v0.17.6

Highlights

• Script execution mode (RunScript, on by default) — the model can run one JavaScript program in a sandbox that calls the other tools from inside; the nested calls never enter the conversation, only the script's own output does. Batch work across dozens of files or pages now finishes in a single round trip. Toggle per session with /script (no argument toggles).
• Context projection edits — every history message now has a stable id, and a persisted edit record can drop or replace individual messages for the model-facing view only. UI, export, trace and resume keep reading the raw record; compaction cut points and token estimates follow the same projection.
• Paper search tool — search four literature sources at once (arXiv, OpenAlex, Crossref, Europe PMC) with interleaved merging, per-source throttling and partial-failure tolerance.

Improvements

• Message links in results open on click (no modifier key, no copy-paste).
• Reasoning rows and tool bodies share one content column across activity blocks.
• Docs: command table, both READMEs and the code-adjacent guides updated for the new mode and projection invariants.

Fixes

• Nested tool calls stay out of the same-step dedup ledger (a duplicate model call could previously deadlock the turn), emit no model-visible events and consume no block ordinals.
• Nested-call recording caps now only degrade the record instead of refusing to run the call.
• Re-editing the same message no longer double-subtracts from the token gauge; edits on deferred messages survive replay.
• Truncated script output never embeds the untruncated text when the output budget is zero.

---

中文补充

• 脚本执行模式 RunScript（默认开启）：模型可在沙箱里写一段 JavaScript 批量调用其他工具，中间过程不进对话、只回最终结果；/script 输入即切换开关。
• 上下文投影层：消息带稳定身份，可持久化地逐条剔除/替换"发给模型的内容"；界面、导出、回看不受影响，压缩切点与 token 估算按同一口径。
• 论文检索工具：一次并行查 4 个学术源，结果交错合并、按源限速、单源失败不影响整体。
• 其他：结果里的链接单击即开；推理与工具正文统一列；嵌套调用不再参与去重账本（修复可能挂死回合的问题）、超限只降级记录；文档同步。