package bundle

import (
	"bytes"
	"compress/gzip"
	"compress/zlib"
	"encoding/base64"
	"encoding/hex"
	"io"
	"regexp"
	"strings"
)

// decode.go closes the docx-indirection evasion: an obfuscated payload (hex,
// base64, gzip/zlib+base64, or split-across-runs) buried in an archive member or
// data file is flagged only as an opaque "embedded-encoded-blob" by the shape
// scanners. That structural-only signal does NOT escalate under the precision
// gate, so a skill that hex-encodes `curl ... attacker.example | sh` inside
// docProps/custom.xml slips through. The fix here DECODES the blob (bounded:
// <=~1MiB recovered, <=2 nested layers, never panics) and RE-SCANS the recovered
// bytes with the shared indicator vocabulary. If the recovered content itself
// exhibits a behavioral signal (exfil host, registry rewrite to a non-allowlisted
// host, curl|sh / env-export), we emit a CORROBORATED SevHigh finding so
// aggregate.bundleEscalates fires. Mere decodability is never enough — only a
// behavioral hit in the recovered bytes corroborates (benign base64 assets and
// hashes stay SevMedium structural).

const (
	// maxDecodedBytes caps the recovered payload fed back into the scanner. Mirrors
	// the archive per-member budget so a decode bomb can never blow past it.
	maxDecodedBytes = maxMemberScanBytes // 1 MiB
	// maxDecodeDepth bounds nested decode layers. gzip+base64 is 2 layers; this is
	// the ceiling so a base64(base64(gzip(...))) chain cannot recurse unbounded.
	maxDecodeDepth = 2
	// minBlobLen is the smallest candidate blob worth decoding. Below this the
	// shared regexes would not have matched anyway; it bounds wasted work.
	minBlobLen = 16
)

// splitRunsRe extracts the concatenated payload from the docx "split across XML
// runs" obfuscation: <<PART1>>a<<PART2>>b<<PART3>>c<<END>>. The parts are joined
// (markers stripped) to reconstruct the plaintext payload. Bounded, non-greedy.
var splitRunsRe = regexp.MustCompile(`(?s)<<PART\d+>>(.*?)<<END>>`)

// partMarkerRe strips the inter-part markers so the concatenated payload reads as
// the original plaintext.
var partMarkerRe = regexp.MustCompile(`<<PART\d+>>`)

// decodeAndRescan inspects text for obfuscated-payload carriers (hex, base64,
// gzip/zlib+base64, split-runs), decodes any it finds (bounded + panic-safe), and
// re-runs sharedIndicatorScan over the recovered bytes. For each candidate whose
// RECOVERED content exhibits a behavioral signal it returns a single corroborated
// SevHigh Finding (decoded-blob-exfil / decoded-registry-rewrite /
// decoded-blob-rce). Benign-but-decodable blobs (no behavioral hit) yield nothing
// here — the caller's existing SevMedium structural "embedded-encoded-blob" stands.
//
// fileRel labels the findings; analyzer is the owning analyzer name. line is the
// 1-based line the carrier was found on (0 if unknown), surfaced on the finding.
func decodeAndRescan(text string, fileRel string, analyzer string, line int) []Finding {
	// Defense in depth: recover from any pathological decode so a single corrupt
	// blob can never crash a scan (the analyzers run under safeAnalyze too).
	defer func() { _ = recover() }()

	var out []Finding
	for _, cand := range extractEncodedCandidates(text) {
		recovered, ok := decodeBounded(cand, 0)
		if !ok || len(recovered) == 0 {
			continue
		}
		recText := string(recovered)
		// Re-scan recovered bytes with the full shared vocabulary. decode=false:
		// the bounded multi-layer decode is handled here in decode.go, so the
		// scanner must not re-trigger another decode pass (recursion guard).
		sub := sharedIndicatorScanInner(recText, fileRel, analyzer, false)
		sig, detail, hit := classifyRecovered(sub, recText)
		if !hit {
			continue
		}
		out = append(out, Finding{
			Analyzer:     analyzer,
			File:         fileRel,
			Signal:       sig,
			Severity:     SevHigh,
			Detail:       detail,
			Line:         line,
			Corroborated: true,
		})
	}
	return dedupeFindings(out)
}

