codex-tool-use-trajectories

"""Redaction for the Codex rollout cleaning pipeline.
Design
------
* All finders run on the ORIGINAL text and return spans (start, end, kind, replacement).
  Overlaps are resolved left-to-right (earliest start wins, then longest), so every
  replacement has exact offsets in the redacted text. That is what lets the pipeline
  log only replacements that survive tool-output truncation.
* Finders are linear-time: nothing starts with an unbounded character class followed by
  a distant anchor (e.g. e-mail search is anchored on '@', not on the local part).
* A second "verification" loop re-scans the redacted text with the *secret* finders and
  re-applies them if anything is left (belt and braces).
* Nothing here ever writes an original value anywhere; callers only receive sha256 digests.
"""
import hashlib
import re

SECRET = "<SECRET>"
EMAIL = "<EMAIL>"
IP = "<IP>"

# ----------------------------------------------------------------------------- secrets
_PEM_FULL = re.compile(
    r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY(?: BLOCK)?-----[\s\S]{0,20000}?-----END [A-Z0-9 ]*PRIVATE KEY(?: BLOCK)?-----"
)
# Unterminated block (output was cut or the END line was lost): redact BEGIN line + base64 body.
_PEM_OPEN = re.compile(
    r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY(?: BLOCK)?-----(?:[A-Za-z0-9+/=\s]|\\n|\\r){0,8000}"
)
# Token patterns are applied exactly as written in the task spec (no left word-boundary), so a plain grep for the
# same regex over the output finds nothing. Matches that are glued to a preceding word character (for example the
# "sk-" inside a hyphenated file name such as "...task-board-...") are logged with kind suffix ":embedded_in_word"
# so the collateral can be audited from redactions.jsonl.
_AKIA = re.compile(r"AKIA[0-9A-Z]{16}")
_GH = re.compile(r"gh[pousr]_[A-Za-z0-9]{20,}")        # ghp_ (+ gho_/ghu_/ghs_/ghr_)
_GH_PAT = re.compile(r"github_pat_[A-Za-z0-9_]{20,}")
_SK = re.compile(r"sk-[A-Za-z0-9_-]{20,}")
_BEARER = re.compile(
    r"(?i)\bAuthorization\s*[:=]\s*\\*[\"']?\s*Bearer\s+(?P<val>[A-Za-z0-9._~+/=-]{8,})"
)

# key = value secrets.  password|passwd|token|api_key|apikey|secret, optionally prefixed
# (access_token, client_secret, x-api-key, DB_PASSWORD ...) but NOT suffixed (token_count, secrets...)
_KV = re.compile(
    r"(?i)(?<![A-Za-z0-9])(?P<key>(?:[A-Za-z0-9]+[_.-])*(?:password|passwd|token|api[_-]?key|apikey|secret))"
    r"(?![A-Za-z0-9])\\*[\"']?[ \t]*(?::|=(?!=))[ \t]*\\*(?:"
    r"(?P<q>[\"'])(?P<qval>[^\"'\r\n\\]{3,200})\\*(?P=q)"
    r"|(?P<bval>[^\s\"'\\,;&(){}\[\]<>|`]{6,200}))"
)
_IDENT = re.compile(r"^[A-Za-z_][A-Za-z_]*$")  # code identifier with no digits: not a literal secret
_HEX = re.compile(r"^0[xX][0-9A-Fa-f]+$")        # decompiler metadata tokens: "// Token: 0x06000001"
_DOTTED = re.compile(r"^[A-Za-z_]\w*(?:\.[A-Za-z_]\w*)+$")  # member access: this.token_value (JWTs are far longer)


def _kv_value_ok(val, quoted):
    if not val or val[0] in "$%<{(@#*":
        return False
    if val.lower() in ("null", "none", "true", "false", "undefined", "nullptr", "redacted"):
        return False
    if quoted:
        return True
    if _IDENT.match(val) or _HEX.match(val):
        return False
    if len(val) <= 64 and _DOTTED.match(val):
        return False
    return True


