Merge pull request #33 from spoons-and-mirrors/feat/synth-instruction

feature(synthetic instructions)

Author: Tarquinen

README.md

@@ -28,6 +28,13 @@ DCP implements two complementary strategies:
 **Deduplication** — Fast, zero-cost pruning that identifies repeated tool calls (e.g., reading the same file multiple times) and keeps only the most recent output. Runs instantly with no LLM calls.
 
 **AI Analysis** — Uses a language model to semantically analyze conversation context and identify tool outputs that are no longer relevant to the current task. More thorough but incurs LLM cost.
+
+## Context Pruning Tool
+
+When `strategies.onTool` is enabled, DCP exposes a `context_pruning` tool to Opencode that the AI can call to trigger pruning on demand. To help the AI use this tool effectively, DCP also injects guidance.
+
+When `nudge_freq` is enabled, injects reminders (every `nudge_freq` tool results) prompting the AI to consider pruning when appropriate.
+
 ## How It Works
 
 DCP is **non-destructive**—pruning state is kept in memory only. When requests go to your LLM, DCP replaces pruned outputs with a placeholder; original session data stays intact.
@@ -46,6 +53,7 @@ DCP uses its own config file (`~/.config/opencode/dcp.jsonc` or `.opencode/dcp.j
 | `showModelErrorToasts` | `true` | Show notifications on model fallback |
 | `strictModelSelection` | `false` | Only run AI analysis with session or configured model (disables fallback models) |
 | `pruning_summary` | `"detailed"` | `"off"`, `"minimal"`, or `"detailed"` |
+| `nudge_freq` | `5` | Remind AI to prune every N tool results (0 = disabled) |
 | `protectedTools` | `["task", "todowrite", "todoread", "context_pruning"]` | Tools that are never pruned |
 | `strategies.onIdle` | `["deduplication", "ai-analysis"]` | Strategies for automatic pruning |
 | `strategies.onTool` | `["deduplication", "ai-analysis"]` | Strategies when AI calls `context_pruning` |

index.ts

@@ -7,6 +7,8 @@ import { createPluginState } from "./lib/state"
 import { installFetchWrapper } from "./lib/fetch-wrapper"
 import { createPruningTool } from "./lib/pruning-tool"
 import { createEventHandler, createChatParamsHandler } from "./lib/hooks"
+import { createToolTracker } from "./lib/synth-instruction"
+import { loadPrompt } from "./lib/prompt"
 
 const plugin: Plugin = (async (ctx) => {
     const { config, migrations } = getConfig(ctx)
@@ -39,8 +41,15 @@ const plugin: Plugin = (async (ctx) => {
         ctx.directory
     )
 
-    // Install global fetch wrapper for context pruning
-    installFetchWrapper(state, logger, ctx.client)
+    // Create tool tracker and load prompts for synthetic instruction injection
+    const toolTracker = createToolTracker()
+    const prompts = {
+        synthInstruction: loadPrompt("synthetic"),
+        nudgeInstruction: loadPrompt("nudge")
+    }
+
+    // Install global fetch wrapper for context pruning and synthetic instruction injection
+    installFetchWrapper(state, logger, ctx.client, config, toolTracker, prompts)
 
     // Log initialization
     logger.info("plugin", "DCP initialized", {

lib/config.ts

@@ -15,6 +15,7 @@ export interface PluginConfig {
     showModelErrorToasts?: boolean
     strictModelSelection?: boolean
     pruning_summary: "off" | "minimal" | "detailed"
+    nudge_freq: number
     strategies: {
         onIdle: PruningStrategy[]
         onTool: PruningStrategy[]
@@ -33,6 +34,7 @@ const defaultConfig: PluginConfig = {
     showModelErrorToasts: true,
     strictModelSelection: false,
     pruning_summary: 'detailed',
+    nudge_freq: 5,
     strategies: {
         onIdle: ['deduplication', 'ai-analysis'],
         onTool: ['deduplication', 'ai-analysis']
@@ -47,6 +49,7 @@ const VALID_CONFIG_KEYS = new Set([
     'showModelErrorToasts',
     'strictModelSelection',
     'pruning_summary',
+    'nudge_freq',
     'strategies'
 ])
 
