pyproject.toml

[build-system]
requires = ["setuptools>=77"]
build-backend = "setuptools.build_meta"

[project]
name = "claude-ctx"
version = "1.0.8"
description = "Skill and agent recommendation system for Claude Code — knowledge graph, wiki, and intake quality gates"
authors = [{ name = "Steve Solun" }]
license = "MIT"
readme = "README.md"
requires-python = ">=3.11"

dependencies = [
    "markdown>=3.6,<4",
    "networkx>=3,<4",
    "numpy>=1.24,<3",
    "plotly>=5,<7",
    "pymdown-extensions>=10,<11",
    "pyyaml>=6,<7",
]

[project.urls]
Homepage = "https://stevesolun.github.io/ctx/"
Repository = "https://github.com/stevesolun/ctx"
Documentation = "https://stevesolun.github.io/ctx/"
Changelog = "https://github.com/stevesolun/ctx/blob/main/CHANGELOG.md"

[project.scripts]
# Plan 001 phase R5: entrypoints for modules moved in R1-R4 now point at
# their canonical package path. Entrypoints for modules still at the
# flat layout (ctx_init, ctx_monitor, etc.) stay unchanged — they will
# move in a later R-phase. The legacy flat shims continue to exist so
# anyone invoking the CLIs via `python path/to/module.py` keeps working.
ctx-init = "ctx_init:main"
ctx-install-hooks = "ctx.adapters.claude_code.inject_hooks:main"
ctx-monitor = "ctx_monitor:main"
ctx-scan-repo = "scan_repo:main"
ctx-skill-quality = "skill_quality:main"
ctx-skill-health = "ctx.adapters.claude_code.skill_health:main"
ctx-toolbox = "toolbox:main"
ctx-lifecycle = "ctx_lifecycle:main"
ctx-skill-add = "skill_add:main"
ctx-skill-install = "ctx.adapters.claude_code.install.skill_install:main"
ctx-skill-mirror = "skill_mirror:main"
ctx-bundle-suggest = "ctx.adapters.claude_code.hooks.bundle_orchestrator:main"
ctx-recommend = "ctx.cli.recommend:main"
ctx-agent-add = "agent_add:main"
ctx-agent-mirror = "agent_mirror:main"
ctx-agent-install = "ctx.adapters.claude_code.install.agent_install:main"
ctx-harness-add = "harness_add:main"
ctx-harness-install = "harness_install:main"
ctx-mcp-add = "mcp_add:main"
ctx-mcp-fetch = "mcp_fetch:main"
ctx-mcp-ingest = "mcp_ingest:main"
ctx-mcp-enrich = "mcp_enrich:main"
ctx-mcp-install = "ctx.adapters.claude_code.install.mcp_install:install_main"
ctx-mcp-uninstall = "ctx.adapters.claude_code.install.mcp_install:uninstall_main"
ctx-mcp-quality = "mcp_quality:main"
ctx-mcp-rebuild-index = "mcp_rebuild_index:main"
ctx-wiki-graphify = "ctx.core.wiki.wiki_graphify:main"
ctx-wiki-worker = "ctx.core.wiki.wiki_queue_worker:main"
ctx-incremental-attach = "ctx.core.graph.incremental_attach:main"
ctx-incremental-shadow = "ctx.core.graph.incremental_shadow:main"
ctx-source-registry = "ctx.core.source_registry:main"
# Pre-ship dedup gate: flag entity pairs at or above the given cosine
# similarity threshold (default 0.85). Cross-type by design — finds
# duplicates across skills, agents, and MCPs in any combination.
ctx-dedup-check = "ctx.core.quality.dedup_check:main"
# Pre-ship tag-completeness gate: scan skills/agents for empty
# `tags:` frontmatter and propose backfills from slug tokens + body
# keywords + the existing tag vocabulary. Report-only by default;
# `--apply` to write.
ctx-tag-backfill = "ctx.core.quality.tag_backfill:main"
# Plan 001 phase H7: the generic harness CLI — `ctx run`,
# `ctx resume`, `ctx sessions`. Ships v1 of the model-agnostic
# harness; requires the [harness] optional dep for LiteLLM.
ctx = "ctx.cli.run:main"
# Plan 001 phase H8: standalone MCP server exposing ctx-core.
# Any MCP-aware host (Cline, Goose, Claude Code itself, Claude
# Agent SDK, custom harness) can attach this server and gain
# access to the skill/agent/MCP recommendation surface WITHOUT
# using our ``ctx run`` harness loop.
ctx-mcp-server = "ctx.mcp_server.server:main"

