local-codebase-indexer 0.3.2__tar.gz → 0.4.0__tar.gz

{local_codebase_indexer-0.3.2/local_codebase_indexer.egg-info → local_codebase_indexer-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: local-codebase-indexer
-Version: 0.3.2
+Version: 0.4.0
 Summary: Local codebase indexer and context provider for AI coding assistants: MCP server and CLI with SQL data-access and C# call relations. No cloud, no server.
 Author: sfy257000
 Project-URL: Homepage, https://github.com/sfy257000/local-codebase-indexer
@@ -88,18 +88,20 @@ claude mcp add local-indexer -- local-codebase-indexer --root /path/to/project s
 | Topology | `using` / `import` resolved to file dependencies, inheritance (base lists) and type usage (`invocation`), stored in the `dependencies` table |
 | Semantic | 384-dim chunk embeddings in `vss_chunks`; queries fuse FTS5 keyword, symbol-name and vector rankings with Reciprocal Rank Fusion, then expand to the dependency neighbours of the top 3 files |
 
-Encodings: UTF-8 first, then Big5, then cp1252. Files over 5MB are skipped. Root and nested `.gitignore` files are respected, and `.git`, `bin`, `obj`, `node_modules` and similar directories are always skipped.
+Encodings: UTF-8 first, then Big5, then cp1252. Files over 5MB are skipped. Root and nested `.gitignore` files are respected, and `.git`, `bin`, `obj`, `node_modules` and similar directories are always skipped. `*.min.js` / `*.min.css` are skipped too. Third-party or generated code that is committed to the repo can be excluded with `.ai_context/ignore` (gitignore syntax), e.g. `Scripts/*`, `!Scripts/Custom/`, `Web References/` (keep generated EF database-first models: their DbContext declares the DbSets).
 
 ## Differences from the spec
 
 - **Vector storage**: `sqlite-vss` needs a native extension and is no longer maintained, so `vss_chunks` stores float32 BLOBs and cosine similarity is computed with numpy. No extension to install.
-- **Embedding model**: the default is a built-in hashing embedder (identifier sub-words, bigrams, CJK character bigrams). It is fast and reproducible but **does not understand synonyms**. For real transformer embeddings, export `all-MiniLM-L6-v2` or `bge-small-en-v1.5` to ONNX, install `onnxruntime` and `tokenizers`, and set `INDEXER_ONNX_MODEL=<folder with model.onnx and tokenizer.json>` (the loading, padding, pooling and re-index logic is covered by a test with a tiny synthetic model; retrieval quality with a real model has not been measured, so run `python benchmarks/run_benchmark.py --compare benchmarks/baseline.json` with the variable set to check it; mean pooling suits MiniLM, bge models recommend CLS pooling). Switching embedders triggers a full re-index.
-- **Call relations** for C# are resolved with limited type inference (using / namespaces, field and constructor-injected types, locals, interface to implementation), and every edge carries a confidence: `confirmed` (the receiver type is known), `probable` (reached through an interface or a single candidate) or `candidate` (name match only). There is no Roslyn: no overload resolution by argument type, no delegates / lambdas / extension methods, and call chains are followed one level. Inheritance edges come from the AST and are reliable.
+- **Embedding model**: the default is a built-in hashing embedder (identifier sub-words, bigrams, CJK character bigrams). It is fast and reproducible but **does not understand synonyms**. For real transformer embeddings, export `all-MiniLM-L6-v2` or `bge-small-en-v1.5` to ONNX, install `onnxruntime` and `tokenizers`, and set `INDEXER_ONNX_MODEL=<folder with model.onnx and tokenizer.json>` (mean pooling suits MiniLM, bge models recommend CLS pooling). Switching embedders triggers a full re-index. Measured (0.4.0): `all-MiniLM-L6-v2` raises benchmark MRR by 0.08-0.13, but on a real 30k-chunk C# / Chinese-comment codebase it scored lower than hashing on paraphrased queries (MRR 0.29 vs 0.48; multilingual MiniLM-L12: 0.41) while a full index took 35-43 minutes instead of 2. Hashing stays the default; for domain jargon, `.ai_context/synonyms.json` helps more than a general model.
+- **Call relations** for C# are resolved with limited type inference (using / namespaces, field and constructor-injected types, `Lazy<T>.Value`, locals, `foreach (var x in items.Where(..))` element types, interface to implementation), and every edge carries a confidence: `confirmed` (the receiver type is known), `probable` (reached through an interface or a single candidate) or `candidate` (name match only). There is no Roslyn: no overload resolution by argument type, no delegates / lambdas / extension methods, and call chains are followed one level. Inheritance edges come from the AST and are reliable.
 - Extra schema additions: `symbols.qualified_name`, `chunk_metadata.symbol_id`, and the `raw_refs`, `chunks_fts` and `meta` tables.
 
 ## Data access (SQL)
 
-`data-access <table-or-procedure>` answers "what writes / reads / executes X". It is built from T-SQL (INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC, with alias resolution; CTEs, temp tables and system objects excluded), from C# SQL strings (literals, `+`, interpolation, `string.Format`, consts, locals) passed to Dapper, `SqlCommand` or EF `FromSql`, from stored-procedure names used with `CommandType.StoredProcedure`, and from EF `ToTable` / `[Table]` mappings. A procedure's callers are indirect writers of the tables it writes (followed up to 4 hops). Each entry has a confidence; **dynamic SQL (`EXEC(@sql)`, `sp_executesql`, SQL built from unknown values) cannot be resolved and is listed separately as `uncertain`**, never mixed into the confirmed lists. Known gaps: `FROM a JOIN b ON ..., c` loses `c`, and a stored-procedure name passed to a non-database method is not seen.
+`data-access <table-or-procedure>` answers "what writes / reads / executes X". It is built from T-SQL (INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC, with alias resolution; CTEs, temp tables and system objects excluded), from C# SQL strings (literals, `+`, interpolation, `string.Format`, consts, locals, `StringBuilder.Append` / `AppendLine` / `AppendFormat`) passed to Dapper, `SqlCommand` or EF `FromSql`, from stored-procedure names used with `CommandType.StoredProcedure`, from EF LINQ on `DbSet<T>` members (`ctx.Orders.Where(..)` / `.Add(..)` / `.Remove(..)`, `from o in ctx.Orders`; `probable`, since only the member name is matched) and from EF `ToTable` / `[Table]` mappings. A procedure's callers are indirect writers of the tables it writes (followed up to 4 hops). `programs` (USED BY PROGRAMS) follows confirmed / probable C# calls upward from those accessors to `*Controller` methods (up to 8 hops) (including from an implementation to the callers of the interface method it implements, for `foreach (var step in steps) step.Run()` pipelines) and groups them per controller file, with the Razor pages (`.cshtml`) whose `Url.Action("Action"[, "Controller"])` or AngularJS `ControllerPath + 'Action'` (controller from `$scope.Controller = 'X'`) targets those actions, so a table maps to the screens that touch it. A method that reads a DbSet and calls `SaveChanges` without an explicit `Add` / `Remove` may update tracked entities; it is listed as a possible (`uncertain`) writer. Each entry has a confidence; **dynamic SQL (`EXEC(@sql)`, `sp_executesql`, SQL built from unknown values) cannot be resolved and is listed separately as `uncertain`**, never mixed into the confirmed lists. Known gaps: `FROM a JOIN b ON ..., c` loses `c`, and a stored-procedure name passed to a non-database method is not seen.
+
+Stored procedures that live only in the database can be exported to a folder outside the repo and listed in `.ai_context/extra_roots` (one folder per line, absolute or relative to the root); their files are indexed under their absolute paths.
 
