#!/usr/bin/env python3
"""PreToolUse hook (Bash): enforce the no-commit / no-push contract structurally.

settings.json's deny rules (`Bash(git push:*)`, ...) match a literal *prefix* of the
command text, so `git -C . push`, `git -c k=v commit`, `/usr/bin/git push`,
`env X=1 git push`, `timeout 5 git push` and `sh -c 'git push'` all slip past them.
This hook parses each shell segment instead: finds the git/gh token wherever it sits,
skips git's global options (`-C dir`, `-c k=v`, `--git-dir=...`) and blocks

    git commit | push | revert | cherry-pick | merge (not --abort/--quit/--continue)
    git rebase -i / --interactive / --autostash
    git clean | reset (any form that moves HEAD or unstages everything; `reset -- <paths>`,
    `reset HEAD <path>` and `reset <existing path>` are fine) | restore <paths> (--staged
    alone is fine, except `--staged .` / `:/`) | checkout -- / checkout . / -f
    git stash drop | pop | clear      git branch -D / -f / -M
    gh pr merge | close | create without --head | edit beyond --title/--body

i.e. anything that commits, rewrites history, or discards working-tree changes -- the last
group matters most with several sessions sharing one checkout. Mirrors pi's permission-gate.
On GitHub Claude may open a PR (settings.json asks first) and retitle / redescribe one, nothing
more: `gh pr create` must name its branch with --head, without which gh pushes it itself (even
under --dry-run), and `gh pr edit` must not touch reviewers, labels, base or the like.
A plain `git rebase` (and its --continue / --abort) passes: Claude rebases and fixes the
conflicts when the user asks (global CLAUDE.md), and settings.json asks before each one, as
pi's gate does. `-i` stays blocked (the squash/reword tool, and it needs an editor), and so
does --autostash: it stashes and re-pops other sessions' uncommitted edits, where a plain
rebase refuses to start on a dirty tree.

Exit 2 + one stderr line = the call is blocked and Claude reads the reason (stage and
stop; /commit-msg drafts the message). Mirrors pi's permission-gate.ts.

Tokenizing is quote-aware first (shlex in punctuation mode), and only then split at the
separators `;`, `|`, `&&`, `(`, `)`: a quoted word that merely *contains* a separator and a
git verb (a grep pattern, a commit message) stays one word and is not a command. `$(...)`
and backtick bodies are scanned on their own and then collapsed, so `git -C $(pwd) push`
still parses as a push. `sh -c '...'` and `eval` bodies recurse.

Deliberate gaps: heredoc bodies are skipped -- they are file/script *contents*, and a
README that documents pushing must stay writable -- so a script piped in through a heredoc
is not inspected. Aliases (`git ci`) are not expanded. The deny list, the auto-mode
classifier and the global CLAUDE.md still cover those; this is the belt for the prefix
rules' braces. Fails open: any parse error lets the command through (exit 0).

Exempt repositories (one main checkout per line in ~/.claude/git-guard-exempt, host-local and
outside the dots repo on purpose: it names projects): there `git commit` and `git merge` pass
-- the user's decision for repos where Claude commits and merges itself. The repo is the one the call
actually targets: the hook input's `cwd`, moved by every `cd` before it in the command and by
`git -C`, resolved to its main checkout (so its worktrees count). Anything that hides the
target (`cd -`, `cd $VAR`, a substitution, `--git-dir`, `GIT_DIR=`) is judged not exempt.
`commit --amend` and `commit -a`/`--all` stay blocked there (they rewrite the last commit of
a shared checkout, or sweep other sessions' edits in), as do push, `rebase -i` and the discard
group. Known gap: a `cd` inside `( ... )` leaks into the commands after the parenthesis.
"""
import json
import os
import re
import shlex
import subprocess
import sys

