@farukada/aws-langgraph-dynamodb-ts 0.9.0 → 1.0.0-rc.2

package/dist/shared/logging/secret-patterns.js

@@ -1,26 +1,64 @@
 "use strict";
+/**
+ * Hides what a secret looks like.
+ *
+ * Which key names mark their value as a secret, which credential formats are
+ * recognised in free text, how a match is replaced while the field's name is
+ * kept, and how a caught value's message is redacted before another error
+ * quotes it are decided here. The redacting logger and every error that
+ * quotes its cause share these rules, so a new credential shape is one edit.
+ */
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.DEFAULT_SECRET_VALUE_PATTERNS = exports.DEFAULT_SECRET_KEY_PATTERNS = exports.REDACTED = void 0;
+exports.normaliseKey = normaliseKey;
 exports.isSecretKey = isSecretKey;
+exports.redactedMessage = redactedMessage;
+exports.isRegExp = isRegExp;
 exports.redactText = redactText;
 exports.binaryLabel = binaryLabel;
 exports.redactErrorText = redactErrorText;
+const base_error_1 = require("../errors/base-error");
+const truncate_1 = require("./truncate");
 /** Marker substituted for anything recognised as secret. */
 exports.REDACTED = '[REDACTED]';
-/** Key names (matched case-insensitively as substrings) whose value is a secret. */
+/**
+ * Key names whose value is a secret. Each entry is a *normalised* name — see
+ * {@link normaliseKey} — and matches a key whose normalised form equals it or
+ * ends with it. Suffix matching is what catches `secretAccessKey`, `x-api-key`,
+ * `client_secret` and `AUTH_TOKEN`; requiring the pattern to be a suffix, not a
+ * substring, is what spares `maxTokens`, `total_tokens`, `tokenizer` and
+ * `secretary` — a bare substring match would redact all four for merely
+ * containing `token` or `secret`.
+ */
 exports.DEFAULT_SECRET_KEY_PATTERNS = [
     'accesskey',
-    'secretkey',
+    'accesskeyid',
     'secret',
     'sessiontoken',
     'securitytoken',
     'authorization',
     'password',
+    'passwd',
+    'passphrase',
     'apikey',
     'bearer',
     'token',
     'privatekey',
 ];
+/**
+ * Canonical form of a key name for matching.
+ *
+ * Accepts: any key name, in any case and with any separators.
+ *
+ * Returns: it lower-cased with every character that is not a letter or digit
+ * removed, so `api_key`, `x-api-key`, `ApiKey` and `API KEY` all become
+ * `apikey` and one pattern covers every spelling.
+ *
+ * Throws: nothing.
+ */
+function normaliseKey(key) {
+    return key.toLowerCase().replace(/[^a-z0-9]/g, '');
+}
 /**
  * Secret shapes recognisable in free text, where key-name matching cannot
  * reach: an error's `message`/`stack`, or any other string value. Deliberately
@@ -59,49 +97,172 @@ exports.DEFAULT_SECRET_VALUE_PATTERNS = [
     /\beyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+\b/g,
     /((?:aws_)?(?:secret_access_key|secretaccesskey|password|passwd|api_?key|token)["']?\s*[=:]\s*)(?:"(?:\\.|[^"\\])*"|'(?:\\.|[^'\\])*'|(?:-?\d+(?:\.\d+)?(?:[eE][+-]?\d+)?|true|false|null)(?=\s*[,}\]]|\s*$)|[^\r\n]+)/gi,
 ];
-/** True when `key` matches any secret-key pattern. */
+/**
+ * Whether a key name says its value is a secret.
+ *
+ * Accepts: `key` — any name. `patterns` — already-normalised names; an empty
+ * list matches nothing, which is how a caller turns key matching off.
+ *
+ * Returns: whether the normalised key equals or *ends with* a pattern. Suffix,
+ * not substring: that is what catches `secretAccessKey` and `x-api-key` while
+ * sparing `maxTokens`, `tokenizer` and `secretary`.