 Natural-language search routes questions such as "who modifies MES_ScrapRecord" or "哪些程式會寫入 MES_Lot" to this index first.
 

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/README.md

@@ -53,18 +53,20 @@ claude mcp add local-indexer -- local-codebase-indexer --root /path/to/project s
 | Topology | `using` / `import` resolved to file dependencies, inheritance (base lists) and type usage (`invocation`), stored in the `dependencies` table |
 | Semantic | 384-dim chunk embeddings in `vss_chunks`; queries fuse FTS5 keyword, symbol-name and vector rankings with Reciprocal Rank Fusion, then expand to the dependency neighbours of the top 3 files |
 
-Encodings: UTF-8 first, then Big5, then cp1252. Files over 5MB are skipped. Root and nested `.gitignore` files are respected, and `.git`, `bin`, `obj`, `node_modules` and similar directories are always skipped.
+Encodings: UTF-8 first, then Big5, then cp1252. Files over 5MB are skipped. Root and nested `.gitignore` files are respected, and `.git`, `bin`, `obj`, `node_modules` and similar directories are always skipped. `*.min.js` / `*.min.css` are skipped too. Third-party or generated code that is committed to the repo can be excluded with `.ai_context/ignore` (gitignore syntax), e.g. `Scripts/*`, `!Scripts/Custom/`, `Web References/` (keep generated EF database-first models: their DbContext declares the DbSets).
 
 ## Differences from the spec
 
 - **Vector storage**: `sqlite-vss` needs a native extension and is no longer maintained, so `vss_chunks` stores float32 BLOBs and cosine similarity is computed with numpy. No extension to install.
-- **Embedding model**: the default is a built-in hashing embedder (identifier sub-words, bigrams, CJK character bigrams). It is fast and reproducible but **does not understand synonyms**. For real transformer embeddings, export `all-MiniLM-L6-v2` or `bge-small-en-v1.5` to ONNX, install `onnxruntime` and `tokenizers`, and set `INDEXER_ONNX_MODEL=<folder with model.onnx and tokenizer.json>` (the loading, padding, pooling and re-index logic is covered by a test with a tiny synthetic model; retrieval quality with a real model has not been measured, so run `python benchmarks/run_benchmark.py --compare benchmarks/baseline.json` with the variable set to check it; mean pooling suits MiniLM, bge models recommend CLS pooling). Switching embedders triggers a full re-index.
-- **Call relations** for C# are resolved with limited type inference (using / namespaces, field and constructor-injected types, locals, interface to implementation), and every edge carries a confidence: `confirmed` (the receiver type is known), `probable` (reached through an interface or a single candidate) or `candidate` (name match only). There is no Roslyn: no overload resolution by argument type, no delegates / lambdas / extension methods, and call chains are followed one level. Inheritance edges come from the AST and are reliable.
+- **Embedding model**: the default is a built-in hashing embedder (identifier sub-words, bigrams, CJK character bigrams). It is fast and reproducible but **does not understand synonyms**. For real transformer embeddings, export `all-MiniLM-L6-v2` or `bge-small-en-v1.5` to ONNX, install `onnxruntime` and `tokenizers`, and set `INDEXER_ONNX_MODEL=<folder with model.onnx and tokenizer.json>` (mean pooling suits MiniLM, bge models recommend CLS pooling). Switching embedders triggers a full re-index. Measured (0.4.0): `all-MiniLM-L6-v2` raises benchmark MRR by 0.08-0.13, but on a real 30k-chunk C# / Chinese-comment codebase it scored lower than hashing on paraphrased queries (MRR 0.29 vs 0.48; multilingual MiniLM-L12: 0.41) while a full index took 35-43 minutes instead of 2. Hashing stays the default; for domain jargon, `.ai_context/synonyms.json` helps more than a general model.
+- **Call relations** for C# are resolved with limited type inference (using / namespaces, field and constructor-injected types, `Lazy<T>.Value`, locals, `foreach (var x in items.Where(..))` element types, interface to implementation), and every edge carries a confidence: `confirmed` (the receiver type is known), `probable` (reached through an interface or a single candidate) or `candidate` (name match only). There is no Roslyn: no overload resolution by argument type, no delegates / lambdas / extension methods, and call chains are followed one level. Inheritance edges come from the AST and are reliable.
 - Extra schema additions: `symbols.qualified_name`, `chunk_metadata.symbol_id`, and the `raw_refs`, `chunks_fts` and `meta` tables.
 
