ph-code-graph 0.4.0__tar.gz → 0.6.0__tar.gz

{ph_code_graph-0.4.0 → ph_code_graph-0.6.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: ph-code-graph
-Version: 0.4.0
+Version: 0.6.0
 Summary: pH plugin: `code_index` and `code_graph`, a Python-native tree-sitter code graph with line-accurate answers.
 Project-URL: Homepage, https://github.com/chastabor/pH
 Project-URL: Repository, https://github.com/chastabor/pH
@@ -21,7 +21,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Classifier: Topic :: Software Development
 Classifier: Typing :: Typed
 Requires-Python: >=3.12
-Requires-Dist: ph-core==0.4.0
+Requires-Dist: ph-core==0.6.0
 Requires-Dist: tree-sitter-language-pack>=1.16
 Requires-Dist: tree-sitter>=0.25
 Description-Content-Type: text/markdown
@@ -206,10 +206,12 @@ would make `callers` return every place that merely *mentions* a type.
 
 ## Where things live
 
-- **the index**: `$PH_CACHE/code-graph/<digest of the workspace root>.db` — the
-  cache root because it is rebuildable from the source, and keyed by root so two
-  checkouts do not answer each other's questions. `path:` in the row's config
-  overrides it.
+- **the index**: `$PH_CACHE/code-graph/<digest of the workspace root>/graph.db` —
+  the cache root because it is rebuildable from the source, and keyed by root so
+  two checkouts do not answer each other's questions. The directory records its
+  workspace, and one whose workspace has been gone a week is removed on the next
+  index, so a profile that runs children in worktrees does not leave a database
+  per tree forever. `path:` in the row's config names one database instead.
 - **the grammars**: `$PH_CACHE/tree-sitter`, set by the row at mount.
 
 That second one is not housekeeping. The pack materializes even its *bundled*

{ph_code_graph-0.4.0 → ph_code_graph-0.6.0}/README.md

@@ -178,10 +178,12 @@ would make `callers` return every place that merely *mentions* a type.
 
 ## Where things live
 
-- **the index**: `$PH_CACHE/code-graph/<digest of the workspace root>.db` — the
-  cache root because it is rebuildable from the source, and keyed by root so two
-  checkouts do not answer each other's questions. `path:` in the row's config
-  overrides it.
+- **the index**: `$PH_CACHE/code-graph/<digest of the workspace root>/graph.db` —
+  the cache root because it is rebuildable from the source, and keyed by root so
+  two checkouts do not answer each other's questions. The directory records its
+  workspace, and one whose workspace has been gone a week is removed on the next
+  index, so a profile that runs children in worktrees does not leave a database
+  per tree forever. `path:` in the row's config names one database instead.
 - **the grammars**: `$PH_CACHE/tree-sitter`, set by the row at mount.
 
 That second one is not housekeeping. The pack materializes even its *bundled*

{ph_code_graph-0.4.0 → ph_code_graph-0.6.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ph-code-graph"
-version = "0.4.0"
+version = "0.6.0"
 description = "pH plugin: `code_index` and `code_graph`, a Python-native tree-sitter code graph with line-accurate answers."
 readme = "README.md"
 requires-python = ">=3.12"
@@ -29,7 +29,7 @@ classifiers = [
 # in (371 available, the rest fetched on demand — which this row never does
 # implicitly; see `_extract.indexable`). No Rust toolchain, no Node, no
 # submodule: the graph itself is stdlib `sqlite3`.
-dependencies = ["ph-core==0.4.0", "tree-sitter>=0.25", "tree-sitter-language-pack>=1.16"]
+dependencies = ["ph-core==0.6.0", "tree-sitter>=0.25", "tree-sitter-language-pack>=1.16"]
 
 [project.urls]
 Homepage = "https://github.com/chastabor/pH"

{ph_code_graph-0.4.0 → ph_code_graph-0.6.0}/src/ph_code_graph/__init__.py

@@ -41,7 +41,6 @@ answers most of what an agent actually asks and reports where it cannot.
 