GIT_BLOCK = {"commit", "push", "clean", "revert", "cherry-pick"}
MERGE_OK = {"--abort", "--quit", "--continue"}
REBASE_BLOCK = {"-i", "--interactive", "--autostash"}
STASH_BLOCK = {"drop", "pop", "clear"}
BRANCH_FORCE = {"-D", "-f", "--force", "-M"}
RESET_MODES = {"--hard", "--soft", "--merge", "--keep"}
UNSTAGE_ALL = {".", ":/", "*", ":"}
_HEX = re.compile(r"^[0-9a-f]{7,40}$")


def looks_like_ref(a):
    return (a in ("HEAD", "ORIG_HEAD", "FETCH_HEAD", "MERGE_HEAD") or a.startswith("refs/")
            or "~" in a or "^" in a or "@{" in a or bool(_HEX.match(a)))
GH_PR_BLOCK = {"merge", "close"}
# `gh pr edit` may only retitle and redescribe; these take their value as the next word (or as
# `--title=x` / `-tx`).
GH_EDIT_OK = {"-t", "--title", "-b", "--body", "-F", "--body-file", "-R", "--repo"}
# Repositories (main checkouts; their worktrees count) where Claude commits and merges itself:
# one path per line, `#` comments; a missing file exempts nothing. GIT_GUARD_EXEMPT
# (colon-separated) replaces the file for the test matrix; it is read from the hook's own
# environment (Claude Code's), which a prefix on the checked command cannot reach.
EXEMPT_FILE = "~/.claude/git-guard-exempt"
EXEMPT_VERBS = {"commit", "merge"}
# `git commit` options whose value is the next word (or, for the short ones, the rest of the word).
COMMIT_ARG_SHORT = set("mFCct")
COMMIT_ARG_LONG = {"--message", "--file", "--author", "--date", "--reuse-message", "--reedit-message",
                   "--fixup", "--squash", "--template", "--trailer", "--cleanup", "--pathspec-from-file"}
# global options that take a *separate* argument; the `--opt=value` spelling is one token.
GIT_OPT_ARG = {"-C", "-c", "--git-dir", "--work-tree", "--exec-path", "--namespace",
               "--super-prefix", "--config-env"}
GH_OPT_ARG = {"-R", "--repo"}
SHELLS = {"sh", "bash", "dash", "zsh", "ksh", "fish"}

# `<<'EOF' ... EOF` / `<<-EOF ... EOF`: drop the body, keep the command around it.
_HEREDOC = re.compile(r"<<-?\s*(['\"]?)(\w+)\1[^\n]*\n.*?\n[ \t]*\2[ \t]*(?=\n|$)", re.S)
# innermost `$(...)` / `...` command substitutions (peeled repeatedly for nesting).
_SUBST = re.compile(r"\$\(([^()]*)\)|`([^`]*)`")
# tokens that end one simple command and start the next.
_SEP = {";", ";;", "|", "||", "|&", "&", "&&", "(", ")"}


def _skip_opts(args, with_arg):
    j = 0
    while j < len(args) and args[j].startswith("-"):
        j += 2 if args[j] in with_arg else 1
    return j


def exempt_repos():
    env = os.environ.get("GIT_GUARD_EXEMPT")
    if env is not None:
        paths = env.split(":")
    else:
        try:
            with open(os.path.expanduser(EXEMPT_FILE)) as f:
                paths = [ln.split("#", 1)[0].strip() for ln in f]
        except OSError:
            return set()
    return {os.path.realpath(os.path.expanduser(p)) for p in paths if p}


def main_checkout(d):
    """The main checkout of the repository `d` lies in (a worktree resolves to it), or None."""
    try:
        out = subprocess.run(["git", "-C", d, "rev-parse", "--path-format=absolute", "--git-common-dir"],
                             capture_output=True, text=True, timeout=3)
    except (OSError, subprocess.SubprocessError):
        return None
    common = out.stdout.strip().rstrip("/")
    if out.returncode != 0 or not common:
        return None
    return os.path.realpath(os.path.dirname(common))


def in_exempt_repo(opts, cwd, pre):
    """Whether a git call with global options `opts`, run from `cwd` after the words `pre` of
    its own command (`env X=1`, ...), targets an exempt repository."""
    exempt = exempt_repos()
    if not exempt or cwd is None or any(t.startswith(("GIT_DIR=", "GIT_WORK_TREE=", "GIT_COMMON_DIR="))
                                        for t in pre):
        return False
    d, j = cwd, 0
    while j < len(opts):
        a = opts[j]
        if a in ("--git-dir", "--work-tree") or a.startswith(("--git-dir=", "--work-tree=")):
            return False
        if a == "-C":
            if j + 1 >= len(opts) or "SUBST" in opts[j + 1] or "$" in opts[j + 1]:
                return False
            d = os.path.join(d, os.path.expanduser(opts[j + 1]))
        j += 2 if a in GIT_OPT_ARG else 1
    return main_checkout(d) in exempt


