gstack/lib/redact-engine.ts

/**
 * redact-engine — pure scanning + auto-redaction over the shared taxonomy.
 *
 * No I/O. Deterministic given its one ambient input: `scan()` reads the
 * credential-named variables of `process.env` by value unless `processEnv` is
 * passed (GSTA-104). The CLI shim (`bin/gstack-redact`), the pre-push hook
 * (`bin/gstack-redact-prepush`), and tests all import from here.
 *
 * Key behaviors (locked in /plan-eng-review + two Codex passes):
 *   - Normalization BEFORE matching (NFKC + strip zero-width + decode a small
 *     set of HTML entities) so Unicode-confusable / zero-width evasion fails.
 *     Findings map back to ORIGINAL offsets via an index map.
 *   - ReDoS safety: a hard input-size cap that fails CLOSED (oversize input
 *     returns a single synthetic HIGH "input too large to scan safely" finding,
 *     so callers block rather than skip). Patterns are linear-time (lint-tested).
 *   - NO visibility-based tier mutation. `repoVisibility` is recorded on each
 *     finding (drives sterner AUQ wording in the skill) but never promotes a
 *     MEDIUM to HIGH. (TENSION-2-followup.)
 *   - Placeholder suppression is per-matched-span.
 *   - Tool-attributed fences (``` ```codex-review ``` / ``` ```greptile ```)
 *     degrade credential findings to a non-blocking WARN — UNLESS the span is a
 *     live-format credential the doc-example heuristic can't excuse. No nonce,
 *     no trust exemption (the marker scheme was dropped as theater).
 */

import {
  PATTERNS,
  PATTERNS_BY_ID,
  isPlaceholderSpan,
  type RedactPattern,
  type Tier,
  type Category,
} from "./redact-patterns";

export type RepoVisibility = "public" | "private" | "unknown";

/** A WARN is a finding that does not block but is surfaced (tool-fence degrade). */
export type Severity = Tier | "WARN";

export interface Finding {
  id: string;
  tier: Tier;
  /** Effective severity after tool-fence degrade. HIGH/MEDIUM/LOW or WARN. */
  severity: Severity;
  category: Category;
  description: string;
  /** 1-based line in the ORIGINAL (un-normalized) text. */
  line: number;
  /** 1-based column in the ORIGINAL text. */
  col: number;
  /** Safe-masked preview (never more than 4 leading chars of the secret). */
  preview: string;
  /** Whether this finding offers one-keystroke auto-redact (PII subset). */
  autoRedactable: boolean;
  /** Repo visibility at scan time — drives sterner AUQ wording, not the tier. */
  repoVisibility: RepoVisibility;
  /** True when degraded to WARN because it sat in a tool-attributed fence. */
  toolFenceDegraded?: boolean;
}

export interface ScanOptions {
  repoVisibility?: RepoVisibility;
  /** Extra allowlist entries (exact strings) that suppress a matched span. */
  allowlist?: string[];
  /** The invoking user's own email (from `git config user.email`) — allowlisted. */
  selfEmail?: string;
  /**
   * Emails already public in the repo (git log authors, package.json, CODEOWNERS).
   * Suppressed for `pii.email` since they're not a new leak.
   */
  repoPublicEmails?: string[];
  /** Hard byte cap. Oversize input fails CLOSED. Default 1 MiB. */
  maxBytes?: number;
  /**
   * Path of the file the input came from, when known (the pre-push hook's
   * `+++ b/` path, `--from-file`). Lets a pattern tell source syntax from a
   * literal: in a TypeScript/JSX file `session: SessionState,` is a type.
   */
  sourcePath?: string;
  /**
   * The environment whose credential-named variables (a name whose last
   * `_`-separated word is KEY, TOKEN, SECRET, PASSWORD, PASSWD, PASS, PAT or
   * APIKEY, or that ends in _PWD, any case, carrying no PUBLIC word; see
   * `credentialEnvValues`) are scanned for BY VALUE: wherever one of those
   * values appears in the text it is an `env.value` HIGH finding, whatever its
   * shape, so a value that reached the text (a dumped environment in a tool
   * output, GSTA-96) never reaches a PR body, receipt, report or Drive file.
   * Defaults to `process.env`; pass `{}` to scan without one.
   */
  processEnv?: Record<string, string | undefined>;
}

