ci: gate docstring quality and coverage in CI (#616)#689
ci: gate docstring quality and coverage in CI (#616)#689planetf1 merged 18 commits intogenerative-computing:mainfrom
Conversation
|
The PR description has been updated. Please fill out the template for your PR to be reviewed. |
|
@psschwei @HendrikStrobelt @jakelorocco We are finally at a state where the docs build is clean, and reporting no issues. I recommend we now make this a hard-fail. We're had at least one PR merged which regressed some docstrings (albeit minor), and another pr in the queue with quite a few changes needed. By ensuring we validate as part of CI we'll maintain the docs in a clean state Note this does also add a slightly looser pre-commit. That is optional if we think too annoying |
psschwei
left a comment
psschwei
left a comment
There was a problem hiding this comment.
I'm in favor of adding, but one question: is the output of our docstring checker the kind of thing that could be easily fed into an agent to produce the appropriate docstrings / make the check pass? (thinking of (a) keeping contribution process easy for new folks, and (b) how we can use agents at some point to do this for us)
Merge ProtectionsYour pull request matches the following merge protections and will not be merged until they are valid. 🟢 Enforce conventional commitWonderful, this rule succeeded.Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/
|
|
Yes, much of the docs have been written and restructured by AI - long term the ability to generate too is on the cards. I think there will be some incremental improvements to make scan results easier... |
planetf1
force-pushed
the
ci/doc-quality-gate-616
branch
from
364fb81 to
71b4bde
Compare
psschwei
left a comment
psschwei
left a comment
There was a problem hiding this comment.
I am in favor, but would be good for @jakelorocco to weigh in too
|
A good time to consider forcing 
|
planetf1
force-pushed
the
ci/doc-quality-gate-616
branch
2 times, most recently
from
58a7f44 to
dfb95c4
Compare
|
An intrinsics updated has introduced some documentation errors - not fatal: missing args, and missing returns (10 omissions) -- but having seen 2 other PRs potentially doing this too, this is why we MUST start enforcement, not just the checks, which were the facilitator of improving the docs. @jakelorocco @HendrikStrobelt What do you think? Are you ok with this in principle? If so I will add the fix-ups into this PR tomorrow and get it approved/merged ASAP. Any concerns? |
|
vv This was the regression 
|
jakelorocco
left a comment
jakelorocco
left a comment
There was a problem hiding this comment.
I think this is my bad for not understanding this earlier, but it looks like we are checking the docstring quality only in terms of our generated docs. I wonder if this is over-complicating things and we should instead be checking docstring quality by looking at our code. Is there an advantage to only checking the quality of things in our generated docs? If we switched to ensuring quality at the code level, could we rely on third party tools / packages to do the checks?
| | `missing_param_type` | `Args:` section exists but one or more parameters have no Python type annotation — the type column is absent from the generated API docs | Add a type annotation to each listed parameter in the function signature (e.g. `def f(x: int)`). Only fires when `no_args` is already satisfied; `*args`/`**kwargs` are excluded. | | ||
| | `missing_return_type` | `Returns:` section is documented but the function has no return type annotation — the return type is absent from the generated API docs | Add a return annotation to the function signature (e.g. `-> str`). Only fires when `no_returns` is already satisfied. | | ||
|
|
||
| #### Class docstrings (Option C) |
There was a problem hiding this comment.
I think Option C should be removed here? It looks like this was originally a list of fix options but we now accept all of them?
There was a problem hiding this comment.
