pyproject.toml

[build-system]
requires = ["setuptools>=82.0.1"]
build-backend = "setuptools.build_meta"

[project]
name = "podcast-scraper"
version = "2.7.0.dev0"
description = "Download podcast transcripts from RSS feeds with optional Whisper fallback"
readme = "README.md"
authors = [{name = "Podcast Scraper Maintainers"}]
license = "MIT"
requires-python = ">=3.10"
dependencies = [
  "requests>=2.31.0,<3.0.0",
  # Server HTTP client, required unconditionally by the create_app import chain (consumer
  # OAuth app_oauth, audio-bridge HEAD validation, reference client, job webhooks) + Ollama
  # health checks + the [observability] control plane. A true base dep: the api image installs
  # only `.[search,sentry]`, so httpx must live in core — not only in extras.
  "httpx>=0.28.1,<1.0.0",
  "urllib3>=2.6.0,<3.0.0",
  "rich>=15.0.0,<16.0.0",
  "defusedxml>=0.7.1,<1.0.0",
  "platformdirs>=3.11.0,<5.0.0",
  "pydantic>=2.13.4,<3.0.0",
  "PyYAML>=6.0.3,<7.0.0",
  "python-dotenv>=1.2.2,<2.0.0",  # For .env file support (RFC-013)
  "zipp>=4.1.0,<5.0.0",  # Security fix: CVE-2024-50208 (infinite loop vulnerability)
  "jinja2>=3.1.6,<4.0.0",  # For prompt templating (RFC-017)
  "filelock>=3.29.7,<4.0.0",  # Security fix: CVE-2025-68146 (TOCTOU race condition)
  # CVE-2026-4539 (ReDoS in AdlLexer, NVD/GHSA: 2.19.0–2.19.2): stay on 2.18.x until 2.19.3+; pip-audit still needs --ignore-vuln until OSV range matches
  # TODO(CVE-2026-4539): After pygments ships a fixed release and pip-audit/OSV match NVD ranges, bump the cap and drop Makefile ignore.
  "pygments>=2.20.0,<2.21.0",
  "psutil>=5.9.0,<8.0.0",  # RFC-065 monitor / sampler (imported with orchestration); Docker base image
  # Pure-Python VADER lexicon sentiment for the deterministic `insight_sentiment` enricher
  # (conversation-timeline sentiment colour). Bundles its own lexicon → no network / no model
  # download → CI-safe + airgapped-safe; a true deterministic-tier dep, so it lives in core.
  "vaderSentiment>=3.3.2,<4.0.0",
  # NOTE: no 3rd-party observability SaaS SDKs here — Sentry/Langfuse are fully
  # optional opt-in extras ([sentry] / [langfuse]); the core no-ops without them.
  # See the [project.optional-dependencies] header + docs/guides/OBSERVABILITY_EXTENSIONS.md.
]

[project.optional-dependencies]
# Canonical extras (no separate [gemini] / [openai] extras):
#   dev       — tests, lint, typecheck, FastAPI/uvicorn (viewer API), text eval metrics; no torch/transformers
#   ml        — Whisper, spaCy (+models), torch, transformers, sentence-transformers, lancedb,
#               llama-cpp-python (GGUF hybrid REDUCE, RFC-042), etc.
#   llm       — API SDKs: openai, google-genai (Gemini), anthropic, mistralai, httpx (Ollama checks)
#   search    — Minimum local ML for the vector index (sentence-transformers + lancedb + torch CPU).
#               Pulled into the cloud-API pipeline image alongside [llm] so cloud_balanced can
#               build its LanceDB index without dragging in spaCy/Whisper/Pegasus/llama-cpp from [ml].
#   compare   — Streamlit compare UI (RFC-047)
#   monitor   — py-spy + memray optional profiling (RFC-065; `--monitor` needs no extra)
#   (FastAPI / uvicorn / metrics / scheduler for ``podcast serve`` live in [dev], not a separate extra.)
#   observability — standalone prod control plane (#803, ``podcast_obs``). LIGHT by design.
# Standalone prod observability control plane (#803). Intentionally LIGHT so it runs as a small
# container anywhere on the tailnet — no torch/spaCy/pipeline deps. (httpx + PyYAML are already
# base deps; listed here so a future slim ``podcast_obs``-only install is self-describing.)
observability = [
  "httpx>=0.28.1,<1.0.0",
  "PyYAML>=6.0.3,<7.0.0",
  "mcp>=1.2.0,<2.0.0",  # FastMCP server for the agent-facing layer (stdio + sse/http)
]

