package bundle

import (
	"fmt"
	"path/filepath"
	"sort"

	"huggingface.co/turenlabs/Vigil/source/pkg/types"
)

// BundleResult is the raw cross-file analysis output for a bundle: every
// per-file/cross-file Finding plus the aggregated BundleSignals derived from
// them. It is independent of the SKILL.md scorer Verdict.
type BundleResult struct {
	Findings []Finding     `json:"findings"`
	Signals  BundleSignals `json:"signals"`
}

// FileVerdict is the per-sibling view surfaced in bundle output: which file,
// its sniffed kind, whether SKILL.md referenced it, whether it is hidden, and
// the findings raised against it.
type FileVerdict struct {
	File       string    `json:"file"`
	Kind       string    `json:"kind"`
	Referenced bool      `json:"referenced"`
	Hidden     bool      `json:"hidden"`
	Findings   []Finding `json:"findings,omitempty"`
}

// BundleVerdict is the authoritative result for a whole skill directory. The
// Verdict field is the unchanged SKILL.md sub-verdict produced by the existing
// scorer; Label is the AUTHORITATIVE final bundle label, which is the stronger
// of the scorer sub-verdict and the precision-aware bundleEscalates decision.
type BundleVerdict struct {
	Dir         string         `json:"dir"`
	SkillMd     string         `json:"skill_md"`
	Verdict     *types.Verdict `json:"verdict"`
	Files       []FileVerdict  `json:"files"`
	Bundle      BundleResult   `json:"bundle"`
	Label       string         `json:"label"`
	EscalatedBy string         `json:"escalated_by,omitempty"`

	// Capability/risk surfacing layer (host-agnostic). These AUGMENT the Label:
	// the label is still derived as malicious iff RiskTier==REVIEW, and REVIEW is
	// pinned to the exact (scorerMalicious || bundleEscalates) decision that sets
	// the label today — so the label is byte-identical to the pre-capability era.
	Capabilities []Capability `json:"capabilities,omitempty"`
	BigNasties   []BigNasty   `json:"big_nasties,omitempty"`
	RiskTier     string       `json:"risk_tier"`
}

// Analyze runs the full analyzer pipeline over a bundle and aggregates the
// resulting cross-file signals.
//
// Dispatch is two-phase by design. Phase 1 runs every sibling File through the
// file-content analyzers (Shell, Python, ScriptOther, Data, Pyc, Archive,
// Binary, Image). Phase 2 then runs the SKILL.md-level analyzers
// (IndirectionAnalyzer, NLDirectiveAnalyzer), which must see the full set of
// sibling findings already collected so they can decide whether a structural
// "delegates-to-X" signal is corroborated by a behavioral finding on the
// target. AnalyzeFile handles the per-File analyzer selection via Handles().
func Analyze(b *Bundle) BundleResult {
	var findings []Finding
	if b == nil {
		return BundleResult{Signals: AggregateSignals(nil, nil)}
	}

	// Phase 1: every sibling file through its matching content analyzers.
	for _, f := range b.Files {
		if f == nil {
			continue
		}
		findings = append(findings, AnalyzeFile(f, b)...)
	}

	// Phase 2: SKILL.md-level analyzers (indirection / NL-directive). These
	// Handle KindSkillMd; the parsed SKILL.md is not part of b.Files (which is
	// siblings only), so we synthesize a File entry pointing at it. Running
	// these last means any sibling finding from phase 1 is already on b for
	// the analyzers' corroboration checks.
	if skillFile := skillMdFile(b); skillFile != nil {
		findings = append(findings, AnalyzeFile(skillFile, b)...)
	}

	findings = dedupeFindings(findings)

	return BundleResult{
		Findings: findings,
		Signals:  AggregateSignals(b, findings),
	}
}

// skillMdFile builds a synthetic *File for the bundle's SKILL.md so the
// SKILL.md-level analyzers (which Handle KindSkillMd) receive a File argument
// with the right Kind. Returns nil when the bundle has no SKILL.md.
func skillMdFile(b *Bundle) *File {
	if b == nil || b.SkillMdPath == "" {
		return nil
	}
	rel := b.SkillMdPath
	if b.Skill != nil && b.Skill.FilePath != "" {
		rel = b.Skill.FilePath
	}
	return &File{
		RelPath: filepath.Base(rel),
		AbsPath: b.SkillMdPath,
		Kind:    KindSkillMd,
	}
}

// MakeBundleVerdict combines the unchanged SKILL.md scorer Verdict with the
// cross-file analysis result into the authoritative BundleVerdict. The final
// Label is the stronger of the scorer sub-verdict and bundleEscalates: a
// SevHigh+ corroborated (or critical) sibling finding makes the whole bundle
// malicious regardless of any benign early-return inside the scorer (defeating
// SKILL.md name-spoofing). When the scorer already said malicious, that stands.
func MakeBundleVerdict(b *Bundle, skillVerdict *types.Verdict, res BundleResult) *BundleVerdict {
	bv := &BundleVerdict{
		Verdict: skillVerdict,
		Bundle:  res,
		Files:   fileVerdicts(b, res.Findings),
	}
	if b != nil {
		bv.Dir = b.Dir
		bv.SkillMd = b.SkillMdPath
	}

	scorerSaysMalicious := skillVerdict != nil && skillVerdict.Label == "malicious"
	escalates, reason := bundleEscalates(res)

