diff --git a/crates/mind-body/src/persona.rs b/crates/mind-body/src/persona.rs index 34a6bd7..6c5b924 100644 --- a/crates/mind-body/src/persona.rs +++ b/crates/mind-body/src/persona.rs @@ -1939,10 +1939,22 @@ fn refused_note(why: Failure) -> String { format!("Your last reply was refused ({why}), so nothing in it was said or kept: answer the same messages again, as one reply.") } +/// Every string of `value` through the one URL-credential redaction, so JSON structure is never inside a candidate run. +fn redact_strings(value: &mut Value) { + match value { + Value::String(text) => *text = crate::self_run::redact_url_userinfo(text), + Value::Array(items) => items.iter_mut().for_each(redact_strings), + Value::Object(map) => map.values_mut().for_each(redact_strings), + _ => {} + } +} + /// The structured output of a result line, as the text `decode_reply` reads; none when it has none. pub(crate) fn structured_output(line: &str) -> Option { let event: Value = serde_json::from_str(line).ok()?; - event.get("structured_output").filter(|output| !output.is_null()).map(Value::to_string) + let mut output = event.get("structured_output").filter(|output| !output.is_null())?.clone(); + redact_strings(&mut output); + Some(output.to_string()) } impl PersonaInbox<'_, C, S> { diff --git a/crates/mind-body/src/self_run.rs b/crates/mind-body/src/self_run.rs index 9e6a4de..754b93c 100644 --- a/crates/mind-body/src/self_run.rs +++ b/crates/mind-body/src/self_run.rs @@ -825,6 +825,7 @@ impl<'a> SelfInbox<'a> { fn reply_to(&mut self, line: &str, ok: bool) { let Ok(Value::Object(event)) = serde_json::from_str::(line) else { return }; let text = result_text(&event); + let text = text.as_str(); let flush = std::mem::take(&mut self.flushing); // The flush result is the session's note to its successor, in the journal: it is no reply. if ok && !flush && !text.trim().is_empty() { @@ -923,8 +924,109 @@ fn result_head(event: &serde_json::Map) -> String { head } -fn result_text(event: &serde_json::Map) -> &str { - event.get("result").and_then(Value::as_str).unwrap_or("") +/// A result's words as they leave the actuator, for the journal and for her mind (the reply): redacted by `redact_url_userinfo`. +fn result_text(event: &serde_json::Map) -> String { + redact_url_userinfo(event.get("result").and_then(Value::as_str).unwrap_or("")) +} + +/// The one redaction of URL credentials. Call it at each place a string the CLI produced enters the journal, her mind or chat: today +/// the Self result (`result_text`: journal line and Reply) and the structured output of a Persona or Sleep result +/// (`persona::structured_output`, string by string). It does not cover claude's stderr (inherited) or the CLI's transcript files. +/// +/// A forge token rides in a remote URL, so every `scheme://` and every protocol-relative `//` opens an authority, read as git, curl and +/// RFC 3986 read it: it ends at the first `/`, `?`, `#`, backslash, quote, backtick, parenthesis, bracket, angle bracket or ASCII whitespace +/// (`is_authority_end`). Its userinfo is everything before the authority's LAST `@` (a raw `@` in a password still redacts), and non-ASCII +/// characters, NBSP and U+2028 included, stay inside it. The userinfo becomes ``: `scheme://@host`. With a colon it is +/// always redacted; without one it is a bare token on http, https, ftp, ftps, ws, wss, any `+http`/`+https` scheme (`git+https`), protocol-relative `//` and an unparseable