package beads

import (
	"context"
	"fmt"
	"net/url"
	"regexp"
	"sort"
	"strings"
	"time"

	"sourcecraft.dev/bigbes/sr-ht-dolt/browse"
)

// --- the memories bd remember writes ------------------------------------------

// memoryTable is where bd keeps them: the beads config table, as ordinary
// key/value rows. memoryPrefix marks the keys that are memories; the rest of the
// table is the tracker's settings (issue_prefix, compact_tier2_days, …).
const (
	memoryTable  = "config"
	memoryPrefix = "kv.memory."
)

// memoryWalkMax bounds the revision walk (see walkMemories): at most this many
// commits back from the head of the ref. The most active tracker on this
// instance had 225 commits three weeks in, so this is roughly two months of
// headroom at that rate. A memory not attributed inside the walk carries no date
// at all rather than one the walk cannot support.
const memoryWalkMax = 500

// MemoryStaleAfter is how old a memory has to be before the page questions it.
// It is one constant and not a per-request knob, and what it produces is a
// question ("stale?") rather than a verdict: some memories are meant to be
// permanent, and only the reader knows which.
const MemoryStaleAfter = 60 * 24 * time.Hour

// MemorySession is the seam the memory projection reads through. It is wider
// than BrowseSession because a memory has no timestamp — the config row is
// (key, value) and nothing else — so the date has to come out of the history:
// Log for the commits and TableHash to skip the ones that did not touch config.
// It is still only what this projection calls, and web's BrowseSession and
// mcpsrv's (and *browse.DB itself) satisfy it structurally.
type MemorySession interface {
	BrowseSession
	// Log lists commits from the head of refStr, newest first.
	Log(ctx context.Context, refStr, fromHash string, limit int) ([]browse.CommitInfo, string, error)
	// TableHash is the content hash of a table at a ref, ok=false when the table
	// does not exist there. It reads no rows, which is what makes the walk cheap.
	TableHash(ctx context.Context, refStr, table string) (string, bool, error)
}

// MemoryRevision is when a memory's value last changed: the commit that wrote
// it, taken from the history rather than from the row.
type MemoryRevision struct {
	Commit string
	Date   time.Time
	Author string
}

// Memory is one `bd remember` entry.
type Memory struct {
	Slug string // the config key with the "kv.memory." prefix stripped
	Text string // the value, with both newline spellings normalised
	// Paragraphs is Text split on blank lines. Values are typed into shell
	// strings, so they arrive with a mix of real newlines and literal "\n"
	// escapes; a page that dumped the raw value would render either as one blob.
	Paragraphs []string
	// Revision is the commit that last changed this key's value, or nil when the
	// walk could not reach it (see MemoryView.WalkTruncated).
	Revision *MemoryRevision
	// Stale reports that the memory is older than MemoryStaleAfter. For a memory
	// with no Revision it is set only when the walk's own oldest commit is
	// already past the threshold — that much is known even without a date.
	Stale bool
}

// MemoryView is the opaque .Data handed to memory.html.
type MemoryView struct {
	Memories []Memory
	Total    int    // memories in the tracker, before ?q= / ?key= narrowing
	Search   string // ?q=, sticky form state
	Key      string // ?key=<slug>, a single memory
	Sort     string // MemorySortSlug (default) | MemorySortAge
	// Query is the request's query as parsed, carried so the sort toggle can
	// rebuild this exact URL with one key replaced (web's withQuery) instead of
	// re-listing the parameters it happens to know about.
	Query url.Values
	// WalkTruncated says the revision walk stopped at WalkMax commits with the
	// history still going. It is the only way a Memory can carry no Revision, and
	// it is what the page says instead of a date.
	WalkTruncated bool
	WalkMax       int // memoryWalkMax, so the page can name the number it hit
}

// Memory sort orders. Slug is the default; Age is the review queue.
const (
	MemorySortSlug = "slug"
	MemorySortAge  = "age"
)