export interface ScanResult {
  findings: Finding[];
  counts: { HIGH: number; MEDIUM: number; LOW: number; WARN: number };
  repoVisibility: RepoVisibility;
  /** True when the input-size cap tripped (caller should BLOCK). */
  oversize: boolean;
}

const DEFAULT_MAX_BYTES = 1024 * 1024; // 1 MiB

const EMAIL_ALLOW_DOMAINS = [/@example\.(com|org|net)$/i, /@example\.[a-z]{2,}$/i];
const EMAIL_ALLOW_LOCALPARTS = [/^noreply@/i, /^no-reply@/i, /^donotreply@/i];
/**
 * Hosts whose `git@<host>` is a git transport endpoint, never a person's
 * mailbox. Matched EXACTLY — a domain that merely starts with "git"
 * (gitmail.com) is a normal domain and must keep firing.
 */
const SSH_GIT_HOSTS = new Set([
  "github.com",
  "gitlab.com",
  "bitbucket.org",
  "ssh.dev.azure.com",
]);

// ── Normalization ─────────────────────────────────────────────────────────────

const ZERO_WIDTH = /[​‌‍⁠]/g;
const HTML_ENTITIES: Record<string, string> = {
  "&amp;": "&",
  "&lt;": "<",
  "&gt;": ">",
  "&quot;": '"',
  "&#39;": "'",
  "&apos;": "'",
};

/**
 * Normalize text for matching while producing an index map back to the original.
 * Returns the normalized string and a function mapping a normalized offset to
 * the corresponding original offset.
 *
 * Strategy: walk the original char-by-char, applying NFKC per char, dropping
 * zero-width chars, and expanding a small fixed set of HTML entities. Each
 * emitted normalized char records the original offset it came from. This keeps
 * the map exact for the transformations we apply (which are all local).
 */
export function normalizeWithMap(input: string): {
  normalized: string;
  map: number[];
} {
  return normalizeOriginal(input);
}

function normalizeOriginal(input: string, spanEnds?: number[]): { normalized: string; map: number[] } {
  const out: string[] = [];
  const map: number[] = [];
  let i = 0;
  while (i < input.length) {
    // HTML entity expansion (fixed small set; longest first).
    let matchedEntity = false;
    for (const ent in HTML_ENTITIES) {
      if (input.startsWith(ent, i)) {
        const rep = HTML_ENTITIES[ent];
        for (const ch of rep) {
          out.push(ch);
          map.push(i);
          spanEnds?.push(i + ent.length);
        }
        i += ent.length;
        matchedEntity = true;
        break;
      }
    }
    if (matchedEntity) continue;

    const ch = input[i];
    if (ZERO_WIDTH.test(ch)) {
      ZERO_WIDTH.lastIndex = 0;
      i += 1;
      continue;
    }
    ZERO_WIDTH.lastIndex = 0;

    const norm = ch.normalize("NFKC");
    for (let j = 0; j < norm.length; j++) {
      out.push(norm[j]);
      map.push(i);
      spanEnds?.push(i + 1);
    }
    i += 1;
  }
  // Sentinel so an offset == length maps to the original length.
  map.push(input.length);
  return { normalized: out.join(""), map };
}

// ── Offset → line/col on the ORIGINAL text ────────────────────────────────────

/** Start offset of every line, built once per scan and only when a finding needs it. */
function lineStarts(original: string): number[] {
  const starts = [0];
  for (let i = 0; i < original.length; i++) if (original[i] === "\n") starts.push(i + 1);
  return starts;
}

/**
 * Binary search over lineStarts: O(log lines) per finding. The previous walk
 * from offset 0 per finding made a match-dense input (a pasted log full of
 * emails and IPs) cost O(findings x bytes) — seconds for a few hundred KiB.
 */
function lineColAt(starts: number[], original: string, offset: number): { line: number; col: number } {
  const at = Math.min(Math.max(0, offset), original.length);
  let lo = 0;
  let hi = starts.length - 1;
  while (lo < hi) {
    const mid = (lo + hi + 1) >> 1;
    if (starts[mid] <= at) lo = mid;
    else hi = mid - 1;
  }
  return { line: lo + 1, col: at - starts[lo] + 1 };
}

// ── Safe preview masking ──────────────────────────────────────────────────────