 ## Data access (SQL)
 
-`data-access <table-or-procedure>` answers "what writes / reads / executes X". It is built from T-SQL (INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC, with alias resolution; CTEs, temp tables and system objects excluded), from C# SQL strings (literals, `+`, interpolation, `string.Format`, consts, locals) passed to Dapper, `SqlCommand` or EF `FromSql`, from stored-procedure names used with `CommandType.StoredProcedure`, and from EF `ToTable` / `[Table]` mappings. A procedure's callers are indirect writers of the tables it writes (followed up to 4 hops). Each entry has a confidence; **dynamic SQL (`EXEC(@sql)`, `sp_executesql`, SQL built from unknown values) cannot be resolved and is listed separately as `uncertain`**, never mixed into the confirmed lists. Known gaps: `FROM a JOIN b ON ..., c` loses `c`, and a stored-procedure name passed to a non-database method is not seen.
+`data-access <table-or-procedure>` answers "what writes / reads / executes X". It is built from T-SQL (INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC, with alias resolution; CTEs, temp tables and system objects excluded), from C# SQL strings (literals, `+`, interpolation, `string.Format`, consts, locals, `StringBuilder.Append` / `AppendLine` / `AppendFormat`) passed to Dapper, `SqlCommand` or EF `FromSql`, from stored-procedure names used with `CommandType.StoredProcedure`, from EF LINQ on `DbSet<T>` members (`ctx.Orders.Where(..)` / `.Add(..)` / `.Remove(..)`, `from o in ctx.Orders`; `probable`, since only the member name is matched) and from EF `ToTable` / `[Table]` mappings. A procedure's callers are indirect writers of the tables it writes (followed up to 4 hops). `programs` (USED BY PROGRAMS) follows confirmed / probable C# calls upward from those accessors to `*Controller` methods (up to 8 hops) (including from an implementation to the callers of the interface method it implements, for `foreach (var step in steps) step.Run()` pipelines) and groups them per controller file, with the Razor pages (`.cshtml`) whose `Url.Action("Action"[, "Controller"])` or AngularJS `ControllerPath + 'Action'` (controller from `$scope.Controller = 'X'`) targets those actions, so a table maps to the screens that touch it. A method that reads a DbSet and calls `SaveChanges` without an explicit `Add` / `Remove` may update tracked entities; it is listed as a possible (`uncertain`) writer. Each entry has a confidence; **dynamic SQL (`EXEC(@sql)`, `sp_executesql`, SQL built from unknown values) cannot be resolved and is listed separately as `uncertain`**, never mixed into the confirmed lists. Known gaps: `FROM a JOIN b ON ..., c` loses `c`, and a stored-procedure name passed to a non-database method is not seen.
+
+Stored procedures that live only in the database can be exported to a folder outside the repo and listed in `.ai_context/extra_roots` (one folder per line, absolute or relative to the root); their files are indexed under their absolute paths.
 
 Natural-language search routes questions such as "who modifies MES_ScrapRecord" or "哪些程式會寫入 MES_Lot" to this index first.
 

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/README.zh-CN.md

@@ -49,18 +49,20 @@ claude mcp add local-indexer -- local-codebase-indexer --root /path/to/project s
 | Topology | `using` / `import` 解析为文件依赖；继承（base list）；类型使用（`invocation`）。存放在 `dependencies` 表 |
 | Semantic | 代码块嵌入 384 维，存入 `vss_chunks`；查询时用 FTS5 关键字、符号名称、向量三路做 Reciprocal Rank Fusion，再展开前 3 名文件的依赖邻居 |
 
-编码：先尝试 UTF-8，再尝试 Big5、cp1252。单个文件超过 5MB 不索引；遵守根目录与子目录的 `.gitignore`；固定跳过 `.git`、`bin`、`obj`、`node_modules` 等。
+编码：先尝试 UTF-8，再尝试 Big5、cp1252。单个文件超过 5MB 不索引；遵守根目录与子目录的 `.gitignore`；固定跳过 `.git`、`bin`、`obj`、`node_modules` 等。`*.min.js` / `*.min.css` 也跳过。已提交进仓库的第三方或生成的代码可用 `.ai_context/ignore`（gitignore 语法）排除，例如 `Scripts/*`、`!Scripts/Custom/`、`Web References/`（EF database-first 生成的 model 不要排除，DbSet 声明在它的 DbContext 里）。
 
 ## 与规格的差异
 
 - **向量存储**：规格中的 `sqlite-vss` 需要原生扩展且已停止维护，这里 `vss_chunks` 改为存 float32 BLOB，用 numpy 暴力计算 cosine，不需要额外安装扩展。
-- **嵌入模型**：默认是内置的 hashing 嵌入（标识符拆词、双词组、中文双字），快速、可复现，但**不理解同义词**。要使用真正的 transformer，把 `all-MiniLM-L6-v2` 或 `bge-small-en-v1.5` 导出为 ONNX，安装 `onnxruntime`、`tokenizers`，并设置 `INDEXER_ONNX_MODEL=<包含 model.onnx 与 tokenizer.json 的文件夹>`（加载、补齐、池化与重建索引的逻辑已用一个小型合成模型测试过；真实模型的检索质量尚未度量，设置好变量后请运行 `python benchmarks/run_benchmark.py --compare benchmarks/baseline.json` 验证；目前使用平均池化，适合 MiniLM，bge 系列建议使用 CLS 池化）。切换嵌入后会自动重建索引。
-- **C# 调用关系**以有限度的类型推断解析（using / namespace、字段与构造函数注入的类型、局部变量、接口对应实现），每条关系都标置信等级：`confirmed`（接收者类型已知）、`probable`（经由接口或唯一候选）、`candidate`（仅名称相符）。没有 Roslyn，所以不做按参数类型的重载解析、不处理委托 / lambda / 扩展方法，调用链只追一层。继承关系来自 AST，可靠。
+- **嵌入模型**：默认是内置的 hashing 嵌入（标识符拆词、双词组、中文双字），快速、可复现，但**不理解同义词**。要使用真正的 transformer，把 `all-MiniLM-L6-v2` 或 `bge-small-en-v1.5` 导出为 ONNX，安装 `onnxruntime`、`tokenizers`，并设置 `INDEXER_ONNX_MODEL=<包含 model.onnx 与 tokenizer.json 的文件夹>`（目前使用平均池化，适合 MiniLM，bge 系列建议使用 CLS 池化）。切换嵌入后会自动重建索引。实测（0.4.0）：`all-MiniLM-L6-v2` 让 benchmark MRR 提高 0.08–0.13，但在真实的 3 万 chunk C# / 中文注释项目上，换个说法的查询反而比 hashing 差（MRR 0.29 对 0.48；多语 MiniLM-L12 为 0.41），完整建索引从 2 分钟变成 35–43 分钟。因此默认维持 hashing；领域术语用 `.ai_context/synonyms.json` 比通用模型有效。
+- **C# 调用关系**以有限度的类型推断解析（using / namespace、字段与构造函数注入的类型、`Lazy<T>.Value`、局部变量、`foreach (var x in items.Where(..))` 的元素类型、接口对应实现），每条关系都标置信等级：`confirmed`（接收者类型已知）、`probable`（经由接口或唯一候选）、`candidate`（仅名称相符）。没有 Roslyn，所以不做按参数类型的重载解析、不处理委托 / lambda / 扩展方法，调用链只追一层。继承关系来自 AST，可靠。
 - 额外增加了 `symbols.qualified_name`、`chunk_metadata.symbol_id`、`raw_refs`、`chunks_fts` 与 `meta` 表。
 