# ── Optional 3rd-party observability extensions ──────────────────────────────────
# o11y SaaS integrations are FULLY OPTIONAL, one extra per service. The core carries
# ZERO of these deps and no-ops without them (every touchpoint is a lazy guarded import
# + config-gated). Two distinct kinds of o11y dep:
#   • READ an o11y backend (Grafana / Loki / Sentry API / Langfuse API) → httpx only, in
#     [observability] (the light control plane). No vendor SDK — uniform across the 5
#     podcast_obs sources.
#   • EMIT from the app → the service's vendor SDK, one extra each below. Each is a
#     config-gated no-op (active only when its secret is present) AND degrades to a no-op
#     when the SDK isn't installed. Install only what you use; compose them into the prod
#     images you deploy (docker/{api,pipeline}); [dev] pulls all of them so tests/CI
#     exercise them. Adding a new one → docs/guides/OBSERVABILITY_EXTENSIONS.md.
sentry = [
  # Error reporting for the api (FastAPI) + pipeline (cli) via
  # podcast_scraper.utils.sentry_init.init_sentry(); no-op without PODCAST_SENTRY_DSN_*.
  "sentry-sdk>=2.60.0,<3.0.0",
]
langfuse = [
  # LLM tracing emitted at the provider cost choke point (#1052); no-op without
  # LANGFUSE_PUBLIC_KEY + LANGFUSE_SECRET_KEY. Pulls OpenTelemetry.
  "langfuse>=4.9.0,<5.0.0",
]
otel = [
  # Distributed tracing -> OTLP (ADR-119 "traces" signal; VictoriaTraces reference
  # backend). `opentelemetry-instrument` auto-wraps the process and
  # `opentelemetry-bootstrap -a install` (run in the Dockerfile) adds per-library
  # instrumentations. No-op unless OTEL_TRACES_EXPORTER=otlp + an OTLP endpoint are
  # set — safe to bake into the image. See docs/guides/OBSERVABILITY_ARCHITECTURE.md.
  "opentelemetry-distro>=0.65b0",
  "opentelemetry-exporter-otlp-proto-http>=1.44.0",
]

