codecompass-mcp 3.2.0__tar.gz → 4.0.0__tar.gz

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codecompass-mcp
-Version: 3.2.0
+Version: 4.0.0
 Summary: Structural code context for AI coding agents — a local code knowledge graph for blast radius, impact, deps, dead code, and flow tracing
 Author: CodeCompass contributors
 License: MIT
@@ -20,10 +20,8 @@ License-File: LICENSE
 Requires-Dist: anthropic>=0.40.0
 Requires-Dist: fastmcp>=2.0.0
 Requires-Dist: networkx>=3.0
-Requires-Dist: pypdf2>=3.0.0
 Requires-Dist: python-dotenv>=1.0.0
 Requires-Dist: rich>=13.0.0
-Requires-Dist: tqdm>=4.66.0
 Requires-Dist: tree-sitter>=0.21.0
 Requires-Dist: tree-sitter-python>=0.21.0
 Requires-Dist: tree-sitter-javascript>=0.21.0
@@ -69,10 +67,8 @@ The graph turns navigation into a cheap, deterministic loop:
 
    | You have… | Use |
    |---|---|
-   | a feature request that names a concept ("session timeout") | `grep` the concept |
-   | a regex / name pattern | `grep` |
-   | keywords | `search` |
-   | a vague, nameless need | `map` (compact index to reason over) |
+   | a concept, name, or pattern | `grep` (regex over graph entities) |
+   | the full layout | `tree` |
 
 2. **Trace** — a relationship around a known symbol/file:
 
@@ -142,8 +138,6 @@ The server speaks stdio MCP and defaults to the working directory.
 ```bash
 # discover
 codecompass query --grep "^get_"            # regex over graph entities
-codecompass query --search "session cookie" # keyword search
-codecompass query --map                     # compact {file: [symbols]} index
 
 # trace
 codecompass query --impact "Session.send"   # callers (disambiguated), with file:line
@@ -157,6 +151,21 @@ codecompass query --tree                     # full hierarchy
 
 Add `--hops N` for traversal depth (start at 1 and follow the one path you need). Add `--rich` for tables.
 
+### Enrichment (agent-in-the-loop)
+
+```bash
+codecompass enrich                 # stage entities for an agent swarm: one-line descriptions + missing call edges
+codecompass enrich --apply         # merge the swarm's results into the graph
+codecompass add-entity <name> --file src/a.py --line 9 --description "Async helper"
+codecompass add-call caller callee --line 2
+```
+
+`enrich` is a bulk, user-triggered pass. `add-entity`/`add-call` are the
+opportunistic version: as an agent reads code and spots something the parser
+missed, it records it immediately. Everything agent-written is marked
+`agent_inferred` and **preserved across re-ingests** — the graph gets better
+with use. Ambiguous call targets are skipped, never guessed.
+
 ### Flow: `flow` vs `flow-summary`
 
 - **`flow`** — lean structure only (node name/kind/file/depth, edge from/to/order/line). What an agent needs to navigate; no embedded source.
@@ -169,8 +178,6 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | Tool | Returns |
 |---|---|
 | `grep(pattern, field, ignore_case)` | Regex search over graph entities |
-| `search(query, kind)` | Keyword search, ranked |
-| `map(include_tests)` | Compact `{file: [symbols]}` index |
 | `impact(symbol, hops)` | Callers/importers, disambiguated, with `resolved` + `line` |
 | `blast_radius(target, hops)` | Files reachable from a file or symbol |
 | `batch_impact(targets, hops)` | Union of blast radii for a multi-file change |
@@ -181,6 +188,9 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | `dead_code(include_entrypoints)` | Entities with no inbound caller |
 | `styles(element)` | CSS selectors that style an element |
 | `tree()` | Full project hierarchy |
+| `enrich(apply, batch_size)` | Stage/merge agent-written descriptions + missing call edges |
+| `add_entity(name, kind, file, line, description)` | Record a parser-missed entity (`agent_inferred`) |
+| `add_call(caller, callee, line)` | Record a parser-missed CALLS edge (`agent_inferred`) |
 | `set_repo` / `get_repo` / `init` / `ingest` | Project selection & indexing |
 
 ---
@@ -208,11 +218,14 @@ are language-agnostic.
 `AGENTS.md` guides any agent through the discover→trace→read→edit loop. For
 Claude Code and [pi](https://pi.dev), `init` also installs a `PreToolUse` hook
 that **blocks code *search*** (`grep`/`rg`, the `Grep`/`Glob` tools) and
-**whole-file `cat`**, routing discovery through the graph — while leaving
-**targeted reads free** (the `Read` tool, `sed -n`, `head`/`tail`). The point is
-to change the default reflex to graph-first, not to remove reads. Both pick the
-files up after a one-time trust prompt (they execute shell commands). The hook
-is a plain file under `.claude/hooks/` — edit or delete it to adjust.
+**whole-file `cat` — but only inside a codecompass-registered repo** (tracked in
+`~/.codecompass/repos`, one line per `init`'d project). Reads outside any
+registered repo pass through: no graph exists there, so nothing is blocked.
+**Targeted reads stay free** (the `Read` tool, `sed -n`, `head`/`tail`). The
+point is to change the default reflex to graph-first, not to remove reads.
+Each project's Claude hook lives under its own `.claude/hooks/` with the
+project root baked in — edit or delete it to adjust. Block messages point the
+agent at the right repo's graph: `codecompass query \"<repo>\" --grep …`.
 
 ---
 
@@ -224,8 +237,11 @@ Source files
    ▼  code_parser         tree-sitter extraction (no API calls) → typed CodeTriples
    ▼  graph.json          NetworkX MultiDiGraph as JSON; file+class-qualified nodes,
                           typed edges (CALLS/IMPORTS/INHERITS/STYLES/…), resolved calls
-   ▼  code_query_cli      traversal: grep / search / map / impact / blast-radius /
+   ▼  code_query_cli      traversal: grep / impact / blast-radius /
                           deps / flow / dead-code / tree
+   ▼  enricher            agent-in-the-loop: enrich batches (descriptions +
+                          missing calls) and opportunistic add_entity/add_call
+                          writes — agent_inferred, preserved across re-ingest
 ```
 