// classifyRecovered inspects the findings produced by re-scanning recovered bytes
// (and the recovered text itself for registry-rewrite directives) and decides
// whether a behavioral signal is present. Returns the decoded-* signal name, a
// human detail, and whether a hit occurred.
func classifyRecovered(sub []Finding, recText string) (signal, detail string, hit bool) {
	var sawExfil, sawRegistry, sawRCE bool
	for _, f := range sub {
		switch f.Signal {
		case "exfil-host-reference":
			sawExfil = true
		case "registry-rewrite":
			// Only a non-downgraded (SevHigh+) registry rewrite corroborates; a
			// known-benign mirror is lowered to SevLow by sharedIndicatorScan.
			if f.Severity >= SevHigh {
				sawRegistry = true
			}
		case "remote-code-execution", "exfil-env-to-network", "destructive-command":
			sawRCE = true
		}
	}
	switch {
	case sawExfil:
		return "decoded-blob-exfil", "decoded an obfuscated payload that references a known exfiltration host", true
	case sawRegistry:
		return "decoded-registry-rewrite", "decoded an obfuscated payload that rewrites a package-manager registry/index to a non-allowlisted host", true
	case sawRCE:
		return "decoded-blob-rce", "decoded an obfuscated payload that pipes-to-shell or exfiltrates environment/credentials", true
	default:
		return "", "", false
	}
}

// extractEncodedCandidates pulls decode candidates out of a text block. It
// returns: each base64 blob match, each hex blob match, and (if present) the
// reassembled split-runs payload. Candidates are deduped and length-bounded.
func extractEncodedCandidates(text string) []string {
	seen := make(map[string]bool)
	var cands []string
	add := func(s string) {
		s = strings.TrimSpace(s)
		if len(s) < minBlobLen || seen[s] {
			return
		}
		seen[s] = true
		cands = append(cands, s)
	}

	// split-runs: reassemble the parts (markers removed) into the plaintext payload.
	for _, m := range splitRunsRe.FindAllStringSubmatch(text, 8) {
		if len(m) < 2 {
			continue
		}
		joined := partMarkerRe.ReplaceAllString(m[1], "")
		add(joined)
	}

	// base64 / hex blobs. Cap the number of candidates so a file packed with
	// thousands of short blobs cannot blow up the work.
	for _, m := range base64BlobRe.FindAllString(text, 64) {
		add(m)
	}
	for _, m := range longHexRe.FindAllString(text, 64) {
		add(m)
	}
	return cands
}

// decodeBounded attempts to decode a single candidate string into recovered bytes,
// trying hex, base64, and (on the decoded result) gzip/zlib decompression. depth
// bounds nested decode layers (gzip+base64 = 2). It never panics and returns
// ok=false when nothing plausibly decoded. Recovered output is capped at
// maxDecodedBytes.
func decodeBounded(s string, depth int) (recovered []byte, ok bool) {
	if depth > maxDecodeDepth {
		return nil, false
	}
	s = strings.TrimSpace(s)
	if len(s) < minBlobLen {
		return nil, false
	}

	// split-runs payloads arrive here already reassembled and may BE the plaintext
	// (the most common docx case): if the raw candidate already carries a
	// behavioral marker, surface it directly without requiring a transform.
	if depth == 0 && looksBehavioral(s) {
		return capBytes([]byte(s)), true
	}

	// Try hex first (hex alphabet is a strict subset of base64's, so a pure-hex
	// blob would also "succeed" as base64 and yield garbage — prefer hex).
	if dec, hexOk := tryHex(s); hexOk {
		if out, refined := refineDecoded(dec, depth); refined {
			return out, true
		}
	}
	// Then base64 (covers base64 and gzip/zlib+base64 via refineDecoded).
	if dec, b64Ok := tryBase64(s); b64Ok {
		if out, refined := refineDecoded(dec, depth); refined {
			return out, true
		}
	}
	return nil, false
}

// refineDecoded takes raw decoded bytes and either (a) decompresses them if they
// are gzip/zlib, recursing one decode layer deeper, or (b) returns them directly
// when they already look like a behavioral payload (or recurses on a still-encoded
// inner blob). Returns ok=true only when the final bytes look meaningful.
func refineDecoded(dec []byte, depth int) (recovered []byte, ok bool) {
	if len(dec) == 0 {
		return nil, false
	}
	// gzip / zlib magic -> decompress, then recurse one layer on the result.
	if decompressed, decOk := tryDecompress(dec); decOk {
		if looksBehavioral(string(decompressed)) {
			return capBytes(decompressed), true
		}
		// The decompressed bytes might themselves be a further-encoded blob.
		if inner, innerOk := decodeBounded(string(decompressed), depth+1); innerOk {
			return inner, true
		}
		// Decompressed to text that isn't behavioral and isn't a nested blob.
		return capBytes(decompressed), looksTexty(decompressed)
	}
	// Not compressed: accept if the decoded bytes look like a behavioral payload.
	if looksBehavioral(string(dec)) {
		return capBytes(dec), true
	}
	// Decoded to text that may carry a still-encoded inner blob (e.g. base64 of
	// base64). Recurse one layer.
	if looksTexty(dec) {
		if inner, innerOk := decodeBounded(string(dec), depth+1); innerOk {
			return inner, true
		}
		// Plain decoded text with no behavioral marker: return it so the caller's
		// re-scan can make the final SevHigh decision (keeps the decision in one place).
		return capBytes(dec), true
	}
	return nil, false
}

