X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/522d7fd117285e4e4ce3ea5ec9b5a1722521b681..d8f3785bb69288663f4cc46bf4d609c6770452d5:/memory.go?ds=inline

diff --git a/memory.go b/memory.go
index 7cbcce6..49215ac 100644
--- a/memory.go
+++ b/memory.go
@@ -5,7 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
-	"log"
+	"log/slog"
 	"os"
 	"path/filepath"
 	"strings"
@@ -22,10 +22,68 @@ var (
 	errMemoryUnsettled = errors.New("a summary of the view failed, and it is tried again at the next message")
 )
 
+// Summarizer builds the text of one node with a model.
+type Summarizer interface {
+	Summarize(ctx context.Context, task Task) (string, error)
+}
+
+// Memory is the memory of one chat: the message log, the summary tree, the chat
+// view, the compaction view, and the queue of the builds. The compaction view
+// is the chat view merged further, at the limits of 16 to 32 KB. It takes the
+// same new lines, and it merges again in one batch when it is too large or when
+// the chat view merges. So the compactions share its prefix in the prompt
+// cache. It saves the views to view.json after each change and loads them at
+// start. It never builds the views again from the log, because a new view
+// differs from the live one and so loses the prompt cache.
+//
+// Each write goes to the operating system at once, and a syncer makes the
+// writes durable by group commit in the background. Flush waits until the
+// writes so far are durable.
+//
+// The summarizer builds the nodes in the queue in the background. The start
+// reads the indexes once to fill the queue again. After that, only a new
+// message or a new node adds work. Memory keeps the nodes of the views and the
+// tasks of the running builds. It reads all other messages and nodes from their
+// day files through the indexes.
+type Memory struct {
+	mutex      sync.Mutex
+	root       string
+	log        *Log
+	syncer     *Syncer
+	tree       *Tree
+	chat       *View
+	cview      *View
+	queue      *BuildQueue
+	summarizer Summarizer
+	logger     *slog.Logger
+	ctx        context.Context
+	cancel     context.CancelFunc
+	workers    sync.WaitGroup
+	changed    chan struct{}
+}
+
+// Task is the build of the node (Level, Index). Context is the compaction view
+// for the build. Input holds the source "kind: text" of a leaf, or the two
+// child lines of a merge with their newlines as spaces. Images holds the leaf
+// image filenames and is empty for merges.
+type Task struct {
+	Level   int
+	Index   int
+	Context string
+	Input   []string
+	Images  []string
+}
+
+// MemoryViewFile is the content of view.json.
+type MemoryViewFile struct {
+	Chat       ViewState `json:"chat"`
+	Compaction ViewState `json:"compaction"`
+}
+
 // openMemory opens the chat in root and starts the builds that are ready. The
 // clock gives the dates of new messages and the days of new nodes. logger gets
 // one line for each failed build. The builds stop when ctx ends or at Close.
-func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer, logger *log.Logger) (*Memory, error) {
+func openMemory(ctx context.Context, root string, clock func() time.Time, summarizer Summarizer, logger *slog.Logger) (*Memory, error) {
 	log, err := openLog(ctx, root, clock)
 	if err != nil {
 		return nil, err
@@ -75,46 +133,6 @@ func openMemory(ctx context.Context, root string, clock func() time.Time, summar
 	return m, nil
 }
 