+ *
+ * Throws: nothing.
+ */
 function isSecretKey(key, patterns) {
-    const lower = key.toLowerCase();
-    return patterns.some((pattern) => lower.includes(pattern));
+    const normalised = normaliseKey(key);
+    return patterns.some((pattern) => normalised === pattern || normalised.endsWith(pattern));
+}
+/**
+ * An error's message with recognised credential shapes redacted.
+ *
+ * Accepts: any error — and, since this is what a `catch` block binds, any
+ * other value a `throw` can produce: `null`, `undefined`, a string, a number,
+ * a symbol, a plain object. Such a value is described through {@link toError}
+ * first and the description is what is redacted, so a secret it carries in its
+ * own text is still caught.
+ *
+ * Returns: the message, redacted with the default value patterns and then
+ * bounded by `truncateRelayedText` — for embedding in another error's
+ * message, since a wrapper that quotes its cause must not leak a `password=`
+ * or token the cause happened to carry. That text reaches `err.message`, which
+ * an application may print without a redacting logger.
+ *
+ * The bound is here rather than at the sites that quote the result, so no call
+ * site can get it wrong and no future one has to remember: a `RETRY_EXHAUSTED`
+ * report, a wrapped AWS failure, a compensation failure and both S3 transfer
+ * errors all inherit it from this one line. Redaction runs **before** the cut,
+ * because cutting first could split a credential shape past the pattern that
+ * would have caught it. Its own cap rather than the identifier cap: this is
+ * prose, and `MAX_RELAYED_MESSAGE_CHARS` records what that number is
+ * measured against. The value worth bounding is not the AWS SDK's own text but
+ * a **caller's** — a `serde` refusal or a `vectorBackend` rejection, whose
+ * length the caller controls entirely and which some paths quote once per row.
+ *
+ * Throws: **nothing**, for any value. Reading `.message` off a thrown
+ * primitive yielded `undefined` and the redaction then raised a `TypeError` —
+ * from inside the `catch` that was reporting the real failure, and from the
+ * one function a static guard funnels every one of this package's `catch`
+ * blocks into. Its promise has to hold for what a `catch` actually binds, or
+ * the guard concentrates every site into a function that is not safe.
+ */
+function redactedMessage(error) {
+    return (0, truncate_1.truncateRelayedText)(redactText((0, base_error_1.toError)(error).message, exports.DEFAULT_SECRET_VALUE_PATTERNS));
+}
+/**
+ * A fresh, global copy of `pattern`.
+ *
+ * Rebuilt per call so a `g` flag's `lastIndex` never leaks between
+ * invocations, and **forced** global: `String.prototype.replace` without `g`
+ * substitutes only the first match, so a caller-supplied pattern written
+ * without the flag would redact the first occurrence of a secret and print
+ * every later one verbatim.
+ */
+function globalCopyOf(pattern) {
+    const flags = pattern.flags.includes('g') ? pattern.flags : `${pattern.flags}g`;
+    return new RegExp(pattern.source, flags);
+}
+/**
+ * Whether `value` is a regular expression.
+ *
+ * Accepts: anything, including a `RegExp` from another realm.
+ *
+ * Returns: whether its tag says so. By tag rather than `instanceof`, which is
+ * banned repo-wide because it answers "no" across a realm or a duplicated
+ * module.
+ *
+ * Throws: nothing.
+ */
+function isRegExp(value) {
+    return Object.prototype.toString.call(value) === '[object RegExp]';
 }
 /**
  * Replace every recognised secret shape inside `value` with {@link REDACTED},
  * leaving the surrounding text intact so a redacted message stays readable.
- * Each pattern is rebuilt per call so a `g` flag's `lastIndex` never leaks
- * between invocations.
- *
- * A pattern may capture a leading group it wants **preserved**: only the rest
- * of the match is replaced, which is what keeps `apiKey=[REDACTED]` saying
- * which field was redacted instead of collapsing to a bare marker. A pattern
- * with no group is replaced whole, as before. `String.prototype.replace`
- * passes the match *offset* — a number — as the second callback argument when
- * the pattern has no group, hence the `typeof` test rather than an
- * `undefined` check.
