Write private data archives with safe extraction

[onechat] / compactor.go

diff --git a/compactor.go b/compactor.go
index 11497b8ddd6aeb919a75cf37c7ae08e8f2c25c59..0089002a4838b057303e210c62a34b434b486ae0 100644 (file)
--- a/compactor.go
+++ b/compactor.go
@@ -9,75 +9,78 @@ import (
        "unicode/utf8"
 )
 
        "unicode/utf8"
 )
 

-const (
-       // compactionTries is the most calls for one node.
-       compactionTries = 5
+// compactorTries is the most calls for one node.
+const compactorTries = 5

 
 

-       // viewBlock is the number of view lines in one input part.
-       viewBlock = 4
-)
+// compactorRuler shows the model the size limit of a line. Models cannot count
+// bytes, and a sample line as the ruler gets its content copied.
+var compactorRuler = strings.Repeat("-", nodeBytes)

 
 

-// ruler shows the model the size limit of a line. Models cannot count bytes,
-// and a sample line as the ruler gets its content copied.
-var ruler = 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
+// 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 {
+       return &Compactor{
+               client:       client,
+               model:        model,
+               instructions: instructions,
+               tools:        promptTools(),
+       }
+}

 
 // 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, in the
-// parts that compactionParts gives. Only the last part holds the end of the
-// view, so compactions whose views end at different lines begin with the same
-// parts up to their last shared block.
+// 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 compactionTries
-// 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. 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 {
 type Compactor struct {

-       client       *Client
+       client       *ModelClient

        model        ModelConfig
        model        ModelConfig

-       instructions string
-       tools        []Tool
-}
-
-// newCompactor returns the compactor of the agent name, which sends its calls
-// through client with model.
-func newCompactor(client *Client, name string, model ModelConfig) *Compactor {
-       return &Compactor{
-               client:       client,
-               model:        model,
-               instructions: prompt(name),
-               tools:        agentTools(),
-       }
+       instructions func() (string, error)
+       tools        []ModelTool

 }
 
 // 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) {

-       text, err := compactionTask(task)
+       text, err := compactorTask(task)
+       if err != nil {
+               return "", err
+       }
+
+       instructions, err := c.instructions()

        if err != nil {
                return "", err
        }
 
        if err != nil {
                return "", err
        }
 

-       first := compactionParts(task.Context, text)
+       first := []any{modelUserMessage(task.Context + "\n" + text)}
+

        var lines []string
        var lines []string

+

        note := ""
        note := ""

-       for len(lines) < compactionTries {
-               parts := first
+       for len(lines) < compactorTries {
+               input := first

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

-                       parts = append(slices.Clip(first), note)
+                       input = append(slices.Clip(first), modelUserMessage(note))

                }
 
                }
 

-               var response Response
-               response, err = c.client.Create(ctx, Request{
+               var response ModelResponse
+
+               response, err = c.client.Create(ctx, ModelRequest{

                        Model:        c.model.Model,
                        Effort:       c.model.Effort,
                        Model:        c.model.Model,
                        Effort:       c.model.Effort,

-                       Instructions: c.instructions,
+                       Instructions: instructions,

                        Tools:        c.tools,
                        Tools:        c.tools,

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

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


@@ -94,41 +97,25 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
                        break
                }
 
                        break
                }
 

-               note = tooLong(line)
+               note = compactorTooLong(line)

        }
 
        if len(lines) == 0 {
        }
 
        if len(lines) == 0 {

-               return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
-       }
-
-       return shortest(lines), nil
-}
-
-// compactionParts returns the input parts of a first try for the compaction
-// view context and the task text: the opening tag, one part for each whole
-// block of viewBlock lines, and then one part with the rest of the lines, the
-// closing tag, a newline, and the task. Joined, the parts are the view, a
-// newline, and the task.
-func compactionParts(context, task string) []string {
-       lines := strings.SplitAfter(strings.TrimPrefix(context, "<chat>\n"), "\n")
-       whole := len(lines) - 1 - (len(lines)-1)%viewBlock
-       parts := []string{"<chat>\n"}
-       for n := 0; n < whole; n += viewBlock {
-               parts = append(parts, strings.Join(lines[n:n+viewBlock], ""))
+               return "", err

        }
 
        }
 

-       return append(parts, strings.Join(lines[whole:], "")+"\n"+task)
+       return compactorShortest(lines), nil

 }
 
 }
 

-// compactionTask returns the task of a build, with its input.
-func compactionTask(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 {
                return 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
        if task.Level == 0 && len(task.Input) == 1 {
                return 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

-</input>`, task.Index, nodeBytes, ruler, task.Input[0]), nil
+</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0]), nil

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


@@ -143,24 +130,36 @@ of them from there too.
 <input>
 %s
 %s
 <input>
 %s
 %s

-</input>`, first, half, first+half, half, nodeBytes, ruler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil
+</input>`, first, half, first+half, half, nodeBytes, compactorRuler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil

        }
 
        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
 }
 
        }
 
        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
 }
 

-// tooLong returns the note for a line over nodeBytes: its size and its first
-// nodeBytes bytes, which show the model where the limit cuts it.
-func tooLong(line string) string {
+// compactorTooLong returns the note for a line over nodeBytes: its size and its
+// first nodeBytes bytes, which show the model where the limit cuts it.
+func compactorTooLong(line string) string {

        return fmt.Sprintf(`Too long: your last line for this <input> was %d bytes,
 over the %d-byte limit. Write the whole line again, cutting just
 enough of the least valuable items to fit before this cut:
        return fmt.Sprintf(`Too long: your last line for this <input> was %d bytes,
 over the %d-byte limit. Write the whole line again, cutting just
 enough of the least valuable items to fit before this cut:

-%s| <- LIMIT`, len(line), nodeBytes, cutLine(line))
+%s| <- LIMIT`, len(line), nodeBytes, compactorCutLine(line))

 }
 
 }
 

-// cutLine returns the first nodeBytes bytes of line, without a part of a
-// character at the end.
-func cutLine(line string) string {
+// compactorShortest returns the first of the compactorShortest lines.
+func compactorShortest(lines []string) string {
+       best := lines[0]
+       for _, line := range lines[1:] {
+               if len(line) < len(best) {
+                       best = line
+               }
+       }
+
+       return best
+}
+
+// compactorCutLine returns the first nodeBytes bytes of line, without a part of
+// a character at the end.
+func compactorCutLine(line string) string {

        if len(line) <= nodeBytes {
                return line
        }
        if len(line) <= nodeBytes {
                return line
        }


@@ -172,15 +171,3 @@ func cutLine(line string) string {
 
        return cut
 }
 
        return cut
 }

-
-// shortest returns the first of the shortest lines.
-func shortest(lines []string) string {
-       best := lines[0]
-       for _, line := range lines[1:] {
-               if len(line) < len(best) {
-                       best = line
-               }
-       }
-
-       return best
-}