nanocode-cli 0.5.4__tar.gz → 0.5.9__tar.gz

nanocode_cli-0.5.9/MANIFEST.in

@@ -0,0 +1,4 @@
+include LICENSE
+include README.zh-CN.md
+prune demo*
+prune tests

{nanocode_cli-0.5.4/nanocode_cli.egg-info → nanocode_cli-0.5.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nanocode-cli
-Version: 0.5.4
+Version: 0.5.9
 Summary: A small terminal coding agent written in Python
 Author-email: hit9 <hit9@icloud.com>
 License-Expression: BSD-3-Clause
@@ -23,6 +23,7 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: anthropic>=0.64.0
 Requires-Dist: code-symbol-index>=0.1.13
+Requires-Dist: json-repair
 Requires-Dist: openai>=2.37.0
 Requires-Dist: prompt-toolkit>=3.0
 Requires-Dist: rich>=13.0
@@ -36,18 +37,22 @@ Dynamic: license-file
 
 A small terminal coding agent written in Python.
 
+[简体中文](README.zh-CN.md)
+
 nanocode is pre-1.0 software. Commands, configuration, and tool behavior may change before a stable release.
 
 ![nanocode screenshot](snapshots/nanocode-snapshot.png)
 
 ## Features
 
-- **Snapshot context**: Every turn starts from a clean workspace snapshot instead of a bloated chat replay.
-- **File memory**: Recent reads and edits stay available as focused, line-numbered file ranges.
-- **Anchored edits**: `line:hash` anchors catch stale edits before they touch the wrong code.
-- **Symbol index**: Jump from names to outlines, references, and changed files without searching blindly.
-- **Tool recall**: Big outputs are bounded in the prompt but still recallable later by `tr.N`.
-- **Terminal-native UI**: Model picking, history search, confirmations, live command output, and status all stay in the terminal.
+- **Live turn control**: Add follow-up input while the agent is still working, without losing the current tool flow.
+- **File-state brain**: Reads and edits build a current, line-numbered view of the files that matter now.
+- **Stale-edit protection**: `line:hash` anchors reject edits when the target code has drifted.
+- **Project-aware navigation**: Use the symbol index to jump through outlines, references, and changed files quickly.
+- **Recoverable context**: Tool output stays bounded in the prompt, while raw `tr.N` results remain recallable.
+- **Cache-aware context**: Stable sections stay early and noisy working state stays late to improve prompt-cache reuse.
+- **Focused working memory**: `Note` separates goal, plan, and known facts from noisy execution logs.
+- **Terminal-first workflow**: Model selection, history search, confirmations, live command output, appended input, and status all stay in one CLI.
 
 ## Install
 
@@ -55,6 +60,12 @@ nanocode is pre-1.0 software. Commands, configuration, and tool behavior may cha
 uv tool install nanocode-cli
 ```
 
+Upgrade:
+
+```sh
+uv tool upgrade nanocode-cli
+```
+
 For local development:
 
 ```sh
@@ -77,6 +88,8 @@ Useful arguments:
 - `--yolo`: skip confirmations for mutating tools.
 - `-v`, `--version`: show the version.
 
+During a running turn, the `+>` prompt accepts follow-up input for the next model request.
+
 ## Commands
 
 - `/help`: show commands and tools.
@@ -99,9 +112,10 @@ Interactive selectors support `j`/`k`, arrows, `/` search, Enter, and Esc. Input
 
 - File: `Read`, `LineCount`, `List`, `Find`, `Search`.
 - Code index: `InspectCode`.
-- Edit: `CreateFile`, `Edit`.
+- Edit: `Edit` creates or patches file content.
 - Shell: `Bash`, `Git`.
-- Tool results: `Recall`, `Forget`.
+- Tool results: `Recall`.
+- Working notes: `Note`.
 
 `Read`, `Search`, and `InspectCode` return line anchors where useful. `Edit` uses current `line:hash` anchors to reject stale edits.
 
