fix: context trim remove tools first #1196

zerob13 · 2025-12-19T08:23:40Z

Summary by CodeRabbit

New Features
- Enhanced conversation context handling with support for function-call and vision-aware prompts and smarter pruning when context limits are approached.
Style
- Updated AppBar navigation icon (globe → compass).
Tests
- Added comprehensive tests validating context selection, token-aware pruning, and handling of tool-call content.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

coderabbitai · 2025-12-19T08:23:50Z

Note

Other AI code review bot(s) detected

CodeRabbit has detected other AI code review bot(s) in this pull request and will avoid duplicating their findings in the review comments. This may lead to a less comprehensive review.

Walkthrough

Reworks prompt builder's context selection to be token-aware: adds token-calculation helpers, removes/compresses assistant tool_call blocks, and iteratively prunes user–assistant pairs while supporting function-call and vision flags. Adds tests for these flows and a single UI icon swap.

Changes

Cohort / File(s)	Change Summary
Token-aware context selection & pruning `src/main/presenter/threadPresenter/utils/promptBuilder.ts`	Replaced/expanded context-selection logic: updated `selectContextMessages` signature to accept `supportsFunctionCall` and `vision`; added token helpers (`calculateToolCallTokens`, `calculateMessageTokens`, `calculateMessagesTokens`, `calculateToolCallBlockTokens`), message utilities (`cloneMessageWithContent`, `removeToolCallsFromAssistant`, `compressToolCallsFromContext`), and iterative pruning/compression that strips tool_call blocks, removes oldest user–assistant pairs (with parent/variant-aware logic), eliminates orphan assistants, and reorders context chronologically.
Test coverage for context selection `test/main/presenter/threadPresenter/promptBuilder.test.ts`	Added comprehensive tests and helpers to simulate token usage and verify: dropping oldest pairs, removing tool_call blocks before pruning, and combined removal+pruning behavior when over token budget.
UI icon update `src/renderer/shell/components/AppBar.vue`	Replaced AppBar browser action icon from `lucide:globe` to `lucide:compass`.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Files needing extra attention:
- promptBuilder.ts: verify token-count math, correctness of tool_call stripping/compression, parentId/variant pairing logic, and chronological/orphan handling.
- Tests: ensure token mocks and approximations align with production token calculations and that edge cases (all-tool-call contexts, mixed vision/function flags) are covered.
- Public API change: callers of selectContextMessages must be updated to supply new flags—verify usages across codebase.

Possibly related PRs

fix: toolcall failed on deepseek #1155 — Related changes to tool_call handling and ID/response generation in promptBuilder.ts.
fix: Multiple Issues Affecting Conversation Flow and Model Settings #1166 — Related modifications to context-building logic and use of buildUserMessageContext in promptBuilder.ts.
fix: add tool call context for better conv #1081 — Related updates to how assistant tool_call content is represented and processed in the prompt builder.

Suggested labels

codex

Poem

🐰 I nibble tokens, tidy the trail,
Strip the calls where answers pale,
Prune old pairs, keep the fresh and bright,
Vision and compass guide the night,
Hoppity hops — prompts now sail!

Pre-merge checks and finishing touches

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	You can run `@coderabbitai generate docstrings` to improve docstring coverage.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the main change: prioritizing tool call removal during context trimming when token budget is exceeded, which aligns with the core functionality in promptBuilder.ts.

✨ Finishing touches

📝 Generate docstrings

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment
Commit unit tests in branch feat/remove-tools-on-context

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

coderabbitai

Actionable comments posted: 1

🧹 Nitpick comments (2)

test/main/presenter/threadPresenter/promptBuilder.test.ts (1)

83-83: Consider adding tests for additional edge cases.

The current tests cover the main scenarios well. For more comprehensive coverage, consider adding tests for:

remainingContextLength <= 0 returns empty array

Orphaned assistant message removal (assistant without corresponding user)

is_variant prioritization when multiple assistants match the same user

vision=true parameter behavior
src/main/presenter/threadPresenter/utils/promptBuilder.ts (1)
506-511: Minor: Redundant token recalculation.