# Heavy ML dependencies - only install when needed
ml = [
  "openai-whisper>=20250625",
  # spaCy 3.8 / thinc 8.3 for numpy 2.x compatibility (required by pyannote.audio 4.x).
  "spacy>=3.8.0,<4.0.0",
  # spaCy's CLI imports ``click`` via typer (``spacy/cli/_util.py``).
  # typer-0.26.0+ dropped click from its hard deps in favour of having
  # consumers depend on it directly. The cloud-thin / API images still
  # get click transitively (uvicorn / fastapi), but the [ml] pipeline
  # image has no such fallback — ``import spacy`` fails with
  # ``ModuleNotFoundError: No module named 'click'`` if click isn't
  # explicit here. Surfaced 2026-05-27 when typer-0.26.1 first appeared
  # in resolution.
  "click>=8.0.0,<9.0.0",
  # spaCy language models are distributed via GitHub releases (not PyPI) due to size constraints.
  # This is the official and standard method recommended by spaCy.
  # Version compatibility: spaCy 3.8.x requires en_core_web_sm 3.8.x and en_core_web_trf 3.8.x
  # See: https://spacy.io/usage/models for model version compatibility
  # Keep scripts/spacy_model_wheels_requirements.txt in sync (make download-spacy-wheels).
  "en-core-web-sm @ https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl",
  "en-core-web-trf @ https://github.com/explosion/spacy-models/releases/download/en_core_web_trf-3.8.0/en_core_web_trf-3.8.0-py3-none-any.whl",
  "torch>=2.11.0,<3.0.0",
  # numpy 2: spaCy 3.8 + thinc 8.3 (above) support it — the migration the previous
  # numpy<2 pin (#899) was waiting for — and pyannote.audio 4.x REQUIRES numpy 2.
  # The old thinc 8.2 ABI crash is gone with thinc 8.3.
  "numpy>=2.2.6,<3.0",
  # transformers v5 (#382): pipeline() removed for summarization / text2text-generation /
  # question-answering — call sites rewritten to AutoModel*.from_pretrained + generate() /
  # QA-head forward. Also picks up the CVE-2026-1839 fix (Trainer rng_state / torch.load
  # weights_only) that shipped in >=5.0.0rc3. See docs/wip/ISSUE-382-*.md + ADR-068 post-impl.
  # Native-wheel drift guard (2026-07-17): the acceptance job was green on
  # 84b17ab (07-11, transformers 5.13.1 / sentencepiece 0.2.1 / torchcodec
  # 0.14.0) and started SIGSEGV/SIGBUS-ing at ML-stack teardown on 2174be4
  # (07-16) with ZERO code cause — the only change was these floating native
  # wheels re-resolving to newer builds (transformers 5.14.1, sentencepiece
  # 0.2.2, torchcodec 0.15.0). Cap below the crashing builds until the upstream
  # teardown regression is isolated + fixed. macOS/arm64 can't mirror the linux
  # crash (it segfaults even on the green stack), so CI is the only valid oracle.
  "transformers>=5.0.0,<5.14",
  "sentencepiece>=0.2.1,<0.2.2",
  # sentence-transformers 5.6.0 fixes silent causal-LM reranker scoring bug + restores
  # TSDAE on transformers v5. Dual v4/v5 compat since ST 5.2.1 — safe floor for #382.
  "sentence-transformers>=5.6.0,<6.0.0",
  # NOTE: lancedb is intentionally NOT here — it is search-only (used solely under
  # ``search/`` + the index-rebuild route + upgrade migration, never by transcription /
  # summarization / GI). It lives in [search]; an ML pipeline that also indexes composes
  # ``.[ml,search]``. The torch/transformers/sentence-transformers/sentencepiece/protobuf
  # overlap with [search] above is genuine — both ML (whisper, summarization, GI grounding +
  # NLI) and search (embeddings) import them — not a subset relationship (#1019).
  "accelerate>=1.13.0,<2.0.0",
  "protobuf>=3.20.0,!=6.33.4,<8.0.0",  # Exclude 6.33.4 due to CVE-2026-0994 (fixed in 6.33.5+)
  # GGUF in-process inference for hybrid_ml Tier 2 REDUCE (RFC-042, Issue #352); Ollama needs no extra
  "llama-cpp-python>=0.3.26,<0.4.0",
