Add ellipse on token limit exceed and flex layout options in PromptTemplateString

packages/core/src/chat.ts

@@ -469,17 +469,17 @@
                 await generator(ctx, structuredClone(messages))
                 const node = ctx.node
                 checkCancelled(cancellationToken)
                 // expand template
-                const { errors, prompt } = await renderPromptNode(
+                const { errors, userPrompt } = await renderPromptNode(
                     options.model,
                     node,
                     {
                         trace,
                     }
                 )
-                if (prompt?.trim().length) {
-                    trace.detailsFenced(`💬 message`, prompt, "markdown")
-                    messages.push({ role: "user", content: prompt })
+                if (userPrompt?.trim().length) {
+                    trace.detailsFenced(`💬 message`, userPrompt, "markdown")
+                    messages.push({ role: "user", content: userPrompt })
                     needsNewTurn = true
                 } else trace.item("no message")
                 if (errors?.length) {

packages/core/src/constants.ts

@@ -227,3 +227,4 @@ export const CONSOLE_COLOR_WARNING = 95
 export const CONSOLE_COLOR_ERROR = 91
 
 export const PLAYWRIGHT_DEFAULT_BROWSER = "chromium"
+export const MAX_TOKENS_ELLIPSE = "..."

packages/core/src/expander.ts

@@ -73,7 +73,7 @@ export async function callExpander(
         const node = ctx.node
         if (provider !== MODEL_PROVIDER_AICI) {
             const {
-                prompt,
+                userPrompt,
                 assistantPrompt,
                 images: imgs,
                 errors,
@@ -84,7 +84,7 @@ export async function callExpander(
                 chatParticipants: cps,
                 fileOutputs: fos,
             } = await renderPromptNode(model, node, { trace })
-            text = prompt
+            text = userPrompt
             assistantText = assistantPrompt
             images = imgs
             schemas = schs

packages/core/src/promptdom.ts

@@ -6,7 +6,11 @@
 import { MarkdownTrace, TraceOptions } from "./trace"
 import { arrayify, assert, toStringList, trimNewlines } from "./util"
 import { YAMLStringify } from "./yaml"
-import { MARKDOWN_PROMPT_FENCE, PROMPT_FENCE } from "./constants"
+import {
+    MARKDOWN_PROMPT_FENCE,
+    MAX_TOKENS_ELLIPSE,
+    PROMPT_FENCE,
+} from "./constants"
 import { parseModelIdentifier } from "./models"
 import { toChatCompletionUserMessage } from "./chat"
 import { errorMessage } from "./error"
@@ -417,11 +421,11 @@
     await visitor.afterNode?.(node)
 }
 
 export interface PromptNodeRender {
-    prompt: string
+    userPrompt: string
     assistantPrompt: string
     images: PromptImage[]
     errors: unknown[]
     schemas: Record<string, JSONSchema>
     functions: ToolCallback[]
     fileMerges: FileMergeHandler[]
@@ -550,20 +554,20 @@
         resolved?: string
         tokens?: number
         maxTokens?: number
+        preview?: string
     }) => {
         if (
             !n.error &&
             n.resolved !== undefined &&
             n.maxTokens !== undefined &&
             n.tokens > n.maxTokens
         ) {
-            const value = n.resolved.slice(
-                0,
-                Math.floor((n.maxTokens * n.resolved.length) / n.tokens)
-            )
-            n.resolved = value
+            const end = Math.floor((n.maxTokens * n.resolved.length) / n.tokens)
+            const value = n.resolved.slice(0, end) + MAX_TOKENS_ELLIPSE
+            n.resolved = n.preview = value
             n.tokens = estimateTokens(value, encoder)
             truncated = true
+            trace.log(`truncated text to ${n.tokens} tokens`)
         }
     }
 