// looksBehavioral is a cheap pre-filter: does the text carry any marker the
// re-scan would act on? Avoids returning megabytes of benign decoded asset bytes
// for a full re-scan when there is plainly nothing actionable.
func looksBehavioral(text string) bool {
	if exfilHostRe.MatchString(text) {
		return true
	}
	low := strings.ToLower(text)
	return matchedAny(low, rceTerms) ||
		matchedAny(low, destructiveTerms) ||
		matchedAny(low, registryRewriteTerms) ||
		matchedAny(low, networkSinkTerms)
}

// tryHex decodes an optionally-0x-prefixed hex string. Requires an even length
// and a strict hex alphabet so arbitrary base64 is not mis-decoded as hex.
func tryHex(s string) ([]byte, bool) {
	s = strings.TrimPrefix(strings.TrimPrefix(s, "0x"), "0X")
	if len(s) < minBlobLen || len(s)%2 != 0 {
		return nil, false
	}
	for i := 0; i < len(s); i++ {
		c := s[i]
		if !((c >= '0' && c <= '9') || (c >= 'a' && c <= 'f') || (c >= 'A' && c <= 'F')) {
			return nil, false
		}
	}
	dec, err := hex.DecodeString(s)
	if err != nil {
		return nil, false
	}
	return capBytes(dec), true
}

// tryBase64 decodes standard or URL-safe base64 (with or without padding).
func tryBase64(s string) ([]byte, bool) {
	for _, enc := range []*base64.Encoding{
		base64.StdEncoding, base64.RawStdEncoding,
		base64.URLEncoding, base64.RawURLEncoding,
	} {
		if dec, err := enc.DecodeString(s); err == nil && len(dec) > 0 {
			return capBytes(dec), true
		}
	}
	return nil, false
}

// tryDecompress decompresses gzip or zlib data, bounded by maxDecodedBytes. It
// never panics and returns ok=false for non-compressed input.
func tryDecompress(data []byte) ([]byte, bool) {
	// gzip magic 1f 8b
	if len(data) >= 2 && data[0] == 0x1f && data[1] == 0x8b {
		zr, err := gzip.NewReader(bytes.NewReader(data))
		if err != nil {
			return nil, false
		}
		defer zr.Close()
		out, err := io.ReadAll(io.LimitReader(zr, maxDecodedBytes+1))
		if err != nil && len(out) == 0 {
			return nil, false
		}
		return capBytes(out), len(out) > 0
	}
	// zlib magic: 0x78 followed by 0x01/0x9c/0xda (common) — let the reader judge.
	if len(data) >= 2 && data[0] == 0x78 {
		zr, err := zlib.NewReader(bytes.NewReader(data))
		if err != nil {
			return nil, false
		}
		defer zr.Close()
		out, err := io.ReadAll(io.LimitReader(zr, maxDecodedBytes+1))
		if err != nil && len(out) == 0 {
			return nil, false
		}
		return capBytes(out), len(out) > 0
	}
	return nil, false
}

// looksTexty reports whether decoded bytes are plausibly text (so re-scanning /
// further-decoding them is worthwhile). Rejects mostly-binary blobs to avoid
// scanning random decoded asset bytes.
func looksTexty(data []byte) bool {
	if len(data) == 0 {
		return false
	}
	printable := 0
	n := len(data)
	if n > 4096 {
		n = 4096
	}
	for i := 0; i < n; i++ {
		c := data[i]
		if c == '\t' || c == '\n' || c == '\r' || (c >= 0x20 && c < 0x7f) {
			printable++
		}
	}
	return float64(printable)/float64(n) >= 0.85
}

// capBytes truncates recovered bytes to the decode budget.
func capBytes(data []byte) []byte {
	if len(data) > maxDecodedBytes {
		return data[:maxDecodedBytes]
	}
	return data
}