 ## 数据访问（SQL）
 
-`data-access <数据表或存储过程>` 回答“谁写入 / 读取 / 执行 X”。数据来源：T-SQL（INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC，含别名解析，排除 CTE、临时表、系统对象）、C# 的 SQL 字符串（字面值、`+`、字符串插值、`string.Format`、常量、局部变量）经 Dapper、`SqlCommand` 或 EF `FromSql` 发出、配合 `CommandType.StoredProcedure` 的存储过程名称，以及 EF 的 `ToTable` / `[Table]` 映射。调用某存储过程的代码视为该过程所写数据表的间接写入者（最多追 4 层）。每条都有置信等级；**动态 SQL（`EXEC(@sql)`、`sp_executesql`、由未知值组成的 SQL）无法解析，会单独列为 `uncertain`**，不会混进确定的清单。已知缺口：`FROM a JOIN b ON ..., c` 会漏掉 `c`；把存储过程名称传给非数据库方法时看不到。
+`data-access <数据表或存储过程>` 回答“谁写入 / 读取 / 执行 X”。数据来源：T-SQL（INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC，含别名解析，排除 CTE、临时表、系统对象）、C# 的 SQL 字符串（字面值、`+`、字符串插值、`string.Format`、常量、局部变量、`StringBuilder.Append` / `AppendLine` / `AppendFormat`）经 Dapper、`SqlCommand` 或 EF `FromSql` 发出、配合 `CommandType.StoredProcedure` 的存储过程名称、对 `DbSet<T>` 成员的 EF LINQ（`ctx.Orders.Where(..)` / `.Add(..)` / `.Remove(..)`、`from o in ctx.Orders`；只比对成员名称，所以是 `probable`），以及 EF 的 `ToTable` / `[Table]` 映射。调用某存储过程的代码视为该过程所写数据表的间接写入者（最多追 4 层）。`programs`（USED BY PROGRAMS）沿 confirmed / probable 的 C# 调用往上追到 `*Controller` 方法（最多 8 层），（实现方法也会接到它所实现接口方法的调用者，涵盖 `foreach (var step in steps) step.Run()` 这种管线），按 controller 文件分组，并列出 `Url.Action("Action"[, "Controller"])` 或 AngularJS `ControllerPath + 'Action'`（controller 取自 `$scope.Controller = 'X'`）指向这些 action 的 Razor 页面（`.cshtml`），查一张表就知道哪些画面会动到它。方法读了某 DbSet 又调用 `SaveChanges`、但没有明确 `Add` / `Remove` 时，可能是在更新跟踪中的实体，列为可能（`uncertain`）写入者。每条都有置信等级；**动态 SQL（`EXEC(@sql)`、`sp_executesql`、由未知值组成的 SQL）无法解析，会单独列为 `uncertain`**，不会混进确定的清单。已知缺口：`FROM a JOIN b ON ..., c` 会漏掉 `c`；把存储过程名称传给非数据库方法时看不到。
+
+只存在数据库里的存储过程，可导出到仓库外的文件夹，再写进 `.ai_context/extra_roots`（一行一个文件夹，绝对路径或相对于项目根目录）；这些文件以绝对路径索引。
 
 自然语言搜索遇到“谁修改 MES_ScrapRecord”“哪些程序会写入 MES_Lot”这类问题，会先走这个索引。
 

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/README.zh-TW.md

@@ -49,18 +49,20 @@ claude mcp add local-indexer -- local-codebase-indexer --root /path/to/project s
 | Topology | `using` / `import` 解析成檔案相依；繼承（base list）；型別使用（`invocation`）。存在 `dependencies` 表 |
 | Semantic | 區塊嵌入 384 維，存 `vss_chunks`；查詢時以 FTS5 關鍵字、符號名稱、向量三路做 Reciprocal Rank Fusion，再展開前 3 名檔案的相依鄰居 |
 
-編碼：先試 UTF-8，再試 Big5、cp1252。單檔超過 5MB 不索引；遵守根目錄與子目錄的 `.gitignore`；固定略過 `.git`、`bin`、`obj`、`node_modules` 等。
+編碼：先試 UTF-8，再試 Big5、cp1252。單檔超過 5MB 不索引；遵守根目錄與子目錄的 `.gitignore`；固定略過 `.git`、`bin`、`obj`、`node_modules` 等。`*.min.js` / `*.min.css` 也略過。已 commit 進 repo 的第三方或產生的程式可用 `.ai_context/ignore`（gitignore 語法）排除，例如 `Scripts/*`、`!Scripts/Custom/`、`Web References/`（EF database-first 產生的 model 不要排除，DbSet 宣告在它的 DbContext 裡）。
 
 ## 與規格的差異
 
 - **向量儲存**：規格的 `sqlite-vss` 需要原生擴充且已停止維護，這裡 `vss_chunks` 改存 float32 BLOB，用 numpy 暴力計算 cosine，不需要額外安裝擴充。
-- **嵌入模型**：預設是內建的 hashing 嵌入（識別字拆詞、雙詞組、中文雙字），快速、可重現，但**不理解同義詞**。要用真正的 transformer，匯出 `all-MiniLM-L6-v2` 或 `bge-small-en-v1.5` 為 ONNX，裝 `onnxruntime`、`tokenizers`，並設 `INDEXER_ONNX_MODEL=<含 model.onnx 與 tokenizer.json 的資料夾>`（載入、補齊、池化與重建索引的邏輯已用一個小型合成模型測試過；真實模型的檢索品質尚未量測，設好變數後請執行 `python benchmarks/run_benchmark.py --compare benchmarks/baseline.json` 驗證；目前用平均池化，適合 MiniLM，bge 系列建議用 CLS 池化）。切換嵌入後會自動重建索引。
-- **C# 呼叫關係**以有限度的型別推斷解析（using / namespace、欄位與建構式注入的型別、區域變數、介面對應實作），每條關係都標信心等級：`confirmed`（接收者型別已知）、`probable`（經由介面或唯一候選）、`candidate`（僅名稱相符）。沒有 Roslyn，所以不做依引數型別的多載解析、不處理委派 / lambda / 擴充方法，呼叫鏈只追一層。繼承關係來自 AST，可靠。
+- **嵌入模型**：預設是內建的 hashing 嵌入（識別字拆詞、雙詞組、中文雙字），快速、可重現，但**不理解同義詞**。要用真正的 transformer，匯出 `all-MiniLM-L6-v2` 或 `bge-small-en-v1.5` 為 ONNX，裝 `onnxruntime`、`tokenizers`，並設 `INDEXER_ONNX_MODEL=<含 model.onnx 與 tokenizer.json 的資料夾>`（目前用平均池化，適合 MiniLM，bge 系列建議用 CLS 池化）。切換嵌入後會自動重建索引。實測（0.4.0）：`all-MiniLM-L6-v2` 讓 benchmark MRR 提高 0.08–0.13，但在真實的 3 萬 chunk C# / 中文註解專案上，換句話說的查詢反而比 hashing 差（MRR 0.29 對 0.48；多語 MiniLM-L12 為 0.41），完整建索引從 2 分鐘變成 35–43 分鐘。因此預設維持 hashing；領域行話用 `.ai_context/synonyms.json` 比通用模型有效。
+- **C# 呼叫關係**以有限度的型別推斷解析（using / namespace、欄位與建構式注入的型別、`Lazy<T>.Value`、區域變數、`foreach (var x in items.Where(..))` 的元素型別、介面對應實作），每條關係都標信心等級：`confirmed`（接收者型別已知）、`probable`（經由介面或唯一候選）、`candidate`（僅名稱相符）。沒有 Roslyn，所以不做依引數型別的多載解析、不處理委派 / lambda / 擴充方法，呼叫鏈只追一層。繼承關係來自 AST，可靠。
 - 額外加了 `symbols.qualified_name`、`chunk_metadata.symbol_id`、`raw_refs`、`chunks_fts` 與 `meta` 表。
 
