Dependency inventory

Status: state after implementation first, then the audit baseline for spec 042. Re-inspected 2026-11-10 at commit 00d1ed4, after the dependency upgrades in #419–#431. The first audit was 2026-09-06 at 3e59e1d. This is a declaration, import and lockfile audit. It is not a measured clean-install footprint. Versions come from uv.lock; upstream "latest" claims are dated and sourced in the selection evidence.

State after implementation (2026-11-10)

Written from pyproject.toml and uv.lock at commit 536cad9, the tip of the 042 stack after T23. The sections below this one record the audit that led to the work and are not rewritten. Where they disagree with this section, this section is current.

Result: Semantic Kernel, azure-ai-evaluation and evaluate are in no install profile. The lock has 329 [[package]] entries (376 before the work, 347 after the Semantic Kernel removal). The base install resolves 143 packages (264 before). Measured sizes and start-up times are in the baseline measurements. requires-python is >=3.11.

Base install ([project].dependencies, 35 declarations)

Dependency	Purpose
pydantic, pyyaml, jinja2, python-dotenv, click, inquirerpy	Agent models and validation, YAML, templates, environment files, the CLI and its prompts
jsonschema[format-nongpl], referencing, lark, bkflow-feel, python-ulid	Schema gates, FEEL decision tables, identifiers
claude-agent-sdk (pinned)	Claude backend (Anthropic and Ollama providers)
litellm	Embeddings and contextual-retrieval chat
mcp (>=1.30.0,<3)	MCP servers and bridges
fastapi, starlette, uvicorn[standard], ag-ui-protocol, werkzeug	holodeck serve and the AG-UI protocol
opentelemetry-sdk, opentelemetry-exporter-otlp-proto-grpc, opentelemetry-exporter-otlp-proto-http	Telemetry and OTLP export
markitdown[pdf,docx,pptx,xlsx], openpyxl, pandas, pdfminer.six, python-pptx, pypdf	File parsing for retrieval and multimodal inputs
tiktoken	Token counts for the plain text splitter
rank-bm25	The in-process BM25 index used for lexical and hybrid search
httpx, aiohttp, urllib3, authlib, azure-core, cryptography	HTTP clients, authentication and transport (several carry security lower bounds)

The base install carries no vector store client, no evaluation library, no optimizer and no docs or release tooling.

Extras ([project.optional-dependencies])

Extra	Installs	Feature
eval	deepeval, anthropic, ollama	DeepEval metrics and judge models
nlp	nltk, rouge-score, sacrebleu	BLEU, ROUGE and METEOR
optimizer	optuna	holodeck test optimize
opensearch	opensearch-py	OpenSearch keyword index
qdrant, chromadb, postgres, pinecone, weaviate, azure-ai-search, azure-cosmos	One native client each (qdrant-client, chromadb, psycopg[binary,pool] and pgvector, pinecone 10, weaviate-client, azure-search-documents and azure-identity, azure-cosmos)	Vector store providers
vectorstores	All seven store extras	All vector stores
openai-agents	openai-agents, openai 3	OpenAI and Azure OpenAI backend
dashboard	dash, plotly, pandas	holodeck test view
temporal	temporalio	Temporal worker and activities
deploy, deploy-aws, deploy-gcp, deploy-azure, deploy-all	docker, boto3, google-cloud-run, azure-mgmt-appcontainers and azure-identity	holodeck deploy and its targets
s3, azure-blob, all-sources	boto3, azure-storage-blob	Remote file sources
otel-prometheus, otel-azmon, otel-all, claude-otel	Exporters and the Claude Agent SDK instrumentation	Extra telemetry
all	Every runtime extra above	Everything except dev
dev	Test, lint and type tools, mkdocs-material, mkdocstrings, twine, audit tools	Development only

A command that needs a missing extra stops early and prints the install command (MissingExtraError).

Removed

• Semantic Kernel and what only it needed: azure-ai-agents, azure-ai-projects, aiortc, av, scipy, openapi-core, prance and others. The faiss, sql-server and azure-cosmos-mongo providers went with it.
• Microsoft AI evaluation (azure-ai-evaluation, ruamel-yaml) and Hugging Face evaluate (with datasets, pyarrow, dill, multiprocess, xxhash).
• markitdown[all]: ten packages, among them azure-ai-documentintelligence, azure-ai-contentunderstanding, youtube-transcript-api, speechrecognition and pydub.
• Unused or duplicate direct declarations: requests (HTTPX replaces it), aiofiles, python-dateutil, absl-py, ml-dtypes, the stale sentry-sdk constraint.
• Moved out of the base install: DeepEval, anthropic, ollama, NLTK, rouge-score, sacrebleu, optuna, opensearch-py (now extras); mkdocs-material and twine (now dev only).
• Dependabot ignore for Pinecone majors, and the tool.uv.prerelease setting.

