pyproject.toml

[build-system]
requires = [
    "hatchling",
    "hatch-vcs",
    "pydantic>=2.11.4,<3.0.0",
    "pint>=0.24.4,<0.25.0",
    "rich>=13.7.0,<14.0.0",
    "click>=8.0.0,<9.0.0",
    "networkx>=3.0,<4.0",
    "PyYAML>=6.0,<7.0",
    "ruamel-yaml>=0.18.17",
    "imas-data-dictionary>=4.1.0",
    "imas-data-dictionaries>=4.1.0",
    "imas-python>=2.0.1",
    "linkml>=1.9.3",
    "linkml-runtime>=1.9.5",
    "neo4j>=5.28.0,<6.0.0",
    "imas-standard-names @ git+https://github.com/Simon-McIntosh/IMAS-Standard-Names.git@v0.7.0rc37",
]
build-backend = "hatchling.build"

[project]
name = "imas-codex"
dynamic = ["version"]
description = "An IMAS Data Dictionary MCP server"
readme = "README.md"
requires-python = ">=3.12,<3.13"
authors = [{ name = "Simon McIntosh", email = "simon.mcintosh@iter.org" }]
license = { text = "CC BY-ND 4.0" }
keywords = [
    "IMAS",
    "MCP",
    "Model Context Protocol",
    "Data Dictionary",
    "Fusion",
    "Plasma Physics",
    "ITER",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Science/Research",
    "Intended Audience :: Developers",
    "License :: Other/Proprietary License",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Physics",
    "Topic :: Software Development :: Libraries :: Python Modules",
    "Topic :: Database :: Database Engines/Servers",
]
dependencies = [
    # --- Core: MCP server, graph queries, search, CLI ---
    "fastmcp>=3.2.0",
    "pydantic>=2.11.4,<3.0.0",
    "pint>=0.24.4,<0.25.0",
    "nest-asyncio>=1.5.0,<2.0.0",
    "click>=8.0.0,<9.0.0",
    "python-dotenv>=1.0.0",
    "PyYAML>=6.0,<7.0",
    "numpy>=2.3.1",
    "cachetools>=5.3.0,<6.0.0",
    "imas-data-dictionaries>=4.1.0",
    "anyio>=4.0.0,<5.0.0",
    "rapidfuzz>=3.0.0",
    "neo4j>=5.28.0,<6.0.0",
    "ruamel-yaml>=0.18.17",
    "rich>=13.7.0,<14.0.0",
    "httpx>=0.27.0",
    "requests>=2.25.0",
    "filelock>=3.0.0",
    # Security: CVE fix requires lxml>=6.1.0 (indirect via python-pptx/python-docx)
    "lxml>=6.1.0",
    "breame>=0.1.2",
]

[project.optional-dependencies]
test = [
    "torch",
    "sentence-transformers>=5.0.0",
    "pytest>=8.3.5,<9.0.0",
    "pytest-cov>=6.1.1,<7.0.0",
    "pytest-asyncio>=0.21.0,<1.0.0",
    "pytest-xdist>=3.0.0,<4.0.0",
    "pytest-benchmark>=4.0.0,<5.0.0",
    "pytest-timeout>=2.1.0,<3.0.0",
    "coverage>=7.0.0",
]
bench = ["asv>=0.6.0,<1.0.0"]
serve = [
    # Embedding server deps (GPU inference)
    "fastapi>=0.115.0",
    "uvicorn>=0.34.0",
    "uvloop>=0.21.0",
]
# PyTorch backend selection - choose one:
#   uv sync --extra cpu  (default, ~2GB, no CUDA)
#   uv sync --extra gpu  (CUDA 12.1, ~6GB, torch 2.5.1 max due to iter driver)
# Both include serve deps (embedding server) for single-extra deploys.
cpu = [
    "torch",
    "sentence-transformers>=5.0.0",
    "imas-codex[serve]",
]
# LLM proxy runs via `uv tool run litellm[proxy]` — no venv deps needed.
gpu = [
    "torch==2.5.1",
    "accelerate>=1.0.0",
    "sentence-transformers>=5.0.0",
    "imas-codex[serve]",
]