Line 508 calculates totalTokens = Math.max(0, totalTokens - removedTokens), but line 510 immediately recalculates totalTokens = calculateMessagesTokens(chatMessages). The line 508 calculation is redundant.
🔎 Proposed simplification
     if (supportsFunctionCall) {
       const { removedTokens } = compressToolCallsFromContext(
         selectedMessages,
         excessTokens,
         supportsFunctionCall
       )

-      totalTokens = Math.max(0, totalTokens - removedTokens)
       chatMessages = addContextMessages(selectedMessages, vision, supportsFunctionCall)
       totalTokens = calculateMessagesTokens(chatMessages)
     }

📜 Review details

Configuration used: defaults

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 6166012 and ca7afeb.

📒 Files selected for processing (3)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (3 hunks)
src/renderer/shell/components/AppBar.vue (1 hunks)
test/main/presenter/threadPresenter/promptBuilder.test.ts (1 hunks)

🧰 Additional context used

📓 Path-based instructions (21)

**/*.{ts,tsx,js,jsx,vue}

📄 CodeRabbit inference engine (CLAUDE.md)

Use English for logs and comments (Chinese text exists in legacy code, but new code should use English)

Files:

src/renderer/shell/components/AppBar.vue
test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.vue

📄 CodeRabbit inference engine (CLAUDE.md)

**/*.vue: Use Vue 3 Composition API for all components instead of Options API
Use Tailwind CSS with scoped styles for component styling

Files:

src/renderer/shell/components/AppBar.vue

src/renderer/**/*.vue

📄 CodeRabbit inference engine (CLAUDE.md)

src/renderer/**/*.vue: All user-facing strings must use i18n keys via vue-i18n for internationalization
Ensure proper error handling and loading states in all UI components
Implement responsive design using Tailwind CSS utilities for all UI components

src/renderer/**/*.vue: Use composition API and declarative programming patterns; avoid options API
Structure files: exported component, composables, helpers, static content, types
Use PascalCase for component names (e.g., AuthWizard.vue)
Use Vue 3 with TypeScript, leveraging defineComponent and PropType
Use template syntax for declarative rendering
Use Shadcn Vue, Radix Vue, and Tailwind for components and styling
Implement responsive design with Tailwind CSS; use a mobile-first approach
Use Suspense for asynchronous components
Use <script setup> syntax for concise component definitions
Prefer 'lucide:' icon family as the primary choice for Iconify icons
Import Icon component from '@iconify/vue' and use with lucide icons following pattern '{collection}:{icon-name}'

Files:

src/renderer/shell/components/AppBar.vue

src/**/*

📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)

New features should be developed in the src directory

Files:

src/renderer/shell/components/AppBar.vue
src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/renderer/**/*.{vue,js,ts}

📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)

Renderer process code should be placed in src/renderer (Vue 3 application)

Files:

src/renderer/shell/components/AppBar.vue

src/renderer/**/*.{ts,tsx,vue}

📄 CodeRabbit inference engine (.cursor/rules/vue-shadcn.mdc)

src/renderer/**/*.{ts,tsx,vue}: Write concise, technical TypeScript code with accurate examples
Use descriptive variable names with auxiliary verbs (e.g., isLoading, hasError)
Avoid enums; use const objects instead
Use arrow functions for methods and computed properties
Avoid unnecessary curly braces in conditionals; use concise syntax for simple statements

Vue 3 app code in src/renderer/src should be organized into components/, stores/, views/, i18n/, lib/ directories with shell UI in src/renderer/shell/

Files:

src/renderer/shell/components/AppBar.vue

src/renderer/**

📄 CodeRabbit inference engine (.cursor/rules/vue-shadcn.mdc)

Use lowercase with dashes for directories (e.g., components/auth-wizard)

Files:

src/renderer/shell/components/AppBar.vue

src/renderer/**/*.{ts,vue}

📄 CodeRabbit inference engine (.cursor/rules/vue-shadcn.mdc)

src/renderer/**/*.{ts,vue}: Use useFetch and useAsyncData for data fetching
Leverage ref, reactive, and computed for reactive state management
Use provide/inject for dependency injection when appropriate
Use Iconify/Vue for icon implementation

Files:

src/renderer/shell/components/AppBar.vue

src/**/*.{ts,tsx,vue,js,jsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use Prettier with single quotes, no semicolons, and 100 character width

Files:

src/renderer/shell/components/AppBar.vue
src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.{ts,tsx}

📄 CodeRabbit inference engine (CLAUDE.md)