 Everything runs locally, in-process. No network, no database, no API keys.
@@ -242,7 +258,9 @@ your-project/
 
 ## Limitations
 
-- **Structure, not semantics** — the graph knows what calls what, not what it means.
+- **Structure first, semantics optional** — the parser knows what calls what,
+  not what it means. `enrich` closes that gap with agent-written descriptions
+  and missed edges, marked `agent_inferred`.
 - **Static analysis** — dynamic dispatch, reflection, and string-based invocation
   can't be fully resolved. `impact` surfaces those flagged `resolved: false`, and
   `dead_code` results are always candidates to verify.

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/README.md

@@ -31,10 +31,8 @@ The graph turns navigation into a cheap, deterministic loop:
 
    | You have… | Use |
    |---|---|
-   | a feature request that names a concept ("session timeout") | `grep` the concept |
-   | a regex / name pattern | `grep` |
-   | keywords | `search` |
-   | a vague, nameless need | `map` (compact index to reason over) |
+   | a concept, name, or pattern | `grep` (regex over graph entities) |
+   | the full layout | `tree` |
 
 2. **Trace** — a relationship around a known symbol/file:
 
@@ -104,8 +102,6 @@ The server speaks stdio MCP and defaults to the working directory.
 ```bash
 # discover
 codecompass query --grep "^get_"            # regex over graph entities
-codecompass query --search "session cookie" # keyword search
-codecompass query --map                     # compact {file: [symbols]} index
 
 # trace
 codecompass query --impact "Session.send"   # callers (disambiguated), with file:line
@@ -119,6 +115,21 @@ codecompass query --tree                     # full hierarchy
 
 Add `--hops N` for traversal depth (start at 1 and follow the one path you need). Add `--rich` for tables.
 
+### Enrichment (agent-in-the-loop)
+
+```bash
+codecompass enrich                 # stage entities for an agent swarm: one-line descriptions + missing call edges
+codecompass enrich --apply         # merge the swarm's results into the graph
+codecompass add-entity <name> --file src/a.py --line 9 --description "Async helper"
+codecompass add-call caller callee --line 2
+```
+
+`enrich` is a bulk, user-triggered pass. `add-entity`/`add-call` are the
+opportunistic version: as an agent reads code and spots something the parser
+missed, it records it immediately. Everything agent-written is marked
+`agent_inferred` and **preserved across re-ingests** — the graph gets better
+with use. Ambiguous call targets are skipped, never guessed.
+
 ### Flow: `flow` vs `flow-summary`
 
 - **`flow`** — lean structure only (node name/kind/file/depth, edge from/to/order/line). What an agent needs to navigate; no embedded source.
@@ -131,8 +142,6 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | Tool | Returns |
 |---|---|
 | `grep(pattern, field, ignore_case)` | Regex search over graph entities |
-| `search(query, kind)` | Keyword search, ranked |
-| `map(include_tests)` | Compact `{file: [symbols]}` index |
 | `impact(symbol, hops)` | Callers/importers, disambiguated, with `resolved` + `line` |
 | `blast_radius(target, hops)` | Files reachable from a file or symbol |
 | `batch_impact(targets, hops)` | Union of blast radii for a multi-file change |