scheme, and a plain +/// username on any other scheme (`ssh://git@host`), which stays. A `@` after the authority (`github.com/@user`, `pkg@1.2.3`) is path, not +/// userinfo, and is left alone. Scheme-less credentials are covered only in the shape `word:secret@host.tld/` at the start of an +/// ASCII-whitespace-delimited word. Not covered: a password holding `/ ? # ` a quote or ASCII whitespace (no parser reads it as userinfo), +/// scp-style `user:secret@host:path`, tokens in query strings, claude's stderr, the kept transcripts. Syntax only; no reading of meaning. +pub(crate) fn redact_url_userinfo(text: &str) -> String { + let mut out = String::with_capacity(text.len()); + let mut rest = text; + while let Some(at) = rest.find("//") { + let (before, after) = rest.split_at(at); + let body = &after[2..]; + out.push_str(before); + out.push_str("//"); + // The scheme is the ASCII run before `:`. Without a colon, `//` opens an authority only at a word start (`a//b@c` is not one). + let scheme = match before.strip_suffix(':') { + // Only ASCII bytes match, so the byte after the last non-match is a character boundary. + Some(head) => head[head.bytes().rposition(|c| !(c.is_ascii_alphanumeric() || matches!(c, b'+' | b'-' | b'.'))).map_or(0, |i| i + 1)..].to_ascii_lowercase(), + None if before.chars().next_back().is_none_or(|c| is_break(c) || matches!(c, '"' | '\'' | '`' | '(' | '[' | '{' | '<' | '=' | ',')) => String::new(), + None => { + rest = body; + continue; + } + }; + let end = body.find(is_authority_end).unwrap_or(body.len()); + let authority = &body[..end]; + match authority.rfind('@') { + Some(i) if i > 0 && (authority[..i].contains(':') || carries_bare_token(&scheme)) => { + out.push_str(""); + out.push_str(&authority[i..]); + } + _ => out.push_str(authority), + } + rest = &body[end..]; + } + out.push_str(rest); + redact_schemeless(&out) +} + +/// Schemes whose lone userinfo is a token, not a username: web and ftp schemes, `ws`/`wss`, and `x+http`/`x+https` wrappers; `scheme` is lowercase. +fn carries_bare_token(scheme: &str) -> bool { + matches!(scheme, "" | "http" | "https" | "ftp" | "ftps" | "ws" | "wss") || scheme.ends_with("+http") || scheme.ends_with("+https") +} + +/// ASCII whitespace; NBSP, U+2028 and other Unicode whitespace are not, so a secret holding them is redacted whole. +fn is_break(c: char) -> bool { + matches!(c, ' ' | '\t' | '\n' | '\r' | '\x0b' | '\x0c') +} + +/// The end of a URL's authority. +fn is_authority_end(c: char) -> bool { + is_break(c) || matches!(c, '/' | '?' | '#' | '\\' | '"' | '\'' | '`' | '(' | ')' | '[' | ']' | '{' | '}' | '<' | '>') +} + +/// `word:secret@host.tld/` at the start of a whitespace-delimited word that holds no `//` becomes `@host.tld/`. +fn redact_schemeless(text: &str) -> String { + let mut out = String::with_capacity(text.len()); + let mut word = String::new(); + let flush = |word: &mut String, out: &mut String| { + out.push_str(&redact_word(word)); + word.clear(); + }; + for c in text.chars() { + if is_break(c) { + flush(&mut word, &mut out); + out.push(c); + } else { + word.push(c); + } + } + flush(&mut word, &mut out); + out +} + +fn redact_word(word: &str) -> String { + let shaped = (|| { + if word.contains("//") { + return None; + } + let at = word.rfind('@')?; + let colon = word.find(':')?; + let (user, host) = (&word[..colon], &word[at + 1..]); + let host_name = &host[..host.find('/')?]; + let user_ok = !user.is_empty() && user.chars().all(|c| c.is_ascii_alphanumeric() || matches!