Files
chanora/poc/diagnostics-redaction-spike/src/redactor.rs
T
EdisonJwa 06ec6f2965 feat(poc/diagnostics): add diagnostics-redaction spike
Proof-of-concept proving the diagnostics-redaction exit criterion from
docs/architecture/proof-of-concept-plan.md §2:
  "Password and identity-secret samples are redacted."

Full coverage of the audit-report test matrix in
docs/security/diagnostic-redaction-audit-report.md §4
(REDACT-TC-001..010), plus two sanity tests.

The spike ships:
  - RedactionPolicy: typed catalogue of regex rules
    (identity-base64-blob, password-kv, ts3server-url-password,
     authorization-bearer, linux/windows/macos user-path) with
    optional capture-group narrowing.
  - Structured-field redaction keyed on case-insensitive name
    substrings (password, secret, token, ...).
  - Bundle-level switches: chat and channel tree excluded by
    default per audit-report §5.
  - KnownSecretRegistry: literal-substring scrub for secrets the
    host application has already loaded into memory (defense in
    depth that regexes alone cannot guarantee — closes the gap
    behind REDACT-TC-002).
  - Length cap (MAX_PROTOCOL_STRING_LEN = 256) with truncation
    marker for REDACT-TC-009.
  - UTF-8 preserved in non-sensitive fields per REDACT-TC-010 /
    ADR-008.

Test suite (12/12 PASS on 2026-05-13):
  REDACT-TC-001 server password in connection data
  REDACT-TC-002 identity secret in storage error (via KnownSecretRegistry)
  REDACT-TC-003 server URL with password field
  REDACT-TC-004 chat text excluded by default
  REDACT-TC-005 channel name with Unicode excluded by default
  REDACT-TC-006 nickname with Unicode preserved in safe field
  REDACT-TC-007 local file path user segment minimized
  REDACT-TC-008 mixed sensitive bundle (whole-bundle JSON scan)
  REDACT-TC-009 long hostile protocol string truncated
  REDACT-TC-010 multilingual safe text preserved
  + known-secret literal scrub
  + empty registered secret ignored

Out of scope: tracing-subscriber integration, diagnostic export
file format, memory/core dumps, performance, adversarial regex
evasion beyond trivial cases. These belong to chanora_diagnostics.

Authority: PoC plan §2, docs/security/diagnostic-redaction-audit-report.md,
SRS-093, SysRS-152/154/155.
Not product code; not promoted into chanora_diagnostics.
2026-05-14 12:26:49 +08:00

128 lines
4.0 KiB
Rust

//! The redactor — applies the policy to free text and to structured
//! diagnostic data.
use std::collections::HashSet;
use std::sync::{Arc, RwLock};
use crate::policy::{RedactionPolicy, MAX_PROTOCOL_STRING_LEN};
use crate::REDACTION_MARKER;
/// Runtime registry of *literal* known secrets. The host application
/// (e.g. the secure-storage layer) registers a secret here when it
/// loads one into memory, so even if it slips into a log line verbatim
/// it gets scrubbed.
///
/// This is the strongest defense for SS-AUD-003 and REDACT-TC-002 —
/// regexes can miss; literal substring matches cannot.
#[derive(Debug, Clone, Default)]
pub struct KnownSecretRegistry {
inner: Arc<RwLock<HashSet<String>>>,
}
impl KnownSecretRegistry {
pub fn new() -> Self {
Self::default()
}
/// Register a literal secret string. Empty strings are ignored
/// (to avoid degenerate `s.replace("", _)` behaviour).
pub fn register(&self, secret: &str) {
if secret.is_empty() {
return;
}
self.inner.write().unwrap().insert(secret.to_string());
}
pub fn forget(&self, secret: &str) {
self.inner.write().unwrap().remove(secret);
}
pub fn snapshot(&self) -> Vec<String> {
self.inner.read().unwrap().iter().cloned().collect()
}
}
pub struct Redactor {
policy: RedactionPolicy,
known: KnownSecretRegistry,
}
impl Redactor {
pub fn new(policy: RedactionPolicy, known: KnownSecretRegistry) -> Self {
Self { policy, known }
}
pub fn with_default_policy() -> Self {
Self::new(RedactionPolicy::default_policy(), KnownSecretRegistry::new())
}
pub fn known_secrets(&self) -> &KnownSecretRegistry {
&self.known
}
pub fn policy(&self) -> &RedactionPolicy {
&self.policy
}
/// Redact `input` according to the policy + known secret registry.
///
/// Order:
/// 1. Known literal secrets (defence in depth — never miss a value).
/// 2. Regex rules (with optional capture-group narrowing).
/// 3. Length cap for hostile / oversized strings (REDACT-TC-009).
pub fn redact_text(&self, input: &str) -> String {
// 1. Literal known secrets.
let mut text = input.to_string();
for known in self.known.snapshot() {
if !known.is_empty() {
text = text.replace(&known, REDACTION_MARKER);
}
}
// 2. Regex rules.
for rule in &self.policy.rules {
text = match rule.redact_group {
None => rule
.pattern
.replace_all(&text, REDACTION_MARKER)
.into_owned(),
Some(group) => rule
.pattern
.replace_all(&text, |caps: &regex::Captures<'_>| {
let full = caps.get(0).map(|m| m.as_str()).unwrap_or("");
match caps.get(group) {
Some(target) => full.replace(target.as_str(), REDACTION_MARKER),
None => full.to_string(),
}
})
.into_owned(),
};
}
// 3. Length cap.
if text.len() > MAX_PROTOCOL_STRING_LEN {
// Find the last char boundary inside the budget so we
// never split a UTF-8 codepoint.
let mut cut = MAX_PROTOCOL_STRING_LEN;
while cut > 0 && !text.is_char_boundary(cut) {
cut -= 1;
}
let mut truncated = text[..cut].to_string();
truncated.push_str("…[truncated]");
text = truncated;
}
text
}
/// True if the policy classifies `field_name` as a sensitive
/// structured field that must be redacted wholesale.
pub fn is_sensitive_field(&self, field_name: &str) -> bool {
let lower = field_name.to_lowercase();
self.policy
.redact_field_substrings
.iter()
.any(|needle| lower.contains(needle))
}
}