  # Neural speaker diarization (RFC-058): pyannote after local Whisper transcribe.
  # 4.x switched audio I/O to torchcodec (no torchaudio.AudioMetaData) so it works
  # with torch>=2.11/torchaudio>=2.9; 3.x is import-broken against that torchaudio.
  "pyannote.audio>=4.0.7,<5.0",
  "torchaudio",
  # torchcodec is transitive (torchaudio 4.x / pyannote 4.x). Cap <0.15: 0.15.0
  # is part of the 2026-07-17 native teardown SIGSEGV drift (see transformers
  # note above). Explicit line so the cap survives transitive re-resolution.
  "torchcodec<0.15",
]
# Streamlit run comparison UI (RFC-047, Issue #373)
compare = [
  "streamlit>=1.41.0,<2.0.0",
  "plotly>=6.7.0,<7.0.0",
  "pandas>=2.3.3,<3.0.0",
  "rouge-score>=0.1.2,<1.0.0",
]
# LLM API providers - grouped for CI/test convenience
llm = [
  "openai>=1.0.0,<3.0.0",  # OpenAI API provider (RFC-013, Issue #405)
  # Google Gemini (GeminiProvider); no [gemini] extra — install via .[llm]
  # 1.x: ThinkingConfig.thinking_budget for gemini-2.5-flash (Issue #572); 0.x lacked the field
  "google-genai>=2.16.0,<3.0.0",
  "google-api-core>=2.30.3,<3.0.0",  # Required by google-genai for some operations
  "anthropic>=0.116.0,<1.0.0",  # Anthropic API provider (Issue #106)
  # Mistral API provider (Issue #106). PyPI ``mistralai`` is quarantined (no installable
  # files on https://pypi.org/simple/mistralai/ as of 2026-05). Pin the official v2.4.5
  # tree (commit matches tag v2.4.5) until PyPI restores the project.
  "mistralai @ https://github.com/mistralai/client-python/archive/7ebe84b5a9eee0714f81c687c8cf636eaa12476f.tar.gz",
  "deepgram-sdk>=7.3.1,<8.0.0",
  "httpx>=0.28.1,<1.0.0",  # Ollama provider (Issue #196) - for health checks
]
# Minimum local ML to build the LanceDB vector index when running an otherwise cloud-only
# pipeline (cloud_balanced on the pipeline-llm image). No spaCy / Whisper / Pegasus /
# llama-cpp — those stay in [ml]. Image size delta on top of [llm]: ~700 MB → ~1.5 GB.
search = [
  "torch>=2.11.0,<3.0.0",
  # Kept aligned with [ml] pin (#382): transformers v5 + sentence-transformers 5.6+.
  # Native-wheel drift caps mirror [ml] (2026-07-17 teardown SIGSEGV; see note there).
  "transformers>=5.0.0,<5.14",
  "sentencepiece>=0.2.1,<0.2.2",
  "sentence-transformers>=5.6.0,<6.0.0",
  # RFC-090 two-tier hybrid backend: embedded vector + native FTS (BM25) in one store.
  "lancedb>=0.33.0,<1.0.0",
  "protobuf>=3.20.0,!=6.33.4,<8.0.0",  # Exclude 6.33.4 due to CVE-2026-0994 (fixed in 6.33.5+)
]
# Live pipeline monitor profiling: py-spy + memray (RFC-065; core `--monitor` uses stdlib + psutil + rich)
monitor = [
  "py-spy>=0.4.2",
  "memray>=1.19.3",
]
# Development tools plus GI/KG viewer HTTP stack (FastAPI, RFC-062). The server-runtime pins
# (fastapi/uvicorn/prometheus-fastapi-instrumentator/apscheduler) are mirrored by
# docker/api/Dockerfile, which installs them directly rather than via ``.[dev]`` (to keep
# pytest/black/etc out of the production image). KEEP THE TWO IN SYNC — they drifted once (#1019).
dev = [
  # All optional o11y emit-SDKs — so every dev/CI/`make install` venv exercises them
  # (the tests import the SDKs to mock them) and the core-without-them path is covered
  # via simulated absence. Runtime stays a no-op unless each service's secrets are set.
  "podcast-scraper[sentry]",
  "podcast-scraper[langfuse]",
  # Cap <0.137 lifted (#1029 → Dependabot #1375): prometheus-fastapi-instrumentator 8.0.1
  # added handling for FastAPI's internal ``_IncludedRouter`` route type (the 0.137
  # regression that returned /api/health 500 when metrics are enabled). Instrumentator
  # floor is raised to >=8.0.1 below to guarantee that fix is present. Verified
  # /api/health green on fastapi 0.141 + instrumentator 8.0.2.
  "fastapi>=0.141.1,<0.142.0",