@@ -126,7 +140,7 @@ Main fields:
 
 ## Context Design
 
-Each model request is built manually as one system message and one user message. The user message is a structured context snapshot, ordered from stable sections to volatile sections so provider prompt caching can reuse the prefix.
+Each model request is built manually from explicit messages. Stable context comes first, conversation stays as messages, working memory follows, and the latest file state is appended at the end.
 
 ```text
 model request
@@ -136,27 +150,27 @@ model request
 +--------------------------------------------------+
 | user                                             |
 |   Environment                                   |
-|   State                                         |
-|   Summary                                       |
-|   Recent Conversation                           |
-|   Tool Result Index                             |
-|   File Context                                  |
-|   Discovery Context                             |
-|   Error Feedback                                |
-|   Latest Tool Results                           |
-|   Current Turn Conversation                     |
++--------------------------------------------------+
+| user/assistant                                  |
+|   conversation, compacted summaries, tools      |
++--------------------------------------------------+
+| user                                             |
+|   Memory: goal, plan, known, date               |
++--------------------------------------------------+
+| user                                             |
+|   FILE STATE: latest Read/Edit file view        |
 +--------------------------------------------------+
 ```
 
 Core rules:
 
-- File Context is rebuilt dynamically from active `Read` and `Edit` results.
+- Mid-turn assistant text and appended user input are kept as conversation.
+- Earlier conversation is compacted into an explicit summary when the context grows too large.
+- FILE STATE is updated by successful `Read` and `Edit` tools and shows current listed file ranges, with recent files first.
 - Newer file lines overwrite older lines; edit invalidations clear stale ranges.
 - File lines are checked against current file stat or line hash before being shown.
-- Discovery Context contains `Find`, `Search`, and `InspectCode` leads, not source truth.
-- Large tool outputs are bounded in context and can be recalled by `tr.N`.
-- Error Feedback keeps only recent failed tool calls.
-- `Forget` removes stale result keys from the active tool result store.
+- Successful `Read` and `Edit` tool messages point to FILE STATE instead of repeating file bodies.
+- Other tool outputs are bounded in conversation messages and can be recalled by `tr.N`.
 
 ## Safety
 

{nanocode_cli-0.5.4 → nanocode_cli-0.5.9}/README.md

@@ -2,18 +2,22 @@
 
 A small terminal coding agent written in Python.
 
+[简体中文](README.zh-CN.md)
+
 nanocode is pre-1.0 software. Commands, configuration, and tool behavior may change before a stable release.
 
 ![nanocode screenshot](snapshots/nanocode-snapshot.png)
 
 ## Features
 
-- **Snapshot context**: Every turn starts from a clean workspace snapshot instead of a bloated chat replay.
-- **File memory**: Recent reads and edits stay available as focused, line-numbered file ranges.
-- **Anchored edits**: `line:hash` anchors catch stale edits before they touch the wrong code.
-- **Symbol index**: Jump from names to outlines, references, and changed files without searching blindly.
-- **Tool recall**: Big outputs are bounded in the prompt but still recallable later by `tr.N`.
-- **Terminal-native UI**: Model picking, history search, confirmations, live command output, and status all stay in the terminal.
+- **Live turn control**: Add follow-up input while the agent is still working, without losing the current tool flow.
+- **File-state brain**: Reads and edits build a current, line-numbered view of the files that matter now.
+- **Stale-edit protection**: `line:hash` anchors reject edits when the target code has drifted.
+- **Project-aware navigation**: Use the symbol index to jump through outlines, references, and changed files quickly.
+- **Recoverable context**: Tool output stays bounded in the prompt, while raw `tr.N` results remain recallable.
+- **Cache-aware context**: Stable sections stay early and noisy working state stays late to improve prompt-cache reuse.
+- **Focused working memory**: `Note` separates goal, plan, and known facts from noisy execution logs.
+- **Terminal-first workflow**: Model selection, history search, confirmations, live command output, appended input, and status all stay in one CLI.
 
 ## Install
 
@@ -21,6 +25,12 @@ nanocode is pre-1.0 software. Commands, configuration, and tool behavior may cha
 uv tool install nanocode-cli
 ```
 
+Upgrade:
+
+```sh
+uv tool upgrade nanocode-cli
+```
+
 For local development:
 
 ```sh
@@ -43,6 +53,8 @@ Useful arguments:
 - `--yolo`: skip confirmations for mutating tools.
 - `-v`, `--version`: show the version.
 