that section has gone - just a single reference to the documentation
|
Improved checks - clarification 
Looking at other comments: |
It's a journey.. the deliverable was docs so my original plan was to minimize code changes -- at least for the first pass - hence the validation focusses on code But.. as a strategy, yes I agree, it is more of a code-based check (the current tool builds the AST). There may be options to use other tools/plugins to achieve similar checks in a more structured way -- but this first version was a pragmatic way to get started and improve quality. Ideal to then reflect/review on how we can do it better in future. Also related - #456 : our current mypy configuration isn't checking for missing types -- yet that is confusing for external classes. ruff checks some things, but not everything. Ruff can do the following -
but what it doesn't easily do is address Public vs Private (the current parser looks at what is exported) - so we'd need to move to cleanup all code including internal, or we'd have to restructure and/or be very careful in creating an exclusion list. I think this would best be tracked in a new issue? What we have focusses on external |
I think that's fair. I am all for getting these checks merged and then evaluating if there's versions of this that can work directly on the code. After running the doc generation pipeline locally though, I'm not too concerned about the fact that it runs on the docs since that code is quick enough. Thank you. |
Add a hard-fail docstring quality gate to the docs-publish workflow: - New 'Docstring quality gate' step runs --quality --fail-on-quality --threshold 100; fails if any quality issue is found or coverage drops below 100% (both currently pass in CI) - Existing audit_coverage step (soft-fail, threshold 80) retained for the summary coverage metric Add typeddict_mismatch checks to audit_coverage.py: - typeddict_phantom: Attributes: documents a field not declared in the TypedDict - typeddict_undocumented: declared field absent from Attributes: section - Mirrors the existing param_mismatch logic for functions Pre-commit: enable --fail-on-quality on the manual-stage hook (CI is the hard gate; hook remains stages: [manual] as docs must be pre-built). Update CONTRIBUTING.md and docs/docs/guide/CONTRIBUTING.md with TypedDict docstring requirements and the two new audit check kinds.
… and fix hints audit_coverage.py: - Add file/line fields to every issue dict (repo-relative path + def line) - _print_quality_report now shows [file:line] per issue, per-kind Fix:/Ref: hints linking to CONTRIBUTING.md anchors, and emits ::error file=...,line=... GHA annotations so issues appear inline in PR diffs - Cap GHA annotations at 10 per check kind with "N more in job log" notice - Add _KIND_FIX_HINTS and _gha_file_annotation helpers; _CONTRIB_DOCS_URL constant validate.py: - Convert all check functions from list[str] to list[dict] errors (file/line/message) - Add line-number tracking to validate_source_links, validate_internal_links, validate_anchor_collisions, and validate_doc_imports - Emit per-error GHA annotations with file/line; shared 20-annotation budget across all checks so every category gets representation in PR diff - Fix icon bug: summary rows now use correct pass/fail icon - Group detailed errors by check type with section headers docs-publish.yml: - Add --orphans and --output /tmp/quality_report.json to quality gate step - Upload quality_report.json as docstring-quality-report artifact (30-day retention) pyproject.toml: - cli/**/*.py: suppress only D2/D3/D4xx style rules; enable D1xx (missing docstrings) as a ruff-level complement to the audit_coverage quality gate docs/docs/guide/CONTRIBUTING.md: - Add CI docstring checks reference section with per-kind tables (fix instructions + anchors) for all 11 check kinds across 4 categories - Add callout explaining GHA annotation cap (10 per kind) and where to find the full list (job log + JSON artifact)
…y links
audit_coverage.py (audit_nav_orphans):
- Probe docs/docs/docs.json before docs/mint.json so both Mintlify v1
and v2 nav configs are supported
- Extend _extract to handle plain string page entries used by docs.json
(v2 uses "pages": ["api/..."] strings; v1 used {"page": "api/..."} dicts)
- Previously mint.json was never found, nav_refs stayed empty, and every
MDX file was reported as an orphan
docs-publish.yml (Write job summary):
- When the quality gate fails, render a prominent markdown callout with a
direct link to the CI docstring checks reference section in CONTRIBUTING.md
- Add a per-kind fix reference table with clickable anchor links to each
category section (missing/short, args/returns, class Option C, TypedDict)
- Per-kind Ref: URLs in the raw log are inside a text block and do
not render as links in the step summary; this table surfaces them rendered
…ped notice docs-publish.yml: - Parse per-kind counts from _print_quality_report section headers in the quality gate log (e.g. "Missing docstrings (12)") and show them as a comma-separated breakdown in the Docstring Quality table cell instead of just the total — gives developers an immediate view of which categories are failing without expanding the log audit_coverage.py: - Remove the "skipped (pass --quality to enable)" GHA notice emitted by the coverage-only step; there is always a dedicated quality gate step immediately after so the notice was misleading and redundant
