#!/usr/bin/env python3
"""
Stop hook: enforces CLAUDE.md's "Verify Before Claiming" rule.

Blocks the turn if the final assistant message contains hedging language
that signals unverified factual claims, without either:
  (a) an inline [unverified] tag, or
  (b) a verifying tool call (WebSearch, WebFetch, Read, Bash, Grep, Glob) in the turn.

Improvements over v1 (2026-04-18, informed by Critic + Research Scout agents):
  - Uses `assistant_message` from payload instead of parsing transcript
  - Removed false-positive hedges: "should be", "must be", "seems", "appears to"
  - Added real hedging signals: "I believe", "I think", "IIRC", "might be", etc.
  - Removed Edit/Write/NotebookEdit from VERIFYING_TOOLS (mutation != verification)
  - Narrowed MCP bypass to search/fetch/query tools only
  - Skips short responses (<100 chars) as trivial
  - Skips mid-turn stops (stop_reason == "tool_use")
"""

import json
import os
import re
import sys


# Hedging language that signals unverified factual claims.
# These are phrases people use when citing from memory, not from evidence.
HEDGES = [
    r"I believe",
    r"I think (?:it|the|this|that)",
    r"if I recall",
    r"IIRC",
    r"as far as I know",
    r"AFAIK",
    r"from memory",
    r"from what I remember",
    r"might be",
    r"could be",
    r"probably",
    r"likely",
    r"presumably",
    r"plausibly",
    r"my understanding is",
]

# Tools that constitute actual verification of a claim.
# Mutation tools (Edit, Write) do NOT verify anything.
VERIFYING_TOOLS = {
    "Bash", "Read", "Grep", "Glob", "WebSearch", "WebFetch",
}

# MCP tools that count as verification (must contain one of these substrings)
MCP_VERIFY_SUBSTRINGS = {"search", "fetch", "query", "get", "list", "read"}

# Minimum response length to check. Short responses are trivial.
MIN_RESPONSE_LENGTH = 100


def main() -> int:
    try:
        payload = json.load(sys.stdin)
    except json.JSONDecodeError:
        return 0

    # Bypass: env var override
    if os.environ.get("CLAUDE_SKIP_VERIFY_HOOK"):
        return 0

    # Bypass: already in forced-continuation (prevents infinite loops)
    if payload.get("stop_hook_active"):
        return 0

    # Bypass: Claude is mid-turn (making a tool call, not presenting final answer)
    if payload.get("stop_reason") == "tool_use":
        return 0

    # Get the assistant message — prefer payload field over transcript parsing
    text = payload.get("assistant_message") or payload.get("last_assistant_message") or ""

    # Fallback: parse transcript if payload doesn't have the message
    if not text:
        transcript_path = payload.get("transcript_path")
        if transcript_path and os.path.isfile(transcript_path):
            text = _extract_last_assistant_text(transcript_path)

    if not text or len(text.strip()) < MIN_RESPONSE_LENGTH:
        return 0

    # Strip code blocks and inline code before scanning
    text_clean = re.sub(r"```.*?```", "", text, flags=re.DOTALL)
    text_clean = re.sub(r"`[^`\n]+`", "", text_clean)
    # Strip blockquotes (lines starting with >)
    text_clean = re.sub(r"^>.*$", "", text_clean, flags=re.MULTILINE)

    # Bypass: explicit [unverified] tag present
    if re.search(r"\[unverified\]", text_clean, re.IGNORECASE):
        return 0

    # Check for hedging language
    pattern = re.compile(
        r"(?:" + "|".join(HEDGES) + r")",
        re.IGNORECASE,
    )
    matches = sorted({m.group(0).lower() for m in pattern.finditer(text_clean)})

    if not matches:
        return 0

    # Check if any verifying tool was called in this turn
    if _turn_has_verification(payload):
        return 0

    # Hedging found, no verification, no [unverified] tag — block
    reason = (
        "Hedging language detected without [unverified] tags or tool "
        f"verification in this turn: {', '.join(matches)}. "
        "Per your CLAUDE.md 'Verify Before Claiming' rule: either "
        "(a) verify the claim via WebSearch/Bash/Read, or (b) tag it "
        "inline, e.g. 'X supports Y [unverified]'. Revise before "
        "ending the turn."
    )

    out = {"decision": "block", "reason": reason}
    print(json.dumps(out))
    return 0


def _turn_has_verification(payload: dict) -> bool:
    """Check if any verifying tool was called in the current turn."""
    transcript_path = payload.get("transcript_path")
    if not transcript_path or not os.path.isfile(transcript_path):
        return False

    try:
        with open(transcript_path, "r", encoding="utf-8") as f:
            entries = [json.loads(line) for line in f if line.strip()]
    except (json.JSONDecodeError, OSError):
        return False

    # Find the last user message index
    last_user_idx = -1
    for i, entry in enumerate(entries):
        if entry.get("type") == "user":
            last_user_idx = i

    # Check assistant messages after the last user message for tool calls
    for entry in entries[last_user_idx + 1:]:
        if entry.get("type") != "assistant":
            continue
        content = entry.get("message", {}).get("content", [])
        if not isinstance(content, list):
            continue
        for block in content:
            if not isinstance(block, dict) or block.get("type") != "tool_use":
                continue
            name = block.get("name", "")
            # Direct tool match
            if name in VERIFYING_TOOLS:
                return True
            # MCP tool match — only if the tool name suggests verification
            if name.startswith("mcp__"):
                name_lower = name.lower()
                if any(sub in name_lower for sub in MCP_VERIFY_SUBSTRINGS):
                    return True

    return False


def _extract_last_assistant_text(transcript_path: str) -> str:
    """Fallback: extract last assistant message text from transcript JSONL."""
    try:
        with open(transcript_path, "r", encoding="utf-8") as f:
            entries = [json.loads(line) for line in f if line.strip()]
    except (json.JSONDecodeError, OSError):
        return ""

    for entry in reversed(entries):
        if entry.get("type") != "assistant":
            continue
        content = entry.get("message", {}).get("content", [])
        if isinstance(content, str):
            return content
        if isinstance(content, list):
            return "\n".join(
                b.get("text", "")
                for b in content
                if isinstance(b, dict) and b.get("type") == "text"
            )
    return ""


if __name__ == "__main__":
    sys.exit(main())