+During a running turn, the `+>` prompt accepts follow-up input for the next model request.
+
 ## Commands
 
 - `/help`: show commands and tools.
@@ -65,9 +77,10 @@ Interactive selectors support `j`/`k`, arrows, `/` search, Enter, and Esc. Input
 
 - File: `Read`, `LineCount`, `List`, `Find`, `Search`.
 - Code index: `InspectCode`.
-- Edit: `CreateFile`, `Edit`.
+- Edit: `Edit` creates or patches file content.
 - Shell: `Bash`, `Git`.
-- Tool results: `Recall`, `Forget`.
+- Tool results: `Recall`.
+- Working notes: `Note`.
 
 `Read`, `Search`, and `InspectCode` return line anchors where useful. `Edit` uses current `line:hash` anchors to reject stale edits.
 
@@ -92,7 +105,7 @@ Main fields:
 
 ## Context Design
 
-Each model request is built manually as one system message and one user message. The user message is a structured context snapshot, ordered from stable sections to volatile sections so provider prompt caching can reuse the prefix.
+Each model request is built manually from explicit messages. Stable context comes first, conversation stays as messages, working memory follows, and the latest file state is appended at the end.
 
 ```text
 model request
@@ -102,27 +115,27 @@ model request
 +--------------------------------------------------+
 | user                                             |
 |   Environment                                   |
-|   State                                         |
-|   Summary                                       |
-|   Recent Conversation                           |
-|   Tool Result Index                             |
-|   File Context                                  |
-|   Discovery Context                             |
-|   Error Feedback                                |
-|   Latest Tool Results                           |
-|   Current Turn Conversation                     |
++--------------------------------------------------+
+| user/assistant                                  |
+|   conversation, compacted summaries, tools      |
++--------------------------------------------------+
+| user                                             |
+|   Memory: goal, plan, known, date               |
++--------------------------------------------------+
+| user                                             |
+|   FILE STATE: latest Read/Edit file view        |
 +--------------------------------------------------+
 ```
 
 Core rules:
 
-- File Context is rebuilt dynamically from active `Read` and `Edit` results.
+- Mid-turn assistant text and appended user input are kept as conversation.
+- Earlier conversation is compacted into an explicit summary when the context grows too large.
+- FILE STATE is updated by successful `Read` and `Edit` tools and shows current listed file ranges, with recent files first.
 - Newer file lines overwrite older lines; edit invalidations clear stale ranges.
 - File lines are checked against current file stat or line hash before being shown.
-- Discovery Context contains `Find`, `Search`, and `InspectCode` leads, not source truth.
-- Large tool outputs are bounded in context and can be recalled by `tr.N`.
-- Error Feedback keeps only recent failed tool calls.
-- `Forget` removes stale result keys from the active tool result store.
+- Successful `Read` and `Edit` tool messages point to FILE STATE instead of repeating file bodies.
+- Other tool outputs are bounded in conversation messages and can be recalled by `tr.N`.
 
