Refactor: Inject prunable list into last assistant message instead of new user message

Tarquinen · Tarquinen · commit 1b6ef9bce168 · 2025-12-08T20:27:53.000-05:00
This change addresses models consistently responding to the injected prunable
list even when instructed not to. By appending to the last assistant message
instead of creating a new user message, the list appears as part of the model's
own context state awareness rather than user input to respond to.

- Rename injectUserMessage -&gt; appendToLastAssistantMessage across all formats
- Each format handler now finds and appends to the last assistant/model message
- Update synthetic.txt prompt to reflect the new assistant message framing
- Rename buildUserInjection -&gt; buildAssistantInjection
diff --git a/lib/fetch-wrapper/formats/anthropic.ts b/lib/fetch-wrapper/formats/anthropic.ts
@@ -33,13 +33,29 @@ export const anthropicFormat: FormatDescriptor = {
         return true
     },
 
-    injectUserMessage(body: any, injection: string): boolean {
-        if (!injection || !body.messages) return false
-        body.messages.push({
-            role: 'user',
-            content: [{ type: 'text', text: injection }]
-        })
-        return true
+    appendToLastAssistantMessage(body: any, injection: string): boolean {
+        if (!injection || !body.messages || body.messages.length === 0) return false
+        
+        // Find the last assistant message
+        for (let i = body.messages.length - 1; i >= 0; i--) {
+            const msg = body.messages[i]
+            if (msg.role === 'assistant') {
+                // Append to existing content array
+                if (Array.isArray(msg.content)) {
+                    msg.content.push({ type: 'text', text: injection })
+                } else if (typeof msg.content === 'string') {
+                    // Convert string content to array format
+                    msg.content = [
+                        { type: 'text', text: msg.content },
+                        { type: 'text', text: injection }
+                    ]
+                } else {
+                    msg.content = [{ type: 'text', text: injection }]
+                }
+                return true
+            }
+        }
+        return false
     },
 
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[] {
diff --git a/lib/fetch-wrapper/formats/bedrock.ts b/lib/fetch-wrapper/formats/bedrock.ts
@@ -32,13 +32,21 @@ export const bedrockFormat: FormatDescriptor = {
         return true
     },
 
-    injectUserMessage(body: any, injection: string): boolean {
-        if (!injection || !body.messages) return false
-        body.messages.push({
-            role: 'user',
-            content: [{ text: injection }]
-        })
-        return true
+    appendToLastAssistantMessage(body: any, injection: string): boolean {
+        if (!injection || !body.messages || body.messages.length === 0) return false
+
+        for (let i = body.messages.length - 1; i >= 0; i--) {
+            const msg = body.messages[i]
+            if (msg.role === 'assistant') {
+                if (Array.isArray(msg.content)) {
+                    msg.content.push({ text: injection })
+                } else {
+                    msg.content = [{ text: injection }]
+                }
+                return true
+            }
+        }
+        return false
     },
 
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[] {
diff --git a/lib/fetch-wrapper/formats/gemini.ts b/lib/fetch-wrapper/formats/gemini.ts
@@ -31,13 +31,21 @@ export const geminiFormat: FormatDescriptor = {
         return true
     },
 
-    injectUserMessage(body: any, injection: string): boolean {
-        if (!injection || !body.contents) return false
-        body.contents.push({
-            role: 'user',
-            parts: [{ text: injection }]
-        })
-        return true
+    appendToLastAssistantMessage(body: any, injection: string): boolean {
+        if (!injection || !body.contents || body.contents.length === 0) return false
+
+        for (let i = body.contents.length - 1; i >= 0; i--) {
+            const content = body.contents[i]
+            if (content.role === 'model') {
+                if (Array.isArray(content.parts)) {
+                    content.parts.push({ text: injection })
+                } else {
+                    content.parts = [{ text: injection }]
+                }
+                return true
+            }
+        }
+        return false
     },
 
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[] {
diff --git a/lib/fetch-wrapper/formats/openai-chat.ts b/lib/fetch-wrapper/formats/openai-chat.ts
@@ -27,10 +27,23 @@ export const openaiChatFormat: FormatDescriptor = {
         return true
     },
 
-    injectUserMessage(body: any, injection: string): boolean {
-        if (!injection || !body.messages) return false
-        body.messages.push({ role: 'user', content: injection })
-        return true
+    appendToLastAssistantMessage(body: any, injection: string): boolean {
+        if (!injection || !body.messages || body.messages.length === 0) return false
+
+        for (let i = body.messages.length - 1; i >= 0; i--) {
+            const msg = body.messages[i]
+            if (msg.role === 'assistant') {
+                if (typeof msg.content === 'string') {
+                    msg.content = msg.content + '\n\n' + injection
+                } else if (Array.isArray(msg.content)) {
+                    msg.content.push({ type: 'text', text: injection })
+                } else {
+                    msg.content = injection
+                }
+                return true
+            }
+        }
+        return false
     },
 