// AppliesMemories fingerprints a tracker that can carry memories: the beads
// fingerprint plus a config table with key and value columns. Like every
// Applies, it sees table shapes and never rows, so a tracker whose config holds
// no memory at all still gets the tab and renders an empty state.
func AppliesMemories(tables []browse.TableInfo) bool {
	if !Applies(tables) {
		return false
	}
	for _, t := range tables {
		if t.Name != memoryTable {
			continue
		}
		var haveKey, haveValue bool
		for _, c := range t.Columns {
			switch c.Name {
			case "key":
				haveKey = true
			case "value":
				haveValue = true
			}
		}
		return haveKey && haveValue
	}
	return false
}

// BuildMemories reads the memories at ref and dates each one from the history.
// now is the clock staleness is measured against, passed in rather than read
// here: this package renders nothing and reads no hidden clock, and a caller
// that pins its own clock (the web view, its tests) gets a deterministic answer.
//
// A missing config table degrades to no memories — the same treatment the other
// optional tables get — but a history that cannot be read is an error, not an
// empty answer: the date is what this view is for, and a page that silently
// dropped every date would look exactly like a tracker whose memories are all
// older than the walk.
func BuildMemories(ctx context.Context, sess MemorySession, ref string, query url.Values, now time.Time) (*MemoryView, error) {
	view := &MemoryView{
		Search:  strings.TrimSpace(query.Get("q")),
		Key:     strings.TrimSpace(query.Get("key")),
		Sort:    parseMemorySort(query.Get("sort")),
		Query:   query,
		WalkMax: memoryWalkMax,
	}

	rows, _, err := readRowsOptional(ctx, sess, ref, memoryTable)
	if err != nil {
		return nil, err
	}
	if rows == nil {
		return view, nil
	}

	// The whole config table, narrowed to the memory keys. raw keeps the stored
	// value under its full key: the walk compares values as they are stored, and
	// normalising first would make two spellings of the same text look equal.
	cols := indexCols(rows.Columns)
	raw := map[string]string{}
	texts := map[string]string{}
	for _, r := range rows.Rows {
		key := cell(cols, r, "key")
		if !strings.HasPrefix(key, memoryPrefix) {
			continue
		}
		slug := strings.TrimPrefix(key, memoryPrefix)
		if slug == "" {
			continue
		}
		raw[key] = cell(cols, r, "value")
		texts[key] = normalizeMemoryText(raw[key])
	}
	view.Total = len(raw)

	// Narrow before walking: the walk is per-key work, and a ?key= page has no
	// business dating the other eight memories.
	tracked := map[string]string{}
	for key, value := range raw {
		slug := strings.TrimPrefix(key, memoryPrefix)
		if view.Key != "" && slug != view.Key {
			continue
		}
		if !matchesMemorySearch(view.Search, slug, texts[key]) {
			continue
		}
		tracked[key] = value
	}
	if len(tracked) == 0 {
		return view, nil
	}

	walk, err := walkMemories(ctx, sess, ref, tracked)
	if err != nil {
		return nil, err
	}
	view.WalkTruncated = walk.truncated

	view.Memories = make([]Memory, 0, len(tracked))
	for key := range tracked {
		m := Memory{
			Slug:       strings.TrimPrefix(key, memoryPrefix),
			Text:       texts[key],
			Paragraphs: memoryParagraphs(texts[key]),
		}
		if rev, ok := walk.revisions[key]; ok {
			m.Revision = &rev
			m.Stale = now.Sub(rev.Date) > MemoryStaleAfter
		} else if !walk.oldest.IsZero() {
			// No date, but a floor: the memory is at least as old as the oldest
			// commit the walk examined. When that alone is past the threshold the
			// question is supportable; otherwise it is not asked.
			m.Stale = now.Sub(walk.oldest) > MemoryStaleAfter
		}
		view.Memories = append(view.Memories, m)
	}
	sortMemories(view.Memories, view.Sort)

	return view, nil
}

// memoryWalk is what walkMemories learned: the commit each key was last written
// by, whether the walk ran out of budget before the history ran out, and the
// date of the oldest commit it examined.
type memoryWalk struct {
	revisions map[string]MemoryRevision
	truncated bool
	oldest    time.Time
}