[dependency-groups]
dev = [
    # --- Code quality ---
    "ruff>=0",
    "mypy>=1.15.0",
    "pre-commit>=4.2.0",
    "tqdm-stubs>=0.2.1",
    # --- Interactive dev ---
    "ipython>=9.2.0",
    "ipykernel>=6.29.5",
    # --- Standard Names catalog (install manually: uv pip install -e ../imas-standard-names) ---
    # --- LLM & Discovery ---
    "litellm>=1.81.0",
    # --- Graph build & schema ---
    "linkml>=1.9.3",
    "linkml-runtime>=1.9.5",
    "imas-python>=2.0.1",
    "networkx>=3.0,<4.0",
    "scikit-learn>=1.7.2",
    "hdbscan>=0.8.41",
    # --- Wiki & document parsing ---
    "beautifulsoup4>=4.14.3",
    "Pillow>=11.0.0",
    "python-docx>=1.1.2",
    "python-pptx>=1.0.2",
    "openpyxl>=3.1.5",
    "nbformat>=5.10.4",
    "xlrd>=2.0.2",
    # --- Auth & remote ---
    "keyring>=25.7.0",
    "secretstorage>=3.5.0",
    "fabric>=3.0.0",
    "pykakasi>=2.3.0",
    "jellyfish>=1.2.1",
    # huggingface_hub is a transitive dep of sentence-transformers.
    # Only the cpu/gpu extras need it directly (for model downloads).
    # The [hf_xet] extra installs native Rust bindings that crash on some
    # environments (ITER SDCC). HF_HUB_DISABLE_XET=1 is set in __init__.py.
    # --- Tree-sitter ---
    "tree-sitter>=0.25.2",
    "tree-sitter-language-pack>=0.13.0",
    "tree-sitter-gdl>=0.2.0",
    # --- Testing ---
    "pytest>=8.4.2",
    "pytest-asyncio>=0.26.0",
    "pytest-cov>=6.1.1",
    "pytest-timeout>=2.1.0",
    "xlwt>=1.3.0",
    # --- Serve (embedding + LLM proxy) ---
    "fastapi>=0.115.0",
    "uvicorn>=0.31.1",
    # --- Catalog preview (mkdocs-driven via 'standard-names serve') ---
    "mkdocs>=1.6.1,<2.0.0",
    "mkdocs-material>=9.6.5,<10.0.0",
    "imas-standard-names @ git+https://github.com/Simon-McIntosh/IMAS-Standard-Names.git@v0.7.0rc37",
]

[project.urls]
Homepage = "https://github.com/iterorganization/imas-codex"
Repository = "https://github.com/iterorganization/imas-codex"
Documentation = "https://github.com/iterorganization/imas-codex#readme"
"Bug Tracker" = "https://github.com/iterorganization/imas-codex/issues"

[project.scripts]
imas-codex = "imas_codex.cli:main"
dd-version = "scripts.dd_version:main"
build-schemas = "scripts.build_schemas:build_schemas"
map-ids-domains = "scripts.map_ids_domains:map_ids_domains"
build-clusters = "scripts.build_clusters:build_clusters"
build-embeddings = "scripts.build_embeddings:build_embeddings"
build-database = "scripts.build_database:build_database"
build-path-map = "scripts.build_path_map:build_path_map_cli"
build-models = "scripts.build_models:build_models"
gen-physics-domains = "scripts.gen_physics_domains:gen_physics_domains"
ingest-mdsplus = "scripts.ingest_mdsplus:ingest_mdsplus"


[tool.hatch.version]
source = "vcs"

[tool.hatch.build.targets.wheel]
packages = ["imas_codex", "scripts"]
exclude = ["imas_codex/resources/.gitkeep"]
artifacts = [
    "imas_codex/resources/**",
    "imas_codex/core/physics_domain.py",
    "imas_codex/graph/models.py",
    "imas_codex/graph/schema_context_data.py",
]

[tool.hatch.build.hooks.custom]
path = "hatch_build_hooks.py"
verbose = true
ids-filter = ""
imas-dd-version = ""

[tool.imas-codex]

# Facility locations — list position determines port offsets for all services.
# Neo4j: bolt = 7687 + index, http = 7474 + index.
# Embed: server = 18765 + index.
# LLM proxy: port = 18400 + index.
# Tunneled connections add +10000 (e.g. iter tunnel → bolt 17687).
locations = ["iter", "tcv", "jt-60sa", "jet", "west", "mast-u", "asdex-u", "east", "diii-d", "kstar"]