 from __future__ import annotations
 
-import hashlib
 import importlib.util
 import logging
 from collections.abc import Mapping
@@ -59,7 +58,7 @@ from ph.llm.types import ContentBlock
 from ph.paths import default_cache_path, resolve_roots
 from ph.seams._registry import contribute_item
 from ph.seams.changes import tree_state
-from ph.seams.commands import CommandContext, CommandDefinition
+from ph.seams.commands import Provisioning, install_or_status
 from ph.seams.diagnostics import Diagnostic, contribute
 from ph.seams.fs import FsService
 from ph.seams.skills import discover_skills
@@ -68,6 +67,7 @@ from ph.tools.definition import ToolModel, ToolOutput, ToolRunContext, define_to
 from ph.tools.errors import HarnessError
 from ph.tools.presentation import simple_views
 from ph.wire import WireModel
+from ph.workspace_caches import digest, use
 
 from ._extract import (
     cache_release,
@@ -127,7 +127,8 @@ class Config(WireModel):
     """Row config for `code-graph`."""
 
     path: str = ""
-    """Where the index lives. Defaults to `$PH_CACHE/code-graph/<root digest>.db`.
+    """Where the index lives. Defaults to `$PH_CACHE/code-graph/<root digest>/graph.db`;
+    a path set here is one database, used as given.
 
     Under the cache root because the index is **rebuildable** — the source is the
     truth and this is derived, which is the lifecycle `$PH_CACHE` names (Q1).
@@ -178,21 +179,22 @@ class CodeGraphSeam:
     _lock: anyio.Lock = field(default_factory=anyio.Lock)
 
     def store_for(self, root: Path) -> CodeGraphStore:
-        """This workspace's index. One file per root, and **nothing prunes them.**
+        """This workspace's index: a directory per root, or the row's `path` as given.
 
         Keyed per root on purpose — a worktree at another revision holds
         different line numbers, so sharing one index would hand out pointers
         that are quietly stale. The cost is that under a profile whose children
-        run in worktrees, every throwaway tree leaves a database behind, and
-        `phern doctor` names only the current one. Stated here rather than left to
-        be discovered, which is this codebase's rule for a cache nothing
-        collects (`ph.seams.uploads` says the same about attachments): deleting
-        `$PH_CACHE/code-graph` reclaims all of them and costs a re-index.
+        run in worktrees, every throwaway tree leaves a database behind, so the
+        directory records its workspace and one whose workspace is gone is swept
+        on a later index (`ph.workspace_caches`).
         """
-        digest = hashlib.sha256(str(root).encode("utf-8")).hexdigest()[:16]
-        return CodeGraphStore(
-            path=default_cache_path(self.config.path, "code-graph", f"{digest}.db")
-        )
+        if self.config.path:
+            return CodeGraphStore(path=default_cache_path(self.config.path, "code-graph"))
+        return CodeGraphStore(path=self.base() / digest(str(root)) / "graph.db")
+
+    def base(self) -> Path:
+        """Where the per-workspace indexes live, when no `path` names one database."""
+        return default_cache_path(None, "code-graph")
 
     def locked(self) -> anyio.Lock:
         return self._lock
@@ -493,7 +495,7 @@ def offer_skills(ctx: Context) -> None:
         contribute_item(ctx, SKILLS, skill, label=f"skill({skill.name})")
 
 
-@plugin("code-graph", inject=[TOOLS, FS], config=Config)
+@plugin("code-graph", affects="environment", inject=[TOOLS, FS], config=Config)
 async def apply(ctx: Context, config: Config) -> None:
     """Mount the seam and register both tools."""
     for module in ("tree_sitter", "tree_sitter_language_pack"):
@@ -560,6 +562,10 @@ async def apply(ctx: Context, config: Config) -> None:
         skipped: list[dict[str, str]] = []
         indexed = unchanged = symbols = removed = 0
 
+        if not config.path:
+            # Before the lock: the record and the sweep touch nothing an index
+            # holds. Not for a configured `path`, one database the row was given.
+            await anyio.to_thread.run_sync(use, seam.base(), fs.root_for(run.agent))
         async with seam.locked():
             await anyio.to_thread.run_sync(book.prepare)
             if args.forget:
@@ -808,7 +814,7 @@ async def apply(ctx: Context, config: Config) -> None:
         )
     )
 
-    async def install(argument: str, _invocation: CommandContext) -> str:
+    async def grammar_command(verb: Provisioning) -> str:
         """`/code-graph install|status` — make the grammars ready, on purpose.
 