/** Show ≤4 leading chars, mask the rest. Never reconstructable. */
export function maskPreview(span: string): string {
  const visible = span.slice(0, 4);
  const masked = span.length > 4 ? "*".repeat(Math.min(span.length - 4, 8)) : "";
  return `${visible}${masked}${span.length > 12 ? "…" : ""}`;
}

// ── Tool-attributed fence detection ───────────────────────────────────────────

const TOOL_FENCE_INFO = /^```(codex-review|greptile|eval|codex|tool-output)\b/;

/**
 * Returns a sorted list of [start, end) offset ranges (in normalized text) that
 * sit inside a tool-attributed fenced code block. Credential findings inside
 * these ranges degrade to WARN (unless the doc-example heuristic says the span
 * is live-format and must still block).
 */
function toolFenceRanges(normalized: string): Array<[number, number]> {
  const ranges: Array<[number, number]> = [];
  const lines = normalized.split("\n");
  let offset = 0;
  let inFence = false;
  let fenceStart = 0;
  for (const ln of lines) {
    const isFenceMarker = ln.startsWith("```");
    if (isFenceMarker) {
      if (!inFence && TOOL_FENCE_INFO.test(ln)) {
        inFence = true;
        fenceStart = offset + ln.length + 1; // content starts after this line
      } else if (inFence) {
        ranges.push([fenceStart, offset]); // up to start of closing fence
        inFence = false;
      }
    }
    offset += ln.length + 1; // +1 for the \n
  }
  if (inFence) ranges.push([fenceStart, normalized.length]); // unterminated → still degrade its own body
  return ranges;
}

function inRanges(offset: number, ranges: Array<[number, number]>): boolean {
  for (const [s, e] of ranges) if (offset >= s && offset < e) return true;
  return false;
}

/**
 * Doc-example heuristic: a credential span inside a tool fence still BLOCKS if
 * it looks like a LIVE credential (not an obvious placeholder/example). We only
 * downgrade-to-WARN spans that are clearly illustrative.
 */
function isObviousDocExample(span: string): boolean {
  return isPlaceholderSpan(span);
}

// ── Proximity check ───────────────────────────────────────────────────────────

function hasNear(
  normalized: string,
  matchStart: number,
  matchEnd: number,
  nearRegex: RegExp,
  window: number,
): boolean {
  const from = Math.max(0, matchStart - window);
  const to = Math.min(normalized.length, matchEnd + window);
  const slice = normalized.slice(from, to);
  const re = new RegExp(nearRegex.source, nearRegex.flags.replace(/g/g, ""));
  return re.test(slice);
}

// ── Email allowlist ───────────────────────────────────────────────────────────

/**
 * True when the matched "email" is really the user@host of a git SSH remote.
 *
 * `pii.email` matches the `git@github.com` inside
 * `git@github.com:org/repo.git` — a transport identity, not PII. This keys on
 * the surrounding URL SHAPE, not on the `git` local part: allowlisting `git@`
 * outright would also suppress a genuine address at a domain that merely
 * begins with "git" (e.g. git@gitmail.com), turning a false positive into a
 * false negative.
 *
 * Two accepted shapes:
 *   - scp-like `<user>@<host>:<path>` where the path ends in `.git`, for any
 *     host — this covers self-hosted remotes.
 *   - `git@<known-host>` for the major forges, whose bare form appears in docs
 *     and in `ssh -T git@github.com` connectivity checks with no path at all.
 */
/** Lookahead window for the scp-path suffix check — long enough for any real
 *  remote path, bounded so a pathological unbroken line can't grow the scan. */
const SSH_REMOTE_PATH_LOOKAHEAD_CHARS = 512;

function isSshGitRemote(email: string, text: string, spanStart: number): boolean {
  const at = email.lastIndexOf("@");
  if (at < 0) return false;
  const local = email.slice(0, at).toLowerCase();
  const host = email.slice(at + 1).toLowerCase();

  // Major forges: `git@host`, with or without a trailing path.
  if (local === "git" && SSH_GIT_HOSTS.has(host)) return true;