+ *
+ * Accepts: `value` — any text. `patterns` — applied in order, each against the
+ * result of the last; an empty list returns the text unchanged. An entry that
+ * is not a `RegExp` is skipped: reading `.source` off one produced
+ * `new RegExp(undefined)` — that is `/(?:)/`, which matches the empty string
+ * and prefixed the marker to every value while catching no secret at all.
+ * Both public entry points — `redactSecrets` and {@link redactLogger} — now
+ * refuse such an entry where it is supplied, since a skipped pattern protects
+ * nothing while its caller believes it does; skipping remains the last-ditch
+ * guard for a list this package assembles internally.
+ *
+ * Returns: the text with every match replaced. A pattern may capture a leading
+ * group it wants **preserved**: only the rest of the match is replaced, which
+ * keeps `apiKey=[REDACTED]` saying which field was redacted instead of
+ * collapsing to a bare marker. A pattern with no group is replaced whole.
+ * `String.prototype.replace` passes the match *offset* — a number — as the
+ * second callback argument when the pattern has no group, hence the `typeof`
+ * test rather than an `undefined` check.
+ *
+ * Throws: nothing.
+ *
+ * Guarantees: every pattern is applied globally and from a fresh copy, so
+ * neither a missing `g` flag nor a leftover `lastIndex` can leave a later
+ * occurrence in the clear.
  */
 function redactText(value, patterns) {
-    return patterns.reduce((text, pattern) => text.replace(new RegExp(pattern.source, pattern.flags), (_match, prefix) => typeof prefix === 'string' ? `${prefix}${exports.REDACTED}` : exports.REDACTED), value);
+    return patterns.reduce((text, pattern) => {
+        if (!isRegExp(pattern))
+            return text;
+        return text.replace(globalCopyOf(pattern), (_match, prefix) => typeof prefix === 'string' ? `${prefix}${exports.REDACTED}` : exports.REDACTED);
+    }, value);
 }
 /**
- * A short label for a binary view. Recursing one would explode it into a
- * per-index numeric map, both unreadable and far larger than the value itself.
+ * A short label for a binary view.
+ *
+ * Accepts: any `ArrayBufferView`.
+ *
+ * Returns: its type and byte length, e.g. `[Uint8Array(4096)]`. Recursing into
+ * one would explode it into a per-index numeric map, both unreadable and far
+ * larger than the value itself — and a payload is exactly the thing a log must
+ * not carry.
+ *
+ * Throws: nothing.
  */
 function binaryLabel(value) {
     return `[${value.constructor.name}(${value.byteLength})]`;
 }
 /**
- * Redact an Error's `name`/`message`/`stack`. `changed` reports whether any
- * secret was actually found, which is what decides between passing a bare
- * Error through by reference (preserving its identity and stack trace) and
- * rebuilding it so the secret cannot escape.
+ * Redact an Error's `name`, `message` and `stack`.
+ *
+ * Accepts: `error` — any error; a missing `stack` stays missing. Any other
+ * value a `throw` can produce is described through {@link toError} first, on
+ * the same reasoning as {@link redactedMessage}: the walk that calls this only
+ * reaches it for a value whose tag says `Error`, but the promise below is
+ * written in this function's own contract and is this function's to keep.
+ *
+ * Returns: the redacted text, plus `changed`: whether any secret was actually
+ * found. That flag is what decides between passing a bare Error through by
+ * reference — preserving its identity and stack trace — and rebuilding it so
+ * the secret cannot escape. It compares against the described error, so a
+ * value that had no text of its own is never reported as changed by the
+ * describing.
+ *
+ * None of the three is cut, and that is a decision rather than an omission.
