Source note
KISS Sorcar: A Stupidly-Simple General-Purpose and Software Engineering AI Assistant
Summary
KISS Sorcar is a local, open-source VS Code assistant for software engineering and general tasks built on a small five-layer agent stack. The paper claims that a simple layered design plus strict continuation, tool use, and git isolation can match or beat stronger commercial coding agents on Terminal Bench 2.0.
Problem
- LLM coding assistants fail on long tasks because context windows run out, one bad step can derail the session, and agents loop in dead ends.
- Generated code often needs cleanup because agents stop before running checks such as linters, type checkers, and tests.
- Changes are hard to inspect or undo when an agent edits the live working tree directly, which matters in real software development.
Approach
- The system uses a five-layer inheritance stack, with each layer adding one job: ReAct execution with budget tracking, continuation across context-limited sub-sessions, coding and browser tools, persistent chat history, and git worktree isolation.
- Continuation is handled by forcing the agent to end a sub-session with a chronological summary of what it did, why it did it, and relevant code snippets; the next sub-session starts from those summaries.
- The software engineering layer gives the model shell, read, edit, write, browser, user-question, Docker, and optional parallel sub-agent tools.
- The outer worktree layer creates a separate git branch and worktree per task, preserves dirty local state through a baseline commit, and supports crash recovery from git metadata.
- The authors keep the framework small, about 1,850 lines of code across the core five agent classes, and report building the system with the system itself over 4.5 months.
Results
- On Terminal Bench 2.0, KISS Sorcar reports a 62.2% overall pass rate with Claude Opus 4.6, equal to 277/445 trial runs across 89 tasks with 5 trials per task.
- It reports 78.7% pass@any (70/89 tasks solved at least once) and 43.8% pass@all (39/89 tasks solved in all five trials).
- The paper compares this to Claude Code at 58% and Cursor Composer 2 at 61.7% on the same benchmark, claiming a small edge over both.
- Trial cost and runtime are reported as $0.45 median and $0.90 mean per trial, with 202 s median and 446 s mean duration.
- Task stability is mixed: 39 always-pass tasks, 19 always-fail tasks, and 31 mixed-result tasks.
- Evaluation details matter for the score: the authors hard-skip 9 tasks they judged infeasible after prior attempts, and skipped tasks still count as failures.