# SSH host aliases for remote services (graph, embedding, etc.).
# When omitted, the location name is used as the SSH alias.
# Only add entries where the SSH alias differs from the location name.
[tool.imas-codex.hosts]
# iter, tcv, jt-60sa use location name as SSH alias (implicit)

[tool.imas-codex.graph]
# Location = where Neo4j runs. Use a facility name ("iter") for the login node,
# or a compute location ("titan") for SLURM-managed services.
# Compute locations are defined in {facility}.yaml under compute_locations.
# Override: IMAS_CODEX_GRAPH_LOCATION=local
location = "titan"
neo4j-version = "2026.01.4-community"  # Docker tag for Apptainer image
username = "neo4j"
password = "imas-codex"

# Explicit profile overrides (convention handles standard cases):
# [tool.imas-codex.graph.profiles.staging]
# location = "staging-server"      # Where Neo4j runs (SSH alias)
# bolt-port = 7700
# http-port = 7701
# data-dir = "/custom/path/neo4j-staging"

[tool.imas-codex.data-dictionary]
version = "4.1.0"
include-ggd = true
include-error-fields = false

[tool.imas-codex.embedding]
model = "Qwen/Qwen3-Embedding-0.6B"
dimension = 256
location = "titan"       # login node T4 (or "titan" for compute P100 via SLURM)

[tool.imas-codex.discovery]
# Minimum combined score for high-value path processing.
# Used by: enrichment auto-threshold, refinement gate, code CLI default.
# Override: IMAS_CODEX_DISCOVERY_THRESHOLD
threshold = 0.90
# Offset subtracted from threshold to get the triage gate.
# triage_threshold = threshold - triage-offset (default: 0.90 - 0.15 = 0.75).
# Override: IMAS_CODEX_TRIAGE_THRESHOLD
triage-offset = 0.15

[tool.imas-codex.logs]
# Where CLI commands run and write logs.
# MCP log tools (get_logs, list_logs, tail_logs) read from this location.
# "local" = read from local filesystem (default)
# "<facility>" = SSH to that host to read logs (e.g. "iter")
# Override: IMAS_CODEX_LOG_LOCATION
location = "iter"

[tool.imas-codex.language]
# openrouter/ prefix required so direct-path bypass preserves cache_control
# and response_cost. Without it, calls silently route through the proxy,
# stripping cache discounts (often 80%+) and zeroing cost telemetry.
# Powers generate_docs (and any other bulk-content non-refine fanouts).
# refine_name and refine_docs were peeled off into [tool.imas-codex.refine]
# on 2026-05-03 after E3 telemetry showed flash-lite refines accept at
# ~5% vs ~42% for Sonnet-4.6 compose.
# 2026-05-03 (post-E3 audit): per user mandate, NO flash-lite in the SN
# pipeline. generate_docs is now on Sonnet 4.6 — bulk doc gen quality
# matters because it drives the docs-axis review and downstream YAML.
model = "openrouter/anthropic/claude-sonnet-4.6"
batch-size = 50

[tool.imas-codex.refine]
# openrouter/ prefix — see [language] note above.
# Powers process_refine_name_batch and process_refine_docs_batch.
# Routed onto Sonnet 4.6 (matches [sn-run] compose tier) on 2026-05-03
# because flash-lite refines accepted at ~5% vs ~42% for Sonnet compose
# in the E3 acceptance audit. Kept as a separate section so the refine
# tier can diverge from compose later (e.g. always-on Opus) without
# affecting the compose pass.
model = "openrouter/anthropic/claude-sonnet-4.6"

[tool.imas-codex.dd-enrichment]
# openrouter/ prefix required so direct-path bypass preserves cache_control
# and response_cost. Without it, calls silently route through the proxy,
# stripping cache discounts (often 80%+) and zeroing cost telemetry.
model = "openrouter/anthropic/claude-sonnet-4.6"

[tool.imas-codex.vision]
# openrouter/ prefix — see [language] note above.
# flash-lite is INTENTIONAL here: this section drives image OCR only
# (wiki/document page image captioning via discovery/wiki and related
# doc-ingestion paths). It is NOT used in the SN compose/review/refine
# pipeline. Flash-lite is adequate for OCR-quality transcription and
# keeps image-processing costs minimal. Any SN-pipeline model must go
# through [sn-run], [sn-enrich], [refine], or [sn.review.*] — never vision.
model = "openrouter/google/gemini-3.1-flash-lite-preview"