-// Summarizer builds the text of one node with a model.
-type Summarizer interface {
-	Summarize(ctx context.Context, task Task) (string, error)
-}
-
-// Memory is the memory of one chat: the message log, the summary tree, the chat
-// view, the compaction view, and the queue of the builds. The compaction view
-// is the chat view merged further, at the limits of 16 to 32 KB. It takes the
-// same new lines, and it merges again in one batch when it is too large or when
-// the chat view merges. So the compactions share its prefix in the prompt
-// cache. It saves the views to view.json after each change and loads them at
-// start. It never builds the views again from the log, because a new view
-// differs from the live one and so loses the prompt cache.
-//
-// Each write goes to the operating system at once, and a syncer makes the
-// writes durable by group commit in the background. Flush waits until the
-// writes so far are durable.
-//
-// The summarizer builds the nodes in the queue in the background. The start
-// reads the indexes once to fill the queue again. After that, only a new
-// message or a new node adds work. Memory keeps the nodes of the views and the
-// tasks of the running builds. It reads all other messages and nodes from their
-// day files through the indexes.
-type Memory struct {
-	mutex      sync.Mutex
-	root       string
-	log        *Log
-	syncer     *Syncer
-	tree       *Tree
-	chat       *View
-	cview      *View
-	queue      *BuildQueue
-	summarizer Summarizer
-	logger     *log.Logger
-	ctx        context.Context
-	cancel     context.CancelFunc
-	workers    sync.WaitGroup
-	changed    chan struct{}
-}
-
 // Close stops the builds, waits for them, makes the writes durable, closes the
 // files of the chat, and releases its lock.
 func (m *Memory) Close() error {
@@ -137,149 +155,6 @@ func (m *Memory) Close() error {
 	return errors.Join(err, m.log.Close())
 }
 
-// Append logs a new message, adds it to the tree and the views, and starts the
-// builds that are ready, also the failed ones.
-func (m *Memory) Append(kind, text string) (Message, error) {
-	return m.AppendDated(kind, text, "")
-}
-
-// AppendDated appends a message with its RFC 3339 date and builds its nodes. An
-// empty date uses the clock, as Append does.
-func (m *Memory) AppendDated(kind, text, date string) (Message, error) {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
-
-	message, err := m.log.AppendDated(kind, text, date)
-	if err != nil {
-		return Message{}, fmt.Errorf("append dated log message: %w", err)
-	}
-
-	defer m.syncer.Written()
-
-	err = m.addMessage(message)
-	if err != nil {
-		return Message{}, fmt.Errorf("add dated message to memory: %w", err)
-	}
-
-	err = m.save()
-	if err != nil {
-		return Message{}, fmt.Errorf("save dated message view: %w", err)
-	}
-
-	m.queue.Retry()
-	m.dispatch()
-
-	return message, nil
-}
-
-// children reads the two children of the node key.
-func (m *Memory) children(key NodeKey) (Node, Node, error) {
-	var children [2]Node
-
-	for n := range children {
-		child, built, err := m.tree.Node(key.level-1, 2*key.index+n)
-		if err != nil {
-			return Node{}, Node{}, err
-		}
-
-		if !built {
-			return Node{}, Node{}, fmt.Errorf("node (%d, %d) is not built", key.level-1, 2*key.index+n)
-		}
-
-		children[n] = child
-	}
-
-	return children[0], children[1], nil
-}
-
-// Wait returns when no build runs. The failed builds wait for the next message.
-func (m *Memory) Wait(ctx context.Context) error {
-	for {
-		m.mutex.Lock()
-
-		running := m.queue.Running()
-		changed := m.changed
-		m.mutex.Unlock()
-
-		if running == 0 {
-			return nil
-		}
-
-		select {
-		case <-changed:
-		case <-ctx.Done():
-			return ctx.Err()
-		}
-	}
-}
-
-// Settle returns when the chat view has no pending line, so that every message
-// so far is summarized. A pending line with no build that runs belongs to a
-// failed build, which waits for the next message, so Settle returns
-// errMemoryUnsettled in place of a wait that never ends.
-func (m *Memory) Settle(ctx context.Context) error {
-	for {
-		m.mutex.Lock()
-
-		pending := m.chat.Pending()
-		running := m.queue.Running()
-		changed := m.changed
-		m.mutex.Unlock()
-
-		if !pending {
-			return nil
-		}
-
-		if running == 0 {
-			return errMemoryUnsettled
-		}
-
-		select {
-		case <-changed:
-		case <-ctx.Done():
-			return ctx.Err()
-		}
-	}
-}
-
-// Flush returns when every write before the call is durable.
-func (m *Memory) Flush(ctx context.Context) error {
-	return m.syncer.Flush(ctx)
-}
-
-// ViewLines returns the lines of the chat view, each with its newline, without
-// the leaf lines of the messages in skip.
-func (m *Memory) ViewLines(skip map[int]bool) []string {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
-
-	return m.chat.LinesWithout(skip)
-}
-
-// Len returns the number of messages in the log.
-func (m *Memory) Len() int {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
-
-	return m.log.Len()
-}
-
-// Message reads the message with the given ID.
-func (m *Memory) Message(id int) (Message, error) {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
-
-	return m.log.Message(id)
-}
-
-// Node reads the node (l, i) if it is built.
-func (m *Memory) Node(level, index int) (Node, bool, error) {
-	m.mutex.Lock()
-	defer m.mutex.Unlock()
-
-	return m.tree.Node(level, index)
-}
-
 // load reads the saved views.
 func (m *Memory) load() error {
 	data, err := os.ReadFile(m.viewPath())
@@ -415,16 +290,66 @@ func (m *Memory) dispatch() {
 			continue
 		}
 