@@ -143,6 +152,9 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | `dead_code(include_entrypoints)` | Entities with no inbound caller |
 | `styles(element)` | CSS selectors that style an element |
 | `tree()` | Full project hierarchy |
+| `enrich(apply, batch_size)` | Stage/merge agent-written descriptions + missing call edges |
+| `add_entity(name, kind, file, line, description)` | Record a parser-missed entity (`agent_inferred`) |
+| `add_call(caller, callee, line)` | Record a parser-missed CALLS edge (`agent_inferred`) |
 | `set_repo` / `get_repo` / `init` / `ingest` | Project selection & indexing |
 
 ---
@@ -170,11 +182,14 @@ are language-agnostic.
 `AGENTS.md` guides any agent through the discover→trace→read→edit loop. For
 Claude Code and [pi](https://pi.dev), `init` also installs a `PreToolUse` hook
 that **blocks code *search*** (`grep`/`rg`, the `Grep`/`Glob` tools) and
-**whole-file `cat`**, routing discovery through the graph — while leaving
-**targeted reads free** (the `Read` tool, `sed -n`, `head`/`tail`). The point is
-to change the default reflex to graph-first, not to remove reads. Both pick the
-files up after a one-time trust prompt (they execute shell commands). The hook
-is a plain file under `.claude/hooks/` — edit or delete it to adjust.
+**whole-file `cat` — but only inside a codecompass-registered repo** (tracked in
+`~/.codecompass/repos`, one line per `init`'d project). Reads outside any
+registered repo pass through: no graph exists there, so nothing is blocked.
+**Targeted reads stay free** (the `Read` tool, `sed -n`, `head`/`tail`). The
+point is to change the default reflex to graph-first, not to remove reads.
+Each project's Claude hook lives under its own `.claude/hooks/` with the
+project root baked in — edit or delete it to adjust. Block messages point the
+agent at the right repo's graph: `codecompass query \"<repo>\" --grep …`.
 
 ---
 
@@ -186,8 +201,11 @@ Source files
    ▼  code_parser         tree-sitter extraction (no API calls) → typed CodeTriples
    ▼  graph.json          NetworkX MultiDiGraph as JSON; file+class-qualified nodes,
                           typed edges (CALLS/IMPORTS/INHERITS/STYLES/…), resolved calls
-   ▼  code_query_cli      traversal: grep / search / map / impact / blast-radius /
+   ▼  code_query_cli      traversal: grep / impact / blast-radius /
                           deps / flow / dead-code / tree
+   ▼  enricher            agent-in-the-loop: enrich batches (descriptions +
+                          missing calls) and opportunistic add_entity/add_call
+                          writes — agent_inferred, preserved across re-ingest
 ```
 
 Everything runs locally, in-process. No network, no database, no API keys.
@@ -204,7 +222,9 @@ your-project/
 
 ## Limitations
 
-- **Structure, not semantics** — the graph knows what calls what, not what it means.
+- **Structure first, semantics optional** — the parser knows what calls what,
+  not what it means. `enrich` closes that gap with agent-written descriptions
+  and missed edges, marked `agent_inferred`.
 - **Static analysis** — dynamic dispatch, reflection, and string-based invocation
   can't be fully resolved. `impact` surfaces those flagged `resolved: false`, and
   `dead_code` results are always candidates to verify.