Enable and maintain strict TypeScript type checking for all files

**/*.{ts,tsx}: Always use try-catch to handle possible errors in TypeScript code
Provide meaningful error messages when catching errors
Log detailed error logs including error details, context, and stack traces
Distinguish and handle different error types (UserError, NetworkError, SystemError, BusinessError) with appropriate handlers in TypeScript
Use structured logging with logger.error(), logger.warn(), logger.info(), logger.debug() methods from logging utilities
Do not suppress errors (avoid empty catch blocks or silently ignoring errors)
Provide user-friendly error messages for user-facing errors in TypeScript components
Implement error retry mechanisms for transient failures in TypeScript
Avoid logging sensitive information (passwords, tokens, PII) in logs

Files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

test/**/*.{test,spec}.{ts,tsx,js}

📄 CodeRabbit inference engine (CLAUDE.md)

Use Vitest framework for unit and integration tests

Files:

test/main/presenter/threadPresenter/promptBuilder.test.ts

**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Do not include AI co-authoring information (e.g., 'Co-Authored-By: Claude') in git commits

Files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.{js,ts,jsx,tsx,mjs,cjs}

📄 CodeRabbit inference engine (.cursor/rules/development-setup.mdc)

Write logs and comments in English

Files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

test/**/*.{test,spec}.ts

📄 CodeRabbit inference engine (AGENTS.md)

test/**/*.{test,spec}.ts: Test files should follow the same directory structure as source code under test/main and test/renderer directories
Use Vitest and Vue Test Utils for testing with jsdom configuration
Test files must be named with .test.ts or .spec.ts extension

Files:

test/main/presenter/threadPresenter/promptBuilder.test.ts

src/main/presenter/**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Organize core business logic into dedicated Presenter classes, with one presenter per functional domain

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/main/**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Use EventBus from src/main/eventbus.ts for main-to-renderer communication, broadcasting events via mainWindow.webContents.send()

src/main/**/*.ts: Use EventBus pattern for inter-process communication within the main process to decouple modules
Use Electron's built-in APIs for file system and native dialogs instead of Node.js or custom implementations

src/main/**/*.ts: Electron main process code belongs in src/main/ with presenters in presenter/ (Window/Tab/Thread/Mcp/Config/LLMProvider) and eventbus.ts for app events
Use the Presenter pattern in the main process for UI coordination

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

{src/main/presenter/**/*.ts,src/renderer/**/*.ts}

📄 CodeRabbit inference engine (.cursor/rules/electron-best-practices.mdc)

Implement proper inter-process communication (IPC) patterns using Electron's ipcRenderer and ipcMain APIs

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/main/**/*.{js,ts}

📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)

Main process code for Electron should be placed in src/main

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.{ts,tsx,js,jsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use OxLint for linting JavaScript and TypeScript files

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.{ts,tsx}

📄 CodeRabbit inference engine (AGENTS.md)

src/**/*.{ts,tsx}: Use camelCase for variable and function names in TypeScript files
Use PascalCase for type and class names in TypeScript
Use SCREAMING_SNAKE_CASE for constant names

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.ts

📄 CodeRabbit inference engine (AGENTS.md)

Use EventBus for inter-process communication events

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

🧠 Learnings (13)

📚 Learning: 2025-11-25T05:28:04.454Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/vue-shadcn.mdc:0-0
Timestamp: 2025-11-25T05:28:04.454Z
Learning: Applies to src/renderer/**/*.vue : Prefer 'lucide:' icon family as the primary choice for Iconify icons

Applied to files:

src/renderer/shell/components/AppBar.vue

📚 Learning: 2025-11-25T05:28:04.454Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/vue-shadcn.mdc:0-0
Timestamp: 2025-11-25T05:28:04.454Z
Learning: Applies to src/renderer/**/*.{ts,vue} : Use Iconify/Vue for icon implementation

Applied to files:

src/renderer/shell/components/AppBar.vue

📚 Learning: 2025-11-25T05:28:04.454Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/vue-shadcn.mdc:0-0
Timestamp: 2025-11-25T05:28:04.454Z
Learning: Applies to src/renderer/**/*.vue : Import Icon component from 'iconify/vue' and use with lucide icons following pattern '{collection}:{icon-name}'

Applied to files:

src/renderer/shell/components/AppBar.vue

📚 Learning: 2025-11-25T05:28:04.454Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/vue-shadcn.mdc:0-0
Timestamp: 2025-11-25T05:28:04.454Z
Learning: Applies to src/renderer/**/*.vue : Use PascalCase for component names (e.g., AuthWizard.vue)