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[] {
diff --git a/lib/fetch-wrapper/formats/openai-responses.ts b/lib/fetch-wrapper/formats/openai-responses.ts
@@ -23,14 +23,23 @@ export const openaiResponsesFormat: FormatDescriptor = {
         return true
     },
 
-    injectUserMessage(body: any, injection: string): boolean {
-        if (!injection || !body.input) return false
-        body.input.push({
-            type: 'message',
-            role: 'user',
-            content: injection
-        })
-        return true
+    appendToLastAssistantMessage(body: any, injection: string): boolean {
+        if (!injection || !body.input || body.input.length === 0) return false
+
+        for (let i = body.input.length - 1; i >= 0; i--) {
+            const item = body.input[i]
+            if (item.type === 'message' && item.role === 'assistant') {
+                if (typeof item.content === 'string') {
+                    item.content = item.content + '\n\n' + injection
+                } else if (Array.isArray(item.content)) {
+                    item.content.push({ type: 'output_text', text: injection })
+                } else {
+                    item.content = injection
+                }
+                return true
+            }
+        }
+        return false
     },
 
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[] {
diff --git a/lib/fetch-wrapper/handler.ts b/lib/fetch-wrapper/handler.ts
@@ -1,7 +1,7 @@
 import type { FetchHandlerContext, FetchHandlerResult, FormatDescriptor, PrunedIdData } from "./types"
 import { type PluginState, ensureSessionRestored } from "../state"
 import type { Logger } from "../logger"
-import { buildPrunableToolsList, buildUserInjection } from "./prunable-list"
+import { buildPrunableToolsList, buildAssistantInjection } from "./prunable-list"
 import { syncToolCache } from "../state/tool-cache"
 import { loadPrompt } from "../core/prompt"
 
@@ -96,11 +96,11 @@ export async function handleFormat(
                 modified = true
             }
 
-            const userInjection = buildUserInjection(prunableList, includeNudge)
+            const assistantInjection = buildAssistantInjection(prunableList, includeNudge)
 
-            if (format.injectUserMessage && format.injectUserMessage(body, userInjection)) {
+            if (format.appendToLastAssistantMessage && format.appendToLastAssistantMessage(body, assistantInjection)) {
                 const nudgeMsg = includeNudge ? " with nudge" : ""
-                ctx.logger.debug("fetch", `Injected prunable tools list${nudgeMsg} into user message (${format.name})`, {
+                ctx.logger.debug("fetch", `Appended prunable tools list${nudgeMsg} to last assistant message (${format.name})`, {
                     ids: numericIds,
                     nudge: includeNudge,
                     toolsSincePrune: ctx.toolTracker.toolResultCount
diff --git a/lib/fetch-wrapper/prunable-list.ts b/lib/fetch-wrapper/prunable-list.ts
@@ -42,7 +42,7 @@ export function buildPrunableToolsList(
     }
 }
 
-export function buildUserInjection(
+export function buildAssistantInjection(
     prunableList: string,
     includeNudge: boolean
 ): string {
diff --git a/lib/fetch-wrapper/types.ts b/lib/fetch-wrapper/types.ts
@@ -19,7 +19,7 @@ export interface FormatDescriptor {
     detect(body: any): boolean
     getDataArray(body: any): any[] | undefined
     injectSystemMessage(body: any, injection: string): boolean
-    injectUserMessage?(body: any, injection: string): boolean
+    appendToLastAssistantMessage?(body: any, injection: string): boolean
     extractToolOutputs(data: any[], state: PluginState): ToolOutput[]
     replaceToolOutput(data: any[], toolId: string, prunedMessage: string, state: PluginState): boolean
     hasToolOutputs(data: any[]): boolean
diff --git a/lib/prompts/synthetic.txt b/lib/prompts/synthetic.txt
@@ -1,5 +1,5 @@
 <system-reminder>
-The <prunable-tools> list and any pruning nudges are injected by a plugin and are invisible to the user.
+The <prunable-tools> list and any pruning nudges are injected by a plugin as assistant messages and are invisible to the user. Do NOT repeat, acknowledge, or respond to these in your output - simply use the information when deciding what to prune.
 
 IMPORTANT: Your thinking/reasoning blocks must NOT reference, discuss, or address the <prunable-tools> list or any nudges about pruning. The user can see your thinking blocks, and referencing invisible plugin content will confuse them. In your thinking, focus only on the user's task and your problem-solving approach.
 </system-reminder>

Original file line number	Diff line number	Diff line change
`@@ -42,7 +42,7 @@ export function buildPrunableToolsList(`
`42`	`42`	`}`
`43`	`43`	`}`
`44`	`44`
`45`		`-export function buildUserInjection(`
	`45`	`+export function buildAssistantInjection(`
`46`	`46`	`prunableList: string,`
`47`	`47`	`includeNudge: boolean`
`48`	`48`	`): string {`