# ----------------------------------------------------------------------------- e-mail
_EMAIL_LOCAL = re.compile(r"[A-Za-z0-9._%+-]+$")
_EMAIL_DOMAIN = re.compile(r"[A-Za-z0-9-]+(?:\.[A-Za-z0-9-]+)*\.([A-Za-z]{2,})")
# TLD-looking file extensions: icon@2x.png, foo@bar.cpp ... are not mail addresses.
_FILE_EXT = frozenset(
    "png jpg jpeg gif svg webp bmp ico tga dds tif tiff psd bin dll exe lib obj asm so dylib pdb log xml yml yaml "
    "html htm css js ts json txt md c h cc cpp hpp cs py rs go java rb php sh bat ps1 cmd csv tsv bk2 bik bik2 "
    "mp4 mkv avi mov wav mp3 ogg zip rar 7z gz tar".split()
)


def _find_emails(text):
    out = []
    i = text.find("@")
    n = len(text)
    while i != -1:
        # local part: walk left (bounded)
        j = i
        lo = max(0, i - 64)
        while j > lo and text[j - 1] in "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789._%+-":
            j -= 1
        local = text[j:i]
        if local and any(ch.isalnum() for ch in local):
            m = _EMAIL_DOMAIN.match(text, i + 1)
            if m and m.group(1).lower() not in _FILE_EXT:
                out.append((j, m.end(), "email", EMAIL))
        i = text.find("@", i + 1)
    return out


# ----------------------------------------------------------------------------- windows user paths
_SYSTEM_PROFILES = frozenset(("public", "default", "default user", "all users", "user", "defaultuser0"))
_WIN_USER = re.compile(
    r"(?P<drive>[A-Za-z]:)(?P<s1>\\{1,8}|/{1,2})(?i:Users)(?P<s2>\\{1,8}|/{1,2})(?P<name>[^\\/\s\"'<>|:*?,;(){}\[\]`]+)"
)
_UNIX_USER = re.compile(
    r"(?<![A-Za-z0-9_.-])(?P<pre>/mnt/[a-z]|/[a-z])/(?i:Users)/(?P<name>[^\\/\s\"'<>|:*?,;(){}\[\]`]+)"
)


def _find_user_paths(text):
    out = []
    if "sers" not in text and "SERS" not in text:
        return out
    for m in _WIN_USER.finditer(text):
        name = m.group("name")
        if name.lower() in _SYSTEM_PROFILES:
            continue
        rep = f"{m.group('drive')}{m.group('s1')}Users{m.group('s2')}user"
        out.append((m.start(), m.end(), "win_user_path", rep))
    for m in _UNIX_USER.finditer(text):
        name = m.group("name")
        if name.lower() in _SYSTEM_PROFILES:
            continue
        out.append((m.start(), m.end(), "win_user_path_posix", f"{m.group('pre')}/Users/user"))
    return out


# ----------------------------------------------------------------------------- OneDrive organisation names
# Work/school OneDrive folders are named "OneDrive - <organisation>" and are followed by a path separator,
# e.g. "D:\OneDrive - Some School\Documents". The organisation identifies the account owner, so it is masked.
# The generic consumer folder "OneDrive - Personal" is not an organisation and is kept.
ORG = "<ORG>"
_ONEDRIVE_ORG = re.compile(r"(?<=OneDrive - )(?P<org>[^\\/\r\n\"<>|:*?]{1,80}?)(?=\\|/)")
_ONEDRIVE_KEEP = frozenset(("personal",))


def _find_onedrive_org(text):
    out = []
    if "OneDrive - " not in text:
        return out
    for m in _ONEDRIVE_ORG.finditer(text):
        if m.group("org").strip().lower() in _ONEDRIVE_KEEP:
            continue
        out.append((m.start("org"), m.end("org"), "onedrive_org", ORG))
    return out