-		m.workers.Add(1)
-
-		go m.run(key, task)
+		m.workers.Add(1)
+
+		go m.run(key, task)
+	}
+}
+
+func (m *Memory) viewPath() string {
+	return filepath.Join(m.root, "view.json")
+}
+
+// build adds node to the tree, keeps it in the views when it is a pending line,
+// and settles its parent.
+func (m *Memory) build(node Node) error {
+	err := m.tree.Add(node)
+	if err != nil {
+		return err
+	}
+
+	m.chat.Refresh(node)
+	m.cview.Refresh(node)
+
+	return m.settle(NodeKey{node.Level, node.Index})
+}
+
+// settle looks at the parent of the built node key when both children are
+// built. It builds each parent that needs no model call, up the tree, and adds
+// the first other parent to the queue.
+func (m *Memory) settle(key NodeKey) error {
+	for {
+		parent := NodeKey{key.level + 1, key.index / 2}
+		ready, err := m.ready(parent)
+		if err != nil || !ready {
+			return err
+		}
+
+		left, right, err := m.children(parent)
+		if err != nil {
+			return err
+		}
+
+		node, free := freeMerge(left, right)
+		if !free {
+			m.queue.AddMerge(parent)
+
+			return nil
+		}
+
+		err = m.tree.Add(node)
+		if err != nil {
+			return err
+		}
+
+		key = parent
 	}
 }
 
 // addMessage builds the leaf of message when it needs no model call, or else
-// adds the leaf to the queue. Then it appends the leaf to the views. A crash can
-// stop an append after the leaf build and before the view save, so a leaf that
-// is built already stays as it is, and restoreQueue settles its parent.
+// adds the leaf to the queue. Then it appends the leaf to the views. A crash
+// can stop an append after the leaf build and before the view save, so a leaf
+// that is built already stays as it is, and restoreQueue settles its parent.
 func (m *Memory) addMessage(message Message) error {
 	built, err := m.tree.Built(0, message.ID)
 	if err != nil {
@@ -478,56 +403,6 @@ func (m *Memory) save() error {
 	return nil
 }
 
-func (m *Memory) viewPath() string {
-	return filepath.Join(m.root, "view.json")
-}
-
-// build adds node to the tree, keeps it in the views when it is a pending line,
-// and settles its parent.
-func (m *Memory) build(node Node) error {
-	err := m.tree.Add(node)
-	if err != nil {
-		return err
-	}
-
-	m.chat.Refresh(node)
-	m.cview.Refresh(node)
-
-	return m.settle(NodeKey{node.Level, node.Index})
-}
-
-// settle looks at the parent of the built node key when both children are
-// built. It builds each parent that needs no model call, up the tree, and adds
-// the first other parent to the queue.
-func (m *Memory) settle(key NodeKey) error {
-	for {
-		parent := NodeKey{key.level + 1, key.index / 2}
-		ready, err := m.ready(parent)
-		if err != nil || !ready {
-			return err
-		}
-
-		left, right, err := m.children(parent)
-		if err != nil {
-			return err
-		}
-
-		node, free := freeMerge(left, right)
-		if !free {
-			m.queue.AddMerge(parent)
-
-			return nil
-		}
-
-		err = m.tree.Add(node)
-		if err != nil {
-			return err
-		}
-
-		key = parent
-	}
-}
-
 // task returns the build task of the node key, with its input read from the log
 // or the tree.
 func (m *Memory) task(key NodeKey) (Task, error) {
@@ -544,6 +419,7 @@ func (m *Memory) task(key NodeKey) (Task, error) {
 		}
 
 		task.Input = []string{leafSource(message)}
+		task.Images = imageFiles(message.Text)
 
 		return task, nil
 	}
@@ -560,6 +436,11 @@ func (m *Memory) task(key NodeKey) (Task, error) {
 	return task, nil
 }
 