 ## 資料存取（SQL）
 
-`data-access <資料表或預存程序>` 回答「誰寫入 / 讀取 / 執行 X」。資料來源：T-SQL（INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC，含別名解析，排除 CTE、暫存表、系統物件）、C# 的 SQL 字串（字面值、`+`、字串內插、`string.Format`、常數、區域變數）經 Dapper、`SqlCommand` 或 EF `FromSql` 送出、搭配 `CommandType.StoredProcedure` 的 SP 名稱，以及 EF 的 `ToTable` / `[Table]` 對應。呼叫某 SP 的程式視為該 SP 所寫資料表的間接寫入者（最多追 4 層）。每筆都有信心等級；**動態 SQL（`EXEC(@sql)`、`sp_executesql`、由未知值組成的 SQL）無法解析，會獨立列為 `uncertain`**，不會混進確定的清單。已知缺口：`FROM a JOIN b ON ..., c` 會漏掉 `c`；把 SP 名稱傳給非資料庫方法時看不到。
+`data-access <資料表或預存程序>` 回答「誰寫入 / 讀取 / 執行 X」。資料來源：T-SQL（INSERT / UPDATE / DELETE / MERGE / TRUNCATE / SELECT INTO / FROM / JOIN / EXEC，含別名解析，排除 CTE、暫存表、系統物件）、C# 的 SQL 字串（字面值、`+`、字串內插、`string.Format`、常數、區域變數、`StringBuilder.Append` / `AppendLine` / `AppendFormat`）經 Dapper、`SqlCommand` 或 EF `FromSql` 送出、搭配 `CommandType.StoredProcedure` 的 SP 名稱、對 `DbSet<T>` 成員的 EF LINQ（`ctx.Orders.Where(..)` / `.Add(..)` / `.Remove(..)`、`from o in ctx.Orders`；只比對成員名稱，所以是 `probable`），以及 EF 的 `ToTable` / `[Table]` 對應。呼叫某 SP 的程式視為該 SP 所寫資料表的間接寫入者（最多追 4 層）。`programs`（USED BY PROGRAMS）沿 confirmed / probable 的 C# 呼叫往上追到 `*Controller` 方法（最多 8 層），（實作方法也會接到它所實作介面方法的呼叫者，涵蓋 `foreach (var step in steps) step.Run()` 這種管線），依 controller 檔案分組，並列出 `Url.Action("Action"[, "Controller"])` 或 AngularJS `ControllerPath + 'Action'`（controller 取自 `$scope.Controller = 'X'`）指向這些 action 的 Razor 頁面（`.cshtml`），查一張表就知道哪些畫面會動到它。方法讀了某 DbSet 又呼叫 `SaveChanges`、但沒有明確 `Add` / `Remove` 時，可能是在更新追蹤中的實體，列為可能（`uncertain`）寫入者。每筆都有信心等級；**動態 SQL（`EXEC(@sql)`、`sp_executesql`、由未知值組成的 SQL）無法解析，會獨立列為 `uncertain`**，不會混進確定的清單。已知缺口：`FROM a JOIN b ON ..., c` 會漏掉 `c`；把 SP 名稱傳給非資料庫方法時看不到。
+
+只存在資料庫裡的預存程序，可匯出到 repo 外的資料夾，再寫進 `.ai_context/extra_roots`（一行一個資料夾，絕對路徑或相對於專案根目錄）；這些檔案以絕對路徑索引。
 
 自然語言搜尋遇到「誰修改 MES_ScrapRecord」「哪些程式會寫入 MES_Lot」這類問題，會先走這個索引。
 

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/TODO.md

@@ -5,15 +5,14 @@ Priorities come from the benchmark (`benchmarks/run_benchmark.py`, baseline in `
 ## Next
 
 0. **Publish to the MCP registry** (not done yet). PyPI 0.3.0 is out and carries the `mcp-name` marker; `server.json` is ready. Run `mcp-publisher login github` then `mcp-publisher publish` from the repo folder, then verify with `curl "https://registry.modelcontextprotocol.io/v0.1/servers?search=local-codebase-indexer"`. Repeat after each release (bump both versions in `server.json`).
-1. **Check a real embedding model on your own machine** (this environment cannot reach Hugging Face, so the ONNX path is only tested with a tiny generated model). Export a model such as `all-MiniLM-L6-v2` (mean pooling) or `bge-small-en-v1.5` (`INDEXER_ONNX_POOLING=cls`, `INDEXER_ONNX_QUERY_PREFIX="Represent this sentence for searching relevant passages: "`), set `INDEXER_ONNX_MODEL`, and run `python benchmarks/run_benchmark.py --compare benchmarks/baseline.json`.
-   Evidence: `synonym` Recall@5 is only 0.30 (hashing embedder plus the dictionary, which cannot cover free paraphrase). Whether a real model fixes this is not measured.
-2. **Resolve `FROM a JOIN b ON ..., c`** (a comma list after a JOIN loses `c`), non-sink SQL proc-name literals (`Run("usp_X")`), and EF query tracing (only the table mapping is recorded now).
-3. **C# call graph gaps**: overload resolution by argument type, delegates / lambdas / extension methods, call chains longer than one level. Without Roslyn these stay approximate; edges carry `confirmed` / `probable` / `candidate`.
+1. **Resolve `FROM a JOIN b ON ..., c`** (a comma list after a JOIN loses `c`), non-sink SQL proc-name literals (`Run("usp_X")`), EF `this.ToTable(..)` inside `EntityTypeConfiguration<T>` (entity not taken from the base class yet), and EF updates through tracked entities when the entity was loaded in another method than the one calling `SaveChanges` (only same-method read + `SaveChanges` is listed, as uncertain).
+2. **C# call graph gaps**: overload resolution by argument type, delegates / lambdas / extension methods, call chains longer than one level. Without Roslyn these stay approximate; edges carry `confirmed` / `probable` / `candidate`.
 