# ----------------------------------------------------------------------------- IPv4
_OCT = r"(?:25[0-5]|2[0-4]\d|1\d\d|[1-9]?\d)"
_IPV4 = re.compile(rf"(?<![\w.]){_OCT}(?:\.{_OCT}){{3}}(?!\d|\.\d)")


_VERSION_CTX = re.compile(r"(?i)version[\"']?\s*[=:]?\s*[\"']?$")


def _find_ips(text):
    out = []
    for m in _IPV4.finditer(text):
        v = m.group(0)
        if v == "0.0.0.0" or v.startswith("127."):
            continue
        if _VERSION_CTX.search(text[max(0, m.start() - 16):m.start()]):
            continue  # .NET "Version=4.0.0.0" style assembly versions are not addresses
        out.append((m.start(), m.end(), "ipv4", IP))
    return out


# ----------------------------------------------------------------------------- finders
def _emb(text, start):
    """':embedded_in_word' when the match is glued to a preceding word character."""
    return ":embedded_in_word" if start > 0 and (text[start - 1].isalnum() or text[start - 1] in "_-") else ""


def _find_secrets(text):
    out = []
    if "-----BEGIN" in text:
        for m in _PEM_FULL.finditer(text):
            out.append((m.start(), m.end(), "secret:pem_private_key", SECRET))
        for m in _PEM_OPEN.finditer(text):
            out.append((m.start(), m.end(), "secret:pem_private_key", SECRET))
    if "AKIA" in text:
        for m in _AKIA.finditer(text):
            out.append((m.start(), m.end(), "secret:aws_access_key" + _emb(text, m.start()), SECRET))
    if "gh" in text:
        for m in _GH.finditer(text):
            out.append((m.start(), m.end(), "secret:github_token" + _emb(text, m.start()), SECRET))
    if "github_pat_" in text:
        for m in _GH_PAT.finditer(text):
            out.append((m.start(), m.end(), "secret:github_pat" + _emb(text, m.start()), SECRET))
    if "sk-" in text:
        for m in _SK.finditer(text):
            out.append((m.start(), m.end(), "secret:sk_key" + _emb(text, m.start()), SECRET))
    if "earer" in text:
        for m in _BEARER.finditer(text):
            v = m.group("val")
            if v[0] in "$%<{":
                continue
            out.append((m.start("val"), m.end("val"), "secret:bearer_token", SECRET))
    low = text.lower()
    if ("password" in low) or ("passwd" in low) or ("token" in low) or ("secret" in low) or ("apikey" in low) or ("api_key" in low) or ("api-key" in low):
        for m in _KV.finditer(text):
            if m.group("q"):
                s, e, val, quoted = m.start("qval"), m.end("qval"), m.group("qval"), True
            else:
                s, e, val, quoted = m.start("bval"), m.end("bval"), m.group("bval"), False
            if not _kv_value_ok(val, quoted):
                continue
            kname = re.sub(r"[^a-z]", "", m.group("key").lower().split("_")[-1].split("-")[-1].split(".")[-1])
            out.append((s, e, "secret:kv:" + (kname or "key"), SECRET))
    return out


def _resolve(spans):
    """Earliest start wins; at equal start the longer span wins. Overlapping losers are dropped."""
    spans.sort(key=lambda s: (s[0], -(s[1] - s[0])))
    kept = []
    last_end = -1
    for s in spans:
        if s[0] >= last_end:
            kept.append(s)
            last_end = s[1]
    return kept


def _sha(s):
    return hashlib.sha256(s.encode("utf-8", "surrogatepass")).hexdigest()