audit_coverage.py: - Coverage miss section now shows a structured Fix:/Ref: block with the exact generate-ast.py command and a link to CONTRIBUTING.md#validating-docstrings - Missing symbols listed one per line (symbol indented under module) for scannability instead of comma-joined on one long line - Emit a ::error or ::warning GHA annotation with symbol/module counts when coverage symbols are undocumented validate.py: - Add _CHECK_FIX_HINTS dict mapping each check label to a (fix text, ref URL) pair, covering all 8 check types with specific fix instructions and links into CONTRIBUTING.md (root or guide as appropriate) - _print_check_errors now prints Fix:/Ref: under each section header, matching the pattern established by _print_quality_report
audit_coverage.py: - Add missing_param_type check: fires when Args: section exists but one or more concrete params lack Python type annotations; naturally non-overlapping with no_args (which fires when section is absent) - Add missing_return_type check: fires when Returns: section is documented but the function has no return annotation; naturally non-overlapping with no_returns (annotation exists but section absent) - Add fix hints and CONTRIBUTING.md anchors for both new check kinds - Update kind_labels and iteration order in _print_quality_report generate-ast.py: - Add remove_internal_modules() post-generation filter step - Uses AST-based import analysis: a submodule is internal when the parent __init__.py imports from at least one sibling submodule but not from this one (import-based visibility, not __all__ name-matching) - Conservative: keeps module when parent imports nothing (indeterminate) or __init__.py cannot be parsed - _CONFIRMED_INTERNAL_MODULES hardcoded set for known internals where parent imports nothing (json_util, backend_instrumentation); these should eventually be renamed with _ prefix per Python convention - Package index files (stem == parent dir) are never filtered docs.json: nav regenerated by build; internal modules removed from nav CONTRIBUTING.md: add missing_param_type / missing_return_type to CI docstring checks reference table docs-publish.yml: add both new kinds to summary kind_short and kind_anchors fix-reference table
audit_coverage.py was walking all non-_-prefixed source modules via Griffe, including internal modules (json_util, backend_instrumentation, etc.) whose MDX files were removed by remove_internal_modules() in generate-ast.py. This caused coverage to drop because those symbols were no longer 'documented' but were still counted in the denominator. Apply the same import-based public-API filter in discover_public_symbols(): skip submodules that the parent __init__.py does not import from, mirroring the generate-ast.py logic. _CONFIRMED_INTERNAL_MODULES kept in sync. Also drop the per-kind anchor table from the GHA job summary. Anchors in GitHub Actions summaries only navigate to the top of the referenced document, so the table added noise without working links.
Adds two new quality check kinds that fire when the type explicitly
stated in an Args:/Returns: docstring entry disagrees with the Python
type annotation in the function signature:
param_type_mismatch — 'param (OldType): ...' vs annotation 'NewType'
return_type_mismatch — 'Returns: OldType: ...' vs annotation '-> NewType'
Both checks fire only when BOTH sides have an explicit type; one-sided
absence is already handled by missing_param_type / missing_return_type.
Type comparison uses _types_match() / _normalize_type() which handles:
- typing aliases: List→list, Dict→dict, Optional→X|None, Union→A|B
- typing. prefix stripping
- pipe-union component ordering (str|None == None|str)
- incidental whitespace
Known conservative suppressions (prefer false negatives over false positives,
since there is no per-site suppression mechanism):
- Nested generics not fully expandable by regex (e.g. Optional[list[str]])
are silently skipped — both sides must fully normalise to be compared
- Union with bracket-containing members
- Callable argument ordering
…ubmodule For a module file mellea/pkg/submodule.py, Griffe gives filepath ending in .py (not __init__.py). The parent __init__.py is fp.parent/__init__.py. The previous code used fp.parent.parent which is correct for packages (whose filepath IS the __init__.py) but goes one level too far for plain module files — it was checking the grandparent init instead of the parent. Effect: genslot, react, unit_test_eval and similar non-exported modules in stdlib/components were incorrectly counted as public symbols, inflating the denominator and lowering the reported coverage percentage.
planetf1
force-pushed
the
ci/doc-quality-gate-616
branch
from
48b32b4 to
4bfcd0a
Compare
|
Ok I think we're good to push on closing out this To fix I'll do the docstring updates & add to this PR so we can merge - quickly. I'll take first approval |
- Fix missing_param_type, missing_return_type, param_type_mismatch, return_type_mismatch, no_args, no_returns, and missing docstring issues - Add TYPE_CHECKING imports for HuggingFace types in util.py with type: ignore[union-attr] for pre-existing None-safety gaps - Add Granite3ChatCompletion import to granite32/33 input.py for correct sanitize() parent signature match - Convert reST-style docstrings to Google style in intrinsics/input.py - Document AST single-quote normalization for Literal types in CONTRIBUTING.md
|
Fixed the outstanding omissions/corrections. Analysis: Docstring Quality Gate Fixes — Change ClassificationTier 1: Docstring-Only (Zero Risk)Pure docstring text changes — no code, no imports, no annotations touched.