[project.optional-dependencies]
dev = [
    "pytest>=8",
    "pytest-cov>=5",
    "hypothesis>=6",
    "mypy>=1.8",
    "ruff>=0.4",
]
browser = [
    "playwright>=1.52",
]
embeddings = [
    "sentence-transformers>=2,<4",
    "torch>=2,<3",
]
ann = [
    "hnswlib>=0.8,<0.9; platform_python_implementation == 'CPython'",
]
harness = [
    # LiteLLM is the substrate for the model-agnostic harness
    # (ctx.adapters.generic.providers). Every remote provider —
    # OpenRouter (the tier-1 aggregator), Anthropic, OpenAI, Gemini,
    # Mistral, ... — and local inference servers (Ollama, vLLM) all
    # flow through a single litellm.completion() call. Keeping the
    # dep optional means users who only want the Claude Code
    # integration (ctx-skill-install, the hooks pipeline) don't pay
    # the LiteLLM dependency-tree cost.
    # LiteLLM 1.83.x pins click==8.1.8. Typer 0.17+ requires newer
    # click, so cap Typer for users who install [embeddings,harness]
    # together through sentence-transformers/transformers.
    "click==8.1.8",
    "typer>=0.16,<0.17",
    "litellm>=1.40,<2",
]

[tool.setuptools]
# Incremental package-layout migration.
# During migration phases the codebase lives in TWO shapes simultaneously:
#   - The legacy flat modules listed in ``py-modules`` below.
#   - The target ``ctx`` package tree (ctx.core, ctx.adapters, ctx.cli,
#     ctx.mcp_server, ctx.utils) declared in ``packages`` below.
# Each R-phase moves a small batch from the flat list into the package
# tree, updates imports, and drops the corresponding py-modules entry.
# Running both layouts at once is intentional — it lets the migration
# ship one commit at a time with the suite green at every step.
package-dir = {"" = "src"}
py-modules = [
    "agent_add",
    "agent_mirror",
    "backup_config",
    "backup_mirror",
    "backup_retention",
    "backup_watchdog",
    "batch_convert",
    "behavior_miner",
    "catalog_builder",
    "change_detector",
    "corpus_cache",
    "cosine_ranker",
    "council_runner",
    "ctx_audit_log",
    "ctx_config",
    "ctx_init",
    "ctx_lifecycle",
    "ctx_monitor",
    "harness_add",
    "harness_install",
    "embedding_backend",
    "flatten_agents",
    "import_strix_skills",
    "intake_gate",
    "intake_pipeline",
    "intent_interview",
    "kpi_dashboard",
    "link_conversions",
    "mcp_add",
    "mcp_canonical_index",
    "mcp_enrich",
    "mcp_entity",
    "mcp_fetch",
    "mcp_ingest",
    "mcp_quality",
    "mcp_quality_signals",
    "mcp_rebuild_index",
    "memory_anchor",
    "scan_repo",
    "skill_add",
    "skill_add_detector",
    "skill_category",
    "skill_mirror",
    "skill_quality",
    "skill_telemetry",
    "toolbox",
    "toolbox_config",
    "toolbox_hooks",
    "toolbox_verdict",
    "update_repo_stats",
    "usage_tracker",
    "versions_catalog",
    "wiki_batch_entities",
    "wiki_orchestrator",
    "wiki_visualize",
]
# Do NOT ship the test suite to PyPI. The previous build inadvertently
# bundled src/tests/ because it was the only directory with __init__.py.
# The ``mcp_sources`` subpackage is listed here because setuptools'
# ``py-modules`` is flat-only; subpackages need an explicit ``packages``
# entry to be installed. Plan 001 (phase R0) adds the ``ctx`` tree
# alongside; both layouts coexist until R6 drops the flat modules.
packages = [
    "mcp_sources",
    "ctx",
    "ctx.core",
    "ctx.core.graph",
    "ctx.core.quality",
    "ctx.core.wiki",
    "ctx.core.resolve",
    "ctx.core.bundle",
    "ctx.adapters",
    "ctx.adapters.claude_code",
    "ctx.adapters.claude_code.hooks",
    "ctx.adapters.claude_code.install",
    "ctx.adapters.generic",
    "ctx.adapters.generic.providers",
    "ctx.adapters.generic.tools",
    "ctx.cli",
    "ctx.mcp_server",
    "ctx.utils",
]

[tool.setuptools.package-data]
ctx = ["config.json", "skill-registry.json"]

[tool.pytest.ini_options]
testpaths = ["src/tests"]
markers = [
    "integration: marks tests that require external models or network (deselect with -m 'not integration')",
    "browser: marks tests that drive a real browser with Playwright",
]

[tool.mypy]
python_version = "3.11"
mypy_path = "src"
namespace_packages = true
explicit_package_bases = true
ignore_missing_imports = true

[tool.ruff]
line-length = 100
target-version = "py311"