Applied to files:

src/renderer/shell/components/AppBar.vue

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : In Provider implementations, include helper methods for Provider-specific operations such as `formatMessages`, `convertToProviderTools`, `parseFunctionCalls`, and `prepareFunctionCallPrompt`

Applied to files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/index.ts : In `src/main/presenter/llmProviderPresenter/index.ts` (`startStreamCompletion`), implement the Agent loop that manages the overall conversation flow, including multiple rounds of LLM calls and tool usage, maintaining `conversationMessages` history, calling `provider.coreStream()` on each iteration, and controlling the loop using `needContinueConversation` and `toolCallCount` (compared against `MAX_TOOL_CALLS`)

Applied to files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/index.ts : In `src/main/presenter/llmProviderPresenter/index.ts`, handle `stop` events by checking `stop_reason`: if `'tool_use'`, add the buffered assistant message and prepare for the next loop iteration; otherwise, add the final assistant message and exit the loop

Applied to files:

test/main/presenter/threadPresenter/promptBuilder.test.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Implement separation of concerns where `src/main/presenter/llmProviderPresenter/index.ts` manages the Agent loop and conversation history, while Provider files handle LLM API interactions, Provider-specific request/response formatting, tool definition conversion, and native vs non-native tool call mechanisms

Applied to files:

test/main/presenter/threadPresenter/promptBuilder.test.ts
src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : In Provider implementations, handle native tool support by converting MCP tools to Provider format using `convertToProviderTools` and including them in the API request; for Providers without native function call support, prepare messages using `prepareFunctionCallPrompt` before making the API call

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/index.ts : In `src/main/presenter/llmProviderPresenter/index.ts`, listen for standardized events yielded by `coreStream` and handle them accordingly: buffer text content (`currentContent`), handle `tool_call_start/chunk/end` events by collecting tool details and calling `presenter.mcpPresenter.callTool`, send frontend events via `eventBus` with tool call status, format tool results for the next LLM call, and set `needContinueConversation = true`

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/main/presenter/mcpPresenter/**/*.ts : Register new MCP tools in `mcpPresenter/index.ts` after implementing them in `inMemoryServers/`

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Convert MCP tools to provider-specific formats and normalize streaming responses to standard events in each provider implementation

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/renderer/**/*.ts : Use the `usePresenter.ts` composable for renderer-to-main IPC communication to call presenter methods directly

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

🧬 Code graph analysis (1)

test/main/presenter/threadPresenter/promptBuilder.test.ts (2)

src/main/presenter/threadPresenter/utils/messageContent.ts (1)

buildUserMessageContext (156-165)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (1)

selectContextMessages (477-584)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

GitHub Check: build-check (x64)

🔇 Additional comments (6)

src/renderer/shell/components/AppBar.vue (1)

93-93: LGTM!

Icon change from lucide:globe to lucide:compass follows the project's Iconify/lucide pattern. Based on learnings, using the lucide: icon family is the preferred choice.

test/main/presenter/threadPresenter/promptBuilder.test.ts (1)

83-176: LGTM! Good test coverage for the context pruning scenarios.

The three test cases effectively cover:

Basic pair dropping without tool calls

Tool call block removal before pair dropping

Combined tool call removal and pair dropping

The tests correctly validate that the last user/assistant pair is preserved even when over the token limit, which aligns with the safeguard in the implementation (lines 523-526 in promptBuilder.ts).

src/main/presenter/threadPresenter/utils/promptBuilder.ts (4)

364-408: LGTM!

The token calculation helpers are well-structured with clear separation of concerns:

calculateToolCallTokens for ChatMessage tool calls

calculateToolCallBlockTokens for AssistantMessageBlock tool calls

Correctly handles the distinction between message formats

517-561: LGTM on pair dropping logic.

The algorithm correctly:

Finds oldest user message (first in reversed array)

Matches it with corresponding assistant via parentId

Prefers non-variant assistants when multiple match

Preserves at least one user/assistant pair

Handles orphaned messages in a cleanup pass

568-584: LGTM!

The orphaned assistant cleanup correctly filters out assistants whose parentId doesn't correspond to a user in the selected set. The final reverse ensures chronological order for downstream consumers.