  # Generic MCP server (PRD-034 / RFC-095) — a core dev/server capability, not a separate
  # extra. The official MCP Python SDK bundles FastMCP; the retrieval tools also need
  # [search] (ML deps) at runtime, same as the viewer.
  "mcp>=1.2.0,<2.0.0",
  # Starlette TestClient (tests/conftest.py preload) requires httpx; not always
  # pulled as a transitive install in minimal CI venvs.
  "httpx>=0.28.1,<1.0.0",
  "uvicorn[standard]>=0.49.0,<1.0.0",
  # Prometheus exporter for /metrics (gated on PODCAST_METRICS_ENABLED in app.py).
  "prometheus-fastapi-instrumentator>=8.0.1,<9.0.0",
  # In-process feed-sweep scheduler (#708); no-op unless ``scheduled_jobs:`` in viewer_operator.yaml.
  "apscheduler>=3.11.2,<4.0.0",
  "black>=26.5.1,<27.0.0",
  "isort>=8.0.1,<9.0.0",
  "flake8>=7.3.0,<8.0.0",
  "pytest>=9.0.3,<10.0.0",
  "pytest-cov>=4.1.0,<8.0.0",
  "pytest-xdist>=3.8.0,<4.0.0",  # Parallel test execution (RFC-018)
  "pytest-rerunfailures>=16.4,<17.0",  # Flaky test reruns (RFC-018)
  "pytest-socket>=0.8.0,<1.0.0",  # Network blocking for E2E tests (RFC-019)
  "pytest-json-report>=1.5.0,<2.0.0",  # JSON test reports for metrics (RFC-025)
  "mypy==2.3.0",  # pinned exactly: a floating range let CI drift to 2.2.0 (differing jinja2 inference)
  "types-PyYAML>=6.0.12,<7.0.0",
  "bandit>=1.9.4,<2.0.0",
  "pip-audit>=2.10.1,<3.0.0",
  "build",
  "rouge-score>=0.1.2,<1.0.0",  # Required for ROUGE computation in evaluation loop
  "jiwer>=3.0.0,<5.0.0",  # Required for WER (Word Error Rate) computation
  # Floor >=3.10 (Dependabot #1373, taken): 3.10 fixes 3 download/pathsec CVEs that
  # ``pip_audit`` (make quality) flags against 3.9.4 — CVE-2026-12061 / 12074 / 12075
  # (zip-slip, signature bypass, SSRF in ``nltk.download`` / ``data.load``). 3.10 also
  # ships ``nltk.inisec``, a CWD import-security hook that is incompatible with our
  # in-tree ``.venv`` (it blocks NLTK-initiated dep imports resolving inside the CWD →
  # breaks eval/BLEU collection from repo root). We disable ONLY that hook via
  # ``NLTK_DISABLE_IMPORT_SECURITY=1`` (set in evaluation/scorer.py, the sole nltk import
  # site) — the pathsec CVE fixes are in a different module and stay in effect.
  "nltk>=3.10.0,<4.0.0",  # Required for BLEU computation
  # Code complexity analysis (RFC-031, #424). wily dropped (#1014) — it pinned radon<5.2 and was
  # near-vestigial (CI/dashboard read radon directly; only local make complexity-track used it).
  "radon>=6.0.1,<6.1",
  "vulture>=2.16,<3.0.0",  # Dead code detection (RFC-031)
  "interrogate>=1.7.0,<2.0.0",  # Docstring coverage (RFC-031)
  "codespell>=2.4.2,<3.0.0",  # Spell checking (RFC-031)
  "jsonschema>=4.26.0,<5.0.0",  # Full JSON Schema validation for eval/GI/KG artifacts
  "pydeps>=3.0.6,<4.0.0",  # Module dependency visualization (RFC-038, #170)
  "pyan3==2.4.0",  # Call graph generation (Issue #425); 1.2.0 has CallGraphVisitor root bug
  "code2flow>=2.5.1,<3.0.0",  # Flowcharts from code (Issue #425)
  # Diarization stack pins (RFC-058) — lazy-imported; keeps CI/dev venv aligned with [ml].
  # Must mirror the [ml] pin: pyannote 3.x is import-broken under torchaudio>=2.9
  # (AudioMetaData removed), which is exactly why [ml] moved to 4.x (#901).
  "pyannote.audio>=4.0.7,<5.0",
  "torchaudio",
  # Must ALSO mirror the [ml] torchcodec cap: it is transitive via torchaudio/pyannote, and
  # 0.15.0 is the native-teardown-SIGSEGV build (see [ml] note). CI installs `.[dev]` (not
  # `.[ml]`), so without this line the resolver pulls torchcodec 0.15.0 and a worker segfaults
  # mid-run once 0.15.0 is available on the index (test-unit xdist INTERNALERROR, #1355).
  "torchcodec<0.15",
]
# Documentation - see docs/requirements.txt