[tool.imas-codex.reasoning]
# openrouter/ prefix — see [dd-enrichment] note above.
model = "openrouter/anthropic/claude-sonnet-4.6"

[tool.imas-codex.sn-compose]
# Maximum concurrent LLM requests issued by the compose worker.
# Derived from the 2026-04-24 extended OpenRouter rate-limit probe (Wave 3):
#   All four working models (Haiku 4.5, Sonnet 4.6, Opus 4.6, GPT-5.4)
#   cleared N=128 with zero 429s.  True ceiling is ≥128 per account tier.
#   75% of lowest measured clean level (128): floor(0.75 * 128) = 96.
#   Case A applied: single global semaphore sufficient (variance <2× across models).
# See docs/ops/openrouter-rate-ceilings.md for probe details.
max-concurrency = 96

[tool.imas-codex.sn-run]
# openrouter/ prefix — see [dd-enrichment] note above.
# Compose pool was billing 358 calls on unprefixed name (cache_control
# stripped via proxy). Re-prefixed 2026-05-02 to restore cache discounts.
model = "openrouter/anthropic/claude-sonnet-4.6"
batch-size = 25
# Batch size used when ``sn generate --name-only`` groups paths by
# ``(physics_domain × unit)`` rather than ``(cluster × unit)``.  The
# name-only mode trades per-item cluster context for breadth: wider
# batches (default 50) cut LLM-call count ~70% during bootstrap at
# the cost of richer per-candidate enrichment, which the subsequent
# review / enrichment passes restore.
name-only-batch-size = 50
# Batch size used when ``sn generate --target docs`` routes through the
# five-phase enrich pipeline. Enrichment packs ~800–1200 tokens per name
# (docs, exemplars, DD context) so batches should be smaller than the
# default compose batches. Mirrors ``[tool.imas-codex.sn-enrich].batch-size``
# but allows tuning the docs-generation pass independently when invoked
# through ``sn generate``.
docs-batch-size = 12
# Pre-flight token budget guard.  Batches whose estimated token count
# exceeds this value are binary-split before dispatch.  Set below the
# model's context window (200 k) to leave headroom for the response.
max-tokens = 150000

[tool.imas-codex.sn-enrich]
model = "openrouter/anthropic/claude-opus-4.6"
# Enrich packs ~800–1200 tokens per name (docs, exemplars, DD context)
# so batches should be smaller than generate batches.
batch-size = 12
# Pre-flight token budget guard (same semantics as sn-run).
max-tokens = 150000

[tool.imas-codex.sn]
# Scored-example injection targets: score thresholds for selecting
# exemplar StandardName nodes into compose/review prompts.
example-target-scores = [1.0, 0.8, 0.65, 0.4]
# Tolerance band around each target when selecting the closest example.
example-tolerance = 0.05
# Maximum examples per score bucket.
example-per-bucket = 1
# Compose retry: max retry attempts on grammar/validation failure.
retry-attempts = 1
# Hybrid-search k expansion factor used on retry.
retry-k-expansion = 12

# Default staging directory for export/preview/publish (ephemeral).
staging-dir = "~/.cache/imas-codex/staging"
# Path to ISNC git checkout. Empty = auto-discover from sibling dirs.
isnc-dir = ""

[tool.imas-codex.sn.review]
# Shared review settings (apply to both axes).
# Disagreement detection: true iff N>=2 AND max(score) - min(score) >= threshold.
disagreement-threshold = 0.20
# Maximum RD-quorum cycles (1 = primary only, 2 = blind pair, 3 = full quorum).
max-cycles = 3