+ * The rule that bounds unchecked text bounds what *this package* writes into
+ * its own `err.message` and its own log lines — `redactedMessage` is that
+ * funnel. This function is on the other side of the boundary: it rebuilds an
+ * Error the **consumer** handed to `redactSecrets`, or to the logger
+ * `redactLogger` wrapped, on the way to the consumer's own transport. Cutting
+ * there would change what their transport receives, for a value this package
+ * neither produced nor quotes, and the caller asked for redaction rather than
+ * for truncation. A consumer who wants a bound has their own transport to put
+ * one in.
+ *
+ * Throws: **nothing**, for any value.
  */
 function redactErrorText(error, patterns) {
-    const message = redactText(error.message, patterns);
-    const stack = error.stack === undefined ? undefined : redactText(error.stack, patterns);
+    const described = (0, base_error_1.toError)(error);
+    const message = redactText(described.message, patterns);
+    const stack = described.stack === undefined ? undefined : redactText(described.stack, patterns);
     return {
-        name: error.name,
+        name: described.name,
         message,
         stack,
-        changed: message !== error.message || stack !== error.stack,
+        changed: message !== described.message || stack !== described.stack,
     };
 }
-//# sourceMappingURL=secret-patterns.js.map

package/dist/shared/logging/truncate.d.ts

@@ -0,0 +1,197 @@
+/**
+ * Hides how much of an unchecked string a line may quote.
+ *
+ * A log line or public error message that quotes a string this package did
+ * not length-check passes it through here. The caps for an identifier, for a
+ * relayed cause's prose and for a list of labels, and the mark that states a
+ * cut value's real length, are chosen here, so a call site says what kind of
+ * value it quotes and never how many characters it keeps.
+ */
+/**
+ * Characters of an unchecked string one log line or one public error message
+ * carries, past which it is cut and marked with its real length.
+ *
+ * Most of what these lines quote is a row's sort key or an offloaded object's
+ * S3 key, so the service already caps each at 1024 bytes — the cost is not one
+ * long line but many. `list: skipped a row that is not a checkpoint meta item`
+ * and `left a foreign row in place` fire once per row, and those passes walk a
+ * whole partition, up to `MAX_TOTAL_ROWS_IN_MEMORY`
+ * (`src/shared/dynamodb/paginate.ts`) rows: one call on a shared table could
+ * write megabytes of log. A consumer's `VectorBackend` carries no such service
+ * cap at all.
+ *
+ * 256 is `MAX_KEY_SEGMENT_BYTES` (`src/shared/dynamodb/table-schema.ts`), this
+ * package's own budget for one identifier inside a key, so any key composed
+ * from identifiers it validated is quoted whole in the common case and only a
+ * foreign row, a hand-written one or a backend's own answer — exactly the
+ * cases these lines report — is cut. Nothing is lost by cutting: the line's
+ * job is to say which row to go and look at, and the row holds the rest.
+ */
+export declare const MAX_LOGGED_VALUE_CHARS = 256;
+/**
+ * Labels of an unchecked `string[]` one log line or one public error message
+ * carries, past which the rest are dropped and the real depth is stated.
+ *
+ * An array is two unbounded things — how many labels there are and how long
+ * each one is — so a bound on the labels alone is not a bound: a backend
+ * answering with one label of a megabyte and one answering with a million
+ * labels of a character cost the same line. {@link MAX_LOGGED_VALUE_CHARS}
+ * covers the first, this covers the second.
+ *
+ * 8 is a budget rather than a rule about namespaces: a store namespace is a
+ * path, what identifies which path is its leading labels, and every namespace
+ * this package's own documentation forms is two or three deep. The marker
+ * states the depth it really had, so a deeper one is cut without being
+ * misreported.
+ *
+ * A `namespace` and `key` pair that passed `parseStoreAddress` needs none of
+ * this and goes in whole: that check measures the sort key they *compose*, so
+ * it bounds how many labels there are as well as how long each one is. A
+ * search or listing **prefix** passes no such check — nothing composes it into
+ * a key — so its depth is unchecked however carefully each label was checked,
+ * and a backend's own answer is unchecked in both.