def redact_text(text):
    """Redact `text`. Returns (new_text, events) where events is a list of
    (start, end, kind, sha256_of_original, replacement) with start/end in *new_text* coordinates."""
    if not text:
        return text, []
    spans = _find_secrets(text) + _find_emails(text) + _find_user_paths(text) + _find_onedrive_org(text) + _find_ips(text)
    if not spans:
        return text, []
    kept = _resolve(spans)
    pieces = []
    events = []
    pos = 0
    out_len = 0
    for s, e, kind, rep in kept:
        seg = text[pos:s]
        pieces.append(seg)
        out_len += len(seg)
        events.append((out_len, out_len + len(rep), kind, _sha(text[s:e]), rep))
        pieces.append(rep)
        out_len += len(rep)
        pos = e
    pieces.append(text[pos:])
    new = "".join(pieces)
    # Verification loop: a secret-class match must never survive.
    for _ in range(3):
        left = _find_secrets(new)
        if not left:
            break
        left = _resolve(left)
        pieces = []
        pos = 0
        shift_events = []
        for s, e, kind, rep in left:
            pieces.append(new[pos:s])
            pieces.append(rep)
            shift_events.append((s, e, kind, _sha(new[s:e]), rep))
            pos = e
        pieces.append(new[pos:])
        new = "".join(pieces)
        # offsets of earlier events can shift; recompute coarsely by dropping offsets (rare path)
        events = [(-1, -1, k, h, r) for (_, _, k, h, r) in events] + [(-1, -1, k, h, r) for (_, _, k, h, r) in shift_events]
    return new, events


def redact_json_leaves(obj, sink):
    """Redact every string leaf of a parsed JSON value. `sink` receives (kind, sha, replacement) tuples.
    Returns (new_obj, changed)."""
    if isinstance(obj, str):
        new, ev = redact_text(obj)
        for (_, _, k, h, r) in ev:
            sink.append((k, h, r))
        return new, bool(ev)
    if isinstance(obj, list):
        changed = False
        res = []
        for x in obj:
            y, c = redact_json_leaves(x, sink)
            res.append(y)
            changed = changed or c
        return res, changed
    if isinstance(obj, dict):
        changed = False
        res = {}
        for k, v in obj.items():
            y, c = redact_json_leaves(v, sink)
            res[k] = y
            changed = changed or c
        return res, changed
    return obj, False


# ----------------------------------------------------------------------------- validation regexes
# Plain (un-boundaried) secret regexes used by the *validation* step, deliberately broader than the
# redaction finders so that a miss caused by a boundary rule is still surfaced.
VALIDATION_REGEXES = {
    "AKIA": re.compile(r"AKIA[0-9A-Z]{16}"),
    "ghp_": re.compile(r"ghp_[A-Za-z0-9]{20,}"),
    "github_pat_": re.compile(r"github_pat_[A-Za-z0-9_]{20,}"),
    "sk-": re.compile(r"sk-[A-Za-z0-9_-]{20,}"),
    "PEM PRIVATE KEY": re.compile(r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY"),
    "Authorization: Bearer": re.compile(r"(?i)Authorization\s*[:=]\s*\\*[\"']?\s*Bearer\s+(?!<SECRET>)[A-Za-z0-9._~+/=-]{8,}"),
}

# ----------------------------------------------------------------------------- phone-like candidates
# Reported, never redacted.
PHONE_RE = re.compile(
    r"(?<![\w.:/\\-])(?:\+\d{1,3}[ .-]?)?(?:\(\d{2,4}\)[ .-]?|\d{2,4}[ .-])\d{3,4}[ .-]\d{3,4}(?:[ .-]\d{1,4})?(?![\w.:/\\-])"
)


def phone_candidates(text):
    res = []
    for m in PHONE_RE.finditer(text):
        v = m.group(0)
        digits = re.sub(r"\D", "", v)
        if 9 <= len(digits) <= 15 and re.search(r"[ ().+-]", v):
            res.append(v)
    return res

	"""Redaction for the Codex rollout cleaning pipeline.