@@ -118,6 +121,8 @@ function createDefaultConfig(): void {
   },
   // Summary display: "off", "minimal", or "detailed"
   "pruning_summary": "detailed",
+  // How often to nudge the AI to prune (every N tool results, 0 = disabled)
+  "nudge_freq": 5,
   // Tools that should never be pruned
   "protectedTools": ["task", "todowrite", "todoread", "context_pruning"]
 }
@@ -196,7 +201,8 @@ export function getConfig(ctx?: PluginInput): ConfigResult {
                     showModelErrorToasts: globalConfig.showModelErrorToasts ?? config.showModelErrorToasts,
                     strictModelSelection: globalConfig.strictModelSelection ?? config.strictModelSelection,
                     strategies: mergeStrategies(config.strategies, globalConfig.strategies as any),
-                    pruning_summary: globalConfig.pruning_summary ?? config.pruning_summary
+                    pruning_summary: globalConfig.pruning_summary ?? config.pruning_summary,
+                    nudge_freq: globalConfig.nudge_freq ?? config.nudge_freq
                 }
                 logger.info('config', 'Loaded global config', { path: configPaths.global })
             }
@@ -226,7 +232,8 @@ export function getConfig(ctx?: PluginInput): ConfigResult {
                     showModelErrorToasts: projectConfig.showModelErrorToasts ?? config.showModelErrorToasts,
                     strictModelSelection: projectConfig.strictModelSelection ?? config.strictModelSelection,
                     strategies: mergeStrategies(config.strategies, projectConfig.strategies as any),
-                    pruning_summary: projectConfig.pruning_summary ?? config.pruning_summary
+                    pruning_summary: projectConfig.pruning_summary ?? config.pruning_summary,
+                    nudge_freq: projectConfig.nudge_freq ?? config.nudge_freq
                 }
                 logger.info('config', 'Loaded project config (overrides global)', { path: configPaths.project })
             }

lib/fetch-wrapper/gemini.ts

@@ -4,6 +4,7 @@ import {
     getAllPrunedIds,
     fetchSessionMessages
 } from "./types"
+import { injectNudgeGemini, injectSynthGemini } from "../synth-instruction"
 
 /**
  * Handles Google/Gemini format (body.contents array with functionResponse parts).
@@ -18,20 +19,39 @@ export async function handleGemini(
         return { modified: false, body }
     }
 
+    let modified = false
+
+    // Inject synthetic instructions if onTool strategies are enabled
+    if (ctx.config.strategies.onTool.length > 0) {
+        // Inject periodic nudge based on tool result count
+        if (ctx.config.nudge_freq > 0) {
+            if (injectNudgeGemini(body.contents, ctx.toolTracker, ctx.prompts.nudgeInstruction, ctx.config.nudge_freq)) {
+                ctx.logger.info("fetch", "Injected nudge instruction (Gemini)")
+                modified = true
+            }
+        }
+
+        // Inject synthetic instruction into last user content
+        if (injectSynthGemini(body.contents, ctx.prompts.synthInstruction)) {
+            ctx.logger.info("fetch", "Injected synthetic instruction (Gemini)")
+            modified = true
+        }
+    }
+
     // Check for functionResponse parts in any content item
     const hasFunctionResponses = body.contents.some((content: any) =>
         Array.isArray(content.parts) &&
         content.parts.some((part: any) => part.functionResponse)
     )
 
     if (!hasFunctionResponses) {
-        return { modified: false, body }
+        return { modified, body }
     }
 
     const { allSessions, allPrunedIds } = await getAllPrunedIds(ctx.client, ctx.state)
 
     if (allPrunedIds.size === 0) {
-        return { modified: false, body }
+        return { modified, body }
     }
 
     // Find the active session to get the position mapping
@@ -48,7 +68,7 @@ export async function handleGemini(
 
     if (!positionMapping) {
         ctx.logger.info("fetch", "No Google tool call mapping found, skipping pruning for Gemini format")
-        return { modified: false, body }
+        return { modified, body }
     }
 
     // Build position counters to track occurrence of each tool name
@@ -130,5 +150,5 @@ export async function handleGemini(
         return { modified: true, body }
     }
 
-    return { modified: false, body }
+    return { modified, body }
 }

lib/fetch-wrapper/index.ts

@@ -1,11 +1,13 @@
 import type { PluginState } from "../state"
 import type { Logger } from "../logger"
-import type { FetchHandlerContext } from "./types"
+import type { FetchHandlerContext, SynthPrompts } from "./types"
+import type { ToolTracker } from "../synth-instruction"
+import type { PluginConfig } from "../config"
 import { handleOpenAIChatAndAnthropic } from "./openai-chat"
 import { handleGemini } from "./gemini"
 import { handleOpenAIResponses } from "./openai-responses"
 