  // Any host in `<user>@<host>:<path>.git` position. The path stops at
  // whitespace or a quote so a trailing delimiter never defeats the suffix.
  const rest = text.slice(
    spanStart + email.length,
    spanStart + email.length + SSH_REMOTE_PATH_LOOKAHEAD_CHARS,
  );
  const scp = /^:(?!\/)([^\s'"`<>]*)/.exec(rest);
  if (scp && /\.git\/?$/.test(scp[1])) return true;

  // ssh:// URL form: ssh://<user>@<host>/<path>.git
  const before = text.slice(Math.max(0, spanStart - 16), spanStart);
  if (/(?:git\+)?ssh:\/\/$/i.test(before)) {
    const slash = /^\/([^\s'"`<>]*)/.exec(rest);
    if (slash && /\.git\/?$/.test(slash[1])) return true;
  }

  return false;
}

function emailAllowed(
  email: string,
  opts: ScanOptions,
  text: string,
  spanStart: number,
): boolean {
  const lower = email.toLowerCase();
  if (opts.selfEmail && lower === opts.selfEmail.toLowerCase()) return true;
  if (opts.repoPublicEmails?.some((e) => e.toLowerCase() === lower)) return true;
  if (EMAIL_ALLOW_DOMAINS.some((re) => re.test(email))) return true;
  if (EMAIL_ALLOW_LOCALPARTS.some((re) => re.test(email))) return true;
  if (isSshGitRemote(email, text, spanStart)) return true;
  return false;
}

// ── The scan ──────────────────────────────────────────────────────────────────

export function scan(input: string, opts: ScanOptions = {}): ScanResult {
  return scanInternal(input, opts);
}

type OriginalSpan = { start: number; end: number };

function scanInternal(input: string, opts: ScanOptions, spans?: Map<Finding, OriginalSpan>): ScanResult {
  const repoVisibility: RepoVisibility = opts.repoVisibility ?? "unknown";
  let starts: number[] | null = null; // line index, built on the first finding
  // #1824: ?? only catches null/undefined, not NaN or <= 0. A bad value
  // (NaN from a malformed --max-bytes, or a negative) would make `byteLen >
  // maxBytes` always false and silently disable the fail-closed oversize guard.
  // Guardrail: any non-finite or non-positive value falls back to the default
  // cap. The CLI is the layer that rejects bad args; this is belt-and-suspenders
  // so the engine never silently runs uncapped.
  const maxBytes =
    Number.isFinite(opts.maxBytes) && (opts.maxBytes as number) > 0
      ? (opts.maxBytes as number)
      : DEFAULT_MAX_BYTES;

  // Fail CLOSED on oversize input. Check byte length BEFORE heavy work.
  const byteLen = Buffer.byteLength(input, "utf8");
  if (byteLen > maxBytes) {
    const finding: Finding = {
      id: "engine.input_too_large",
      tier: "HIGH",
      severity: "HIGH",
      category: "secret",
      description: `Input too large to scan safely (${byteLen} > ${maxBytes} bytes) — blocking fail-closed`,
      line: 1,
      col: 1,
      preview: "",
      autoRedactable: false,
      repoVisibility,
    };
    return {
      findings: [finding],
      counts: { HIGH: 1, MEDIUM: 0, LOW: 0, WARN: 0 },
      repoVisibility,
      oversize: true,
    };
  }

  const spanEnds: number[] | undefined = spans ? [] : undefined;
  const { normalized, map } = normalizeOriginal(input, spanEnds);
  const fenceRanges = toolFenceRanges(normalized);
  const allow = new Set(opts.allowlist ?? []);

  const findings: Finding[] = [];
  // Dedup by (id, original-offset) so overlapping global matches don't double-count.
  const seen = new Set<string>();

  for (const pat of PATTERNS) {
    const re = new RegExp(pat.regex.source, withFlags(pat.regex.flags));
    let m: RegExpExecArray | null;
    while ((m = re.exec(normalized)) !== null) {
      // Guard against zero-width matches looping forever.
      if (m.index === re.lastIndex) re.lastIndex++;

      const span = m[1] ?? m[0];
      const normOffset = m.indices?.[1]?.[0] ?? m.index;

      // Per-span placeholder suppression.
      if (isPlaceholderSpan(span)) continue;
      if (allow.has(span)) continue;

      // Pattern-specific validators (Luhn, entropy, RFC1918, etc).
      if (pat.validate && !pat.validate(span, m, opts)) continue;