	Design
	------
	* All finders run on the ORIGINAL text and return spans (start, end, kind, replacement).
	Overlaps are resolved left-to-right (earliest start wins, then longest), so every
	replacement has exact offsets in the redacted text. That is what lets the pipeline
	log only replacements that survive tool-output truncation.
	* Finders are linear-time: nothing starts with an unbounded character class followed by
	a distant anchor (e.g. e-mail search is anchored on '@', not on the local part).
	* A second "verification" loop re-scans the redacted text with the secret finders and
	re-applies them if anything is left (belt and braces).
	* Nothing here ever writes an original value anywhere; callers only receive sha256 digests.
	"""
	import hashlib
	import re

	SECRET = "<SECRET>"
	EMAIL = "<EMAIL>"
	IP = "<IP>"

	# ----------------------------------------------------------------------------- secrets
	_PEM_FULL = re.compile(
	r"-----BEGIN [A-Z0-9 ]PRIVATE KEY(?: BLOCK)?-----[\s\S]{0,20000}?-----END [A-Z0-9 ]PRIVATE KEY(?: BLOCK)?-----"
	)
	# Unterminated block (output was cut or the END line was lost): redact BEGIN line + base64 body.
	_PEM_OPEN = re.compile(
	r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY(?: BLOCK)?-----(?:[A-Za-z0-9+/=\s]\|\\n\|\\r){0,8000}"
	)
	# Token patterns are applied exactly as written in the task spec (no left word-boundary), so a plain grep for the
	# same regex over the output finds nothing. Matches that are glued to a preceding word character (for example the
	# "sk-" inside a hyphenated file name such as "...task-board-...") are logged with kind suffix ":embedded_in_word"
	# so the collateral can be audited from redactions.jsonl.
	_AKIA = re.compile(r"AKIA[0-9A-Z]{16}")
	_GH = re.compile(r"gh[pousr]_[A-Za-z0-9]{20,}") # ghp_ (+ gho_/ghu_/ghs_/ghr_)
	_GH_PAT = re.compile(r"github_pat_[A-Za-z0-9_]{20,}")
	_SK = re.compile(r"sk-[A-Za-z0-9_-]{20,}")
	_BEARER = re.compile(
	r"(?i)\bAuthorization\s[:=]\s\\[\"']?\sBearer\s+(?P<val>[A-Za-z0-9._~+/=-]{8,})"
	)

	# key = value secrets. password\|passwd\|token\|api_key\|apikey\|secret, optionally prefixed
	# (access_token, client_secret, x-api-key, DB_PASSWORD ...) but NOT suffixed (token_count, secrets...)
	_KV = re.compile(
	r"(?i)(?<![A-Za-z0-9])(?P<key>(?:[A-Za-z0-9]+[_.-])*(?:password\|passwd\|token\|api[_-]?key\|apikey\|secret))"
	r"(?![A-Za-z0-9])\\[\"']?[ \t](?::\|=(?!=))[ \t]\\(?:"
	r"(?P<q>[\"'])(?P<qval>[^\"'\r\n\\]{3,200})\\*(?P=q)"
	r"\|(?P<bval>[^\s\"'\\,;&(){}\[\]<>\|`]{6,200}))"
	)
	_IDENT = re.compile(r"^[A-Za-z_][A-Za-z_]*$") # code identifier with no digits: not a literal secret
	_HEX = re.compile(r"^0[xX][0-9A-Fa-f]+$") # decompiler metadata tokens: "// Token: 0x06000001"
	_DOTTED = re.compile(r"^[A-Za-z_]\w(?:\.[A-Za-z_]\w)+$") # member access: this.token_value (JWTs are far longer)


	def _kv_value_ok(val, quoted):
	if not val or val[0] in "$%<{(@#*":
	return False
	if val.lower() in ("null", "none", "true", "false", "undefined", "nullptr", "redacted"):
	return False
	if quoted:
	return True
	if _IDENT.match(val) or _HEX.match(val):
	return False
	if len(val) <= 64 and _DOTTED.match(val):
	return False
	return True