[tool.imas-codex.sn.review.names]
# Reviewer model chain for the name-quality axis (grammar/semantic/convention/completeness).
# Length semantics:
#   1 model  → quorum disabled (single reviewer, mirrors legacy behaviour)
#   2 models → blind primary + blind secondary, no escalator
#   3 models → full RD-quorum: models[0] primary (blind), models[1] secondary (blind),
#              models[2] escalator (sees both, authoritative)
#   4+       → rejected at config load time (ValueError)
#
# 3-model RD-quorum chain (full escalator).
# Sonnet 4.6 primary, GPT-5.4 secondary (cross-vendor independence), Opus 4.6 escalator.
#
# Why Sonnet primary (not Opus): As of 2026-05, OpenRouter's prompt-cache
# discount is broken for `anthropic/claude-opus-4.6` across all upstream
# providers (Anthropic-native, Google Vertex, Amazon Bedrock all return
# cache_creation=0 / cache_read=0 despite cache_control blocks). Sonnet 4.6
# caches correctly (90% discount on warm calls). Probe: research/cache_probe.py.
# Reverting to Opus primary as soon as upstream caching is restored will
# require a single-line edit here.
#
# Gemini 3.1-pro-preview removed: Phase 7 probe showed it returns prose despite
# response_format=json_object — structurally unreliable as a reviewer despite
# _sanitize_content prose-extraction.  GPT-5.4 provides cross-vendor bias signal
# while staying inside a validated structured-output family.
# Temperature: GPT-5.4 rejects temperature=0.0; llm.py auto-clamps to None.
models = [
  "openrouter/anthropic/claude-sonnet-4.6",  # cycle 0 primary (blind, cache-enabled)
  "openrouter/openai/gpt-5.4",               # cycle 1 secondary (blind, cross-vendor)
  "openrouter/anthropic/claude-opus-4.6",    # cycle 2 escalator (disputes only, no cache)
]
disagreement-threshold = 0.20

# ---------------------------------------------------------------------------
# Reviewer profiles — select with IMAS_CODEX_SN_REVIEW_PROFILE env var or
# the --reviewer-profile CLI flag on `sn run` / `sn review`.
# ---------------------------------------------------------------------------

[tool.imas-codex.sn.review.names.profiles.default]
# Mirror of the top-level [sn.review.names] keys; used when no profile is set.
# See top-level [sn.review.names] for rationale on Sonnet-primary ordering.
models = [
  "openrouter/anthropic/claude-sonnet-4.6",  # cycle 0 primary (blind, cache-enabled)
  "openrouter/openai/gpt-5.4",               # cycle 1 secondary (blind, cross-vendor)
  "openrouter/anthropic/claude-opus-4.6",    # cycle 2 escalator (disputes only, no cache)
]
disagreement-threshold = 0.20

[tool.imas-codex.sn.review.names.profiles.quality-cost-balanced]
# Quality-preserving cost reduction: Sonnet 4.6 primary + GPT-5.4 secondary
# (cross-vendor) + Opus 4.6 arbiter (disputes only, authoritative).
# Rationale: Sonnet 4.6 benchmarked at 76.5 avg compose quality vs Opus
# equivalent; reviewer benchmark expected similar.  Opus retained as
# escalator so disputed names still get the strongest model's verdict.
# Expected cost: ~70-80% reduction vs default (Opus on every name → Opus
# only on disagreements). NO Haiku — quality floor preserved.
# Validate on ≥30 already-accepted names (per-dim score divergence ≤0.05,
# tier agreement ≥90%, net cost ≤40% of default) before adopting as default.
models = [
  "openrouter/anthropic/claude-sonnet-4.6",  # cycle 0 primary (blind)
  "openrouter/openai/gpt-5.4",               # cycle 1 secondary (blind, cross-vendor)
  "openrouter/anthropic/claude-opus-4.6",    # cycle 2 escalator (disputes only, authoritative)
]
disagreement-threshold = 0.20

[tool.imas-codex.sn.review.names.profiles.opus-only]
# Single-model mode: no quorum, no escalator. Maximum quality, maximum cost.
models = ["openrouter/anthropic/claude-opus-4.6"]
disagreement-threshold = 1.0  # never escalate (single-model)

[tool.imas-codex.sn.review.docs]
# Reviewer model chain for the documentation-quality axis.
# Same length semantics as [sn.review.names]. Sonnet primary for cache discount —
# see [sn.review.names] block for rationale.
models = [
  "openrouter/anthropic/claude-sonnet-4.6",
  "openrouter/openai/gpt-5.4",
  "openrouter/anthropic/claude-opus-4.6",
]
disagreement-threshold = 0.20