[tool.setuptools]
# Package is now in src/ directory (src-layout)
# Use find: to automatically discover all packages (more reliable than explicit listing)
# Exclude non-package directories that might be detected incorrectly
packages = {find = {where = ["src"], exclude = ["*.egg-info", "examples", "docker"]}}
package-dir = {"" = "src"}

# Ship Jinja prompt templates with the wheel — without this setuptools
# only includes ``.py`` files, and Gemini / cloud paths fail at runtime
# with ``PromptNotFoundError`` (``.j2`` files left behind in the source
# tree). Glob covers the prompts/ subtree which is the only non-Python
# resource we ship today.
[tool.setuptools.package-data]
podcast_scraper = [
    "prompts/**/*.j2",
    "prompts/**/*.md",
    "data/pricing_assumptions.yaml",
    "enrichment/_schema/*.json",
]

[tool.black]
line-length = 100
target-version = ["py310"]
include = '\.(py|pyi)$'
exclude = '''
(
  /(
      \.git
    | \.mypy_cache
    | \.pytest_cache
    | \.venv
    | \.venv-dev
    | \.build
    | \.test_outputs
    | build
    | dist
    | __pycache__
  )/
)
'''

[tool.isort]
profile = "black"
line_length = 100
known_first_party = ["podcast_scraper"]
combine_as_imports = true
force_alphabetical_sort_within_sections = true
# Match Black/flake8: secondary venv lives in-repo for CI-parity unit tests (.venv-dev).
skip_glob = [".venv/**", ".venv-dev/**"]

[tool.mypy]
python_version = "3.10"
ignore_missing_imports = true
warn_return_any = true
warn_unused_configs = true
# Ignore attr-defined errors for Episode, RssFeed, TranscriptionJob dataclass attributes
# These are false positives - attributes exist at runtime but mypy doesn't recognize them
# when imported via TYPE_CHECKING pattern
# annotation-unchecked: bare ``dict`` / untyped bodies in tests; CI treats notes as errors.
disable_error_code = ["attr-defined", "annotation-unchecked"]
exclude = [
    "^scripts/.*",  # Scripts use sys.path manipulation, exclude from mypy
    "^tools/.*",  # Standalone Streamlit tools (not part of src package)
    "^\\.venv/.*",  # Local venv (make type uses ``mypy .``)
    "^\\.venv-dev/.*",  # CI-parity venv (see Makefile VENVDEV)
    "^autoresearch/.*",  # Standalone autoresearch scripts (not part of src package)
    "^infra/dgx/.*-server/app\\.py$",  # Docker-image app.py files (pyannote-server, whisper-server) — both named "app", not part of src package
]

[[tool.mypy.overrides]]
module = ["requests", "requests.*"]
ignore_missing_imports = true