	# ----------------------------------------------------------------------------- e-mail
	_EMAIL_LOCAL = re.compile(r"[A-Za-z0-9._%+-]+$")
	_EMAIL_DOMAIN = re.compile(r"[A-Za-z0-9-]+(?:\.[A-Za-z0-9-]+)*\.([A-Za-z]{2,})")
	# TLD-looking file extensions: icon@2x.png, foo@bar.cpp ... are not mail addresses.
	_FILE_EXT = frozenset(
	"png jpg jpeg gif svg webp bmp ico tga dds tif tiff psd bin dll exe lib obj asm so dylib pdb log xml yml yaml "
	"html htm css js ts json txt md c h cc cpp hpp cs py rs go java rb php sh bat ps1 cmd csv tsv bk2 bik bik2 "
	"mp4 mkv avi mov wav mp3 ogg zip rar 7z gz tar".split()
	)


	def _find_emails(text):
	out = []
	i = text.find("@")
	n = len(text)
	while i != -1:
	# local part: walk left (bounded)
	j = i
	lo = max(0, i - 64)
	while j > lo and text[j - 1] in "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789._%+-":
	j -= 1
	local = text[j:i]
	if local and any(ch.isalnum() for ch in local):
	m = _EMAIL_DOMAIN.match(text, i + 1)
	if m and m.group(1).lower() not in _FILE_EXT:
	out.append((j, m.end(), "email", EMAIL))
	i = text.find("@", i + 1)
	return out


	# ----------------------------------------------------------------------------- windows user paths
	_SYSTEM_PROFILES = frozenset(("public", "default", "default user", "all users", "user", "defaultuser0"))
	_WIN_USER = re.compile(
	r"(?P<drive>[A-Za-z]:)(?P<s1>\\{1,8}\|/{1,2})(?i:Users)(?P<s2>\\{1,8}\|/{1,2})(?P<name>[^\\/\s\"'<>\|:*?,;(){}\[\]`]+)"
	)
	_UNIX_USER = re.compile(
	r"(?<![A-Za-z0-9_.-])(?P<pre>/mnt/[a-z]\|/[a-z])/(?i:Users)/(?P<name>[^\\/\s\"'<>\|:*?,;(){}\[\]`]+)"
	)


	def _find_user_paths(text):
	out = []
	if "sers" not in text and "SERS" not in text:
	return out
	for m in _WIN_USER.finditer(text):
	name = m.group("name")
	if name.lower() in _SYSTEM_PROFILES:
	continue
	rep = f"{m.group('drive')}{m.group('s1')}Users{m.group('s2')}user"
	out.append((m.start(), m.end(), "win_user_path", rep))
	for m in _UNIX_USER.finditer(text):
	name = m.group("name")
	if name.lower() in _SYSTEM_PROFILES:
	continue
	out.append((m.start(), m.end(), "win_user_path_posix", f"{m.group('pre')}/Users/user"))
	return out


	# ----------------------------------------------------------------------------- OneDrive organisation names
	# Work/school OneDrive folders are named "OneDrive - <organisation>" and are followed by a path separator,
	# e.g. "D:\OneDrive - Some School\Documents". The organisation identifies the account owner, so it is masked.
	# The generic consumer folder "OneDrive - Personal" is not an organisation and is kept.
	ORG = "<ORG>"
	_ONEDRIVE_ORG = re.compile(r"(?<=OneDrive - )(?P<org>[^\\/\r\n\"<>\|:*?]{1,80}?)(?=\\\|/)")
	_ONEDRIVE_KEEP = frozenset(("personal",))


	def _find_onedrive_org(text):
	out = []
	if "OneDrive - " not in text:
	return out
	for m in _ONEDRIVE_ORG.finditer(text):
	if m.group("org").strip().lower() in _ONEDRIVE_KEEP:
	continue
	out.append((m.start("org"), m.end("org"), "onedrive_org", ORG))
	return out