477-483: All call sites of selectContextMessages have been properly updated with the new function signature. The function now correctly requires supportsFunctionCall and vision parameters, and all usages in both the main code (line 139) and test files (lines 94, 128, 166) pass the required five arguments in the correct order.

src/main/presenter/threadPresenter/utils/promptBuilder.ts

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2025-12-19T08:28:06Z

src/main/presenter/threadPresenter/utils/promptBuilder.ts

+        const lastUserIndex = selectedMessages.findLastIndex((msg) => msg.role === 'user')
+        if (userIndex === lastUserIndex) {
+          break


Avoid returning context that still exceeds limits

The trimming loop breaks when only one user message remains (userIndex === lastUserIndex), even if totalTokens still exceeds remainingContextLength. This can leave an over‑budget context (e.g., a single long user/assistant pair) and cause the final prompt to exceed the model’s context window, which the previous implementation avoided by skipping oversized messages entirely. You likely need a fallback that drops the remaining pair or truncates content when the last pair alone is too large.

Useful? React with 👍 / 👎.

coderabbitai

Actionable comments posted: 0

♻️ Duplicate comments (1)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (1)

548-592: Edge case: Context may still exceed token limit when final pair is too large.

The loop exits when only one user message remains (lines 554-556), but doesn't handle the case where this final user/assistant pair still exceeds remainingContextLength. This could result in returning an over-budget context to downstream consumers.

Consider adding a fallback to either:

Truncate the last message content

Return an empty array with a warning

Skip the oversized pair entirely

🧹 Nitpick comments (3)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (3)
410-428: Type safety concern: Array content casting assumes AssistantMessageBlock type.

The function casts array elements to AssistantMessageBlock (line 415), but it's called on all message types including user messages. User messages may have different content structures (UserMessageContent), which could lead to incorrect cloning behavior.

Consider adding a role check or handling different content structures:
🔎 Proposed fix
 function cloneMessageWithContent(message: Message): Message {
   const cloned: Message = { ...message }

   if (Array.isArray(message.content)) {
-    cloned.content = message.content.map((block) => {
-      const clonedBlock: AssistantMessageBlock = { ...(block as AssistantMessageBlock) }
-      if (block.type === 'tool_call' && block.tool_call) {
-        clonedBlock.tool_call = { ...block.tool_call }
-      }
-      return clonedBlock
-    })
+    if (message.role === 'assistant') {
+      cloned.content = message.content.map((block) => {
+        const clonedBlock: AssistantMessageBlock = { ...(block as AssistantMessageBlock) }
+        if (block.type === 'tool_call' && block.tool_call) {
+          clonedBlock.tool_call = { ...block.tool_call }
+        }
+        return clonedBlock
+      })
+    } else {
+      // For user or other roles, deep clone the array
+      cloned.content = JSON.parse(JSON.stringify(message.content))
+    }
   } else if (message.content && typeof message.content === 'object') {
     cloned.content = JSON.parse(JSON.stringify(message.content))
   } else {
     cloned.content = message.content
   }

   return cloned
 }
460-503: Minor: In-place array mutation could be clarified.

The function mutates the input messages array directly (line 489) then returns it as compressedMessages. While this is safe because selectContextMessages creates a cloned array at line 518, the function signature suggests immutability. Consider either documenting this behavior or making the return explicit:
🔎 Suggestion for clarity
+/**
+ * Compresses tool calls from context messages to reduce token count.
+ * Note: This function mutates the input messages array in place.
+ */
 function compressToolCallsFromContext(
   messages: Message[],
   excessTokens: number,
   supportsFunctionCall: boolean
 ): { compressedMessages: Message[]; removedTokens: number } {
528-528: Use console.debug for diagnostic logging.

Per coding guidelines, use structured logging with appropriate log levels. This diagnostic log should use console.debug instead of console.log.
🔎 Proposed fix
-  console.log('totalTokens', totalTokens, 'remainingContextLength', remainingContextLength)
+  console.debug('totalTokens', totalTokens, 'remainingContextLength', remainingContextLength)

📜 Review details

Configuration used: defaults

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between ca7afeb and c5c9e97.