[tool.imas-codex.sn.benchmark]
# Models to evaluate for standard name composition.
# Each is benchmarked independently; results are compared side-by-side.
# Format: OpenRouter model IDs (provider/model-slug).
compose-models = [
    "anthropic/claude-sonnet-4.6",
    "anthropic/claude-haiku-4.5",
    "openai/gpt-5.4",
    "openai/gpt-5.4-mini",
    "google/gemini-3.1-pro-preview",
    "google/gemini-3-flash-preview",
    "google/gemini-3.1-flash-lite-preview",
]
# Frontier model used as independent judge for quality scoring.
# Must be a strong reasoning model — evaluates 5 dimensions (0-100).
reviewer-model = "anthropic/claude-opus-4.6"

[tool.imas-codex.sn.fanout]
# Plan 39 — Structured fan-out for the SN compose pipeline.
# Master switch.  Phase 1B shipped at 6f0bb52b; enabled for smoke + rotation.
enabled = true
# Hard cap on Stage A query count (parse-time bound).
max-fan-degree = 3
# Per-runner and total Stage B timeouts.
function-timeout-s = 5.0
total-timeout-s = 12.0
# Per-result hit cap (renderer truncates each FanoutResult.hits).
result-hit-cap = 8
# Total evidence-block token caps (renderer truncates total evidence).
evidence-token-cap-baseline = 2000
evidence-token-cap-escalation = 800
# Stage A LLM model + sampling.
# NOTE: Anthropic's structured-output API rejects schemas with `oneOf`
# (FanoutPlan uses a discriminated union over fn_id). Gemini and OpenAI
# accept oneOf cleanly. Per 2026-05-03 user mandate (NO flash-lite in
# pipeline), proposer is now on Gemini 3.1 Pro Preview which supports
# oneOf and provides production-tier quality.
proposer-model = "openrouter/google/gemini-3.1-pro-preview"
proposer-temperature = 0.1
# Cost padding added to the parent lease at refine-cycle start (tiered, plan §7.3).
fanout-cost-estimate-baseline = 0.005
fanout-cost-estimate-escalation = 0.05
# Per-cycle hard cap on cumulative fan-out sub-event spend (tiered, plan §7.3).
fanout-max-charge-per-cycle-baseline = 0.02
fanout-max-charge-per-cycle-escalation = 0.10
# Refine trigger predicate.
# Keywords match the language reviewers actually use ("decomposition", "absorbed",
# "compound", "awkward") plus disambiguation cues. Case-insensitive substring match.
refine-trigger-keywords = ["unclear", "ambiguous", "duplicate", "consider", "compare", "decomposition", "absorbed", "compound", "awkward"]
# Reviewer-comment dim allow-list for the trigger predicate (plan §5.1 I3).
# MUST match the actual review schema: ('grammar', 'semantic', 'convention', 'completeness').
# We scan grammar + semantic because that is where decomposition/disambiguation
# defects surface (a previous default of ('clarity', 'disambiguation') silently
# disabled the trigger because those dims do not exist in the rubric).
refine-trigger-comment-dims = ["grammar", "semantic"]
# Comment excerpt total length cap (plan §5.1 S3).
refine-trigger-comment-chars = 800

[tool.imas-codex.sn.fanout.sites]
# Per-site enable flags.  Phase 1 flips refine_name to true after 1B
# wires the worker; the master `enabled` switch above gates everything.
refine_name = true

[tool.imas-codex.llm]
# LiteLLM proxy settings. Port base = 18400, offset by location index.
# Runs on login node (needs outbound HTTPS for OpenRouter/Anthropic).
location = "iter"       # facility hosting LLM proxy (or "local" for in-process)

[tool.hatch.envs.test]
dependencies = [
    "pytest>=8.3.5,<9.0.0",
    "pytest-cov>=6.1.1,<7.0.0",
    "coverage>=7.0.0",
]

[tool.hatch.envs.dev]
dependencies = [
    "black>=23.0.0",
    "ruff>=0",
    "ipython>=9.2.0",
    "ipykernel>=6.29.5",
    "mypy>=1.15.0",
    "pre-commit>=4.2.0",
]

[tool.mypy]
python_version = "3.12"
warn_return_any = true
warn_unused_configs = true
disallow_untyped_defs = true
disallow_incomplete_defs = true

[[tool.mypy.overrides]]
module = ["imas.*", "pint.*"]
ignore_missing_imports = true
warn_return_any = true
warn_unused_configs = true
disallow_untyped_defs = true
disallow_incomplete_defs = true

