+```
+{project_tree}
+```
+"""
+
+    # Build dependencies section if available
+    deps_section = ""
+    if dependencies:
+        deps_section = f"""
+### Installed Dependencies (use these exact versions)
+```
+{dependencies}
+```
+"""
+
+    # Build global rules section (from .crosslink/rules/global.md)
+    # Then append/replace the tracking section based on tracking_mode
+    global_section = ""
+    if global_rules:
+        global_section = f"\n{global_rules}\n"
+    else:
+        # Fallback to hardcoded defaults if no rules file
+        global_section = f"""
+### Pre-Coding Grounding (PREVENT HALLUCINATIONS)
+Before writing code that uses external libraries, APIs, or unfamiliar patterns:
+1. **VERIFY IT EXISTS**: Use WebSearch to confirm the crate/package/module exists and check its actual API
+2. **CHECK THE DOCS**: Fetch documentation to see real function signatures, not imagined ones
+3. **CONFIRM SYNTAX**: If unsure about language features or library usage, search first
+4. **USE LATEST VERSIONS**: Always check for and use the latest stable version of dependencies (security + features)
+5. **NO GUESSING**: If you can't verify it, tell the user you need to research it
+
+Examples of when to search:
+- Using a crate/package you haven't used recently → search "[package] [language] docs {current_year}"
+- Uncertain about function parameters → search for actual API reference
+- New language feature or syntax → verify it exists in the version being used
+- System calls or platform-specific code → confirm the correct API
+- Adding a dependency → search "[package] latest version {current_year}" to get current release
+
+### General Requirements
+1. **NO STUBS - ABSOLUTE RULE**:
+   - NEVER write `TODO`, `FIXME`, `pass`, `...`, `unimplemented!()` as implementation
+   - NEVER write empty function bodies or placeholder returns
+   - NEVER say "implement later" or "add logic here"
+   - If logic is genuinely too complex for one turn, use `raise NotImplementedError("Descriptive reason: what needs to be done")` and create a crosslink issue
+   - The PostToolUse hook WILL detect and flag stub patterns - write real code the first time
+2. **NO DEAD CODE**: Discover if dead code is truly dead or if it's an incomplete feature. If incomplete, complete it. If truly dead, remove it.
+3. **FULL FEATURES**: Implement the complete feature as requested. Don't stop partway or suggest "you could add X later."
+4. **ERROR HANDLING**: Proper error handling everywhere. No panics/crashes on bad input.
+5. **SECURITY**: Validate input, use parameterized queries, no command injection, no hardcoded secrets.
+6. **READ BEFORE WRITE**: Always read a file before editing it. Never guess at contents.
+
+### Conciseness Protocol
+Minimize chattiness. Your output should be:
+- **Code blocks** with implementation
+- **Tool calls** to accomplish tasks
+- **Brief explanations** only when the code isn't self-explanatory
+
+NEVER output:
+- "Here is the code" / "Here's how to do it" (just show the code)
+- "Let me know if you need anything else" / "Feel free to ask"
+- "I'll now..." / "Let me..." (just do it)
+- Restating what the user asked
+- Explaining obvious code
+- Multiple paragraphs when one sentence suffices
+
+When writing code: write it. When making changes: make them. Skip the narration.
+
+### Large File Management (500+ lines)
+If you need to write or modify code that will exceed 500 lines:
+1. Create a parent issue for the overall feature: `crosslink issue create "<feature name>" -p high`
+2. Break down into subissues: `crosslink issue subissue <parent_id> "<component 1>"`, etc.
+3. Inform the user: "This implementation will require multiple files/components. I've created issue #X with Y subissues to track progress."
+4. Work on one subissue at a time, marking each complete before moving on.
+
+### Context Window Management
+If the conversation is getting long OR the task requires many more steps:
+1. Create a crosslink issue to track remaining work: `crosslink issue create "Continue: <task summary>" -p high`
+2. Add detailed notes as a comment: `crosslink issue comment <id> "<what's done, what's next>"`
+3. Inform the user: "This task will require additional turns. I've created issue #X to track progress."
+
+Use `crosslink session work <id>` to mark what you're working on.
+"""
+
+    # Inject tracking rules from per-mode markdown file
+    tracking_rules = load_tracking_rules(crosslink_dir, tracking_mode) if crosslink_dir else ""
+    tracking_section = f"\n{tracking_rules}\n" if tracking_rules else ""
+
+    # Build project rules section (from .crosslink/rules/project.md)
+    project_section = ""
+    if project_rules:
+        project_section = f"\n### Project-Specific Rules\n{project_rules}\n"
+
+    # Build knowledge section (from .crosslink/rules/knowledge.md)
+    knowledge_section = ""
+    if knowledge_rules:
+        knowledge_section = f"\n{knowledge_rules}\n"
+
+    # Build quality section (from .crosslink/rules/quality.md)
+    quality_section = ""
+    if quality_rules:
+        quality_section = f"\n{quality_rules}\n"
+
+    reminder = f"""<crosslink-behavioral-guard>
+## Code Quality Requirements
+
+You are working on a {lang_list} project. Follow these requirements strictly:
+{tree_section}{deps_section}{global_section}{tracking_section}{quality_section}{lang_section}{project_section}{knowledge_section}
+</crosslink-behavioral-guard>"""
+
+    return reminder
+
+
+def get_guard_marker_path(crosslink_dir):
+    """Get the path to the guard-full-sent marker file."""