[tool.coverage.run]
branch = true
# Measure the installed ``podcast_scraper`` package (editable install: ``src/podcast_scraper``).
# Unit, integration, and **pytest E2E** all use this config. E2E measures the full ``podcast_scraper``
# tree (no subtree ``omit``); the E2E job’s percentage is expected in the mid-30s until more E2E
# exercises server, search, GI, and related paths. Threshold: ``COVERAGE_THRESHOLD_E2E`` in the Makefile / CI.
source = ["podcast_scraper"]
# pytest-cov 7+ / coverage.py: measure code run inside ``subprocess.run([python, -m, ...])``.
# Many E2E tests invoke ``python -m podcast_scraper.cli`` in a child process; without this
# patch those runs do not contribute hits while all ``podcast_scraper`` lines still count
# toward the denominator, which drags E2E layer coverage below CI thresholds.
patch = ["subprocess"]
# Suppress "Couldn't parse ... MagicMock/join()/..." warning (fake paths from mocks/pytest-xdist)
disable_warnings = ["couldnt-parse"]

[tool.coverage.report]
show_missing = true
skip_covered = true
precision = 2
# Prevent exit 1 when coverage hits unparseable paths (e.g. MagicMock from pytest-xdist)
ignore_errors = true
# Note: fail_under is NOT set globally because different test types have different coverage:
# - Unit tests: ~70-80% (isolated, high coverage)
# - Integration tests: ~40-60% (component interactions, lower coverage)
# - E2E tests: ~30-50% (happy paths only)
# Per-test-type thresholds are enforced via --cov-fail-under in Makefile and CI:
# - Integration: 42% (COVERAGE_THRESHOLD_INTEGRATION)
# - E2E: 40% (COVERAGE_THRESHOLD_E2E; full package denominator)
# - Combined: 70% (COVERAGE_THRESHOLD_COMBINED) - enforced in make ci coverage-enforce
# See .github/workflows/python-app.yml for CI enforcement.