@@ -574,12 +578,15 @@
             n.maxTokens !== undefined &&
             n.tokens > n.maxTokens
         ) {
-            n.resolved.content = n.resolved.content.slice(
-                0,
-                Math.floor((n.maxTokens * n.resolved.content.length) / n.tokens)
+            const end = Math.floor(
+                (n.maxTokens * n.resolved.content.length) / n.tokens
             )
+            n.resolved.content =
+                n.resolved.content.slice(0, end) + MAX_TOKENS_ELLIPSE
             n.tokens = estimateTokens(n.resolved.content, encoder)
+            n.preview = n.resolved.content
             truncated = true
+            trace.log(`truncated def ${n.name} to ${n.tokens} tokens`)
         }
     }
 
@@ -593,6 +600,20 @@
     return truncated
 }
 
+async function flexPromptNode(
+    model: string,
+    node: PromptNode,
+    options?: TraceOptions
+): Promise<boolean> {
+    const FLEX_BASIS_DEFAULT = 1
+    const FLEX_GROW_DEFAULT = Infinity
+    const { trace } = options || {}
+    const encoder = await resolveTokenEncoder(model)
+    let flexed = false
+
+    return flexed
+}
+
 async function tracePromptNode(
     trace: MarkdownTrace,
     root: PromptNode,
@@ -640,7 +661,11 @@
     const truncated = await truncatePromptNode(model, node, options)
     if (truncated) await tracePromptNode(trace, node, { label: "truncated" })
 
-    let prompt = ""
+    const flexed = await flexPromptNode(model, node, options)
+    if (flexed) await tracePromptNode(trace, node, { label: "flexed" })
+
+    let systemPrompt = ""
+    let userPrompt = ""
     let assistantPrompt = ""
     const images: PromptImage[] = []
     const errors: unknown[] = []
@@ -655,12 +680,12 @@
         text: async (n) => {
             if (n.error) errors.push(n.error)
             const value = n.resolved
-            if (value != undefined) prompt += value + "\n"
+            if (value != undefined) userPrompt += value + "\n"
         },
         def: async (n) => {
             if (n.error) errors.push(n.error)
             const value = n.resolved
-            if (value !== undefined) prompt += renderDefNode(n) + "\n"
+            if (value !== undefined) userPrompt += renderDefNode(n) + "\n"
         },
         assistant: async (n) => {
             if (n.error) errors.push(n.error)
@@ -670,7 +695,7 @@
         stringTemplate: async (n) => {
             if (n.error) errors.push(n.error)
             const value = n.resolved
-            if (value != undefined) prompt += value + "\n"
+            if (value != undefined) userPrompt += value + "\n"
         },
         image: async (n) => {
             if (n.error) errors.push(n.error)
@@ -691,8 +716,8 @@
             const value = n.resolved
             if (value) {
                 for (const [filename, content] of Object.entries(value)) {
-                    prompt += content
-                    prompt += "\n"
+                    userPrompt += content
+                    userPrompt += "\n"
                     if (trace)
                         trace.detailsFenced(
                             `📦 import template ${filename}`,
@@ -727,7 +752,7 @@
 ${trimNewlines(schemaText)}
 \`\`\`
 `
-            prompt += text
+            userPrompt += text
             n.tokens = estimateTokens(text, encoder)
             if (trace && format !== "json")
                 trace.detailsFenced(
@@ -771,7 +796,7 @@
 
     const fods = fileOutputs?.filter((f) => !!f.description)
     if (fods?.length > 0) {
-        prompt += `
+        userPrompt += `
 ## File generation rules
 
 When generating files, use the following rules which are formatted as "file glob: description":
@@ -782,15 +807,15 @@
     }
 
     const messages: ChatCompletionMessageParam[] = [
-        toChatCompletionUserMessage(prompt, images),
+        toChatCompletionUserMessage(userPrompt, images),
     ]
     if (assistantPrompt)
         messages.push(<ChatCompletionAssistantMessageParam>{
             role: "assistant",
             content: assistantPrompt,
         })
-    const res = {
-        prompt,
+    const res = <PromptNodeRender>{
+        userPrompt,
         assistantPrompt,
         images,
         schemas,