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/codecompass_mcp.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codecompass-mcp
-Version: 3.2.0
+Version: 4.0.0
 Summary: Structural code context for AI coding agents — a local code knowledge graph for blast radius, impact, deps, dead code, and flow tracing
 Author: CodeCompass contributors
 License: MIT
@@ -20,10 +20,8 @@ License-File: LICENSE
 Requires-Dist: anthropic>=0.40.0
 Requires-Dist: fastmcp>=2.0.0
 Requires-Dist: networkx>=3.0
-Requires-Dist: pypdf2>=3.0.0
 Requires-Dist: python-dotenv>=1.0.0
 Requires-Dist: rich>=13.0.0
-Requires-Dist: tqdm>=4.66.0
 Requires-Dist: tree-sitter>=0.21.0
 Requires-Dist: tree-sitter-python>=0.21.0
 Requires-Dist: tree-sitter-javascript>=0.21.0
@@ -69,10 +67,8 @@ The graph turns navigation into a cheap, deterministic loop:
 
    | You have… | Use |
    |---|---|
-   | a feature request that names a concept ("session timeout") | `grep` the concept |
-   | a regex / name pattern | `grep` |
-   | keywords | `search` |
-   | a vague, nameless need | `map` (compact index to reason over) |
+   | a concept, name, or pattern | `grep` (regex over graph entities) |
+   | the full layout | `tree` |
 
 2. **Trace** — a relationship around a known symbol/file:
 
@@ -142,8 +138,6 @@ The server speaks stdio MCP and defaults to the working directory.
 ```bash
 # discover
 codecompass query --grep "^get_"            # regex over graph entities
-codecompass query --search "session cookie" # keyword search
-codecompass query --map                     # compact {file: [symbols]} index
 
 # trace
 codecompass query --impact "Session.send"   # callers (disambiguated), with file:line
@@ -157,6 +151,21 @@ codecompass query --tree                     # full hierarchy
 
 Add `--hops N` for traversal depth (start at 1 and follow the one path you need). Add `--rich` for tables.
 
+### Enrichment (agent-in-the-loop)
+
+```bash
+codecompass enrich                 # stage entities for an agent swarm: one-line descriptions + missing call edges
+codecompass enrich --apply         # merge the swarm's results into the graph
+codecompass add-entity <name> --file src/a.py --line 9 --description "Async helper"
+codecompass add-call caller callee --line 2
+```
+
+`enrich` is a bulk, user-triggered pass. `add-entity`/`add-call` are the
+opportunistic version: as an agent reads code and spots something the parser
+missed, it records it immediately. Everything agent-written is marked
+`agent_inferred` and **preserved across re-ingests** — the graph gets better
+with use. Ambiguous call targets are skipped, never guessed.
+
 ### Flow: `flow` vs `flow-summary`
 
 - **`flow`** — lean structure only (node name/kind/file/depth, edge from/to/order/line). What an agent needs to navigate; no embedded source.
@@ -169,8 +178,6 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | Tool | Returns |
 |---|---|
 | `grep(pattern, field, ignore_case)` | Regex search over graph entities |
-| `search(query, kind)` | Keyword search, ranked |
-| `map(include_tests)` | Compact `{file: [symbols]}` index |
 | `impact(symbol, hops)` | Callers/importers, disambiguated, with `resolved` + `line` |
 | `blast_radius(target, hops)` | Files reachable from a file or symbol |
 | `batch_impact(targets, hops)` | Union of blast radii for a multi-file change |
@@ -181,6 +188,9 @@ Add `--hops N` for traversal depth (start at 1 and follow the one path you need)
 | `dead_code(include_entrypoints)` | Entities with no inbound caller |
 | `styles(element)` | CSS selectors that style an element |
 | `tree()` | Full project hierarchy |
+| `enrich(apply, batch_size)` | Stage/merge agent-written descriptions + missing call edges |
+| `add_entity(name, kind, file, line, description)` | Record a parser-missed entity (`agent_inferred`) |
+| `add_call(caller, callee, line)` | Record a parser-missed CALLS edge (`agent_inferred`) |
 | `set_repo` / `get_repo` / `init` / `ingest` | Project selection & indexing |
 
 ---
@@ -208,11 +218,14 @@ are language-agnostic.
 `AGENTS.md` guides any agent through the discover→trace→read→edit loop. For
 Claude Code and [pi](https://pi.dev), `init` also installs a `PreToolUse` hook
 that **blocks code *search*** (`grep`/`rg`, the `Grep`/`Glob` tools) and
-**whole-file `cat`**, routing discovery through the graph — while leaving
-**targeted reads free** (the `Read` tool, `sed -n`, `head`/`tail`). The point is
-to change the default reflex to graph-first, not to remove reads. Both pick the
-files up after a one-time trust prompt (they execute shell commands). The hook
-is a plain file under `.claude/hooks/` — edit or delete it to adjust.
+**whole-file `cat` — but only inside a codecompass-registered repo** (tracked in
+`~/.codecompass/repos`, one line per `init`'d project). Reads outside any
+registered repo pass through: no graph exists there, so nothing is blocked.
+**Targeted reads stay free** (the `Read` tool, `sed -n`, `head`/`tail`). The
+point is to change the default reflex to graph-first, not to remove reads.
+Each project's Claude hook lives under its own `.claude/hooks/` with the
+project root baked in — edit or delete it to adjust. Block messages point the
+agent at the right repo's graph: `codecompass query \"<repo>\" --grep …`.
 
 ---
 
@@ -224,8 +237,11 @@ Source files
    ▼  code_parser         tree-sitter extraction (no API calls) → typed CodeTriples
    ▼  graph.json          NetworkX MultiDiGraph as JSON; file+class-qualified nodes,
                           typed edges (CALLS/IMPORTS/INHERITS/STYLES/…), resolved calls
-   ▼  code_query_cli      traversal: grep / search / map / impact / blast-radius /
+   ▼  code_query_cli      traversal: grep / impact / blast-radius /
                           deps / flow / dead-code / tree
+   ▼  enricher            agent-in-the-loop: enrich batches (descriptions +
+                          missing calls) and opportunistic add_entity/add_call
+                          writes — agent_inferred, preserved across re-ingest
 ```
 