 ## Deferred (not worth doing yet)
 
 | Item | Why deferred | Revisit when |
 |---|---|---|
+| Real embedding model as default | Measured 0.4.0 on a real 30k-chunk C# codebase with Chinese comments, 12 paraphrased queries (6 zh, 6 en): hashing MRR 0.48 / hit@5 6, `all-MiniLM-L6-v2` 0.29 / 6, `paraphrase-multilingual-MiniLM-L12-v2` (int8) 0.41 / 6; full index 128 s vs 35 / 43 min. Benchmark: MiniLM +0.08-0.13 MRR, multilingual -0.01..+0.03. Domain jargon ("split a lot into two") was missed by all three | A model trained on code + Chinese, or embedding becomes incremental enough that branch switches stay under a minute |
 | ANN index | Exact search is not the bottleneck. Measured (Python stdlib copies, linear growth): 17k chunks cold CLI search 0.35 s / warm 0.13 s / index 10 s; 52k chunks 0.75 s / 0.36 s / 28 s; 104k chunks 1.09 s / 0.73 s / 54 s, 372 MB RSS, 729 MB database. At 104k chunks the cosine product plus top-k takes about 12 ms; loading the vectors out of SQLite (fetch 0.31 s, join 0.12 s) is about half of a warm query | Exact cosine itself becomes slow, which is far beyond 1M chunks |
 | Vector matrix cache | Cheap and measurable: it removes about half of a warm query at 104k chunks (about 0.8 s -> 0.3 s). Not needed below ~50k chunks | A real project passes ~50k chunks. Keep the matrix in the MCP server process, and/or store it as a memory-mapped `.npy` next to the database so the CLI benefits too (invalidate on any index change) |
 | `codebase_trace_execution` | Needs reliable call / data-flow relations; today's usage edges are name matches | Call graph gaps (item 3) are narrowed; today edges carry a confidence level |
@@ -23,6 +22,8 @@ Priorities come from the benchmark (`benchmarks/run_benchmark.py`, baseline in `
 
 ## Done
 
+- Data access: `StringBuilder` SQL, EF LINQ on `DbSet<T>` members (resolved after indexing against declared DbSets), `programs` roll-up from accessors to controller actions; `relations Type.Method`; call graph resolves `Lazy<T>.Value` and `foreach (var x in ...)` element types; programs roll-up walks from an implementation to its interface method's callers and lists Razor pages (`Url.Action`, AngularJS `ControllerPath + 'Action'`); a typed identifier's scope includes files named after it and ranks names matching the rest of the query first ("WD0096 存檔" -> SetRecord); EF `SaveChanges` on tracked entities listed as possible writes; `.ai_context/extra_roots`; `.ai_context/ignore`, minified files skipped; C# namespace imports only link files whose types are used; piped CLI output is UTF-8
+
 - SQL / data-access relations: `data_access` table and tool (`data-access` CLI, `codebase_get_data_access` MCP); T-SQL read / write / exec, C# SQL strings and Dapper / SqlCommand / EF sinks, EF table mappings, indirect writers through procedures, dynamic SQL reported as `uncertain`; search routes "who writes / reads / calls X" to it. Benchmark `sql_dataflow` Recall@5 0.72 -> 1.00, new `data_access` queries 0.67 -> 1.00
 - C# symbol-level calls: using / field / constructor-injection / local type inference, interface dispatch, confidence levels. `relations Save` no longer shares one reference list across four unrelated methods
 - Domain-term query expansion (`local_codebase_indexer/synonyms.py`, extendable with `.ai_context/synonyms.json`); `abbreviation` Recall@5 0.67 -> 1.00, `cross_lingual` 0.50 -> 1.00 (small, author-written sets: optimistic). ONNX pooling / query-prefix options

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/local_codebase_indexer/__init__.py

@@ -1,3 +1,3 @@
 """Universal Local Codebase Indexer & Context Provider (see index.md)."""
 
-__version__ = "0.3.2"
+__version__ = "0.4.0"

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/local_codebase_indexer/cli.py

@@ -1,4 +1,4 @@
-"""Command line interface: ``bin/indexer <tool_name> --params`` (index.md section 4)."""
+"""Command line interface: ``local-codebase-indexer <tool_name> --params`` (index.md section 4)."""
 
 from __future__ import annotations
 
@@ -59,14 +59,14 @@ def build_parser() -> argparse.ArgumentParser:
 
 
 def _fix_stdio(serving: bool) -> None:
-    """MCP speaks UTF-8 JSON; the Windows locale codec (cp950, cp1252) would garble Chinese queries. On a console,
-    keep its encoding but never crash on a character it cannot show."""
+    """MCP and piped output (an AI assistant calling the CLI) are UTF-8; the Windows locale codec (cp950, cp1252)
+    would garble Chinese. On a console, keep its encoding but never crash on a character it cannot show."""
     for name in ("stdin", "stdout", "stderr"):
         stream = getattr(sys, name, None)
         if stream is None or not hasattr(stream, "reconfigure"):
             continue
         try:
-            if serving:
+            if serving or (name != "stdin" and not stream.isatty()):
                 stream.reconfigure(encoding="utf-8", errors="replace")
             elif name != "stdin":
                 stream.reconfigure(errors="replace")

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/local_codebase_indexer/core.py

@@ -117,6 +117,7 @@ class Indexer:
             dbmod.set_meta(conn, "embedder", self.embedder.name)
             dbmod.set_meta(conn, "index_version", dbmod.INDEX_VERSION)
             conn.commit()
+            conn.execute("VACUUM")  # return the old index's pages to the file system
 
         existing = {r["file_path"]: (r["last_modified"], r["md5_hash"]) for r in
                     conn.execute("SELECT file_path, last_modified, md5_hash FROM files")}
@@ -140,6 +141,7 @@ class Indexer:
                 stats.warnings.append(f"{sf.rel_path}: {exc}")
                 continue
             if b"\x00" in data[:8192]:
+                seen.discard(sf.rel_path)  # drop any earlier entry for it
                 stats.skipped += 1
                 continue
             md5 = hashlib.md5(data).hexdigest()
@@ -173,6 +175,8 @@ class Indexer:
         if changed or force:
             resolve_dependencies(conn)
             resolve_calls(conn)
+            resolve_page_actions(conn)
+            resolve_ef_access(conn)
         self._record_run_meta()
         conn.commit()
         stats.chunks = conn.execute("SELECT COUNT(*) FROM chunk_metadata").fetchone()[0]
@@ -273,6 +277,82 @@ def _resolve_relative(source: str, spec: str, files: set[str]) -> str | None:
     return None
 