Tier 2: Type Annotation Additions (Low Risk)Adding type annotations to previously untyped parameters/returns. No runtime behavior change — Python ignores annotations at runtime.
Tier 3: Requires Justification (Medium Risk)Changes involving imports or type annotations that interact with inheritance or optional dependencies. 3a. Granite 3.2/3.3
|
Adding TYPE_CHECKING annotations to util.py made mypy check function bodies it previously skipped (untyped params = implicit Any = no body checking). This exposed a pre-existing Tensor-not-callable issue and a dict-variance issue in mobject.py. Suppress with targeted type: ignore comments — these are not new bugs, just newly visible ones.
a4b0a71
3 participants
Misc PR
Type of PR
Description
Adds a hard-fail docstring quality gate to the docs-publish workflow (
--quality --fail-on-quality --threshold 100). Both checks currently pass in CI (100% coverage, 0 quality issues).Also adds a
typeddict_mismatchscanner toaudit_coverage.py— flagsAttributes:sections onTypedDictclasses that document phantom fields or omit declared ones (mirrors the existingparam_mismatchlogic for functions).Pre-commit hook updated to use
--fail-on-quality; staysstages: [manual]since it requires pre-built docs. CI is the hard gate.Contribution docs updated with TypedDict docstring requirements and the two new check kinds.
Docstring fixes (to reach 0 quality gate issues)
To pass the new gate, 36 docstring issues across 17 files were fixed. Changes are classified below by risk level.
Docstring text only (zero risk):
cli/decompose/decompose.py,cli/eval/runner.py,mellea/core/sampling.py,mellea/core/utils.py,mellea/stdlib/functional.py,mellea/stdlib/sampling/base.py,mellea/formatters/granite/intrinsics/input.py,cli/alora/intrinsic_uploader.py. Pure text changes — missingReturns:sections, corrected return types in docstrings, reST-to-Google-style conversion, andLiteral['a', 'b']single-quote normalisation (Python's AST normalises string literals, so docstrings must use single quotes to match).Type annotation additions (low risk):
cli/alora/train.py,mellea/backends/tools.py,mellea/core/utils.py,mellea/helpers/openai_compatible_helpers.py,mellea/stdlib/components/mify.py,mellea/stdlib/session.py,mellea/stdlib/sampling/base.py. Adding type annotations to previously untyped parameters and return values. No runtime behaviour change — Python ignores annotations at runtime.Requires justification:
granite32/input.py,granite33/input.py—sanitize()override: Addedfrom ...granite3.types import Granite3ChatCompletionimport and updated thesanitize()signature to use it. Required because the parent class declaressanitize(cls, chat_completion: Granite3ChatCompletion, ...)and mypy flagged an incompatible override.Granite3ChatCompletionis already loaded transitively through the parent import chain; 7 other files use the same import. Method body is unchanged.mellea/formatters/granite/base/util.py— HuggingFace optional-dependency types: Addedfrom __future__ import annotations, aTYPE_CHECKINGguard importingPreTrainedModelandPreTrainedTokenizerBase, and typed 4 previously untyped params.TYPE_CHECKINGis dead code at runtime — the HF types are never imported when the module loads. The| Noneontokenizer/modelis truthful (they have=Nonedefaults).Adding real types caused mypy to check function bodies it previously skipped (untyped params = implicit
Any= no body checking). This exposed pre-existing issues, suppressed with targetedtype: ignorecomments:# type: ignore[union-attr]— pre-existing None-unsafe attribute access ontokenizer/model(e.g. calling.apply_chat_template()without a None guard). Real gaps, out of scope for a docs PR.# type: ignore[operator]ingenerate_with_transformers— HuggingFace stub quality issue: stubs markPreTrainedModel.generatein a way mypy reads as non-callable, but it is callable at runtime.mellea/stdlib/components/mobject.py: 1×# type: ignore[arg-type]added. Same root cause — our annotations changed mypy's import graph enough to trigger checking of this function. The underlying issue is a dict-invariance warning:dict[str, MelleaTool]passed wheredict[str, AbstractMelleaTool] | Noneis expected.MelleaToolis a subtype ofAbstractMelleaTooland the dict is not mutated through the parameter, so this is safe at runtime. Pre-existing issue, not introduced by this PR.All
type: ignoresuppressions are narrow (error-code qualified) and document pre-existing issues that should be addressed in a dedicated type-cleanup issue.Testing