 Everything runs locally, in-process. No network, no database, no API keys.
@@ -242,7 +258,9 @@ your-project/
 
 ## Limitations
 
-- **Structure, not semantics** — the graph knows what calls what, not what it means.
+- **Structure first, semantics optional** — the parser knows what calls what,
+  not what it means. `enrich` closes that gap with agent-written descriptions
+  and missed edges, marked `agent_inferred`.
 - **Static analysis** — dynamic dispatch, reflection, and string-based invocation
   can't be fully resolved. `impact` surfaces those flagged `resolved: false`, and
   `dead_code` results are always candidates to verify.

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/codecompass_mcp.egg-info/SOURCES.txt

@@ -11,32 +11,23 @@ codecompass_mcp.egg-info/entry_points.txt
 codecompass_mcp.egg-info/requires.txt
 codecompass_mcp.egg-info/top_level.txt
 graph/__init__.py
-graph/cli.py
 graph/code_graph_client.py
 graph/code_query_cli.py
-graph/setup.py
 ingestion/__init__.py
-ingestion/chunker.py
 ingestion/code_normalizer.py
 ingestion/code_parser.py
-ingestion/description_enricher.py
-ingestion/entity_resolver.py
+ingestion/enricher.py
 ingestion/file_watcher.py
-ingestion/graph_writer.py
 ingestion/hierarchy_builder.py
-ingestion/reader_agent.py
 ingestion/source_context.py
 models/__init__.py
 models/code_types.py
-models/types.py
 tests/test_accuracy_fixes.py
 tests/test_batch_impact.py
 tests/test_blast_radius.py
+tests/test_block_hook.py
 tests/test_code_parser.py
-tests/test_description_enricher.py
-tests/test_ingestion.py
+tests/test_enricher.py
 tests/test_lit_css_parser.py
 tests/test_mcp_server.py
-tests/test_php_parser.py
-utils/__init__.py
-utils/formatting.py
+tests/test_php_parser.py