[tool.pytest.ini_options]
testpaths = ["tests"]
# Default: run only unit tests (fast feedback)
# Note: -m marker filter removed from addopts to avoid conflicts with explicit -m flags
# Use explicit -m flags in Makefile/CI commands instead
# No global -q: local ``make test*/ci`` recipes and ad-hoc pytest show progress (dots/bar).
# Suites that need quiet pass ``-q`` on the command line (see Makefile / pre-commit hooks).
addopts = "-ra"
# Suppress warnings from external libraries that we can't fix
filterwarnings = [
    # spaCy/Click deprecation warning (external library, will be fixed in future Click version)
    "ignore::DeprecationWarning:spacy.cli._util",
    # Whisper FP16 warning (expected on CPU, harmless)
    "ignore::UserWarning:whisper.transcribe",
    # thinc/spaCy FutureWarning about torch.cuda.amp.autocast — cosmetic,
    # not actionable by us. Suppressed at CLI startup (src/podcast_scraper/cli.py)
    # for prod; this line is the pytest-side equivalent so acceptance tests
    # don't clutter logs with the same message. Track: #436 (upstream fix
    # will let us drop both suppressions). See docs/guides/DEPENDENCIES_GUIDE.md.
    "ignore:.*torch\\.cuda\\.amp\\.autocast.*deprecated.*:FutureWarning:thinc",
]
# Parallel execution: use -n auto to enable, or -n 0 to disable
# Reruns: use --reruns 2 --reruns-delay 1 for flaky tests
markers = [
    "analysis: diagnostic/analysis tools (not run in regular test suites)",
    "analytical: analytical/diagnostic test tools (separate from regular tests, in tests/analytical/)",
    "unit: marks tests as unit tests (deselect with '-m \"not unit\"')",
    "multi_episode: multi-episode tests with multiple short episodes (10-15 seconds each) for testing multi-episode processing logic",
    "data_quality: data quality validation tests with multiple episodes (nightly only)",
    "nightly: marks tests as nightly-only (comprehensive tests with production models, not run in regular CI)",
    "slow: marks tests as slow (deselect with '-m \"not slow\"')",
    "integration: integration tests (slower, test component interactions)",
    "e2e: end-to-end workflow tests (slowest, test full workflows)",
    "network: hits the network (off by default, use -m network to enable)",
    "ml_models: requires ML dependencies and real model loading (requires openai-whisper, spacy, transformers)",
    "diarization: real pyannote speaker-diarization tests (requires pyannote.audio + a gated HF token; RFC-058)",
    "critical_path: marks tests as part of the critical path (RSS → Parse → Download/Transcribe → NER → Summarization → Metadata → Files). These tests run in fast suite regardless of ml_models marker if models are cached.",
    "app: consumer-app / server-route integration tests (auto-applied to tests/integration/server/test_app_*.py). Run in the fast PR suite alongside critical_path so app-route coverage lands on PRs, not only post-merge on main.",
    "llm: tests that use LLM APIs (may incur costs or rate limits, includes openai and future providers)",
    "openai: tests that use OpenAI API specifically (subset of llm marker)",
    "gemini: tests that use Gemini API specifically (subset of llm marker)",
    "mistral: tests that use Mistral API specifically (subset of llm marker)",
    "deepgram: tests that use the Deepgram API specifically (subset of llm marker)",
    "anthropic: tests that use Anthropic API specifically (subset of llm marker)",
    "deepseek: tests that use DeepSeek API specifically (subset of llm marker)",
    "ollama: tests that use Ollama API specifically (subset of llm marker, local/offline)",
    "grok: tests that use Grok API specifically (subset of llm marker)",
    "golden: golden tests with known stable episodes for regression testing (Issue #379)",
    "chaos: chaos tests for error conditions and edge cases (broken RSS, 404, large files, non-English) (Issue #379)",
    "offline: offline tests with networking disabled (verifies air-gapped operation) (Issue #379)",
    "integration_http: HTTP integration tests (use real HTTP client with test server, not external network)",
    "infrastructure: tests for test infrastructure itself (pytest markers, Makefile, CI config)",
    "serial: marks tests that must run sequentially (not in parallel) due to resource conflicts or race conditions",
    "flaky: marks tests as potentially flaky (may fail intermittently in parallel execution, should use reruns)",
    "module_config: tests for config.py and config_constants.py",
    "module_cli: tests for cli.py",
    "module_service: tests for service.py",
    "module_workflow: tests for workflow/*.py",
    "module_downloader: tests for rss/downloader.py",
    "module_rss_parser: tests for rss/parser.py",
    "module_ml_providers: tests for providers/ml/*.py",
    "module_openai_providers: tests for providers/openai/*.py",
    "module_gemini_providers: tests for providers/gemini/*.py",
    "module_anthropic_providers: tests for providers/anthropic/*.py",
    "module_mistral_providers: tests for providers/mistral/*.py",
    "module_deepseek_providers: tests for providers/deepseek/*.py",
    "module_ollama_providers: tests for providers/ollama/*.py",
    "module_grok_providers: tests for providers/grok/*.py",
    "module_summarization: tests for summarization/*.py",
    "module_transcription: tests for transcription/*.py",
    "module_speaker_detection: tests for speaker_detectors/*.py",
    "module_preprocessing: tests for preprocessing/*.py",
    "module_cache: tests for cache/*.py",
    "module_evaluation: tests for evaluation/*.py",
    "module_utils: tests for utils/*.py",
    "module_models: tests for models.py",
    "module_exceptions: tests for exceptions.py",
    "module_prompts: tests for Jinja prompt templates under prompts/ (offline contract tests)",
]
# Parallel execution configuration (pytest-xdist)
# Use -n auto to automatically detect CPU count, or -n N for specific number
# Use -n 0 to disable parallel execution
# Reruns configuration (pytest-rerunfailures)
# Use --reruns N to retry failed tests N times
# Use --reruns-delay N to wait N seconds between retries
# Network blocking configuration (pytest-socket) for E2E tests
# Block all external network calls except localhost/127.0.0.1
# This is configured per-test-directory in conftest.py files

[tool.interrogate]
ignore-init-module = true
ignore-init-method = true
ignore-magic = true
ignore-semiprivate = true
ignore-private = true
ignore-property-decorators = true
ignore-module = true
ignore-nested-functions = true
fail-under = 100
exclude = ["tests", "scripts"]
verbose = 1