def commit_rewrites_or_sweeps(rest):
    """`commit --amend` (rewrites the last commit) or `-a`/`--all` (takes in every modified
    tracked file, other sessions' edits included)."""
    skip = False
    for a in rest:
        if skip:
            skip = False
            continue
        if a == "--":
            break
        if a in ("--amend", "--all"):
            return True
        if a in COMMIT_ARG_LONG:
            skip = True
        elif a.startswith("-") and not a.startswith("--"):
            for k, ch in enumerate(a[1:], 1):
                if ch == "a":
                    return True
                if ch in COMMIT_ARG_SHORT:
                    skip = k == len(a) - 1  # the value is the next word
                    break
    return False


def change_dir(cwd, args):
    """Where `cd args` leaves the shell started in `cwd`; None when that cannot be told."""
    args = [a for a in args if a not in ("-L", "-P", "--")]
    if not args:
        return os.path.expanduser("~")
    t = args[0]
    if t == "-" or "SUBST" in t or "$" in t or (cwd is None and not t.startswith(("/", "~"))):
        return None
    return os.path.normpath(os.path.join(cwd or "/", os.path.expanduser(t)))


def git_reason(args, cwd=None, pre=()):
    j = _skip_opts(args, GIT_OPT_ARG)
    if j >= len(args):
        return None
    sub, rest = args[j], args[j + 1:]
    if (sub in EXEMPT_VERBS and not (sub == "commit" and commit_rewrites_or_sweeps(rest))
            and in_exempt_repo(args[:j], cwd, pre)):
        return None
    if sub in GIT_BLOCK:
        return f"git {sub}"
    if sub == "rebase":
        flag = next((a for a in rest if a in REBASE_BLOCK), None)
        if flag == "--autostash" or any(o.lower().startswith("rebase.autostash") for o in args[:j]):
            return "git rebase --autostash (stashes other sessions' uncommitted edits)"
        return f"git rebase {flag} (interactive history rewrite)" if flag else None
    if sub == "reset":
        opts = [a for a in rest if a.startswith("-") and a != "--"]
        if RESET_MODES & set(opts):
            return "git reset --hard/--soft/--merge/--keep (moves HEAD)"
        if "-p" in opts or "--patch" in opts or "--" in rest:
            return None  # interactive, or the explicit path form
        pos = [a for a in rest if not a.startswith("-")]
        if not pos or pos == ["HEAD"]:
            return "git reset (unstages everything, other sessions' rounds included)"
        if len(pos) == 1 and (looks_like_ref(pos[0]) or not os.path.exists(pos[0])):
            return "git reset <commit> (moves HEAD)"
        return None  # `reset HEAD <paths>` / `reset <paths>`
    if sub == "merge" and not (MERGE_OK & set(rest)):
        return "git merge (creates a commit)"
    if sub == "restore":
        staged = any(a in ("--staged", "-S") for a in rest)
        worktree = any(a in ("--worktree", "-W") for a in rest)
        if not staged or worktree:
            return "git restore (discards working-tree changes)"
        if UNSTAGE_ALL & {a for a in rest if not a.startswith("-")}:
            return "git restore --staged . (unstages everything, other sessions' rounds included)"
    if sub == "checkout" and ("--" in rest or "." in rest or "-f" in rest or "--force" in rest):
        return "git checkout -- <paths> (discards working-tree changes)"
    if sub == "stash" and rest and rest[0] in STASH_BLOCK:
        return f"git stash {rest[0]}"
    if sub == "branch" and BRANCH_FORCE & set(rest):
        return "git branch -D/-f (force-deletes or force-moves a branch)"
    return None


def gh_edit_extra(rest):
    """The first `gh pr edit` option that is neither the title nor the body, else None."""
    k = 0
    while k < len(rest):
        a = rest[k]
        if a == "--":
            break
        name = a.split("=", 1)[0]
        if name in GH_EDIT_OK:
            k += 2 if name == a else 1  # skip the value, which may well start with `-`
            continue
        if a.startswith("-") and not (a[:2] in GH_EDIT_OK and not a.startswith("--")):
            return a
        k += 1
    return None