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/codecompass_mcp.egg-info/entry_points.txt

@@ -1,3 +1,3 @@
 [console_scripts]
-codecompass = graph.cli:main
+codecompass = main:main
 codecompass-mcp = mcp_server:main

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/codecompass_mcp.egg-info/requires.txt

@@ -1,10 +1,8 @@
 anthropic>=0.40.0
 fastmcp>=2.0.0
 networkx>=3.0
-pypdf2>=3.0.0
 python-dotenv>=1.0.0
 rich>=13.0.0
-tqdm>=4.66.0
 tree-sitter>=0.21.0
 tree-sitter-python>=0.21.0
 tree-sitter-javascript>=0.21.0

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/codecompass_mcp.egg-info/top_level.txt

@@ -4,4 +4,3 @@ ingestion
 main
 mcp_server
 models
-utils

{codecompass_mcp-3.2.0 → codecompass_mcp-4.0.0}/graph/code_graph_client.py

@@ -10,7 +10,7 @@ import os
 import re
 import networkx as nx
 from datetime import datetime, timezone
-from typing import Optional, Any
+from typing import Optional
 
 from models.code_types import CodeTriple, FileNode, FolderNode
 
@@ -236,6 +236,11 @@ class LocalGraphClient:
             "project": project,
             "file": triple.source_file,
         })
+        # Definition line comes only from DEFINED_IN triples — a CALLS triple's
+        # line_number is the call site, not the definition. External symbols
+        # (never defined in-project) legitimately have no line.
+        if triple.relation_type == "DEFINED_IN" and triple.line_number:
+            self.graph.nodes[from_id]["line"] = triple.line_number
         # A symbol is public if any of its definitions is exported. Never let a
         # later non-exported definition clear a flag an earlier one set.
         if getattr(triple, "is_exported", False):
@@ -254,6 +259,9 @@ class LocalGraphClient:
             existing["language"] = language
             existing["kind"] = f"{triple.to_type}:{language}"
             existing["description"] = f"{language} {triple.to_type}"
+        # A module's definition is line 1 of its file.
+        if triple.relation_type == "DEFINED_IN" and triple.to_type == "module":
+            existing.setdefault("line", 1)
 
         self.graph.add_edge(
             from_id,
@@ -321,6 +329,13 @@ class LocalGraphClient:
             to_id = self._resolve_to_id(triple, project, defs_by_name, class_file)
             self.write_code_triple(triple, file_id, project, from_id=from_id, to_id=to_id)
 
+        # Module nodes are the to-side of DEFINED_IN and may exist under a
+        # different resolved id — give every in-project module its line (1).
+        for _id, attr in self.graph.nodes(data=True):
+            if (attr.get("type") == "Entity" and attr.get("entity_type") == "module"
+                    and attr.get("file") and "line" not in attr):
+                attr["line"] = 1
+
         self.save()
         return len(triples)
 
@@ -543,6 +558,7 @@ class LocalGraphClient:
                         "name": a.get("name"),
                         "kind": a.get("kind"),
                         "file": a.get("file"),
+                        "line": a.get("line"),
                         "matched_field": f,
                         "match": m.group(0),
                     })
@@ -552,90 +568,6 @@ class LocalGraphClient:
         matches.sort(key=lambda r: (r["file"] or "", r["name"] or ""))
         return {"pattern": pattern, "matches": matches, "count": len(matches)}
 