 
+def _view_route(page_path: str) -> tuple[str, str | None]:
+    """(controller, area) a Razor page belongs to: ``Areas/<area>/Views/<controller>/x.cshtml`` or
+    ``Views/<controller>/x.cshtml``; area None outside Areas."""
+    parts = page_path.split("/")
+    ctrl = parts[parts.index("Views") + 1] if "Views" in parts[:-2] else ""
+    area = parts[parts.index("Areas") + 1] if "Areas" in parts[:-1] else None
+    return ctrl, area
+
+
+def resolve_page_actions(conn: sqlite3.Connection) -> None:
+    """Add ``calls`` edges from Razor pages to the controller actions their ``Url.Action(..)`` targets."""
+    by_action: dict[str, list[sqlite3.Row]] = defaultdict(list)
+    for r in conn.execute("SELECT id, symbol_name, qualified_name, file_path FROM symbols WHERE symbol_type='method' "
+                          "AND qualified_name LIKE '%Controller.%'"):
+        by_action[r["symbol_name"]].append(r)
+    rows = []
+    for r in conn.execute("SELECT r.file_path, r.name, r.line, r.owner FROM raw_refs r WHERE r.kind='action'"):
+        ctrl, action = r["name"].split("|", 1)
+        own_ctrl, area = _view_route(r["file_path"])
+        ctrl = ctrl or own_ctrl
+        if not ctrl:
+            continue
+        cands = [m for m in by_action.get(action, ()) if m["qualified_name"].split(".")[-2] == ctrl + "Controller"]
+        if area is not None:
+            same_area = [m for m in cands if f"/Areas/{area}/" in "/" + m["file_path"]]
+        else:
+            same_area = [m for m in cands if "/Areas/" not in "/" + m["file_path"]]
+        cands = same_area or cands
+        for m in cands:
+            rows.append((r["owner"], m["id"], r["file_path"], r["line"], "confirmed" if len(cands) == 1 else "probable",
+                         "Url.Action in page"))
+    conn.executemany("INSERT INTO calls(caller_id, callee_id, file_path, line, confidence, reason) VALUES(?,?,?,?,?,?)",
+                     rows)
+
+
+_DBSET = re.compile(r"\bI?DbSet<\s*([\w.]+)\s*>")
+
+
+def resolve_ef_access(conn: sqlite3.Connection) -> None:
+    """Turn ``ef_raw`` candidates whose member is a declared ``DbSet<T>`` into ``ef_linq`` table accesses.
+    Table = the EF mapping of T (``ToTable`` / ``[Table]``) when there is one, else the DbSet property name.
+    A method that reads a DbSet and calls ``SaveChanges`` without an explicit Add / Remove on it may update tracked
+    entities: that becomes an ``uncertain`` write (listed as possible, never as a confirmed writer)."""
+    conn.execute("DELETE FROM data_access WHERE via='ef_linq'")
+    dbsets: dict[str, str] = {}
+    for r in conn.execute("SELECT symbol_name, signature FROM symbols WHERE symbol_type='property' AND signature LIKE '%DbSet<%'"):
+        m = _DBSET.search(r[1] or "")
+        if m:
+            dbsets[r[0].lower()] = m.group(1).split(".")[-1]
+    if not dbsets:
+        return
+    mapped = {r[0].split(".")[-1]: r[1] for r in conn.execute(
+        "SELECT detail, target_name FROM data_access WHERE via='ef_mapping' AND detail != ''")}
+    rows = []
+    for r in conn.execute("SELECT * FROM data_access WHERE via='ef_raw'").fetchall():
+        entity = dbsets.get(r["target_key"][3:])
+        if entity is None:
+            continue
+        table = mapped.get(entity, r["target_name"])
+        rows.append((r["file_path"], r["source_id"], r["line"], table.lower(), table, "", "table", r["access"],
+                     r["op"], "ef_linq", "probable", r["evidence"], entity))
+    savers = {r[0]: (r[1], r[2]) for r in conn.execute(
+        "SELECT source_id, line, evidence FROM data_access WHERE via='ef_raw' AND target_key='ef:#savechanges' "
+        "AND source_id IS NOT NULL")}
+    written = {(r[1], r[3]) for r in rows if r[7] == "write"}
+    for r in list(rows):
+        if r[7] == "read" and r[1] in savers and (r[1], r[3]) not in written:
+            written.add((r[1], r[3]))
+            line, evidence = savers[r[1]]
+            rows.append((r[0], r[1], line, r[3], r[4], "", "table", "write", "update", "ef_linq", "uncertain",
+                         f"tracked {r[4]} entities + {evidence}", r[12]))
+    conn.executemany(
+        "INSERT INTO data_access(file_path, source_id, line, target_key, target_name, qualifier, target_kind, access,"
+        " op, via, confidence, evidence, detail) VALUES(?,?,?,?,?,?,?,?,?,?,?,?,?)", rows)
+
+
 def resolve_dependencies(conn: sqlite3.Connection) -> None:
     """Rebuild the file-level dependency graph from raw_refs and symbols."""
     conn.execute("DELETE FROM dependencies")
@@ -289,6 +369,15 @@ def resolve_dependencies(conn: sqlite3.Connection) -> None:
     for r in conn.execute("SELECT symbol_name, file_path FROM symbols WHERE symbol_type='namespace'"):
         ns_files[r[0]].add(r[1])
 
+    # a namespace import only links to the files in it whose types the importing file names
+    file_types: dict[str, set[str]] = defaultdict(set)
+    for name, paths in type_files.items():
+        for p in paths:
+            file_types[p].add(name)
+    used: dict[str, set[str]] = defaultdict(set)
+    for r in conn.execute("SELECT file_path, name FROM raw_refs WHERE kind IN ('ident', 'base')"):
+        used[r[0]].add(r[1])
+
     deps: set[tuple[str, str, str]] = set()
 
     for r in conn.execute("SELECT file_path, name FROM raw_refs WHERE kind='import'"):
@@ -299,7 +388,7 @@ def resolve_dependencies(conn: sqlite3.Connection) -> None:
             if t:
                 targets.add(t)
         elif name in ns_files:
-            targets |= ns_files[name]
+            targets |= {t for t in ns_files[name] if file_types[t] & used[src]}
         else:
             slug = name.replace(".", "/")
             for cand in (slug + ".py", slug + "/__init__.py"):

{local_codebase_indexer-0.3.2 → local_codebase_indexer-0.4.0}/local_codebase_indexer/cscalls.py

@@ -14,8 +14,14 @@ _SCOPES = {
     "method_declaration", "constructor_declaration", "destructor_declaration", "operator_declaration",
     "indexer_declaration", "property_declaration",
 }
-_TASK = re.compile(r"^(?:[\w.]+\.)?(?:Task|ValueTask)\s*<(.+)>$")
+_TASK = re.compile(r"^(?:[\w.]+\.)?(?:Task|ValueTask|Lazy)\s*<(.+)>$")
 _OUTER = re.compile(r"^([\w.]+)\s*(?:<.*>)?$", re.S)
+_ELEM = re.compile(r"^(?:[\w.]+\.)?(?:IEnumerable|IList|List|ICollection|IReadOnlyList|IReadOnlyCollection|HashSet|ISet"
+                   r"|IQueryable)\s*<(.+)>$", re.S)
+# LINQ calls whose result has the same element type as their source
+_KEEP_ELEM = {"Where", "OrderBy", "OrderByDescending", "ThenBy", "ThenByDescending", "ToList", "ToArray", "AsEnumerable",
+              "Skip", "Take", "Distinct", "Reverse"}
+_CLASSES = {"class_declaration", "struct_declaration", "record_declaration", "interface_declaration"}
 