 ## Safety
 

nanocode_cli-0.5.9/README.zh-CN.md

@@ -0,0 +1,142 @@
+# nanocode
+
+一个用 Python 写的小型终端编程代理。
+
+[English](README.md)
+
+nanocode 仍是 1.0 前的软件。稳定版发布前，命令、配置和工具行为都可能变化。
+
+![nanocode screenshot](snapshots/nanocode-snapshot.png)
+
+## 特性
+
+- **实时回合控制**：代理还在工作时，也可以追加输入，不打断当前工具流程。
+- **文件状态大脑**：`Read` 和 `Edit` 会构建当前重要文件的带行号视图。
+- **过期编辑保护**：`line:hash` 锚点会在目标代码漂移后拒绝错误编辑。
+- **项目级导航**：通过符号索引快速查看 outline、references 和变更文件。
+- **可恢复上下文**：prompt 中的工具输出保持有界，原始 `tr.N` 结果仍可按需召回。
+- **缓存友好上下文**：稳定内容靠前，嘈杂的工作状态靠后，提高 prompt cache 复用率。
+- **聚焦工作记忆**：`Note` 把 goal、plan、known facts 从嘈杂执行日志中拆出来。
+- **终端优先工作流**：模型选择、历史搜索、确认、实时命令输出、追加输入和状态展示都在一个 CLI 内完成。
+
+## 安装
+
+```sh
+uv tool install nanocode-cli
+```
+
+升级：
+
+```sh
+uv tool upgrade nanocode-cli
+```
+
+本地开发：
+
+```sh
+uv sync --extra dev
+uv run nanocode
+```
+
+## 使用
+
+启动 CLI：
+
+```sh
+nanocode
+```
+
+常用参数：
+
+- `--config <path>`：使用指定 TOML 配置文件。
+- `--init-config`：创建默认配置文件。
+- `--yolo`：跳过会修改环境的工具确认。
+- `-v`, `--version`：显示版本。
+
+代理运行中，`+>` 提示符可以接收追加输入，并在下一次模型请求中发送。
+
+## 命令
+
+- `/help`：显示命令和工具。
+- `/status`：显示运行状态。
+- `/config`：显示当前配置。
+- `/api [auto|chat|anthropic]`：显示或设置 provider API 格式。
+- `/debug [on|off]`：切换模型 I/O debug trace。
+- `/compact`：立即压缩上下文。
+- `/index [force]`：同步或重建代码符号索引。
+- `/provider [NAME]`：显示或设置 provider。
+- `/model [MODEL]`：显示或设置模型。
+- `/reason`：选择 reasoning effort。
+- `/set KEY VALUE`：设置 provider/runtime 值。
+- `/yolo`：切换工具确认。
+- `/exit`, `/quit`：退出。
+
+交互选择器支持 `j`/`k`、方向键、`/` 搜索、Enter 和 Esc。输入框支持历史、补全和 `Ctrl-R` 历史搜索。
+
+## 工具
+
+- 文件：`Read`, `LineCount`, `List`, `Find`, `Search`。
+- 代码索引：`InspectCode`。
+- 编辑：`Edit` 创建或修改文件内容。
+- Shell：`Bash`, `Git`。
+- 工具结果：`Recall`。
+- 工作笔记：`Note`。
+
+`Read`、`Search` 和 `InspectCode` 会在合适时返回行锚点。`Edit` 使用当前 `line:hash` 锚点拒绝过期编辑。
+
+## 配置
+
+运行：
+
+```sh
+nanocode --init-config
+```
+
+默认配置位置是 `~/.nanocode/config.toml`。
+
+主要字段：
+
+- `[provider] active = "name"`
+- `[provider.<name>]`：`url`, `key`, `model`, `api`, `prompt_cache_key`, `available_models`, `reasoning`, `chat_reasoning`, `temperature`, `timeout`
+- `[paths] data_dir`
+- `[runtime] shell_timeout`, `max_agent_steps`, `max_context_tokens`, `yolo`
+
+`api = "auto"` 会根据 provider/model profile 在 Chat Completions 和 Anthropic Messages 之间选择。`prompt_cache_key = "auto"` 会根据 provider、model、workspace 和工具 schema 名称生成稳定 key。
+
+## 上下文设计
+
+每次模型请求都由 nanocode 手动构建成明确的 messages。稳定上下文在前，会话作为 messages 保留，工作记忆随后，最新文件状态放在末尾。
+
+```text
+model request
++--------------------------------------------------+
+| system                                           |
+|   concise agent contract and tool rules          |
++--------------------------------------------------+
+| user                                             |
+|   Environment                                   |
++--------------------------------------------------+
+| user/assistant                                  |
+|   conversation, compacted summaries, tools      |
++--------------------------------------------------+
+| user                                             |
+|   Memory: goal, plan, known, date               |
++--------------------------------------------------+
+| user                                             |
+|   FILE STATE: latest Read/Edit file view        |
++--------------------------------------------------+
+```
+
+核心规则：
+
+- 回合中的 assistant 文本和用户追加输入都会作为 conversation 保留。
+- 上下文过大时，较早 conversation 会压缩成明确的 summary。
+- FILE STATE 由成功的 `Read` 和 `Edit` 更新，展示当前文件范围，最近文件优先。
+- 更新的文件行会覆盖旧行；edit invalidation 会清理过期范围。
+- 文件行展示前会通过当前文件 stat 或行 hash 校验。
+- 成功的 `Read` 和 `Edit` 工具消息只指向 FILE STATE，不重复塞入文件正文。
+- 其他工具输出在 conversation messages 中保持有界，并可通过 `tr.N` 召回。
+
+## 安全
+
+nanocode 会在启动它的环境中编辑文件和执行 shell 命令。它不提供 sandbox 保护。需要隔离时，请在你自己的 sandbox、容器、虚拟机或其他隔离环境中运行。