Add codemode: typechecked TypeScript execution with tools

[jonastemplestein]

Summary

Adds codemode execution to the agent: the LLM can write TypeScript code blocks that are typechecked and executed, with access to tools for file I/O, shell commands, web fetching, and secrets.

Features

Code Execution

• <codemode>...</codemode> blocks are typechecked with strict TypeScript
• Executed in a Bun subprocess with streamed stdout/stderr
• Type errors trigger agent retry (agent sees errors and can fix)

Available Tools

• t.sendMessage(msg) — show message to USER (no agent turn)
• t.readFile(path) / t.writeFile(path, content) — file I/O
• t.exec(command) — run shell commands
• t.fetch(url) — HTTP requests
• t.getSecret(name) — access secrets (hidden from LLM)

Agent Loop

• console.log() output triggers another agent turn
• t.sendMessage() output goes to user, no turn
• Most tasks complete in one turn
• Max 3 iterations before forced stop

Event Model

• Added triggerAgentTurn property to all events
• LLM triggered by any event with triggerAgentTurn="after-current-turn"
• CodemodeResultEvent persisted with stdout/stderr for conversation history

Files Added

• src/codemode.model.ts — code block parsing, event types
• src/codemode.service.ts — orchestrates parse → typecheck → execute
• src/codemode.repository.ts — stores generated code files
• src/code-executor.service.ts — runs code in Bun subprocess
• src/typechecker.service.ts — TypeScript type checking
• test/codemode.e2e.test.ts — end-to-end tests

All 75 tests pass.

---

Note

Adds codemode enabling LLM-emitted TypeScript to be typechecked and executed via Bun with tools, persisting results to drive an agent loop across CLI/UI/HTTP and a new per-context storage layout.

• Codemode (core):
  • Add codemode pipeline: parse <codemode> blocks, typecheck (TS), execute via Bun subprocess, stream events.
  • New services: CodemodeService, TypecheckService, CodeExecutor, CodemodeRepository.
  • New events: CodemodeResultEvent, CodemodeValidationErrorEvent and codemode streaming events; add triggerAgentTurn to events.
  • LLM prompt building updated to include codemode results/validation.
• Context/Agent:
  • ContextService.addEvents supports codemode and emits codemode events; swaps in codemode system prompt when enabled.
  • Agent continuation driven by triggerAgentTurn with persisted CodemodeResult.
• CLI/UI:
  • Add codemode run subcommand; integrate codemode into chat flow with an agent loop (iteration cap) and colored streaming output.
  • TUI updated to render codemode results/validation and ignore ephemeral codemode stream events.
• HTTP/Adapters:
  • SSE streaming and LayerCode adapter generalized to ContextOrCodemodeEvent.
• Persistence:
  • Contexts now stored per-directory with events.yaml; repository APIs updated (getContextDir, listing by directories).
• Tests/Config:
  • Add E2E and unit tests for codemode; vitest config includes src/**/*.test.ts.
  • ESLint ignores .mini-agent/**.

^{Written by Cursor Bugbot for commit 8f32ec5. This will update automatically on new commits. Configure here.}

[cursor]

Bug: System prompt permanently overwritten when codemode enabled

The ensureCodemodePrompt function replaces the user's original system prompt with CODEMODE_SYSTEM_PROMPT, and this modified event list gets persisted to disk via repo.save(). This permanently overwrites the user's custom system prompt in the context file, even though the codemode prompt replacement is likely intended only for the current LLM request, not for permanent storage. Subsequent loads of the context will return the codemode prompt instead of the user's original.

 

[cursor]

Bug: Multiple codeblocks lose earlier output on later failure

When processing multiple codeblocks in a single assistant response, the tracking variables (stdout, stderr, exitCode, typecheckFailed) accumulate across all blocks. If an early block executes successfully but a later block fails typechecking, the typecheckFailed path at line 175-185 discards all accumulated stdout with stdout: "", losing output from the successful earlier blocks. Similarly, exitCode only captures the last ExecutionComplete event, so if an earlier block fails (exit code 1) but a later block succeeds (exit code 0), the failure is masked.

 

[cursor]

Bug: No execution timeout allows infinite loops to hang agent

The subprocess execution in CodeExecutor has no timeout mechanism. The stream waits indefinitely for process.exitCode to resolve. If LLM-generated code contains an infinite loop (e.g., while(true){}), a blocking operation that never completes, or code that hangs waiting for input, the agent will hang forever with no way to recover except killing the process. This could make the agent unresponsive until manually terminated.

 

[cursor]

Bug: HTTP handler drops valid FileAttachmentEvent inputs

The HTTP event filter only keeps UserMessageEvent using isUserMessage, but InputEvent includes FileAttachmentEvent and CodemodeResultEvent as valid input types. File attachment events would be dropped even if they were properly parsed. Additionally, ScriptInputEvent in server.service.ts only unions UserMessageEvent and SystemPromptEvent, preventing FileAttachmentEvent from being accepted via HTTP at all, making the HTTP endpoint unable to handle file attachments that the CLI supports.

Additional Locations (1)

• src/server.service.ts#L16-L17

 

[cursor]

Bug: Missing iteration limit allows unbounded agent loop recursion

The runAgentContinuation function recursively calls itself at line 252 without any iteration limit, unlike the CLI's runEventStream which uses MAX_AGENT_LOOP_ITERATIONS = 15. This unbounded recursion occurs whenever needsContinuation is true, which happens when CodemodeResultEvent or CodemodeValidationErrorEvent has triggerAgentTurn === "after-current-turn". If the LLM consistently produces typechecking errors, outputs to stdout via console.log(), or fails to include <codemode> tags, the agent loop will recurse indefinitely, potentially causing stack overflow or resource exhaustion in the interactive chat UI.

Additional Locations (1)

• src/cli/chat-ui.ts#L173-L175

 

[cursor]

Bug: CLI agent loop ignores validation error retry trigger

The agent loop in runEventStream only tracks lastCodemodeResult (CodemodeResultEvent) but ignores CodemodeValidationErrorEvent. When the LLM responds without codemode tags, a CodemodeValidationErrorEvent is emitted and persisted with triggerAgentTurn: "after-current-turn", but since lastCodemodeResult remains undefined, the while loop condition fails and the agent never retries. The chat-ui.ts correctly handles this via triggersContinuation checking both event types.