+// failed logs the failed build of the node key with its cause err.
+func (m *Memory) failed(key NodeKey, err error) {
+	m.logger.Warn("summary failed", "level", key.level, "index", key.index, "err", err)
+}
+
 // run builds one node with the summarizer. A failed build stays in the queue
 // until the next message.
 func (m *Memory) run(key NodeKey, task Task) {
@@ -577,6 +458,7 @@ func (m *Memory) run(key NodeKey, task Task) {
 			Text:  text,
 			Size:  len(text),
 		})
+
 		m.syncer.Written()
 	}
 
@@ -596,11 +478,6 @@ func (m *Memory) run(key NodeKey, task Task) {
 	m.dispatch()
 }
 
-// failed logs the failed build of the node key with its cause err.
-func (m *Memory) failed(key NodeKey, err error) {
-	m.logger.Printf("summary failed, retry at next message: node (%d, %d): %v", key.level, key.index, err)
-}
-
 // ready tells if both children of the node key are built and the node is not.
 func (m *Memory) ready(key NodeKey) (bool, error) {
 	for _, index := range []int{2 * key.index, 2*key.index + 1} {
@@ -615,6 +492,26 @@ func (m *Memory) ready(key NodeKey) (bool, error) {
 	return !built, err
 }
 
+// children reads the two children of the node key.
+func (m *Memory) children(key NodeKey) (Node, Node, error) {
+	var children [2]Node
+
+	for n := range children {
+		child, built, err := m.tree.Node(key.level-1, 2*key.index+n)
+		if err != nil {
+			return Node{}, Node{}, err
+		}
+
+		if !built {
+			return Node{}, Node{}, fmt.Errorf("node (%d, %d) is not built", key.level-1, 2*key.index+n)
+		}
+
+		children[n] = child
+	}
+
+	return children[0], children[1], nil
+}
+
 func (m *Memory) context(level, index int) string {
 	end := index
 	if level > 0 {
@@ -624,6 +521,192 @@ func (m *Memory) context(level, index int) string {
 	return m.cview.Context(end)
 }
 