	// Build the host-agnostic capability profile from the already-extracted
	// findings (pure; NO new detection). The tier PINS REVIEW to the exact
	// (scorerMalicious || escalates) decision used below, so the derived label is
	// byte-identical to today's. The combo taxonomy only ranks the non-escalating
	// remainder into ELEVATED/INFO/CLEAN; it is never a new escalation source.
	escalatedOrMalicious := scorerSaysMalicious || escalates
	prof := BuildCapabilityProfile(res.Findings)
	tier := ComputeTier(escalatedOrMalicious, prof)
	bv.Capabilities = prof.Capabilities
	bv.BigNasties = prof.BigNasties
	bv.RiskTier = string(tier)

	// Label is derived from the tier: malicious iff REVIEW. This equals the prior
	// switch (malicious iff scorerMalicious||escalates) because ComputeTier returns
	// REVIEW iff escalatedOrMalicious is true.
	bv.Label = DeriveLabel(tier)

	// EscalatedBy still names the cross-file evidence whenever the bundle
	// independently escalates, exactly as before (set even when the scorer led).
	if escalates {
		bv.EscalatedBy = reason
	}

	return bv
}

// bundleEscalates is the AUTHORITATIVE, precision-aware escalation decision for
// a bundle. It returns true (with a human-readable reason naming the offending
// signal/sibling) when the cross-file evidence is strong enough to mark the
// whole bundle malicious, independent of the SKILL.md scorer.
//
// Escalation fires when ANY of:
//   - a Finding at SevCritical or above (e.g. exfil host inside a native binary)
//   - a Finding at SevHigh that is behaviorally Corroborated OR not Structural
//     (a real behavior was observed, not merely a shape)
//   - Signals.CorroboratedHighRisk (the aggregated precision gate)
//
// A bare Structural SevHigh finding (ships-opaque-executable, or
// delegates-to-bundled-script with no corroborating finding on the target) does
// NOT escalate on its own. This is the precision lever that prevents the bundle
// scanner from becoming a false-positive cannon on skills that legitimately
// ship a .so or merely mention a filename.
func bundleEscalates(res BundleResult) (escalates bool, reason string) {
	const highWeight = 0.85 // SevHigh; single source of truth via severityToWeight

	// Prefer the strongest, most specific finding as the reported reason.
	// Evaluate criticals first, then corroborated/behavioral highs.
	var critical *Finding
	var corroboratedHigh *Finding
	var behavioralHigh *Finding

	for i := range res.Findings {
		f := &res.Findings[i]
		w := severityToWeight(f.Severity)
		if w >= severityToWeight(SevCritical) {
			if critical == nil {
				critical = f
			}
			continue
		}
		if w >= highWeight {
			if f.Corroborated && corroboratedHigh == nil {
				corroboratedHigh = f
			}
			if !f.Structural && behavioralHigh == nil {
				behavioralHigh = f
			}
		}
	}

	switch {
	case critical != nil:
		return true, findingReason(critical)
	case corroboratedHigh != nil:
		return true, findingReason(corroboratedHigh)
	case behavioralHigh != nil:
		return true, findingReason(behavioralHigh)
	case res.Signals.CorroboratedHighRisk:
		return true, "corroborated-high-risk"
	default:
		return false, ""
	}
}

// findingReason renders a stable, compact escalation reason from a finding,
// naming the originating sibling so operators can locate the payload.
func findingReason(f *Finding) string {
	if f == nil {
		return ""
	}
	if f.File != "" {
		return fmt.Sprintf("%s (%s)", f.Signal, f.File)
	}
	return f.Signal
}

// fileVerdicts groups findings by their originating sibling and emits a
// FileVerdict per file in the bundle. Files with no findings are still listed
// (so reviewers see the full inventory, including unreferenced/hidden ones).
func fileVerdicts(b *Bundle, findings []Finding) []FileVerdict {
	if b == nil {
		return nil
	}

	byFile := make(map[string][]Finding, len(findings))
	for _, f := range findings {
		byFile[f.File] = append(byFile[f.File], f)
	}

	out := make([]FileVerdict, 0, len(b.Files))
	for _, f := range b.Files {
		if f == nil {
			continue
		}
		fv := FileVerdict{
			File:       f.RelPath,
			Kind:       f.Kind.String(),
			Referenced: f.Referenced,
			Hidden:     f.Hidden,
			Findings:   byFile[f.RelPath],
		}
		delete(byFile, f.RelPath)
		out = append(out, fv)
	}

	// Findings whose File did not match any sibling RelPath (e.g. SKILL.md-level
	// findings, or archive-member paths) are surfaced under their own synthetic
	// entries so nothing is silently dropped from output.
	leftover := make([]string, 0, len(byFile))
	for file := range byFile {
		leftover = append(leftover, file)
	}
	sort.Strings(leftover)
	for _, file := range leftover {
		out = append(out, FileVerdict{
			File:     file,
			Kind:     KindUnknown.String(),
			Findings: byFile[file],
		})
	}

	return out
}