# Optional dependency overrides (only when installed)
[[tool.mypy.overrides]]
module = ["packaging.*"]
ignore_missing_imports = true

[tool.hatch.metadata]
allow-direct-references = true

[tool.pytest.ini_options]
testpaths = ["tests"]
timeout = 30
faulthandler_timeout = 60
addopts = [
    "--strict-markers",
    "--strict-config",
    "--tb=short",
    "--no-header",
    "-q",
    "--durations=10",
    "-m", "not slow and not sn_health and not graph",
]
asyncio_mode = "auto"
asyncio_default_fixture_loop_scope = "session"
markers = [
    "asyncio: marks tests as async",
    "slow: marks tests as slow (deselect with '-m \"not slow\"')",
    "integration: marks tests as integration tests",
    "unit: marks tests as unit tests",
    "fast: marks tests as fast-running unit tests",
    "performance: marks tests that focus on performance",
    "timeout: marks tests with timeout constraints (requires pytest-timeout)",
    "graph: marks tests requiring a live Neo4j connection (auto-skipped when unavailable)",
    "graph_mcp: marks graph-native MCP tests requiring Neo4j with fixture data",
    "fixture_only: marks tests that require fixture data and should be skipped on production graphs",
    "sn_health: marks standard name health gate tests requiring a populated Neo4j graph (run with -m sn_health)",
    "requires_graph: marks tests requiring a live Neo4j connection with DD content",
]
filterwarnings = [
    "ignore::DeprecationWarning",
    "ignore::PendingDeprecationWarning",
    "ignore::pytest.PytestUnraisableExceptionWarning",
]
# Performance optimizations
cache_dir = ".pytest_cache"
python_files = ["test_*.py", "*_test.py"]
python_classes = ["Test*"]
python_functions = ["test_*"]

[tool.coverage.run]
source = ["imas_codex"]
omit = ["tests/*", "*/tests/*", "**/test_*.py", "**/conftest.py"]

[tool.coverage.report]
exclude_lines = [
    "pragma: no cover",
    "def __repr__",
    "if self.debug:",
    "if settings.DEBUG",
    "raise AssertionError",
    "raise NotImplementedError",
    "if 0:",
    "if __name__ == .__main__.:",
]
show_missing = true
precision = 2

[tool.coverage.html]
directory = "htmlcov"
precision = 2

# Ruff configuration
[tool.ruff]
target-version = "py312"
line-length = 88
indent-width = 4

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
skip-magic-trailing-comma = false
line-ending = "auto"

[tool.ruff.lint]
select = [
    "E",   # pycodestyle errors
    "W",   # pycodestyle warnings
    "F",   # pyflakes
    "I",   # isort
    "B",   # flake8-bugbear
    "C4",  # flake8-comprehensions
    "UP",  # pyupgrade
]
ignore = [
    "C901",  # too complex
    "E501",  # long lines 
]

[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["F401"]  # Allow unused imports in __init__.py
"tests/**/*" = ["B011", "F401", "F811"]  # Allow assert False and unused imports in tests
# Remote scripts run on Python 3.6+ systems - disable type hint and subprocess modernization
"imas_codex/remote/scripts/*" = ["UP006", "UP007", "UP021", "UP022", "UP031", "UP035", "UP038", "UP045"]

[tool.ruff.lint.isort]
known-first-party = ["imas_codex"]
force-single-line = false
combine-as-imports = true

# UV configuration: PyTorch index selection by extra
# Default sync uses cpu extra (see default-groups below)
[[tool.uv.index]]
name = "pytorch-cpu"
url = "https://download.pytorch.org/whl/cpu"
explicit = true

[[tool.uv.index]]
name = "pytorch-gpu"
url = "https://download.pytorch.org/whl/cu121"
explicit = true

[tool.uv.sources]
torch = [
  { index = "pytorch-cpu", extra = "cpu" },
  { index = "pytorch-cpu", extra = "test" },
  { index = "pytorch-gpu", extra = "gpu" },
]


[tool.uv]
prerelease = "if-necessary-or-explicit"
# cpu, gpu, and test extras for torch are mutually exclusive with gpu
conflicts = [
  [
    { extra = "cpu" },
    { extra = "gpu" },
  ],
  [
    { extra = "test" },
    { extra = "gpu" },
  ],
]