+ */
+export declare const MAX_LOGGED_LABELS = 8;
+/**
+ * Characters of a relayed *cause's* text one public error message or one log
+ * line carries, past which it is cut and marked with its real length.
+ *
+ * Its own cap rather than {@link MAX_LOGGED_VALUE_CHARS} because the two bound
+ * different things. That one bounds an **identifier** — a sort key, an S3
+ * object key, a namespace label — and 256 is this package's own budget for one
+ * identifier inside a key, so a value past it is already abnormal and the line
+ * only has to say which row to go and look at. This one bounds **prose**: the
+ * sentence an AWS SDK error, a consumer's `VectorBackend` or a caller's own
+ * `serde` wrote to explain a failure, which `redactedMessage` relays into
+ * `err.message`. Cutting that at an identifier's budget would throw away the
+ * half of a diagnostic that says what to do about it, and a diagnostic is the
+ * entire value of relaying it at all.
+ *
+ * 1024 is measured against the longest text this package actually relays: an
+ * IAM `AccessDenied`, which names the calling principal's ARN, the action and
+ * the resource ARN and then says why no policy allows it, runs to the mid
+ * hundreds of characters, and a role ARN with a long path and a session name
+ * pushes it further. 1024 clears that whole, so the case an operator most
+ * needs to read arrives intact.
+ *
+ * What it is *for* is the other direction. `redactedMessage` also relays a
+ * **caller's own** thrown error — a `serde` refusing a value, a `vectorBackend`
+ * rejecting a query — whose length the caller controls entirely, and those
+ * messages are quoted once per row on paths that walk a whole prefix or table.
+ * Unbounded, one such error fills a log; at 1024 a thousand of them are a
+ * megabyte rather than an unbounded amount.
+ *
+ * Its own literal at the same value as `MAX_SORT_KEY_BYTES`
+ * (`src/shared/dynamodb/table-schema.ts`) and `MAX_S3_KEY_BYTES`
+ * (`src/shared/codec/s3/config.ts`) rather than an alias of either, for the
+ * reason `LIST_SCAN_WARN_THRESHOLD` (`src/shared/dynamodb/paginate.ts`)
+ * records: aliasing two caps would move one whenever the other is retuned,
+ * and these three answer unrelated questions.
+ */
+export declare const MAX_RELAYED_MESSAGE_CHARS = 1024;
+/**
+ * Bound a string this package did not length-check.
+ *
+ * **The rule, in one sentence:** a string goes through here before a log line
+ * or a public error message quotes it, unless this package composed it from
+ * identifiers it length-checked. Where the string came from does not decide
+ * it — a row, a bucket's own configuration, a consumer's `VectorBackend` and
+ * an object the caller handed in are the same thing here, which is that
+ * nothing bounded them. A caller's `sessionId`, `threadId`, `namespace` and
+ * `key`, and every key built from them, are capped by
+ * `MAX_PARTITION_ID_BYTES`, `MAX_KEY_SEGMENT_BYTES` and `MAX_SORT_KEY_BYTES`
+ * before a request is made, so those go in as they are; a row's own `SK`, an
+ * offloaded object's key, a descriptor's `location`, a backend's `namespace`
+ * and `key`, a lifecycle rule's scope and a stored message's `type` are
+ * bounded by nothing this package ran.
+ *
+ * **It says "or a public error message" for one reason:** one value must not
+ * get two answers. An `s3Key` off a row was cut for the `warn` that reports
+ * refusing to delete the object and quoted whole by the error that refuses to
+ * read it — one package, one value, two answers. An error's text is not the
+ * compatibility surface (the README tells callers to branch on `code`, `name`
+ * and the structured fields, never on text) and the `context` those errors
+ * carry keeps the value whole, so what reads an error as data loses nothing
+ * by this.