(c, '.' | '_' | '-')); + let host_ok = host_name.contains('.') && host_name.chars().all(|c| c.is_ascii_alphanumeric() || matches!(c, '.' | '-')); + (colon + 1 < at && user_ok && host_ok).then_some(at) + })(); + match shaped { + Some(at) => format!("{}", &word[at..]), + None => word.to_string(), + } } /// `text` in consecutive pieces of at most `limit` bytes, each cut at a character boundary; their concatenation is `text`. diff --git a/crates/mind-body/src/self_run/tests/reply.rs b/crates/mind-body/src/self_run/tests/reply.rs index 0a4e37c..1e20e4d 100644 --- a/crates/mind-body/src/self_run/tests/reply.rs +++ b/crates/mind-body/src/self_run/tests/reply.rs @@ -118,3 +118,110 @@ fn a_refused_piece_of_a_long_result_holds_back_the_pieces_behind_it() { let ordinals: Vec = (1..=3).map(|n| mind.view(K::Reply, &format!("{INSTANCE}:reply:selfrun-a-r{n}")).unwrap_or_else(|| panic!("piece {n} lost")).admission.ordinal).collect(); assert!(ordinals.windows(2).all(|w| w[0] < w[1]), "the pieces reached the mind out of order: {ordinals:?}"); } + +// A canary stands where a forge token would: no byte of it may reach the journal or her mind. +const CANARY: &str = "CANARY-NOT-A-REAL-TOKEN-0000"; + +#[test] +fn url_userinfo_is_redacted_where_it_is_a_secret_and_kept_where_it_is_a_username() { + let said = format!( + "a https://user:{CANARY}@host.example/r.git b https://{CANARY}@host.example/r.git c ssh://u:{CANARY}@h/x d http://{CANARY}@h e HTTPS://x:{CANARY}@h \ + f https://u:p@ss{CANARY}@h/x g ssh://git@host.example/r.git h git://git@h/x i https://host.example/no-userinfo j (https://u:{CANARY}@h) k ://@x" + ); + let want = "a https://@host.example/r.git b https://@host.example/r.git c ssh://@h/x d http://@h e HTTPS://@h \ + f https://@h/x g ssh://git@host.example/r.git h git://git@h/x i https://host.example/no-userinfo j (https://@h) k ://@x"; + assert_eq!(redact_url_userinfo(&said), want); + assert!(!redact_url_userinfo(&said).contains(CANARY)); +} + +#[test] +fn a_credential_in_a_url_reaches_neither_the_journal_nor_the_admitted_reply() { + let mind = live_mind(); + let said = format!("origin https://user:{CANARY}@host.example/r.git and https://{CANARY}@host.example/s.git"); + let ran = mind_run(&mind, vec![Out(echo()), Out(saying(&said))]); + assert!(ran.outcome.is_ok()); + assert!(!ran.out.contains(CANARY), "{}", ran.out); + assert!(ran.out.contains("https://@host.example/r.git"), "{}", ran.out); + let text = text_of(&mind, 1); + assert!(!text.contains(CANARY), "{text}"); + assert_eq!(text, "origin https://@host.example/r.git and https://@host.example/s.git"); +} + +#[test] +fn a_credential_split_by_the_reply_limit_is_redacted_before_the_cut() { + let mind = live_mind(); + let said = format!("{}https://u:{CANARY}@h/x{}", "a".repeat(REPLY_TEXT_LIMIT - 25), "b".repeat(100)); + mind_run(&mind, vec![Out(echo()), Out(saying(&said))]).outcome.unwrap(); + let first = reply_of(&mind, 1).expect("the long result was admitted in pieces"); + assert!(reply_of(&mind, 2).is_some(), "the credential's place straddles the cut"); + assert!(!first.text.0.contains("CANARY") && !reply_of(&mind, 2).unwrap().text.0.contains("CANARY")); +} + +#[test] +fn a_secret_is_redacted_whole_by_the_authority_rule_and_commonplace_urls_are_left_alone() { + let r = redact_url_userinfo; + // Raw @, percent-encoding and non-ASCII (NBSP, U+2028) stay inside userinfo. + for secret in ["pa@ssT", "pa%40ssT", "pa\u{a0}ssT", "pa\u{2028}ssT", "pa:ssT", "\u{e9}\u{65e5}ssT"] { + let out = r(&format!