 
 @dataclass
@@ -33,7 +39,7 @@ class CallSite:
 
 def norm_type(text: str) -> str:
     """Outer type name without namespace, generic arguments, nullability or array brackets.
-    ``Task<T>`` / ``ValueTask<T>`` become ``T`` (the type a caller works with after ``await``)."""
+    ``Task<T>`` / ``ValueTask<T>`` / ``Lazy<T>`` become ``T`` (the type a caller works with after ``await`` / ``.Value``)."""
     t = (text or "").strip()
     t = re.sub(r"^(?:ref|out|in|readonly|const|params)\s+", "", t)
     t = t.replace("global::", "")
@@ -49,6 +55,15 @@ def norm_type(text: str) -> str:
     return m.group(1).split(".")[-1] if m else ""
 
 
+def elem_type(text: str) -> str:
+    """Element type of a collection type (``IEnumerable<T>``, ``List<T>``, ``T[]`` ...), else ""."""
+    t = re.sub(r"\?$", "", (text or "").strip())
+    if t.endswith("[]"):
+        return norm_type(t[:-2])
+    m = _ELEM.match(t)
+    return norm_type(m.group(1)) if m else ""
+
+
 def _text(n, src: bytes) -> str:
     return src[n.start_byte:n.end_byte].decode("utf-8", errors="replace")
 
@@ -94,10 +109,59 @@ class _Calls:
         spans.sort(key=lambda t: (t[0], t[1].start))
         self.spans = [s for _, s in spans]
         self.out: list[CallSite] = []
+        self._field_types: dict[int, dict[str, str]] = {}
 
     def text(self, n) -> str:
         return _text(n, self.src)
 
+    def _class_member_types(self, scope) -> dict[str, str]:
+        """Declared type text of the fields and properties of the class enclosing ``scope``."""
+        cls = scope.parent
+        while cls is not None and cls.type not in _CLASSES:
+            cls = cls.parent
+        if cls is None:
+            return {}
+        if cls.id not in self._field_types:
+            types: dict[str, str] = {}
+            body = cls.child_by_field_name("body")
+            for m in (body.named_children if body is not None else ()):
+                if m.type == "property_declaration":
+                    t, nm = m.child_by_field_name("type"), m.child_by_field_name("name")
+                    if t is not None and nm is not None:
+                        types[self.text(nm)] = self.text(t)
+                elif m.type == "field_declaration":
+                    for vd in m.named_children:
+                        if vd.type != "variable_declaration":
+                            continue
+                        t = vd.child_by_field_name("type")
+                        for d in vd.named_children:
+                            if d.type != "variable_declarator":
+                                continue
+                            ident = next((c for c in d.named_children if c.type == "identifier"), None)
+                            if t is not None and ident is not None:
+                                types[self.text(ident)] = self.text(t)
+            self._field_types[cls.id] = types
+        return self._field_types[cls.id]
+
+    def _foreach_elem(self, expr, raw: dict, scope) -> str:
+        """Element type of ``foreach (var x in <expr>)`` when <expr> is a collection variable, field or property,
+        optionally filtered by LINQ calls that keep the element type."""
+        n = self._unwrap(expr)
+        while n is not None and n.type == "invocation_expression":
+            func = n.child_by_field_name("function")
+            if func is None or func.type != "member_access_expression" or self._method_name(func, self.text) not in _KEEP_ELEM:
+                return ""
+            n = self._unwrap(func.child_by_field_name("expression"))
+        if n is not None and n.type == "member_access_expression":
+            inner = n.child_by_field_name("expression")
+            if inner is None or inner.type != "this_expression":
+                return ""
+            n = n.child_by_field_name("name")
+        if n is None or n.type != "identifier":
+            return ""
+        name = self.text(n)
+        return elem_type(raw.get(name) or self._class_member_types(scope).get(name, ""))
+
     def locate(self, line: int):
         for s in self.spans:
             if s.start <= line <= s.end:
@@ -140,6 +204,7 @@ class _Calls:
 
     def collect_locals(self, scope) -> dict:
         env: dict[str, str] = {}
+        raw: dict[str, str] = {}  # declared type text, for foreach element types
         plist = scope.child_by_field_name("parameters")
         if plist is not None:
             for p in plist.named_children:
@@ -147,6 +212,7 @@ class _Calls:
                     t, nm = p.child_by_field_name("type"), p.child_by_field_name("name")
                     if t is not None and nm is not None:
                         self._set(env, self.text(nm), norm_type(self.text(t)))
+                        raw[self.text(nm)] = self.text(t)
         stack = list(scope.children)
         while stack:
             n = stack.pop()
@@ -167,10 +233,16 @@ class _Calls:
                             init = c
                     typ = declared if not implicit else (self._infer_initializer(init, env) if init is not None else "")
                     self._set(env, self.text(ident), typ)
+                    if not implicit:
+                        raw[self.text(ident)] = self.text(tnode)
             elif ty == "foreach_statement":
                 tnode, left = n.child_by_field_name("type"), n.child_by_field_name("left")
                 if left is not None and left.type == "identifier":
-                    typ = "" if (tnode is None or tnode.type == "implicit_type") else norm_type(self.text(tnode))
+                    if tnode is None or tnode.type == "implicit_type":
+                        right = n.child_by_field_name("right")
+                        typ = self._foreach_elem(right, raw, scope) if right is not None else ""
+                    else:
+                        typ = norm_type(self.text(tnode))
                     self._set(env, self.text(left), typ)
             elif ty == "catch_declaration":
                 kids = list(n.named_children)
@@ -220,6 +292,10 @@ class _Calls:
             nm = n.child_by_field_name("name")
             if inner is not None and inner.type == "this_expression" and nm is not None:
                 return "member", self.text(nm), "", ""
+            if inner is not None and inner.type == "identifier" and nm is not None and self.text(nm) == "Value":
+                # ponytail: x.Value is read as x (Lazy<T> / Nullable<T> already normalised to T); any other type's
+                # .Value resolves to that type, usually outside the index, so the call gets no edge
+                return self._classify_receiver(inner, env, depth)
             return "expr", "", "", ""
         if t == "object_creation_expression":
             tn = n.child_by_field_name("type")