+ *
+ * Accepts: `value` — the string as whatever produced it carried it. Declared
+ * `string` because a table's own key attributes always are, and because the
+ * interfaces a consumer implements say so; anything else is returned
+ * untouched rather than coerced or refused, since a report of a value that is
+ * already wrong is the last place to raise a `TypeError` of its own.
+ *
+ * Returns: the value unchanged at or under {@link MAX_LOGGED_VALUE_CHARS}
+ * characters, otherwise that many characters followed by `…(len N)` giving the
+ * length it really had. The mark is what keeps a cut value honest: without it
+ * a truncated key reads as a key that simply ends there.
+ *
+ * Throws: nothing.
+ *
+ * Guarantees: a cut never falls between the halves of a surrogate pair, so a
+ * well-formed value stays well-formed. A lone surrogate is what
+ * `assertWellFormed` exists to keep out of this package's strings, and a JSON
+ * log transport rewrites one to U+FFFD without saying so.
+ */
+export declare function truncateForLog(value: string): string;
+/**
+ * Bound the **prose** of a relayed cause, which takes its own cap.
+ *
+ * {@link truncateForLog} bounds an identifier, where 256 characters is this
+ * package's own budget for one and anything past it is already abnormal. The
+ * text an AWS SDK error, a consumer's `VectorBackend` or a caller's own
+ * `serde` wrote is a sentence rather than a name: cut at an identifier's
+ * budget it loses the half that says what to do, which is the only reason to
+ * relay it. {@link MAX_RELAYED_MESSAGE_CHARS} records what the larger number
+ * is measured against, and why the two differ.
+ *
+ * It has exactly one caller — `redactedMessage`, the funnel every `catch` in
+ * this package goes through — so no call site can get the cap wrong and no
+ * future one has to remember it. That also means a site must not cut the
+ * result again: a second cut marks the length of the first cut's output rather
+ * than of the original, which is the one thing the mark exists to prevent.
+ *
+ * Accepts: `value` — the redacted text. Redaction runs first, so a credential
+ * shape can never be half-cut past the pattern that would have caught it.
+ *
+ * Returns: the value unchanged at or under the cap, otherwise that many
+ * characters followed by `…(len N)`.
+ *
+ * Throws: nothing.
+ */
+export declare function truncateRelayedText(value: string): string;
+/**
+ * Bound a list of labels — a store `namespace`, a list of channel names — the
+ * same rule reaches.
+ *
+ * An array is two unbounded things, how many labels it holds and how long each
+ * one is, so a bound on one of them is not a bound. {@link truncateForLog}
+ * covers the labels; the count is covered here.
+ *
+ * A `namespace` and `key` pair that passed `parseStoreAddress` is already
+ * bounded in both — that check measures the sort key they compose — and goes
+ * in whole. A search or listing prefix is not: nothing composes it into a key,
+ * so however carefully each label was checked, how many there are was not.
+ *
+ * The labels stay a list rather than being joined into one string, for two
+ * reasons. A structured transport already carries the field as a list and the
+ * README's Logging table documents it as one, so joining would change the
+ * shape of a line rather than only its size. And a label nothing validated may
+ * hold the `#` a join would put between labels, so a joined line cannot say
+ * whether the backend answered with one label or with two — which is the very
+ * thing these lines report.
+ *
+ * Accepts: `labels` — as the row, the backend or the caller gave it. Declared
+ * `string[]` because that is what the interfaces say; anything else is
+ * returned untouched, for the reason {@link truncateForLog} gives, and that
+ * case is reached rather than defensive — a `namespace` that is not an array
+ * is one of the things `parseStoreAddress` refuses, and the line reporting the
+ * refusal quotes what was refused.
+ *
+ * Returns: at most {@link MAX_LOGGED_LABELS} labels, each bounded by
+ * {@link truncateForLog}, with one further label reading `…(len N)` when some
+ * were dropped, giving the depth the list really had. A list within both
+ * bounds comes back as an equal list.
+ *
+ * Throws: nothing.
+ */
+export declare function truncateLabelsForLog(labels: string[]): string[];