📒 Files selected for processing (1)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (3 hunks)

🧰 Additional context used

📓 Path-based instructions (13)

**/*.{ts,tsx,js,jsx,vue}

📄 CodeRabbit inference engine (CLAUDE.md)

Use English for logs and comments (Chinese text exists in legacy code, but new code should use English)

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.{ts,tsx}

📄 CodeRabbit inference engine (CLAUDE.md)

Enable and maintain strict TypeScript type checking for all files

**/*.{ts,tsx}: Always use try-catch to handle possible errors in TypeScript code
Provide meaningful error messages when catching errors
Log detailed error logs including error details, context, and stack traces
Distinguish and handle different error types (UserError, NetworkError, SystemError, BusinessError) with appropriate handlers in TypeScript
Use structured logging with logger.error(), logger.warn(), logger.info(), logger.debug() methods from logging utilities
Do not suppress errors (avoid empty catch blocks or silently ignoring errors)
Provide user-friendly error messages for user-facing errors in TypeScript components
Implement error retry mechanisms for transient failures in TypeScript
Avoid logging sensitive information (passwords, tokens, PII) in logs

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/main/presenter/**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Organize core business logic into dedicated Presenter classes, with one presenter per functional domain

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/main/**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Use EventBus from src/main/eventbus.ts for main-to-renderer communication, broadcasting events via mainWindow.webContents.send()

src/main/**/*.ts: Use EventBus pattern for inter-process communication within the main process to decouple modules
Use Electron's built-in APIs for file system and native dialogs instead of Node.js or custom implementations

src/main/**/*.ts: Electron main process code belongs in src/main/ with presenters in presenter/ (Window/Tab/Thread/Mcp/Config/LLMProvider) and eventbus.ts for app events
Use the Presenter pattern in the main process for UI coordination

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.ts

📄 CodeRabbit inference engine (CLAUDE.md)

Do not include AI co-authoring information (e.g., 'Co-Authored-By: Claude') in git commits

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

**/*.{js,ts,jsx,tsx,mjs,cjs}

📄 CodeRabbit inference engine (.cursor/rules/development-setup.mdc)

Write logs and comments in English

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

{src/main/presenter/**/*.ts,src/renderer/**/*.ts}

📄 CodeRabbit inference engine (.cursor/rules/electron-best-practices.mdc)

Implement proper inter-process communication (IPC) patterns using Electron's ipcRenderer and ipcMain APIs

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*

📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)

New features should be developed in the src directory

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/main/**/*.{js,ts}

📄 CodeRabbit inference engine (.cursor/rules/project-structure.mdc)

Main process code for Electron should be placed in src/main

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.{ts,tsx,vue,js,jsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use Prettier with single quotes, no semicolons, and 100 character width

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.{ts,tsx,js,jsx}

📄 CodeRabbit inference engine (AGENTS.md)

Use OxLint for linting JavaScript and TypeScript files

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.{ts,tsx}

📄 CodeRabbit inference engine (AGENTS.md)

src/**/*.{ts,tsx}: Use camelCase for variable and function names in TypeScript files
Use PascalCase for type and class names in TypeScript
Use SCREAMING_SNAKE_CASE for constant names

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

src/**/*.ts

📄 CodeRabbit inference engine (AGENTS.md)

Use EventBus for inter-process communication events

Files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

🧠 Learnings (10)

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : In Provider implementations, include helper methods for Provider-specific operations such as `formatMessages`, `convertToProviderTools`, `parseFunctionCalls`, and `prepareFunctionCallPrompt`

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : In Provider implementations, handle native tool support by converting MCP tools to Provider format using `convertToProviderTools` and including them in the API request; for Providers without native function call support, prepare messages using `prepareFunctionCallPrompt` before making the API call

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/index.ts : In `src/main/presenter/llmProviderPresenter/index.ts`, listen for standardized events yielded by `coreStream` and handle them accordingly: buffer text content (`currentContent`), handle `tool_call_start/chunk/end` events by collecting tool details and calling `presenter.mcpPresenter.callTool`, send frontend events via `eventBus` with tool call status, format tool results for the next LLM call, and set `needContinueConversation = true`

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Implement separation of concerns where `src/main/presenter/llmProviderPresenter/index.ts` manages the Agent loop and conversation history, while Provider files handle LLM API interactions, Provider-specific request/response formatting, tool definition conversion, and native vs non-native tool call mechanisms

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/main/presenter/configPresenter/**/*.ts : Store and retrieve custom prompts via `configPresenter.getCustomPrompts()` for config-based data source management