+// Append logs a new message, adds it to the tree and the views, and starts the
+// builds that are ready, also the failed ones.
+func (m *Memory) Append(kind, text string) (Message, error) {
+	return m.AppendDated(kind, text, "")
+}
+
+// AppendDated appends a message with its RFC 3339 date and builds its nodes. An
+// empty date uses the clock, as Append does.
+func (m *Memory) AppendDated(kind, text, date string) (Message, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.appendDated(kind, text, date)
+}
+
+// appendDated appends under the memory lock and updates the views and builds.
+func (m *Memory) appendDated(kind, text, date string) (Message, error) {
+	message, err := m.log.AppendDated(kind, text, date)
+	if err != nil {
+		return Message{}, fmt.Errorf("append dated log message: %w", err)
+	}
+
+	defer m.syncer.Written()
+
+	err = m.addMessage(message)
+	if err != nil {
+		return Message{}, fmt.Errorf("add dated message to memory: %w", err)
+	}
+
+	err = m.save()
+	if err != nil {
+		return Message{}, fmt.Errorf("save dated message view: %w", err)
+	}
+
+	m.queue.Retry()
+	m.dispatch()
+
+	return message, nil
+}
+
+// AppendOnce appends a queued message. prepare saves its log ID before the
+// append, under the memory lock. A saved ID permits recovery after a stop
+// between that save and the append, without a second log entry. prepare must
+// not call Memory. The caller stops delivery after a failed append.
+func (m *Memory) AppendOnce(kind, text string, id *int, prepare func(int) error) (Message, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	if id != nil && *id < m.log.Len() {
+		message, err := m.log.Message(*id)
+		if err != nil {
+			return Message{}, fmt.Errorf("read queued message: %w", err)
+		}
+
+		if message.Kind != kind || message.Text != text {
+			return Message{}, errors.New("queued message does not match its log entry")
+		}
+
+		return message, nil
+	}
+
+	if id != nil && *id != m.log.Len() {
+		return Message{}, errors.New("queued message is beyond the log")
+	}
+
+	err := prepare(m.log.Len())
+	if err != nil {
+		return Message{}, fmt.Errorf("prepare queued message: %w", err)
+	}
+
+	message, err := m.appendDated(kind, text, "")
+	if err != nil {
+		return Message{}, fmt.Errorf("append queued message: %w", err)
+	}
+
+	return message, nil
+}
+
+// Wait returns when no build runs. The failed builds wait for the next message.
+func (m *Memory) Wait(ctx context.Context) error {
+	for {
+		m.mutex.Lock()
+
+		running := m.queue.Running()
+		changed := m.changed
+		m.mutex.Unlock()
+
+		if running == 0 {
+			return nil
+		}
+
+		select {
+		case <-changed:
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+	}
+}
+
+// Settle returns when the chat view has no pending line, so that every message
+// so far is summarized. A pending line with no build that runs belongs to a
+// failed build, which waits for the next message, so Settle returns
+// errMemoryUnsettled in place of a wait that never ends.
+func (m *Memory) Settle(ctx context.Context) error {
+	for {
+		m.mutex.Lock()
+
+		pending := m.chat.Pending()
+		running := m.queue.Running()
+		changed := m.changed
+		m.mutex.Unlock()
+
+		if !pending {
+			return nil
+		}
+
+		if running == 0 {
+			return errMemoryUnsettled
+		}
+
+		select {
+		case <-changed:
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+	}
+}
+
+// Snapshot flushes all memory writes and runs snapshot while no append or
+// summary write can change the log, indexes, tree, or views. snapshot must not
+// call Memory.
+func (m *Memory) Snapshot(ctx context.Context, snapshot func() error) error {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	err := m.Flush(ctx)
+	if err != nil {
+		return fmt.Errorf("flush memory snapshot: %w", err)
+	}
+
+	err = snapshot()
+	if err != nil {
+		return fmt.Errorf("copy memory snapshot: %w", err)
+	}
+
+	return nil
+}
+
+// Flush returns when every write before the call is durable.
+func (m *Memory) Flush(ctx context.Context) error {
+	return m.syncer.Flush(ctx)
+}
+
+// ViewLines returns the lines of the chat view, each with its newline, without
+// the leaf lines of the messages in skip.
+func (m *Memory) ViewLines(skip map[int]bool) []string {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.chat.LinesWithout(skip)
+}
+
+// Len returns the number of messages in the log.
+func (m *Memory) Len() int {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.log.Len()
+}
+
+// Message reads the message with the given ID.
+func (m *Memory) Message(id int) (Message, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.log.Message(id)
+}
+
+// Node reads the node (l, i) if it is built.
+func (m *Memory) Node(level, index int) (Node, bool, error) {
+	m.mutex.Lock()
+	defer m.mutex.Unlock()
+
+	return m.tree.Node(level, index)
+}
+
 // View returns the chat view as the model reads it.
 func (m *Memory) View() string {
 	m.mutex.Lock()
@@ -715,19 +798,3 @@ func (m *Memory) buildBuiltByModel(key NodeKey) (bool, error) {
 
 	return !free, nil
 }
-
-// Task is the build of the node (Level, Index). Context is the compaction view
-// for the build. Input holds the source "kind: text" of a leaf, or the two
-// child lines of a merge with their newlines as spaces.
-type Task struct {
-	Level   int
-	Index   int
-	Context string
-	Input   []string
-}
-
-// MemoryViewFile is the content of view.json.
-type MemoryViewFile struct {
-	Chat       ViewState `json:"chat"`
-	Compaction ViewState `json:"compaction"`
-}