      // Proximity requirement.
      if (
        pat.nearRegex &&
        !hasNear(normalized, m.index, m.index + m[0].length, pat.nearRegex, pat.nearWindow ?? 100)
      ) {
        continue;
      }

      // Email allowlist (layered on top of the pattern).
      if (pat.id === "pii.email" && emailAllowed(span, opts, normalized, normOffset))
        continue;

      const origOffset = map[Math.min(normOffset, map.length - 1)] ?? 0;
      const key = `${pat.id}:${origOffset}`;
      if (seen.has(key)) continue;
      seen.add(key);

      starts ??= lineStarts(input);
      const { line, col } = lineColAt(starts, input, origOffset);

      // Tool-fence degrade: only credential-category, only obvious doc examples.
      let severity: Severity = pat.tier;
      let toolFenceDegraded = false;
      if (
        pat.category === "secret" &&
        inRanges(normOffset, fenceRanges) &&
        isObviousDocExample(span)
      ) {
        severity = "WARN";
        toolFenceDegraded = true;
      }

      const finding: Finding = {
        id: pat.id,
        tier: pat.tier,
        severity,
        category: pat.category,
        description: pat.description,
        line,
        col,
        preview: maskPreview(span),
        autoRedactable: !!pat.autoRedactable,
        repoVisibility,
        ...(toolFenceDegraded ? { toolFenceDegraded } : {}),
      };
      findings.push(finding);
      if (spans) {
        const end = spanEnds?.[normOffset + span.length - 1];
        if (end !== undefined && end > origOffset) spans.set(finding, { start: origOffset, end });
      }
    }
  }

  // Environment values, by exact value: a credential that is already in the
  // process environment needs no shape to be recognized.
  for (const { name, value: raw } of credentialEnvValues(opts.processEnv ?? process.env)) {
    // The text was normalized before matching; the value takes the same form,
    // and the allowlist and placeholder rules see both forms.
    const value = normalizeOriginal(raw).normalized;
    if (allow.has(raw) || allow.has(value) || value.length < MIN_ENV_VALUE_LENGTH || isPlaceholderSpan(value)) continue;
    let from = 0;
    let normOffset: number;
    while ((normOffset = normalized.indexOf(value, from)) !== -1) {
      from = normOffset + value.length;
      const origOffset = map[Math.min(normOffset, map.length - 1)] ?? 0;
      // Keyed on the span, not the offset alone: when one credential's value is
      // a prefix of another's, both findings exist and their spans coalesce.
      const key = `env.value:${origOffset}:${value.length}`;
      if (seen.has(key)) continue;
      seen.add(key);
      starts ??= lineStarts(input);
      const { line, col } = lineColAt(starts, input, origOffset);
      const finding: Finding = {
        id: "env.value",
        tier: "HIGH",
        severity: "HIGH",
        category: "secret",
        description: `Value of the ${name} environment variable`,
        line,
        col,
        // A live credential: the name already says what leaked; show no bytes.
        preview: "********",
        autoRedactable: false,
        repoVisibility,
      };
      findings.push(finding);
      if (spans) {
        const end = spanEnds?.[normOffset + value.length - 1];
        if (end !== undefined && end > origOffset) spans.set(finding, { start: origOffset, end });
      }
    }
  }

  // Stable order: by line, then col, then id.
  findings.sort((a, b) => a.line - b.line || a.col - b.col || a.id.localeCompare(b.id));

  const counts = { HIGH: 0, MEDIUM: 0, LOW: 0, WARN: 0 };
  for (const f of findings) counts[f.severity] += 1;

  return { findings, counts, repoVisibility, oversize: false };
}

const CREDENTIAL_ENV_NAME = /(^|_)(KEY|TOKEN|SECRET|PASSWORD|PASSWD|PASS|PAT|APIKEY)$|_PWD$/i;
/** Framework-public keys (NEXT_PUBLIC_*_KEY, VITE_PUBLIC_*, *_PUBLIC_KEY) are committed on purpose. */
const PUBLIC_ENV_NAME = /(^|_)PUBLIC(_|$)/i;
/** gstack's own credential-named variables whose values are not secrets (bin/gstack-brain-lock.sh sets the lock token). */
export const NON_SECRET_ENV_NAMES: ReadonlySet<string> = new Set(["GSTACK_BRAIN_LOCK_TOKEN"]);
const MIN_ENV_VALUE_LENGTH = 8;

