Totoro

Full-Stack & AI Agent Engineer 💻 | INTJ-A

[Totoro @Totoromsy]

@tianyi https://t.co/geTXmWjIG5 自荐一下我做的 dsh-plugin-bridge：任务做到一半想换模式时，先生成可预览、可编辑的交接摘要，再迁入目标模式的新会话继续。原会话不动，也不搬旧工具调用历史，支持文本／Markdown 编辑，已发布到 npm，在 DSH「插件 → 添加插件」中输入 dsh-plugin-bridge 即可安装。

[Totoro @Totoromsy]

@_LuoFuli Thank you and your team for your contributions to the field of AI and the open-source community. This could be an even more promising direction for reinforcement learning.

[Totoro @Totoromsy]

中文补充：这次在全新空 workspace 里重跑，目标仍是 deepseek-v4-flash-vision-exp/high。Bridge 在 kickoff 前保留源 VLM，并把未解析原图带到 minimal 会话；首答 7.9 秒，5/5，0 次工具调用。完整报告和原始数据都在仓库。

[Totoro @Totoromsy]

Bridge now handles real visual handoffs on DSH 0.1.1-rc.2.

An unresolved PNG moved standard → minimal, kept deepseek-v4-flash-vision-exp, and recovered 5/5 hidden facts on the first response — 0 tool calls.

v0.2.6 · 121/121 tests
https://t.co/geTXmWjIG5

[Totoro @Totoromsy]

26 runs + A/B: pro hits 95% probe recall, 100% summary fidelity. Bare restart: 55% — its wins came from the agent burning 2.2M tokens scavenging disk logs with 25+ tool calls. The summary does 95% at half the cost. Data: https://t.co/geTXmWjIG5

[Totoro @Totoromsy]

How: compress history → open a new session under the target preset → attach the summary as its goal. You preview before anything runs; the new session restates its understanding first; the original is read-only — click back any time. Cost: ~2.4K tokens, 0.4% of a session.