def gh_reason(args):
    j = _skip_opts(args, GH_OPT_ARG)
    if j >= len(args) or args[j] != "pr":
        return None
    rest = args[j + 1:]
    k = _skip_opts(rest, GH_OPT_ARG)
    if k >= len(rest) or "--help" in rest:
        return None
    sub, rest = rest[k], rest[k + 1:]
    if sub in GH_PR_BLOCK:
        return f"gh pr {sub}"
    if sub == "create" and not any(a in ("-H", "--head") or a.startswith("--head=")
                                   or (a.startswith("-H") and not a.startswith("--")) for a in rest):
        return "gh pr create without --head: gh would push the branch itself"
    if sub == "edit":
        extra = gh_edit_extra(rest)
        if extra:
            return f"gh pr edit {extra.split('=', 1)[0]}: only --title / --body edits are allowed"
    return None


def tokenize(text):
    """Quote-aware shell words; the separator characters come out as their own tokens
    (shlex punctuation mode). Comment handling is off: a `#` mid-word must not swallow
    the rest of the line."""
    lex = shlex.shlex(text, posix=True, punctuation_chars=True)
    lex.whitespace_split = True
    lex.commenters = ""
    try:
        return list(lex)
    except ValueError:  # unbalanced quotes: best effort
        return text.split()


def segments(toks):
    seg = []
    for t in toks:
        if t in _SEP:
            if seg:
                yield seg
            seg = []
        else:
            seg.append(t)
    if seg:
        yield seg


def scan(text, depth=0, cwd=None):
    """Reason string for the first blocked invocation in `text` (run from `cwd`), else None."""
    if depth > 4:
        return None
    text = _HEREDOC.sub(" ", text)
    # command substitutions: scan each body on its own, then collapse it to a plain word so
    # the enclosing command still parses (`git -C $(pwd) push` -> `git -C SUBST push`).
    bodies = []

    def _grab(m):
        bodies.append(m.group(1) if m.group(1) is not None else m.group(2))
        return " SUBST "

    prev = None
    while prev != text:
        prev, text = text, _SUBST.sub(_grab, text)
    for body in bodies:
        r = scan(body, depth + 1, cwd)
        if r:
            return r
    for toks in segments(tokenize(text)):
        if toks[0] == "cd":
            cwd = change_dir(cwd, toks[1:])
            continue
        if toks[0] in ("pushd", "popd"):
            cwd = None
            continue
        for i, tok in enumerate(toks):
            base = os.path.basename(tok)
            r = None
            if base == "git":
                r = git_reason(toks[i + 1:], cwd, toks[:i])
            elif base == "gh":
                r = gh_reason(toks[i + 1:])
            elif base == "eval":
                r = scan(" ".join(toks[i + 1:]), depth + 1, cwd)
            elif base in SHELLS and "-c" in toks[i + 1:]:
                k = toks.index("-c", i + 1)  # `sh -c '<cmd>'`: the real command is one word
                if k + 1 < len(toks):
                    r = scan(toks[k + 1], depth + 1, cwd)
            if r:
                return r
    return None


def main():
    try:
        data = json.load(sys.stdin)
        if data.get("tool_name", "Bash") != "Bash":
            return
        cmd = (data.get("tool_input") or {}).get("command") or ""
        reason = scan(cmd, 0, data.get("cwd") or os.getcwd()) if cmd else None
    except Exception:
        return  # fail open: the deny list still applies
    if reason:
        short = cmd.strip().splitlines()[0]
        short = short if len(short) <= 80 else short[:77] + "..."
        hint = ("On a PR Claude reads, opens one when asked (with --head <pushed branch>) and edits its "
                "--title / --body; pushing, merging, closing and every other edit are the user's."
                if reason.startswith("gh ") else
                "Rebase with a plain `git rebase` on a clean tree, and only when the user asked; if the "
                "tree holds other sessions' edits, tell the user instead of stashing them."
                if reason.startswith("git rebase") else
                "The user commits, pushes and merges manually (global CLAUDE.md): stage and stop, "
                "and draft the message with /commit-msg.")
        sys.stderr.write(f"git-guard: blocked `{short}` ({reason}). {hint}\n")
        sys.exit(2)


if __name__ == "__main__":
    main()