         A **command** rather than a tool, per the seam's own rule: a person asks
@@ -817,9 +823,6 @@ async def apply(ctx: Context, config: Config) -> None:
         the long tail fetches — but "is this ready" should have one answer per
         plugin, asked the same way.
         """
-        verb = argument.strip().lower() or "status"
-        if verb not in ("install", "status"):
-            return f"/code-graph takes `install` or `status`, not {argument.strip()!r}."
         wanted = [one for one in (config.languages or DEFAULT_LANGUAGES) if indexable(one)]
         if verb == "status":
             ready, missing = await anyio.to_thread.run_sync(readiness, wanted)
@@ -834,11 +837,10 @@ async def apply(ctx: Context, config: Config) -> None:
             )
         return f"all {count_of(len(ready), 'grammar')} ready under {seam.grammars}."
 
-    command = CommandDefinition(
-        name="code-graph",
-        summary="Make the tree-sitter grammars ready, or report whether they are.",
-        argument_hint="[install|status]",
-        run=install,
+    command = install_or_status(
+        "code-graph",
+        "Make the tree-sitter grammars ready, or report whether they are.",
+        grammar_command,
     )
     contribute_item(
         ctx,

{ph_code_graph-0.4.0 → ph_code_graph-0.6.0}/tests/test_code_graph.py

@@ -47,6 +47,7 @@ from ph.llm.types import text_of
 from ph.testing import FAKE_OPTIONS, MountProfile, report_section, run_tool
 from ph.testing.git import git, git_repo
 from ph.testing.jj import jj_repo
+from ph.workspace_caches import WORKSPACE_FILE
 from ph_code_graph import CODE_GRAPH
 from ph_code_graph._extract import (
     INHERITS,
@@ -309,6 +310,25 @@ async def test_indexing_then_asking_answers_with_a_readable_pointer(
     assert f"{hit['path']}:{hit['start_line']}" in text_of(found.content)
 
 
+async def test_an_index_records_the_workspace_its_graph_serves(
+    mount: MountProfile, tmp_path: Path
+) -> None:
+    """What lets a sweep remove the graph of a worktree that is gone: its directory
+    names the workspace, and a digest cannot (`ph.workspace_caches`).
+
+    Sabotage: drop the `use` call from `code_index`, and nothing is recorded.
+    """
+    _tree(tmp_path)
+    ctx = await mount(ROW)
+    root = ctx.require(FS).root_for(None)
+
+    built = await run_tool(ctx, "code_index", {"paths": ["pkg"]}, agent=_agent(ctx))
+
+    assert not built.is_error, text_of(built.content)
+    recorded = ctx.require(CODE_GRAPH).store_for(root).path.parent / WORKSPACE_FILE
+    assert recorded.read_text(encoding="utf-8") == str(root)
+
+
 async def test_callers_names_the_calling_symbol_and_the_calling_line(
     mount: MountProfile, tmp_path: Path
 ) -> None:
@@ -903,7 +923,7 @@ async def test_every_row_in_the_rlm_indexed_profile_activates(
     the `PROFILES` entry and cannot test it — nothing there depends on the two
     indexing distributions, which is the point of discovering them as bundles.
     """
-    from ph.cordis import Context, Profile, load_profile_documents
+    from ph.cordis import Context, Profile, ProfileDocument, load_profile_documents
     from ph_app.profiles import resolve_profile
 
     # `PH_HOME`/`PH_CACHE` are not pinned here: the suite's autouse
@@ -911,7 +931,7 @@ async def test_every_row_in_the_rlm_indexed_profile_activates(
     # the first comes to be believed and bypassed.
     documents = load_profile_documents(resolve_profile("rlm-indexed"))
     documents.append(
-        (
+        ProfileDocument(
             "test-overlay",
             [
                 {"id": "fs", "config": {"root": str(tmp_path)}},