Carry new images through model request chains

[onechat] / compactor.go

diff --git a/compactor.go b/compactor.go
index c0d1a9c78b95b47c4a92aed51b902310ba475800..2b239015ce763f56d1a23a4621647f056023b07b 100644 (file)
--- a/compactor.go
+++ b/compactor.go
@@ -17,38 +17,43 @@ var compactorRuler = strings.Repeat("-", nodeBytes)
 
 // newCompactor returns the compactor of an agent, which sends its calls through
 // client with model. instructions gives the system prompt of the agent at each
 
 // newCompactor returns the compactor of an agent, which sends its calls through
 // client with model. instructions gives the system prompt of the agent at each

-// build, so a build sends the same prompt as the turns of that time.
-func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig) *Compactor {
+// build. tools are the tools of that agent, so a build sends the same prompt
+// and tools as its turns.
+func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig, tools []ModelTool, images *ImageStore) *Compactor {

        return &Compactor{
                client:       client,
                model:        model,
                instructions: instructions,
        return &Compactor{
                client:       client,
                model:        model,
                instructions: instructions,

-               tools:        promptTools(),
+               tools:        tools,
+               images:       images,

        }
 }
 
 // Compactor builds the nodes of one agent with a model. A compaction is a call
 // like a turn, with the same system prompt and tools, so it reads them from the
 // prompt cache of the turns. The tools are never called. After them comes one
        }
 }
 
 // Compactor builds the nodes of one agent with a model. A compaction is a call
 // like a turn, with the same system prompt and tools, so it reads them from the
 // prompt cache of the turns. The tools are never called. After them comes one

-// user message with the compaction view up to the node and the task.
+// user message with the compaction view up to the node, its closing tag, a
+// newline, and the task.

 //
 //

-// When the line is over nodeBytes, the next try sends the same view and task,
-// plus a note with the size of the newest line and its cut at the limit. A try
-// never sends the model's earlier reply or reasoning back, so every try has the
-// same cached prefix and about the same size. After at most compactorTries
-// tries, the shortest line is the node. A few bytes over the limit is fine,
-// because the view measures the real sizes. A call that fails before any line
-// fails the build, and the core tries it again at the next message.
+// When the line is over nodeBytes, the next try sends the same message, plus
+// one more user message with a note with the size of the newest line and its
+// cut at the limit. Only the first call of an image leaf sends its pixels. A
+// try never sends the model's earlier reply or reasoning back, so every try
+// begins with the same view and task and has about the same size. After at most
+// compactorTries tries, the shortest line is the node. A few bytes over the
+// limit is fine, because the view measures the real sizes. A call that fails
+// before any line fails the build, and the core tries it again at the next
+// message.

 type Compactor struct {
        client       *ModelClient
        model        ModelConfig
        instructions func() (string, error)
        tools        []ModelTool
 type Compactor struct {
        client       *ModelClient
        model        ModelConfig
        instructions func() (string, error)
        tools        []ModelTool

+       images       *ImageStore

 }
 
 // Summarize returns the line of the node of task.
 func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 }
 
 // Summarize returns the line of the node of task.
 func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {

-       view := task.Context + "\n"

        text, err := compactorTask(task)
        if err != nil {
                return "", err
        text, err := compactorTask(task)
        if err != nil {
                return "", err


@@ -56,25 +61,35 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 
        instructions, err := c.instructions()
        if err != nil {
 
        instructions, err := c.instructions()
        if err != nil {

-               return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
+               return "", err

        }
 
        }
 

+       source := task.Context + "\n" + text
+       parts, err := imageParts(c.images, source, task.Images)
+       if err != nil {
+               return "", fmt.Errorf("read compaction images: %w", err)
+       }
+
+       first := []any{modelUserParts(parts)}
+

        var lines []string
        var lines []string

+

        note := ""
        for len(lines) < compactorTries {
        note := ""
        for len(lines) < compactorTries {

-               parts := []string{view, text}
+               input := first

                if note != "" {
                if note != "" {

-                       parts = append(parts, note)
+                       input = []any{modelUserMessage(source), modelUserMessage(note)}

                }
 
                var response ModelResponse
                }
 
                var response ModelResponse

+

                response, err = c.client.Create(ctx, ModelRequest{
                        Model:        c.model.Model,
                        Effort:       c.model.Effort,
                        Instructions: instructions,
                        Tools:        c.tools,
                response, err = c.client.Create(ctx, ModelRequest{
                        Model:        c.model.Model,
                        Effort:       c.model.Effort,
                        Instructions: instructions,
                        Tools:        c.tools,

-                       Input:        []any{modelUserMessage(parts...)},
-                       Node:         &nodeKey{task.Level, task.Index},
+                       Input:        input,
+                       Node:         &NodeKey{task.Level, task.Index},

                })
                if err != nil {
                        break
                })
                if err != nil {
                        break


@@ -95,7 +110,7 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
        }
 
        if len(lines) == 0 {
        }
 
        if len(lines) == 0 {

-               return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
+               return "", err

        }
 
        return compactorShortest(lines), nil
        }
 
        return compactorShortest(lines), nil


@@ -104,12 +119,17 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 // compactorTask returns the task of a build, with its input.
 func compactorTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {
 // compactorTask returns the task of a build, with its input.
 func compactorTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {

-               return fmt.Sprintf(`Compaction: compress message %d into one line of at most %d bytes
+               text := fmt.Sprintf(`Compaction: compress message %d into one line of at most %d bytes

 (about 70 words), the length of this ruler:
 %s
 <input>
 %s
 (about 70 words), the length of this ruler:
 %s
 <input>
 %s

-</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0]), nil
+</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0])
+               if len(task.Images) > 0 {
+                       text += "\nKeep each [image FILE] name and describe its kind, visible text, amounts, names, and context."
+               }
+
+               return text, nil

        }
 
        if task.Level > 0 && len(task.Input) == 2 {
        }
 
        if task.Level > 0 && len(task.Input) == 2 {