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:27:12.209Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: .cursor/rules/llm-agent-loop.mdc:0-0
Timestamp: 2025-11-25T05:27:12.209Z
Learning: Applies to src/main/presenter/llmProviderPresenter/index.ts : In `src/main/presenter/llmProviderPresenter/index.ts` (`startStreamCompletion`), implement the Agent loop that manages the overall conversation flow, including multiple rounds of LLM calls and tool usage, maintaining `conversationMessages` history, calling `provider.coreStream()` on each iteration, and controlling the loop using `needContinueConversation` and `toolCallCount` (compared against `MAX_TOOL_CALLS`)

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-06-21T15:49:17.044Z

Learnt from: neoragex2002
Repo: ThinkInAIXYZ/deepchat PR: 550
File: src/renderer/src/stores/chat.ts:1011-1035
Timestamp: 2025-06-21T15:49:17.044Z
Learning: In src/renderer/src/stores/chat.ts, the user prefers to keep both `text` and `content` properties in the `handleMeetingInstruction` function's `sendMessage` call, even though they are redundant, rather than removing the `content` property.

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/main/presenter/mcpPresenter/**/*.ts : Register new MCP tools in `mcpPresenter/index.ts` after implementing them in `inMemoryServers/`

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/main/presenter/llmProviderPresenter/providers/*.ts : Convert MCP tools to provider-specific formats and normalize streaming responses to standard events in each provider implementation

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

📚 Learning: 2025-11-25T05:26:11.312Z

Learnt from: CR
Repo: ThinkInAIXYZ/deepchat PR: 0
File: CLAUDE.md:0-0
Timestamp: 2025-11-25T05:26:11.312Z
Learning: Applies to src/renderer/**/*.ts : Use the `usePresenter.ts` composable for renderer-to-main IPC communication to call presenter methods directly

Applied to files:

src/main/presenter/threadPresenter/utils/promptBuilder.ts

🧬 Code graph analysis (1)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (1)

src/shared/types/presenters/legacy.presenters.d.ts (1)

ChatMessage (1709-1709)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

GitHub Check: build-check (x64)

🔇 Additional comments (3)

src/main/presenter/threadPresenter/utils/promptBuilder.ts (3)

364-408: LGTM! Token calculation helpers are well-structured.

The helper functions correctly handle different content types and provide consistent token estimation using the tokenx library.

430-458: Previous mutation concern has been addressed.

The function now properly clones the message via cloneMessageWithContent (line 439) before modifying, preventing side effects on the original contextMessages.

599-614: LGTM! Orphan cleanup and ordering are well-implemented.

The orphan assistant removal logic correctly ensures consistency by filtering assistants whose parentId doesn't correspond to a user in the selected messages. The final reverse restores chronological order for downstream consumers.

zerob13 added 2 commits December 19, 2025 16:03

fix: message list reverse

3d59368

chore: change icon to compass

ca7afeb

coderabbitai bot reviewed Dec 19, 2025

View reviewed changes

src/main/presenter/threadPresenter/utils/promptBuilder.ts Show resolved Hide resolved

chatgpt-codex-connector bot reviewed Dec 19, 2025

View reviewed changes

fix: clone message to trim

c5c9e97

coderabbitai bot reviewed Dec 19, 2025

View reviewed changes

zerob13 merged commit 9f4359e into dev Dec 19, 2025
2 checks passed

zerob13 deleted the feat/remove-tools-on-context branch January 6, 2026 12:18

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix: context trim remove tools first #1196

fix: context trim remove tools first #1196

Uh oh!

zerob13 commented Dec 19, 2025 •

edited by coderabbitai bot

Loading

Uh oh!

coderabbitai bot commented Dec 19, 2025 •

edited

Loading

Other AI code review bot(s) detected

Uh oh!

coderabbitai bot left a comment

Uh oh!

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Dec 19, 2025

Uh oh!

coderabbitai bot left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

fix: context trim remove tools first #1196

fix: context trim remove tools first #1196

Uh oh!

Conversation

zerob13 commented Dec 19, 2025 • edited by coderabbitai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Uh oh!

coderabbitai bot commented Dec 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Other AI code review bot(s) detected

Walkthrough

Changes

Estimated code review effort

Possibly related PRs

Suggested labels

Poem

Pre-merge checks and finishing touches

Uh oh!

coderabbitai bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Dec 19, 2025

Choose a reason for hiding this comment

Uh oh!

coderabbitai bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

zerob13 commented Dec 19, 2025 •

edited by coderabbitai bot

Loading

coderabbitai bot commented Dec 19, 2025 •

edited

Loading