/**
 * The (name, value) pairs of `env` worth scanning for by value: credential-named
 * variables (the last `_`-separated word is KEY, TOKEN, SECRET, PASSWORD, PASSWD,
 * PASS, PAT or APIKEY, or the name ends in _PWD; names carrying a PUBLIC word and
 * gstack's own non-secret tokens are exempt) whose trimmed value is long enough to
 * be a credential and is not a placeholder. A short or placeholder value would
 * match ordinary prose. There is no path exemption on purpose: a shape rule
 * exempted base64 secrets that start with `/`, and an existence check let the
 * scanned agent exempt a value by creating a path of that name; a path-valued
 * SSL_KEY that must appear in a text goes through `allowlist` where a caller
 * passes one (`gstack-redact`, the pre-push hook), else the variable is set
 * empty for that one command. The value is
 * trimmed, so a key read from a file with a trailing newline still matches its
 * leaked form; it is matched against the normalized text in its own normalized
 * form, so a value NFKC rewrites still finds its leaked bytes.
 */
export function credentialEnvValues(env: Record<string, string | undefined>): Array<{ name: string; value: string }> {
  const out: Array<{ name: string; value: string }> = [];
  for (const [name, raw] of Object.entries(env)) {
    if (!CREDENTIAL_ENV_NAME.test(name) || PUBLIC_ENV_NAME.test(name) || NON_SECRET_ENV_NAMES.has(name)) continue;
    const value = raw?.trim() ?? "";
    if (value.length < MIN_ENV_VALUE_LENGTH) continue;
    if (isPlaceholderSpan(value)) continue;
    out.push({ name, value });
  }
  return out;
}

function withFlags(flags: string): string {
  let f = flags;
  if (!f.includes("g")) f += "g";
  if (!f.includes("m")) f += "m";
  if (!f.includes("d")) f += "d";
  return f;
}

// ── Auto-redaction ────────────────────────────────────────────────────────────

export interface RedactResult {
  body: string;
  /** ASCII unified-diff preview of the substitutions. */
  diff: string;
  /** Findings that could NOT be auto-redacted (structural-corruption guard). */
  skipped: Finding[];
}

/**
 * Substitute redact tokens for the given finding ids, right-to-left so offsets
 * stay valid. Refuses to redact a span that sits inside a structural token
 * (markdown link target, JSON string value) — those fall back to `skipped` so
 * the skill drops the user to manual edit rather than silently mangling output.
 */
export function applyRedactions(
  input: string,
  findingIds: string[],
  opts: ScanOptions = {},
): RedactResult {
  const ids = new Set(findingIds);
  const spans = new Map<Finding, OriginalSpan>();
  const { findings } = scanInternal(input, opts, spans);
  const targets = findings
    .filter((f) => ids.has(f.id) && f.autoRedactable)
    .map((f) => ({ f, ...(spans.get(f) ?? { start: -1, end: -1 }) }))
    .filter((t) => t.start >= 0);

  // Right-to-left so earlier offsets remain valid after splicing.
  targets.sort((a, b) => b.start - a.start);

  const skipped: Finding[] = [];
  const diffLines: string[] = [];
  let body = input;

  for (const t of targets) {
    const pat = PATTERNS_BY_ID[t.f.id];
    const token = pat?.redactToken ?? "<REDACTED>";
    if (inStructuralToken(body, t.start, t.end)) {
      skipped.push(t.f);
      continue;
    }
    const before = lineContaining(body, t.start);
    body = body.slice(0, t.start) + token + body.slice(t.end);
    const after = lineContaining(body, t.start);
    diffLines.push(`- ${before}`);
    diffLines.push(`+ ${after}`);
  }

  return { body, diff: diffLines.reverse().join("\n"), skipped };
}

/**
 * Patterns whose regex captures only a MARKER, not the secret payload itself
 * (the PEM header line; the GCP JSON key prefix). Span replacement on these
 * would redact the header and forward the key body — so redactFindingSpans
 * drops the whole payload instead.
 */
const MARKER_ONLY_PATTERN_IDS = new Set(["pem.private_key", "gcp.service_account"]);

