Chore/code quanlity 2#3
Merged
Merged
Conversation
Coverage uplift to a hard 90% pre-push gate (pytest-cov, statement+branch): - daemon 76% -> 90.78% (449 tests, was 256) - search 63% -> 96.12% (171 tests, was 62) - hooks 57% -> 91.97% (213 tests, was 38 + 7 broken) Test infrastructure: per-package conftest fixtures (tmp knowledge dirs, fake SDK MagicMock via mock.patch.dict, fake priming sockets, hook runner). Hyphenated hook entrypoints (pre-tool-use.py et al) refactored to 8-line shims importing matching underscore modules — the only way to exercise them in-process since hyphens aren't legal identifiers. Complexity violations cleared (ruff C901/PLR0912/PLR0915): - Extract Method for sequential pipelines (scholar.review, ingest.poll_once, library_tools._move_entries_impl, _priming_client._local_priming, ...) - Strategy chain for parsers (_response_parser.extract_json_blob, _blockers._parse_block_triggers, _events.append_event) - StrEnum dispatch dict for tools/search cli main (Subcommand) and daemon priming_rpc handler (RpcOp) - Rule-list refactor for scholar_prompt.check_invariants (complexity 34 -> four pure check functions composed at the call site) Python floor unified to 3.12 across all three packages (was 3.10/3.11/3.12), pyright pythonVersion and ruff target-version aligned. Authors set on each package pyproject. No test-detection branches added to production code. All helpers underscore-prefixed. Public API unchanged. Behavior preserved — every existing test still passes against the refactored code. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three small typing-hygiene moves, no mode change (still basic):
- venvPath/venv pointed at each package's .venv so pyright's LSP can
resolve claude_agent_sdk imports without uv's VIRTUAL_ENV. Fixes the
editor-only "reportMissingImports" diagnostics that pre-commit's
pyright (running under uv) never saw.
- reportUnusedImport promoted to "error" — baseline was 2 violations,
both fixed in this commit (real dead-import cleanup, not suppression).
- reportPrivateUsage at "warning" — surfaces _underscore imports
crossing module boundaries (mostly tests reaching into internals);
informational only, doesn't block.
Cleanups for the two reportUnusedImport hits:
- priming_rpc.py used "from embeddings import load_or_build as _emb_check"
purely to test importability; switched to __import__("embeddings") which
doesn't bind a name.
- test_events.py had a redundant `import _events` followed by
importlib.import_module("_events") on the next line; dropped the bare
import.
Not adding a claude_agent_sdk .pyi stub — the SDK ships py.typed with
full type info in its own types.py. A local stub would duplicate
upstream and drift over time. The IDE "missing imports" warnings were
about venv discovery, not type info.
Held back on reportArgumentType / reportOptionalMemberAccess /
reportAttributeAccessIssue — those have unknown baselines once Unknown
cascades come in, and turning them on without first sizing the
backlog would brick pre-commit. Follow-up.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The "(bm25 backend unavailable)" / "(embeddings not importable)" silent-fallback shape masked real packaging bugs by degrading at runtime. All 7 sites wrapped imports of HARD-declared deps (bm25, embeddings, markdown_it, sentence_transformers, claude_agent_sdk) so the guards never legitimately fired. Removed the try/except wrappers; imports now fail loudly when the dep is genuinely missing. Where collision exists across the agent-mem-search flat layout (bm25.load_or_build vs embeddings.load_or_build) names are aliased at module top. Three dead tests (asserting the now-removed fallback strings) deleted. Test patches that targeted sys.modules now patch the binding on the importing module — proper monkeypatch pattern. tools/search adds a `fake_sdk` conftest fixture replacing the old `_install_fake_sdk` context manager so cli.py's now-module-top SDK imports are still mockable. Ruff PLC0415 (import-not-at-top-of-file) added to all three pyprojects, with tests/conftests excepted. Hoisted a few additional inline imports the new rule surfaced (yaml in priming, claude_agent_sdk in library_tools, .paths in priming). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-commit pyright runs on staged .py files only, so a toml-only commit doesn't trigger pyright. After this lands, every Python edit touched by pre-commit must pass strict-mode pyright. Baselines to fix in subsequent parallel passes: - daemon: 399 errors across 19 source files - src: 337 errors across 25 source files (hooks + scripts) - tools/search: 106 errors across 5 source files Total: 842 strict-mode errors. Tests are excluded from pyright entirely (still ruff-checked, pytest-run, coverage-counted). Per user direction: "lets not bother with strict types on tests, lets just do standard." Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pre-push was running pyright on every .py file changed since origin, which blocks pushes when strict-mode is on top of an existing baseline of unfixed errors. Pyright already gates at pre-commit; pre-push retains pytest. No loss in coverage — just removes a redundant gate that conflicts with the parallel strict-mode fixup workflow. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Eliminates 133 strict-mode pyright errors across the seven CLI script entrypoints. The core moves: - Add a ``State`` / ``IngestedEntry`` TypedDict in utils.py and thread it through compile.py, query.py, lint.py, flush.py so the on-disk state.json schema is enforced statically. - Add a ``FlushState`` TypedDict in flush.py for the dedup sidecar file. - Add an ``Issue`` TypedDict (severity Literal + NotRequired auto_fixable) in lint.py; replace every ``list[dict]`` return / accumulator with ``list[Issue]``. - Replace scope.py's untyped ``os.PathLike`` with a parameterised ``Union[str, os.PathLike[str]]``. - Replace flush.py's untyped ``kwargs: dict`` Popen spread with explicit per-platform Popen calls — ``creationflags`` (win) vs ``start_new_session`` (POSIX) — keeping every argument concretely typed. - Lift the lazy ``claude_agent_sdk`` / ``subprocess`` / ``traceback`` / ``hashlib`` imports out of function bodies so PLC0415 doesn't trip when ruff re-checks the touched files. No new ``# type: ignore``, no ``Any`` shortcuts, no behaviour changes. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… runs, buffer, main, ...) - scheduler.py: type the tailer (JsonlTailer), event payload, in-flight Future set, and debounce callbacks. Replace the redundant `isinstance(packet, dict)` with a `cast(object, …)` round-trip so the runtime guard against contract-violating librarian callables survives pyright's narrowing (test_librarian_non_dict_return_is_dropped still covers it). - runs.py: parameterise the cost-state dicts (`Dict[str, Any]`) and cast the json.loads result through the isinstance check. - buffer.py: parameterise the dataclass default_factory calls so deque / list / dict fields carry real element types. - __main__.py: signal handler now takes `(int, FrameType | None) -> None`. Hoist `knowledge_dir` to a top-level import and rename the `_prewarm_indexes` parameter to avoid shadowing; the optional `bm25`/`embeddings` imports stay local with documented `noqa: PLC0415` since they live in a sibling search package not always on sys.path. 0 strict-mode errors across all 8 files in the slice; full test suite (447 tests) green. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- frontmatter.py: introduce `Frontmatter = dict[str, object]` as the in-memory YAML shape; tighten `read`/`write`/`set_status` signatures. Cast PyYAML's untyped `represent_scalar` boundary cleanly. - bm25.py: add `BM25Hit` NamedTuple for typed search returns (tuple subclass — daemon's existing `(p, score, _snippet)` destructure keeps working). Type rank_bm25 surface via a local `_BM25Backend` Protocol + cast at the boundary. Expose non-private aliases (`build_snippet`, `frontmatter_search_text`, `iter_markdown`, `strip_and_extract_frontmatter`) so embeddings.py can import them without tripping reportPrivateUsage. - embeddings.py: switch to the new public aliases; annotate the embedding matrix as `npt.NDArray[np.float32]`; parameterise the dataclass `field(default_factory=list[_DocRecord])` so element type isn't lost. - cli.py: thread `Frontmatter`/`Mapping[str, object]` through every fm helper. Add `_as_float` to coerce JSON-typed cost-state values safely. Cast `list` narrowings when iterating `applies-when`. - aliases.py: type the json.loads boundary via `object` + isinstance, then cast the narrowed dict for typed iteration. No new try/except, no `Any` shortcuts (only a single documented cast through Any at the PyYAML `represent_scalar` boundary). 170 tests pass; pyright clean on the slice. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_schemas.py: - Pyright cannot infer the parameter of ``list`` from a bare ``default_factory=list``; spell it out as ``default_factory=list[T]`` for every non-``str`` element type (EvidenceItem, ProposedAction, LibrarianInterrupt, ScholarDecision, ScholarInterruptDecision, Dict[str, object]). Same shape, just an explicit factory generic. - Annotated the ``_coerce_evidence`` field_validator: ``v: object`` in, ``object`` out. Replaced ``isinstance(v, dict)`` with the wider ``Mapping`` check (any user-supplied mapping passes); narrowed via ``cast`` where pyright keeps the type opaque inside ``isinstance`` branches. - Replaced the legacy ``Dict[str, Any]`` shapes on ``LibrarianResponse`` and ``ScholarResponse`` with ``Dict[str, object]`` — semantically the same catch-all but without ``Any``. llm.py: - The Claude Agent SDK ships ``py.typed``. Threaded the proper types through every SDK boundary instead of bare ``dict``: ``ClaudeAgentOptions`` for options, ``McpServerConfig`` for the mcp_servers dict, ``ToolPermissionContext`` on the can_use_tool callback's third arg, and an explicit ``Callable[..., Awaitable[PermissionResult*]]`` alias on ``_make_path_guard``'s return. - Replaced ``opts: dict = dict(...); ClaudeAgentOptions(**opts)`` in ``run_librarian_call`` with a direct ``ClaudeAgentOptions(...)