	# ----------------------------------------------------------------------------- IPv4
	_OCT = r"(?:25[0-5]\|2[0-4]\d\|1\d\d\|[1-9]?\d)"
	_IPV4 = re.compile(rf"(?<![\w.]){_OCT}(?:\.{_OCT}){{3}}(?!\d\|\.\d)")


	_VERSION_CTX = re.compile(r"(?i)version[\"']?\s[=:]?\s[\"']?$")


	def _find_ips(text):
	out = []
	for m in _IPV4.finditer(text):
	v = m.group(0)
	if v == "0.0.0.0" or v.startswith("127."):
	continue
	if _VERSION_CTX.search(text[max(0, m.start() - 16):m.start()]):
	continue # .NET "Version=4.0.0.0" style assembly versions are not addresses
	out.append((m.start(), m.end(), "ipv4", IP))
	return out


	# ----------------------------------------------------------------------------- finders
	def _emb(text, start):
	"""':embedded_in_word' when the match is glued to a preceding word character."""
	return ":embedded_in_word" if start > 0 and (text[start - 1].isalnum() or text[start - 1] in "_-") else ""


	def _find_secrets(text):
	out = []
	if "-----BEGIN" in text:
	for m in _PEM_FULL.finditer(text):
	out.append((m.start(), m.end(), "secret:pem_private_key", SECRET))
	for m in _PEM_OPEN.finditer(text):
	out.append((m.start(), m.end(), "secret:pem_private_key", SECRET))
	if "AKIA" in text:
	for m in _AKIA.finditer(text):
	out.append((m.start(), m.end(), "secret:aws_access_key" + _emb(text, m.start()), SECRET))
	if "gh" in text:
	for m in _GH.finditer(text):
	out.append((m.start(), m.end(), "secret:github_token" + _emb(text, m.start()), SECRET))
	if "github_pat_" in text:
	for m in _GH_PAT.finditer(text):
	out.append((m.start(), m.end(), "secret:github_pat" + _emb(text, m.start()), SECRET))
	if "sk-" in text:
	for m in _SK.finditer(text):
	out.append((m.start(), m.end(), "secret:sk_key" + _emb(text, m.start()), SECRET))
	if "earer" in text:
	for m in _BEARER.finditer(text):
	v = m.group("val")
	if v[0] in "$%<{":
	continue
	out.append((m.start("val"), m.end("val"), "secret:bearer_token", SECRET))
	low = text.lower()
	if ("password" in low) or ("passwd" in low) or ("token" in low) or ("secret" in low) or ("apikey" in low) or ("api_key" in low) or ("api-key" in low):
	for m in _KV.finditer(text):
	if m.group("q"):
	s, e, val, quoted = m.start("qval"), m.end("qval"), m.group("qval"), True
	else:
	s, e, val, quoted = m.start("bval"), m.end("bval"), m.group("bval"), False
	if not _kv_value_ok(val, quoted):
	continue
	kname = re.sub(r"[^a-z]", "", m.group("key").lower().split("_")[-1].split("-")[-1].split(".")[-1])
	out.append((s, e, "secret:kv:" + (kname or "key"), SECRET))
	return out


	def _resolve(spans):
	"""Earliest start wins; at equal start the longer span wins. Overlapping losers are dropped."""
	spans.sort(key=lambda s: (s[0], -(s[1] - s[0])))
	kept = []
	last_end = -1
	for s in spans:
	if s[0] >= last_end:
	kept.append(s)
	last_end = s[1]
	return kept


	def _sha(s):
	return hashlib.sha256(s.encode("utf-8", "surrogatepass")).hexdigest()


