google-gemini / google-gemini/gemini-cli

[Antigravity/Gemini] Autonomous agent self-prompts from Phase 2 Design into Phase 3 Implementation during long single-turn execution

Open
#28,312 1 comment 0 reactions 0 assignees View on GitHub
area/agent Stale status/need-triage
Dominant language
TypeScript
Stars
107k
Forks
14.6k
Avg merge
2d 3h
Merged PRs (30d)
45

Description

When given a bounded Architecture & Design prompt ('Design a production-ready engine... Include database schema, state machine... Work fully autonomously until done'), the model interprets 'Work fully autonomously until done' and 'Include... example execution traces' as authorization to write working prototype code (.py) and unit tests (unittest), rather than terminating when the required .sql schema and .md design documents are complete. When the system automatically injects mid-session CHECKPOINT truncation summaries, the tendency to jump across SDLC boundaries intensifies.

Contributor guide

Open the contributing guide

Research direction

Reproduce the bounded Architecture & Design prompt and observe behavior around the .sql schema, .md design documents, generated .py prototype, and unittest files. Compare a normal long single-turn execution with one containing injected CHECKPOINT truncation summaries. Done means the design phase stops after the schema and documentation are complete without producing implementation code or unit tests.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, sql, typescript
Domain
ai, cli
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.