`` construction. The previous pattern hid every keyword's type behind a generic ``dict`` and produced ~80 of the 111 errors here. ``cwd`` and ``can_use_tool`` are passed as ``None`` when no knowledge root is supplied — semantically identical to omitting them (both default to ``None`` on the dataclass). - ``_recursion_guard_env`` now returns ``dict[str, str]`` (not bare ``dict``); ``_run_with_timeout`` is generic over the coroutine's return type via TypeVar so callers don't lose the inner result type. - Added ``# noqa: PLC0415`` to the existing lazy SDK imports — those are deliberate (keep daemon startup light if the SDK is absent), and ruff's strict in-function-import rule trips once the file is touched. _response_parser.py: had zero strict errors and remains untouched (re-verified post-change). Two ``reportPrivateUsage`` warnings on ``library_tools._SERVER_NAME`` remain — pre-existing and out of slice (would need a public alias added in library_tools.py). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Zero strict-mode pyright errors across all 12 source files in src/hooks/ (was 204). 213 hook tests still pass; coverage holds at 91.99% (>= 90% gate). Shared TypedDicts hoisted into _events.py so every hook in the slice annotates the same stdin shape: - HookPayload (total=False) — session_id, cwd, transcript_path, tool_name, tool_input, tool_response, error, prompt, source, reason. - EventLine — one serialised event-line schema. - EventPayload — the small dict shipped on the wire as the event payload. _priming_client.py also gained PrimingRequest / PrimingResponse TypedDicts matching daemon/agent_mem_daemon/priming_rpc.py's wire shapes, plus a Frontmatter alias for the lite-YAML scan path and a ScoredEntry tuple alias for the rank pipeline. The connection helper was split into _send_request + _parse_response to keep cyclomatic complexity under ruff's C901 cap. scope.current_project_slug is owned by another slice and still uses an untyped os.PathLike parameter, so its inferred signature leaks Unknown into every caller. session_start.py, session_end.py, pre_compact.py and user_prompt_submit.py now re-bind it through getattr + cast to a clean str | None -> str callable at module top — keeps strict mode happy without touching upstream and also resolves ruff's PLC0415 local-import warning these files were already tripping. _nudges.take_nudges's per-session budget loader extracted to a helper (_load_consumed) for the same complexity reason. No new try/except wrappers, no # type: ignore, no Any shortcuts — narrow Any locals appear only at JSON-parse boundaries where the input genuinely is unknown.
Worktree agents fanned out from f25f01e, each fixing a disjoint slice of source files under typeCheckingMode='strict': A (09005fa) daemon: llm + _response_parser + _schemas (133 → 0) B (fe6804e) daemon: priming + priming_rpc ( 89 → 0) C (a228c8f) daemon: librarian + librarian_prompt + library_tools (60 → 0) D (51f7175) daemon: scholar + scholar_prompt + ingest ( 74 → 0) E (8408e16) daemon: scheduler + runs + buffer + main + ... ( 44 → 0) F (5054780) src/hooks/* (12 files) (204 → 0) G (c23b73a) src/scripts/* (7 files) (133 → 0) H (b246143) tools/search source modules (5 files) (106 → 0) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Agent F's worktree pre-merged needed local getattr+cast wrappers in pre_compact.py, session_start.py, session_end.py, and user_prompt_submit.py because scope.current_project_slug's original signature used a bare os.PathLike and leaked Unknown. Agent G's parallel worktree fixed scope.py to use os.PathLike[str], which makes those workarounds redundant. Replace each with a direct `from scope import current_project_slug`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The constant is the MCP server registration key — daemon-wide config that llm.py looks up in two places. Leading underscore was a lie about its scope; dropping it makes the cross-module access legitimate and clears the last two pyright warnings (reportPrivateUsage) in daemon. _BM25_TOOL_NAME and _MOVE_TOOL_NAME remain private (only used through the public fully_qualified_bm25_name / fully_qualified_move_name helpers). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
# Conflicts: # src/hooks/test_user_prompt_submit.py
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.