-    def search_entities(self, query: str, project: str, limit: int = 30,
-                        kind: str | None = None) -> list[dict]:
-        """Find graph entities matching a keyword — the way in when you don't yet
-        know a symbol name.
-
-        Matches the query (case-insensitive) against each entity's name, file
-        path, and description. Any term may match (OR), and entities matching
-        more terms — and matching in the name rather than file/description — rank
-        highest, so the most on-point symbols come first.
-        Optional `kind` filters by entity_type (e.g. "function", "class").
-        Returns a lean list (name/kind/file/entity_type) to then feed into
-        impact/flow/deps.
-        """
-        terms = [t for t in query.lower().split() if t]
-        if not terms:
-            return []
-
-        scored = []
-        for _id, a in self.graph.nodes(data=True):
-            if a.get("type") != "Entity" or a.get("project") != project:
-                continue
-            if not a.get("file"):
-                continue
-            if kind and a.get("entity_type") != kind:
-                continue
-            name = (a.get("name") or "")
-            file = (a.get("file") or "")
-            desc = (a.get("description") or "")
-            hay_name = name.lower()
-            hay_rest = f"{file} {desc}".lower()
-            # any term may match (OR); score rewards name hits and more terms matched
-            score = 0
-            for t in terms:
-                if t == hay_name:
-                    score += 100
-                elif t in hay_name:
-                    score += 10
-                elif t in hay_rest:
-                    score += 1
-            if score == 0:
-                continue
-            scored.append((score, {
-                "name": name,
-                "entity_type": a.get("entity_type", ""),
-                "kind": a.get("kind", ""),
-                "file": file,
-            }))
-
-        scored.sort(key=lambda s: (-s[0], s[1]["file"], s[1]["name"]))
-        return [row for _s, row in scored[:limit]]
-
-    def symbol_map(self, project: str, include_tests: bool = False) -> dict:
-        """A compact index of the codebase for an agent to reason over.
-
-        Groups every source entity under its file as `{file: [names...]}` —
-        names only, no depth/kind/description bloat — so the whole map fits in a
-        few thousand tokens. The agent reads it once and uses its own judgment to
-        pick what's relevant to a vague task ("where would caching go?"), then
-        drills in with impact/flow/deps. This is the semantic-discovery entry
-        point that keyword `search` can't be: an LLM knows caching lives near the
-        request handler and response path even when nothing is named "cache".
-
-        `include_tests=False` (default) drops test/example files so the map shows
-        the actual implementation surface.
-        """
-        by_file: dict[str, list[str]] = {}
-        for _id, a in self.graph.nodes(data=True):
-            if a.get("type") != "Entity" or a.get("project") != project:
-                continue
-            f = a.get("file")
-            if not f:
-                continue
-            if not include_tests and ("test/" in f or "test\\" in f
-                                      or f.startswith("test") or "example" in f):
-                continue
-            name = a.get("name")
-            if name:
-                by_file.setdefault(f, [])
-                if name not in by_file[f]:
-                    by_file[f].append(name)
-        for f in by_file:
-            by_file[f].sort()
-        return dict(sorted(by_file.items()))
-
     def find_dependencies(self, file_path: str, project: str, max_hops: int = 3) -> list[dict]:
         """Return all modules imported (directly or transitively) by file_path."""
         file_node = next((n for n, attr in self.graph.nodes(data=True) 
@@ -662,6 +594,7 @@ class LocalGraphClient:
                         results.append({
                             "dependency": node.get("name"),
                             "dep_type": node.get("entity_type"),
+                            "line": node.get("line"),
                             "depth": depth + 1
                         })
                     
@@ -710,6 +643,7 @@ class LocalGraphClient:
             "kind": start_attr.get("kind", ""),
             "description": start_attr.get("description", ""),
             "file": start_attr.get("file", ""),
+            "line": start_attr.get("line"),
             "depth": 0,
         }
 
@@ -743,6 +677,7 @@ class LocalGraphClient:
                         "kind": succ_attr.get("kind", ""),
                         "description": succ_attr.get("description", ""),
                         "file": succ_attr.get("file", ""),
+                        "line": succ_attr.get("line"),
                         "depth": depth + 1,
                     }
                     queue.append((succ, depth + 1))
@@ -776,6 +711,7 @@ class LocalGraphClient:
                             "callee_name": node.get("name"),
                             "callee_type": node.get("entity_type"),
                             "callee_file": node.get("file"),
+                            "line": node.get("line"),
                             "depth": depth + 1
                         })
                         visited.add(succ)
@@ -847,6 +783,7 @@ class LocalGraphClient:
                 "kind": attr.get("kind", ""),
                 "entity_type": attr.get("entity_type", ""),
                 "file": attr.get("file", ""),
+                "line": attr.get("line"),
             }
             # Exported symbols and name-heuristic entry points are public API:
             # having no in-repo caller is expected, not evidence of dead code.