+    if not crosslink_dir:
+        return None
+    cache_dir = os.path.join(crosslink_dir, '.cache')
+    return os.path.join(cache_dir, 'guard-full-sent')
+
+
+def should_send_full_guard(crosslink_dir):
+    """Check if this is the first prompt (no marker) or marker is stale."""
+    marker = get_guard_marker_path(crosslink_dir)
+    if not marker:
+        return True
+    if not os.path.exists(marker):
+        return True
+    # Re-send full guard if marker is older than 4 hours (new session likely)
+    try:
+        age = datetime.now().timestamp() - os.path.getmtime(marker)
+        if age > 4 * 3600:
+            return True
+    except OSError:
+        return True
+    return False
+
+
+def mark_full_guard_sent(crosslink_dir):
+    """Create marker file indicating full guard has been sent this session."""
+    marker = get_guard_marker_path(crosslink_dir)
+    if not marker:
+        return
+    try:
+        cache_dir = os.path.dirname(marker)
+        os.makedirs(cache_dir, exist_ok=True)
+        with open(marker, 'w') as f:
+            f.write(str(datetime.now().timestamp()))
+    except OSError:
+        pass
+
+
+def load_tracking_rules(crosslink_dir, tracking_mode):
+    """Load the tracking rules markdown file for the given mode.
+
+    Checks rules.local/ first for a local override, then falls back to rules/.
+    """
+    if not crosslink_dir:
+        return ""
+    filename = f"tracking-{tracking_mode}.md"
+    # Check rules.local/ first
+    local_path = os.path.join(crosslink_dir, "rules.local", filename)
+    try:
+        with open(local_path, "r", encoding="utf-8") as f:
+            return f.read().strip()
+    except (OSError, IOError):
+        pass
+    # Fall back to rules/
+    path = os.path.join(crosslink_dir, "rules", filename)
+    try:
+        with open(path, "r", encoding="utf-8") as f:
+            return f.read().strip()
+    except (OSError, IOError):
+        return ""
+
+
+# Condensed reminders kept short — these don't need full markdown files
+CONDENSED_REMINDERS = {
+    "strict": (
+        "- **MANDATORY — Crosslink Issue Tracking**: You MUST create a crosslink issue BEFORE writing ANY code. "
+        "NO EXCEPTIONS. Use `crosslink quick \"title\" -p <priority> -l <label>` BEFORE your first Write/Edit/Bash. "
+        "If you skip this, the PreToolUse hook WILL block you. Do NOT treat this as optional.\n"
+        "- **Session**: ALWAYS use `crosslink session work <id>` to mark focus. "
+        "End with `crosslink session end --notes \"...\"`. This is NOT optional."
+    ),
+    "normal": (
+        "- **Crosslink**: Create issues before work. Use `crosslink quick` for create+label+work. Close with `crosslink close`.\n"
+        "- **Session**: Use `crosslink session work <id>`. End with `crosslink session end --notes \"...\"`."
+    ),
+    "relaxed": "",
+}
+
+
+def build_condensed_reminder(languages, tracking_mode):
+    """Build a short reminder for subsequent prompts (after full guard already sent)."""
+    lang_list = ", ".join(languages) if languages else "this project"
+    tracking_lines = CONDENSED_REMINDERS.get(tracking_mode, "")
+
+    return f"""<crosslink-behavioral-guard>
+## Quick Reminder ({lang_list})
+
+{tracking_lines}
+- **Security**: Use `mcp__crosslink-safe-fetch__safe_fetch` for web requests. Parameterized queries only.
+- **Quality**: No stubs/TODOs. Read before write. Complete features fully. Proper error handling.
+- **Testing**: Run tests after changes. Fix warnings, don't suppress them.
+
+Full rules were injected on first prompt. Use `crosslink issue list -s open` to see current issues.
+</crosslink-behavioral-guard>"""
+
+
+def estimate_prompt_chars(input_data):
+    """Estimate characters consumed by this prompt turn.
+
+    The hook only sees the user prompt, not tool outputs or model responses.
+    We apply a multiplier (5x) to account for the full turn cost:
+    user prompt + tool calls + tool results + model response.
+    """
+    TURN_MULTIPLIER = 5
+    try:
+        prompt_text = input_data.get("prompt", "")
+        if isinstance(prompt_text, str):
+            return len(prompt_text) * TURN_MULTIPLIER
+        return 2000 * TURN_MULTIPLIER
+    except (AttributeError, TypeError):
+        return 2000 * TURN_MULTIPLIER
+
+
+def check_context_budget(crosslink_dir, state, prompt_chars):
+    """Check if estimated context usage has exceeded the budget.