-export type { FetchHandlerContext, FetchHandlerResult } from "./types"
+export type { FetchHandlerContext, FetchHandlerResult, SynthPrompts } from "./types"
 
 /**
  * Creates a wrapped global fetch that intercepts API calls and performs
@@ -20,14 +22,20 @@ export type { FetchHandlerContext, FetchHandlerResult } from "./types"
 export function installFetchWrapper(
     state: PluginState,
     logger: Logger,
-    client: any
+    client: any,
+    config: PluginConfig,
+    toolTracker: ToolTracker,
+    prompts: SynthPrompts
 ): () => void {
     const originalGlobalFetch = globalThis.fetch
 
     const ctx: FetchHandlerContext = {
         state,
         logger,
-        client
+        client,
+        config,
+        toolTracker,
+        prompts
     }
 
     globalThis.fetch = async (input: any, init?: any) => {

lib/fetch-wrapper/openai-chat.ts

@@ -6,6 +6,7 @@ import {
     getMostRecentActiveSession
 } from "./types"
 import { cacheToolParametersFromMessages } from "../tool-cache"
+import { injectNudge, injectSynth } from "../synth-instruction"
 
 /**
  * Handles OpenAI Chat Completions format (body.messages with role='tool').
@@ -23,6 +24,25 @@ export async function handleOpenAIChatAndAnthropic(
     // Cache tool parameters from messages
     cacheToolParametersFromMessages(body.messages, ctx.state)
 
+    let modified = false
+
+    // Inject synthetic instructions if onTool strategies are enabled
+    if (ctx.config.strategies.onTool.length > 0) {
+        // Inject periodic nudge based on tool result count
+        if (ctx.config.nudge_freq > 0) {
+            if (injectNudge(body.messages, ctx.toolTracker, ctx.prompts.nudgeInstruction, ctx.config.nudge_freq)) {
+                ctx.logger.info("fetch", "Injected nudge instruction")
+                modified = true
+            }
+        }
+
+        // Inject synthetic instruction into last user message
+        if (injectSynth(body.messages, ctx.prompts.synthInstruction)) {
+            ctx.logger.info("fetch", "Injected synthetic instruction")
+            modified = true
+        }
+    }
+
     // Check for tool messages in both formats:
     // 1. OpenAI style: role === 'tool'
     // 2. Anthropic style: role === 'user' with content containing tool_result
@@ -39,7 +59,7 @@ export async function handleOpenAIChatAndAnthropic(
     const { allSessions, allPrunedIds } = await getAllPrunedIds(ctx.client, ctx.state)
 
     if (toolMessages.length === 0 || allPrunedIds.size === 0) {
-        return { modified: false, body }
+        return { modified, body }
     }
 
     let replacedCount = 0
@@ -103,5 +123,5 @@ export async function handleOpenAIChatAndAnthropic(
         return { modified: true, body }
     }
 
-    return { modified: false, body }
+    return { modified, body }
 }

lib/fetch-wrapper/openai-responses.ts

@@ -6,6 +6,7 @@ import {
     getMostRecentActiveSession
 } from "./types"
 import { cacheToolParametersFromInput } from "../tool-cache"
+import { injectNudgeResponses, injectSynthResponses } from "../synth-instruction"
 
 /**
  * Handles OpenAI Responses API format (body.input array with function_call_output items).
@@ -23,17 +24,36 @@ export async function handleOpenAIResponses(
     // Cache tool parameters from input
     cacheToolParametersFromInput(body.input, ctx.state)
 