// walkMemories attributes each tracked key to the commit that last changed its
// value, walking the history of ref newest to oldest, at most memoryWalkMax
// commits.
//
// The trick that makes it affordable is the table hash. At each commit the
// content hash of config is O(1) and reads no rows; when it equals the hash at
// the newer neighbour, config is byte-identical across that step and the newer
// commit wrote no memory — skip, read nothing. Only a commit that actually
// touched config costs a row read, and then each still-unresolved key whose
// value differs from the newer neighbour's was written *by that newer commit*.
//
// The commit message is not the signal. `bd remember` does write
// "bd: remember (auto-commit) by <author>", but that is a claim by whoever wrote
// it; the table hash is the fact.
//
// A key still unresolved when the history itself runs out was present with this
// value at the root commit, so the root is what wrote it — that is a date the
// walk supports. A key unresolved because the walk hit its budget gets no date
// at all, and truncated says so.
//
// The walk reads the log's linearization (Log is reverse-topological). Beads
// histories are linear chains of auto-commits, which this is exact for; across a
// merge, attribution is to the nearest commit in that order.
func walkMemories(ctx context.Context, sess MemorySession, ref string, tracked map[string]string) (memoryWalk, error) {
	out := memoryWalk{revisions: map[string]MemoryRevision{}}

	commits, next, err := sess.Log(ctx, ref, "", memoryWalkMax)
	if err != nil {
		return memoryWalk{}, fmt.Errorf("beads: read history of %q: %w", ref, err)
	}
	if len(commits) == 0 {
		return out, nil
	}
	out.oldest = commits[len(commits)-1].Date
	out.truncated = next != ""

	// unresolved carries each key's value at the newer neighbour of the commit
	// being examined; it starts as the value at the head, which is where the
	// memories themselves were read.
	unresolved := make(map[string]string, len(tracked))
	for k, v := range tracked {
		unresolved[k] = v
	}

	newerHash, err := memoryTableHash(ctx, sess, commits[0].Hash)
	if err != nil {
		return memoryWalk{}, err
	}

	for i := 1; i < len(commits) && len(unresolved) > 0; i++ {
		hash, err := memoryTableHash(ctx, sess, commits[i].Hash)
		if err != nil {
			return memoryWalk{}, err
		}
		if hash == newerHash {
			continue // config unchanged across this step: nothing to read
		}

		older, err := memoryValuesAt(ctx, sess, commits[i].Hash)
		if err != nil {
			return memoryWalk{}, err
		}
		newer := commits[i-1]
		for key, newerValue := range unresolved {
			if older[key] == newerValue {
				continue
			}
			out.revisions[key] = MemoryRevision{
				Commit: newer.Hash,
				Date:   newer.Date,
				Author: newer.Author,
			}
			delete(unresolved, key)
		}
		for key := range unresolved {
			unresolved[key] = older[key]
		}
		newerHash = hash
	}

	if !out.truncated {
		// The history ended with these keys never changing: the oldest commit
		// walked is the root, and it carries the value we are looking at.
		root := commits[len(commits)-1]
		for key := range unresolved {
			out.revisions[key] = MemoryRevision{Commit: root.Hash, Date: root.Date, Author: root.Author}
		}
	}

	return out, nil
}

// memoryTableHash is the config table's content hash at one commit. A table that
// does not exist there is "" — an ordinary answer while walking backwards past
// the commit that created it, and one that compares correctly against another
// commit where it is equally absent.
func memoryTableHash(ctx context.Context, sess MemorySession, at string) (string, error) {
	hash, ok, err := sess.TableHash(ctx, at, memoryTable)
	if err != nil {
		return "", fmt.Errorf("beads: hash of %s at %s: %w", memoryTable, at, err)
	}
	if !ok {
		return "", nil
	}
	return hash, nil
}