("x https://u:{secret}@h/x")); + assert_eq!(out, "x https://@h/x", "{secret:?}"); + } + // The authority ends at / ? # \\ quotes, brackets and ASCII whitespace: no parser reads those as userinfo, so what follows is path. + assert_eq!(r("https://u:pa/ssT@h/x"), "https://u:pa/ssT@h/x"); + assert_eq!(r("https://u:pa ssT@h/x"), "https://u:pa ssT@h/x"); + assert_eq!(r("https://u:pa?ssT@h"), "https://u:pa?ssT@h"); + // Ordinary URLs with @ in the path are byte-identical. + for same in [ + "https://github.com/@user/repo", + "https://medium.com/@name/post-1", + "https://registry.npmjs.org/@scope/pkg", + "https://h/pkg@1.2.3", + "a//b@c", + "see //cdn.h/@x/y and https://h/?mail=a@b.c#f@g", + "https://h:8080/p git@github.com:org/r.git ssh://git@host/r.git ssh://git@h:22/r git://git@h/x", + ] { + assert_eq!(r(same), same); + } + // Compact JSON: only the credentialed URL changes. + let json = format!(r#"{{"a":"https://u:{CANARY}@h/x","b":"https://h2/@v","c":"https://h3/p@1"}}"#); + assert_eq!(r(&json), r#"{"a":"https://@h/x","b":"https://h2/@v","c":"https://h3/p@1"}"#); + assert_eq!(r("a/https://u:T@h/x,https://v:T@h2"), "a/https://@h/x,https://@h2"); + assert_eq!(r("HTTPS://TOKEN@h"), "HTTPS://@h"); + assert_eq!(r("[l](https://T@h/x) `https://u:T@h`"), "[l](https://@h/x) @h> `https://@h`"); + assert_eq!(r("a //u:T@host/x b //T@host c \"//T@h\" (//T@h)"), "a //@host/x b //@host c \"//@h\" (//@h)"); + // Token-bearing schemes beyond http: +http/+https wrappers and ws/wss redact a bare token; git+ssh and ssh keep their username. + assert_eq!(r("git+https://T@github.com/a/b GIT+HTTPS://T@h wss://T@h/ ws://T@h/"), "git+https://@github.com/a/b GIT+HTTPS://@h wss://@h/ ws://@h/"); + for same in ["git+ssh://git@host/x", "ssh://git@host/x"] { + assert_eq!(r(same), same); + } + // Scheme-less: only word:secret@host.tld/ . Each guard of the shape has its own case. + assert_eq!(r("user:T@host.tld/path"), "@host.tld/path"); + for same in ["a:b@c", "a:b@host/x", "a:b@ho_st.tld/x", "u$er:T@host.tld/x", "a:@host.tld/x", "mail me@x.org/ and :T@host.tld/x"] { + assert_eq!(r(same), same); + } +} + +#[test] +fn multibyte_text_never_panics_the_redaction() { + let r = redact_url_userinfo; + for case in ["see \u{201c}https://h/x", "\u{2192}https://h/x", "docs\u{ff1a}https://h/x", "\u{e9}://x", "\u{65e5}\u{672c}://u:T@h", "https://x/\u{e9}://", "\u{e9}//u:T@h", "\u{e9}:u:T@h.tld/"] { + let _ = r(case); + } + assert_eq!(r("\u{65e5}\u{672c}://u:T@h"), "\u{65e5}\u{672c}://@h"); + assert_eq!(r("\u{2192}https://u:T@h/x"), "\u{2192}https://@h/x"); + let seed = "\u{201c}\u{2192}\u{ff1a}\u{e9}\u{65e5}\u{a0}https://u:T@h/x//y:z@a.b/ ://@ //"; + let chars: Vec = seed.chars().collect(); + for start in 0..=chars.len() { + for end in start..=chars.len() { + let _ = r(&chars[start..end].iter().collect::()); + } + } +} + +#[test] +fn persona_and_sleep_structured_output_is_redacted_string_by_string() { + let line = json!({"type": "result", "structured_output": {"say": format!("https://u:{CANARY}@h/x,"), "n": 3, "list": [format!("https://{CANARY}@h")]}}).to_string(); + let text = crate::persona::structured_output(&line).unwrap(); + assert!(!text.contains(CANARY), "{text}"); + let back: Value = serde_json::from_str(&text).unwrap(); + assert_eq!(back, json!({"say": "https://@h/x,", "n": 3, "list": ["https://@h"]})); +}