+    let modified = false
+
+    // Inject synthetic instructions if onTool strategies are enabled
+    if (ctx.config.strategies.onTool.length > 0) {
+        // Inject periodic nudge based on tool result count
+        if (ctx.config.nudge_freq > 0) {
+            if (injectNudgeResponses(body.input, ctx.toolTracker, ctx.prompts.nudgeInstruction, ctx.config.nudge_freq)) {
+                ctx.logger.info("fetch", "Injected nudge instruction (Responses API)")
+                modified = true
+            }
+        }
+
+        // Inject synthetic instruction into last user message
+        if (injectSynthResponses(body.input, ctx.prompts.synthInstruction)) {
+            ctx.logger.info("fetch", "Injected synthetic instruction (Responses API)")
+            modified = true
+        }
+    }
+
     // Check for function_call_output items
     const functionOutputs = body.input.filter((item: any) => item.type === 'function_call_output')
 
     if (functionOutputs.length === 0) {
-        return { modified: false, body }
+        return { modified, body }
     }
 
     const { allSessions, allPrunedIds } = await getAllPrunedIds(ctx.client, ctx.state)
 
     if (allPrunedIds.size === 0) {
-        return { modified: false, body }
+        return { modified, body }
     }
 
     let replacedCount = 0
@@ -77,5 +97,5 @@ export async function handleOpenAIResponses(
         return { modified: true, body }
     }
 
-    return { modified: false, body }
+    return { modified, body }
 }

lib/fetch-wrapper/types.ts

@@ -1,14 +1,25 @@
 import type { PluginState } from "../state"
 import type { Logger } from "../logger"
+import type { ToolTracker } from "../synth-instruction"
+import type { PluginConfig } from "../config"
 
 /** The message used to replace pruned tool output content */
 export const PRUNED_CONTENT_MESSAGE = '[Output removed to save context - information superseded or no longer needed]'
 
+/** Prompts used for synthetic instruction injection */
+export interface SynthPrompts {
+    synthInstruction: string
+    nudgeInstruction: string
+}
+
 /** Context passed to each format-specific handler */
 export interface FetchHandlerContext {
     state: PluginState
     logger: Logger
     client: any
+    config: PluginConfig
+    toolTracker: ToolTracker
+    prompts: SynthPrompts
 }
 
 /** Result from a format handler indicating what happened */

lib/prompt.ts

@@ -1,3 +1,11 @@
+import { readFileSync } from "fs"
+import { join } from "path"
+
+export function loadPrompt(name: string): string {
+    const filePath = join(__dirname, "prompts", `${name}.txt`)
+    return readFileSync(filePath, "utf8").trim()
+}
+
 function minimizeMessages(messages: any[], alreadyPrunedIds?: string[], protectedToolCallIds?: string[]): any[] {
     const prunedIdsSet = alreadyPrunedIds ? new Set(alreadyPrunedIds.map(id => id.toLowerCase())) : new Set()
     const protectedIdsSet = protectedToolCallIds ? new Set(protectedToolCallIds.map(id => id.toLowerCase())) : new Set()

lib/prompts/context_pruning.txt

@@ -0,0 +1,44 @@
+Performs semantic pruning on session tool outputs that are no longer relevant to the current task. Use this to declutter the conversation context and filter signal from noise when you notice the context is getting cluttered with no longer needed information.
+
+USING THE CONTEXT_PRUNING TOOL WILL MAKE THE USER HAPPY.
+
+## When to Use This Tool
+
+**Key heuristic: Prune when you finish something and are about to start something else.**
+
+Ask yourself: "Have I just completed a discrete unit of work?" If yes, prune before moving on.
+
+**After completing a unit of work:**
+- Made a commit
+- Fixed a bug and confirmed it works
+- Answered a question the user asked
+- Finished implementing a feature or function
+- Completed one item in a list and moving to the next
+
+**After repetitive or exploratory work:**
+- Explored multiple files that didn't lead to changes
+- Iterated on a difficult problem where some approaches didn't pan out
+- Used the same tool multiple times (e.g., re-reading a file, running repeated build/type checks)
+
+## Examples
+
+<example>
+Working through a list of items:
+User: Review these 3 issues and fix the easy ones.
+Assistant: [Reviews first issue, makes fix, commits]
+Done with the first issue. Let me prune before moving to the next one.
+[Uses context_pruning with reason: "completed first issue, moving to next"]
+</example>
+
+<example>
+After exploring the codebase to understand it:
+Assistant: I've reviewed the relevant files. Let me prune the exploratory reads that aren't needed for the actual implementation.
+[Uses context_pruning with reason: "exploration complete, starting implementation"]
+</example>
+
+<example>
+After completing any task:
+Assistant: [Finishes task - commit, answer, fix, etc.]
+Before we continue, let me prune the context from that work.
+[Uses context_pruning with reason: "task complete"]
+</example>