// memoryValuesAt reads the config table at one commit as key → value. A missing
// table is an empty map, which is what it means here: no key had a value yet.
func memoryValuesAt(ctx context.Context, sess MemorySession, at string) (map[string]string, error) {
	rows, _, err := readRowsOptional(ctx, sess, at, memoryTable)
	if err != nil {
		return nil, err
	}
	out := map[string]string{}
	if rows == nil {
		return out, nil
	}
	cols := indexCols(rows.Columns)
	for _, r := range rows.Rows {
		if key := cell(cols, r, "key"); key != "" {
			out[key] = cell(cols, r, "value")
		}
	}
	return out, nil
}

// memoryEscapedNewline matches the two-character escapes that reach the value
// because the memory was typed into a shell string: "\r\n" and "\n" written out
// as backslash sequences rather than as newlines.
var memoryEscapedNewline = regexp.MustCompile(`\\r\\n|\\n`)

// memoryBlankLine splits a memory into paragraphs on runs of blank lines.
var memoryBlankLine = regexp.MustCompile(`\n[ \t]*\n[ \t\n]*`)

// normalizeMemoryText brings a stored value's two newline spellings together.
// The same memory arrives with real newlines when it was written from a file or
// a heredoc and with literal backslash-n when it was typed into a shell string,
// and both spellings turn up in the same tracker.
func normalizeMemoryText(v string) string {
	v = strings.ReplaceAll(v, "\r\n", "\n")
	v = strings.ReplaceAll(v, "\r", "\n")
	v = memoryEscapedNewline.ReplaceAllString(v, "\n")
	return strings.TrimSpace(v)
}

// memoryParagraphs splits normalised text on blank lines. Line breaks *inside* a
// paragraph are kept — memories are full of bullet lists, and joining those into
// a sentence is not rendering, it is losing the structure the author typed.
func memoryParagraphs(text string) []string {
	if text == "" {
		return nil
	}
	parts := memoryBlankLine.Split(text, -1)
	out := make([]string, 0, len(parts))
	for _, p := range parts {
		if strings.TrimSpace(p) == "" {
			continue
		}
		out = append(out, strings.Trim(p, "\n"))
	}
	return out
}

// matchesMemorySearch is the ?q= rule: a case-insensitive substring of the slug
// or of the text. An empty query matches everything.
func matchesMemorySearch(q, slug, text string) bool {
	if q == "" {
		return true
	}
	q = strings.ToLower(q)
	return strings.Contains(strings.ToLower(slug), q) || strings.Contains(strings.ToLower(text), q)
}

// parseMemorySort reads ?sort=, defaulting (and falling back from an unknown
// value) to slug order.
func parseMemorySort(v string) string {
	if strings.ToLower(strings.TrimSpace(v)) == MemorySortAge {
		return MemorySortAge
	}
	return MemorySortSlug
}

// sortMemories orders the list: by slug, or oldest first for the review queue.
// In age order a memory with no revision leads — it is older than anything the
// walk could date — and ties fall back to the slug, so the order is total.
func sortMemories(ms []Memory, order string) {
	sort.SliceStable(ms, func(i, j int) bool {
		a, b := ms[i], ms[j]
		if order == MemorySortAge {
			switch {
			case a.Revision == nil && b.Revision != nil:
				return true
			case a.Revision != nil && b.Revision == nil:
				return false
			case a.Revision != nil && b.Revision != nil && !a.Revision.Date.Equal(b.Revision.Date):
				return a.Revision.Date.Before(b.Revision.Date)
			}
		}
		return a.Slug < b.Slug
	})
}