/**
 * Replace EVERY finding's span with `<REDACTED-{id}>`, regardless of tier or
 * autoRedactable. For machine egress surfaces (telemetry error_message,
 * #1947) where structure preservation doesn't matter and fail-closed beats
 * fidelity. Returns null — caller must drop the whole payload — when:
 *   - any finding's span cannot be located, or
 *   - any finding matched a marker-only pattern (PEM / GCP service-account
 *     JSON): their regexes capture the header, not the key material, so a
 *     span splice would leak the body that follows the marker.
 * Overlapping spans (e.g. a Bearer token that is also a JWT) are coalesced
 * before splicing so stale offsets never leave partial secret bytes behind.
 * (Contrast applyRedactions, which is the interactive, autoRedactable-only,
 * structure-preserving path.)
 */
export function redactFindingSpans(input: string, opts: ScanOptions = {}): string | null {
  const merged = mergedFindingSpans(input, opts);
  return merged && spliceEdits(input, merged.map((m) => ({ start: m.start, end: m.end, text: marker(m.ids) })));
}

/** One replacement on the original text: [start, end) becomes `text`. */
export interface RedactionEdit {
  start: number;
  end: number;
  text: string;
}

/**
 * redactFindingSpans for a stream declared as JSON (a scanner report: one
 * document, several top-level values, or NDJSON), keeping it parseable. A span
 * inside a string keeps the in-string marker, widened to whole escape
 * sequences; a span inside a number replaces that whole number token with the
 * quoted marker, so `{"max_memory_bytes":1145427776}` stays valid JSON
 * (#3185). A span that touches more than one token, a structural character,
 * whitespace, a literal or a string's quotes returns null (withhold). A
 * stream that does not tokenize as JSON gets the plain splice. Returns the
 * edits (sorted, disjoint, on the original offsets) so a caller can map
 * interleaved runs onto the result.
 */
export function redactJsonFindingSpans(input: string, opts: ScanOptions = {}): { body: string; edits: RedactionEdit[] } | null {
  const merged = mergedFindingSpans(input, opts);
  if (!merged) return null;
  const tokens = jsonTokens(input);
  const spans: Array<{ start: number; end: number; ids: string[]; quote: boolean }> = [];
  for (const m of merged) {
    if (!tokens) {
      spans.push({ ...m, quote: false });
      continue;
    }
    let lo = 0, hi = tokens.length;
    while (lo < hi) {
      const mid = (lo + hi) >> 1;
      if (tokens[mid].end <= m.start) lo = mid + 1; else hi = mid;
    }
    const t = tokens[lo];
    if (!t || t.start > m.start || t.end < m.end) return null;
    let span: { start: number; end: number; quote: boolean };
    if (t.kind === "number") span = { start: t.start, end: t.end, quote: true };
    else if (t.kind === "string" && m.start > t.start && m.end < t.end) {
      span = { start: m.start, end: m.end, quote: false };
      for (const [s, e] of t.escapes) {
        if (s < span.start && span.start < e) span.start = s;
        if (s < span.end && span.end < e) span.end = e;
      }
    } else return null;
    const last = spans[spans.length - 1];
    if (last && span.start < last.end) {
      last.end = Math.max(last.end, span.end);
      for (const id of m.ids) if (!last.ids.includes(id)) last.ids.push(id);
    } else spans.push({ ...span, ids: [...m.ids] });
  }
  const edits = spans.map((x) => ({ start: x.start, end: x.end, text: x.quote ? `"${marker(x.ids)}"` : marker(x.ids) }));
  return { body: spliceEdits(input, edits), edits };
}

const marker = (ids: string[]) => `<REDACTED-${ids.join("+")}>`;

function spliceEdits(input: string, edits: RedactionEdit[]): string {
  // Right-to-left so earlier offsets remain valid after splicing.
  let body = input;
  for (let i = edits.length - 1; i >= 0; i--) body = body.slice(0, edits[i].start) + edits[i].text + body.slice(edits[i].end);
  return body;
}

/** Every finding's span, coalesced; null when one cannot be located or is marker-only. */
function mergedFindingSpans(input: string, opts: ScanOptions): Array<{ start: number; end: number; ids: string[] }> | null {
  const spans = new Map<Finding, OriginalSpan>();
  const { findings } = scanInternal(input, opts, spans);
  if (findings.some((f) => MARKER_ONLY_PATTERN_IDS.has(f.id))) return null;
  const targets = findings.map((f) => ({ f, ...(spans.get(f) ?? { start: -1, end: -1 }) }));
  if (targets.some((t) => t.start < 0)) return null;