	def redact_text(text):
	"""Redact `text`. Returns (new_text, events) where events is a list of
	(start, end, kind, sha256_of_original, replacement) with start/end in new_text coordinates."""
	if not text:
	return text, []
	spans = _find_secrets(text) + _find_emails(text) + _find_user_paths(text) + _find_onedrive_org(text) + _find_ips(text)
	if not spans:
	return text, []
	kept = _resolve(spans)
	pieces = []
	events = []
	pos = 0
	out_len = 0
	for s, e, kind, rep in kept:
	seg = text[pos:s]
	pieces.append(seg)
	out_len += len(seg)
	events.append((out_len, out_len + len(rep), kind, _sha(text[s:e]), rep))
	pieces.append(rep)
	out_len += len(rep)
	pos = e
	pieces.append(text[pos:])
	new = "".join(pieces)
	# Verification loop: a secret-class match must never survive.
	for _ in range(3):
	left = _find_secrets(new)
	if not left:
	break
	left = _resolve(left)
	pieces = []
	pos = 0
	shift_events = []
	for s, e, kind, rep in left:
	pieces.append(new[pos:s])
	pieces.append(rep)
	shift_events.append((s, e, kind, _sha(new[s:e]), rep))
	pos = e
	pieces.append(new[pos:])
	new = "".join(pieces)
	# offsets of earlier events can shift; recompute coarsely by dropping offsets (rare path)
	events = [(-1, -1, k, h, r) for (_, _, k, h, r) in events] + [(-1, -1, k, h, r) for (_, _, k, h, r) in shift_events]
	return new, events


	def redact_json_leaves(obj, sink):
	"""Redact every string leaf of a parsed JSON value. `sink` receives (kind, sha, replacement) tuples.
	Returns (new_obj, changed)."""
	if isinstance(obj, str):
	new, ev = redact_text(obj)
	for (_, _, k, h, r) in ev:
	sink.append((k, h, r))
	return new, bool(ev)
	if isinstance(obj, list):
	changed = False
	res = []
	for x in obj:
	y, c = redact_json_leaves(x, sink)
	res.append(y)
	changed = changed or c
	return res, changed
	if isinstance(obj, dict):
	changed = False
	res = {}
	for k, v in obj.items():
	y, c = redact_json_leaves(v, sink)
	res[k] = y
	changed = changed or c
	return res, changed
	return obj, False


	# ----------------------------------------------------------------------------- validation regexes
	# Plain (un-boundaried) secret regexes used by the validation step, deliberately broader than the
	# redaction finders so that a miss caused by a boundary rule is still surfaced.
	VALIDATION_REGEXES = {
	"AKIA": re.compile(r"AKIA[0-9A-Z]{16}"),
	"ghp_": re.compile(r"ghp_[A-Za-z0-9]{20,}"),
	"github_pat_": re.compile(r"github_pat_[A-Za-z0-9_]{20,}"),
	"sk-": re.compile(r"sk-[A-Za-z0-9_-]{20,}"),
	"PEM PRIVATE KEY": re.compile(r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY"),
	"Authorization: Bearer": re.compile(r"(?i)Authorization\s[:=]\s\\[\"']?\sBearer\s+(?!<SECRET>)[A-Za-z0-9._~+/=-]{8,}"),
	}

	# ----------------------------------------------------------------------------- phone-like candidates
	# Reported, never redacted.
	PHONE_RE = re.compile(
	r"(?<![\w.:/\\-])(?:\+\d{1,3}[ .-]?)?(?:\(\d{2,4}\)[ .-]?\|\d{2,4}[ .-])\d{3,4}[ .-]\d{3,4}(?:[ .-]\d{1,4})?(?![\w.:/\\-])"
	)


	def phone_candidates(text):
	res = []
	for m in PHONE_RE.finditer(text):
	v = m.group(0)
	digits = re.sub(r"\D", "", v)
	if 9 <= len(digits) <= 15 and re.search(r"[ ().+-]", v):
	res.append(v)
	return res