+
+    Returns True if the budget is exceeded and full reinjection is needed.
+    Default budget: 1,000,000 chars ~ 250k tokens.
+    """
+    config = load_config_merged(crosslink_dir) if crosslink_dir else {}
+    budget = int(config.get("context_budget_chars", 1_000_000))
+    if budget <= 0:
+        return False
+
+    current = state.get("estimated_context_chars", 0)
+    current += prompt_chars
+    state["estimated_context_chars"] = current
+
+    return current >= budget
+
+
+def build_context_budget_warning(languages, tracking_mode):
+    """Build the compression directive when context budget is exceeded."""
+    lang_list = ", ".join(languages) if languages else "this project"
+    tracking_lines = CONDENSED_REMINDERS.get(tracking_mode, "")
+
+    return f"""<crosslink-context-budget-exceeded>
+## CONTEXT BUDGET EXCEEDED — COMPRESSION REQUIRED
+
+Your estimated context usage has exceeded 250k tokens. Research shows instruction
+adherence degrades significantly past this point. You MUST take the following steps
+IMMEDIATELY, before doing anything else:
+
+1. **Record your current state**: Run `crosslink session action "Context budget reached. Working on: <current task summary>"`
+2. **Save any in-progress work context** as a crosslink comment: `crosslink issue comment <id> "Progress: <what's done, what's next>" --kind observation`
+3. **The system will compress context automatically.** After compression, re-read any files you need and continue working.
+
+## Re-injected Rules ({lang_list})
+
+{tracking_lines}
+- **Security**: Use `mcp__crosslink-safe-fetch__safe_fetch` for web requests. Parameterized queries only.
+- **Quality**: No stubs/TODOs. Read before write. Complete features fully. Proper error handling.
+- **Testing**: Run tests after changes. Fix warnings, don't suppress them.
+- **Documentation**: Add typed crosslink comments (--kind plan/decision/observation/result) at every step.
+</crosslink-context-budget-exceeded>"""
+
+
+def main():
+    input_data = {}
+    try:
+        # Read input from stdin (Claude Code passes prompt info)
+        input_data = json.load(sys.stdin)
+    except json.JSONDecodeError:
+        pass
+    except Exception:
+        pass
+
+    # Find crosslink directory and load rules
+    crosslink_dir = find_crosslink_dir()
+    tracking_mode = load_tracking_mode(crosslink_dir)
+
+    # Agents always get condensed reminders — skip expensive tree/deps scanning
+    if is_agent_context(crosslink_dir):
+        languages = detect_languages()
+        print(build_condensed_reminder(languages, tracking_mode))
+        sys.exit(0)
+
+    # Check if we should send full or condensed guard
+    if not should_send_full_guard(crosslink_dir):
+        config = load_config_merged(crosslink_dir)
+        interval = int(config.get("reminder_drift_threshold", 3))
+
+        state = load_guard_state(crosslink_dir)
+        state["total_prompts"] = state.get("total_prompts", 0) + 1
+
+        # Check context budget — if exceeded, reinject full guard + compression directive
+        prompt_chars = estimate_prompt_chars(input_data)
+        if check_context_budget(crosslink_dir, state, prompt_chars):
+            languages = detect_languages()
+            language_rules, global_rules, project_rules, knowledge_rules, quality_rules = load_all_rules(crosslink_dir)
+            project_tree = get_project_tree()
+            dependencies = get_dependencies()
+            print(build_reminder(languages, project_tree, dependencies, language_rules, global_rules, project_rules, tracking_mode, crosslink_dir, knowledge_rules, quality_rules))
+            print(build_context_budget_warning(languages, tracking_mode))
+            state["estimated_context_chars"] = 0
+            state["context_budget_reinjections"] = state.get("context_budget_reinjections", 0) + 1
+            save_guard_state(crosslink_dir, state)
+            sys.exit(0)
+
+        # Normal condensed reminder at interval
+        if interval == 0 or state["total_prompts"] % interval == 0:
+            languages = detect_languages()
+            print(build_condensed_reminder(languages, tracking_mode))
+
+        save_guard_state(crosslink_dir, state)
+        sys.exit(0)
+
+    language_rules, global_rules, project_rules, knowledge_rules, quality_rules = load_all_rules(crosslink_dir)
+
+    # Detect languages in the project
+    languages = detect_languages()
+
+    # Generate project tree to prevent path hallucinations
+    project_tree = get_project_tree()
+
+    # Get installed dependencies to prevent version hallucinations
+    dependencies = get_dependencies()
+
+    # Output the full reminder
+    print(build_reminder(languages, project_tree, dependencies, language_rules, global_rules, project_rules, tracking_mode, crosslink_dir, knowledge_rules, quality_rules))
+
+    # Mark that we've sent the full guard this session
+    mark_full_guard_sent(crosslink_dir)
+
+    # Initialize context budget tracking for this session
+    state = load_guard_state(crosslink_dir)
+    state["estimated_context_chars"] = 0
+    save_guard_state(crosslink_dir, state)
+
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()