  // Coalesce overlapping/touching ranges — splicing two intersecting spans
  // independently applies a stale end offset to already-modified text and
  // can leave trailing secret bytes in place.
  targets.sort((a, b) => a.start - b.start);
  const merged: Array<{ start: number; end: number; ids: string[] }> = [];
  for (const t of targets) {
    const last = merged[merged.length - 1];
    if (last && t.start <= last.end) {
      last.end = Math.max(last.end, t.end);
      if (!last.ids.includes(t.f.id)) last.ids.push(t.f.id);
    } else {
      merged.push({ start: t.start, end: t.end, ids: [t.f.id] });
    }
  }
  return merged;
}

type JsonToken = { kind: "string" | "number" | "other"; start: number; end: number; escapes: Array<[number, number]> };
const JSON_NUMBER = /-?\d+(?:\.\d+)?(?:[eE][+-]?\d+)?/y;
const JSON_LITERAL = /true|false|null/y;

/** Lexical JSON tokens of one or more top-level values (NDJSON included); null when the text is not JSON. */
function jsonTokens(text: string): JsonToken[] | null {
  const tokens: JsonToken[] = [];
  for (let i = 0; i < text.length;) {
    const c = text[i];
    if (c === " " || c === "\t" || c === "\n" || c === "\r") { i++; continue; }
    if ("{}[]:,".includes(c)) { tokens.push({ kind: "other", start: i, end: i + 1, escapes: [] }); i++; continue; }
    if (c === '"') {
      const escapes: Array<[number, number]> = [];
      let j = i + 1;
      for (;;) {
        if (j >= text.length) return null;
        if (text[j] === '"') break;
        if (text[j] === "\\") {
          const end = j + (text[j + 1] === "u" ? 6 : 2);
          if (end > text.length) return null;
          escapes.push([j, end]);
          j = end;
        } else j++;
      }
      tokens.push({ kind: "string", start: i, end: j + 1, escapes });
      i = j + 1;
      continue;
    }
    let matched = false;
    for (const [kind, pattern] of [["number", JSON_NUMBER], ["other", JSON_LITERAL]] as const) {
      pattern.lastIndex = i;
      if (!pattern.test(text)) continue;
      tokens.push({ kind, start: i, end: pattern.lastIndex, escapes: [] });
      i = pattern.lastIndex;
      matched = true;
      break;
    }
    if (!matched) return null;
  }
  return tokens;
}

function inStructuralToken(body: string, start: number, end: number): boolean {
  // Markdown link target: [text](...span...). The span may sit anywhere inside
  // the parenthesized target (e.g. an email embedded in a URL). Walk backward
  // from the span: if we reach `](` before hitting `)`/whitespace, and forward
  // we reach `)` before whitespace, the span is inside a link target.
  for (let i = start - 1; i >= 0; i--) {
    const ch = body[i];
    if (ch === ")" || ch === "\n" || ch === " " || ch === "\t") break;
    if (ch === "(" && i > 0 && body[i - 1] === "]") {
      for (let j = end; j < body.length; j++) {
        const c = body[j];
        if (c === " " || c === "\t" || c === "\n") break;
        if (c === ")") return true;
      }
      break;
    }
  }
  // JSON string value: "key": "...span..."  — span is inside a quoted value.
  const before = body.slice(Math.max(0, start - 80), start);
  const after = body.slice(end, Math.min(body.length, end + 4));
  if (/:\s*"$/.test(before) && /^"/.test(after)) return true;
  return false;
}

function lineContaining(body: string, offset: number): string {
  const start = body.lastIndexOf("\n", offset - 1) + 1;
  let end = body.indexOf("\n", offset);
  if (end === -1) end = body.length;
  return body.slice(start, end);
}

// ── Exit-code helper for the CLI shim ─────────────────────────────────────────

/** 0 clean, 2 MEDIUM present (no HIGH), 3 HIGH present. WARN does not gate. */
export function exitCodeFor(result: ScanResult): 0 | 2 | 3 {
  if (result.counts.HIGH > 0) return 3;
  if (result.counts.MEDIUM > 0) return 2;
  return 0;
}