Constraints still in [tool.uv]

• override-dependencies: pytz>=2024. bkflow-feel pins pytz<2024, which would downgrade pytz and optuna; the FEEL conformance tests in tests/unit/workflow/ cover the newer pytz.
• constraint-dependencies (23 entries) are security lower bounds on transitive packages, each with its advisory in a comment in pyproject.toml: filelock, jaraco-context, nltk, pillow, protobuf, python-multipart, markdown, pygments, aiohttp, pyjwt, pip, msgpack, pydantic-settings, joserfc, soupsieve, setuptools, pymdown-extensions, anyio, certifi, fsspec, multidict, oauthlib, opentelemetry-api. Remove one when the packages that need it require a fixed version themselves.
• No cap comes from Semantic Kernel any more. mcp allows 2.x (locked 2.3.0) and pinecone is 10.x. One cap comes from a client: weaviate-client needs grpcio<1.80, so the lock holds grpcio 1.78.2 for it.
• Known issue: the locked pypdfium2 5.15.1 is yanked on PyPI (reason given there: closing of remaining objects on interpreter exit is broken); 5.15.2 exists. See the technical debt tracker.

Audit baseline before the work

The rest of this file is the audit as it stood at 00d1ed4.

What changed since 2026-09-06

• requires-python is now >=3.11 (was >=3.10). CI runs 3.11 and 3.13. FR-006 must be read against 3.11.
• Upgraded: openai-agents 0.17 → 0.24 with openai 3, LiteLLM 1.88 → 1.108, DeepEval 3.8 → 4.3, claude-agent-sdk 0.2.82 → 0.2.179, chromadb → 1.5, temporalio 1.32 → 1.35, opensearch-py 2 → 3, dash 3 → 4, twine 6 → 7, azure-mgmt-appcontainers 4 → 5, pypdf 6 → 7, optuna 4 → 5, MarkItDown 0.1.5b1 → 0.1.9 (stable), Semantic Kernel 1.43 → 1.47.
• SK is now the only package that holds back mcp 2 (it requires mcp<2.0) and newer Pinecone (~=7.0). Both native agent SDKs allow mcp<3.
• The CLI imports each command lazily (#430), so a heavy dependency no longer slows unrelated commands. Profile isolation (FR-005) still needs installed-wheel tests.
• The lock has 370 packages. The base install (core dependencies only) resolves 269; all extras resolve 368.

Measured removal effect on the base install

Computed from uv.lock by removing a root and counting packages that nothing else in the base install requires. These are package counts, not bytes or import time (FR-008 still requires those).

Root removed	Packages that leave the base install
semantic-kernel	26
markitdown[all] → selective extras	up to 26 (depends on the extras kept)
deepeval → evaluation profile	17
mkdocs-material → docs tooling	14
twine → release tooling	14
evaluate	5
azure-ai-evaluation	3
SK + azure-ai-evaluation + evaluate + mkdocs-material + twine together	71 (269 → 198)

Direct runtime declarations

All 50 [project].dependencies declarations are accounted for. "No import" is evidence for review, not proof that a dependency is safe to remove: dynamic loading, upstream requirements and security constraints must be checked. Paths are relative to src/holodeck/ unless stated otherwise.

Dependency constraint	Locked version	Observed role	Proposed disposition
requests>=2.32.5,<3.0.0	2.34.2	lib/file_processor.py, services/mcp_registry.py.	Move both sites to HTTPX and drop the direct declaration. No blocker found; set follow_redirects=True, because HTTPX does not follow redirects by default. Requests stays in the graph through MarkItDown, tiktoken, DeepEval and OpenAI Agents.
pydantic>=2.11.0	2.13.5	46 source files: models, config validation, backend output, workflow and serving.	Retain as HoloDeck schema/contract boundary. SK caps pydantic<2.14; the cap goes away with SK.
pyyaml>=6.0.0	6.0.3	Config loader/manager, templates, workflows, optimizer output.	Retain.
click>=8.3.3,<9.0.0	8.3.3	13 CLI source files. Commands load lazily since #430.	Retain; preserve security lower bound.
jinja2>=3.0.0,<4.0.0	3.1.6	lib/template_engine.py, deploy/dockerfile.py.	Retain.
python-dotenv>=1.0.0	1.2.4	cli/main.py.	Retain.
python-dateutil>=2.8.0	2.9.0.post0	No import in src/ or tests/. Required by bkflow-feel, botocore, opensearch-py and others.	Remove the direct declaration; it stays as a transitive dependency.
urllib3>=2.8.0,<3.0.0	2.8.0	No direct source import; security floor.	Move the floor to resolver constraints and published-wheel policy; do not lose it.
cryptography>=50.0.0,<51.0.0	50.0.2	No direct source import; security floor.	Same as urllib3: keep the floor, review where it is declared.
mkdocs-material>=9.7.7,<10.0.0	9.8.1	Documentation tooling; no runtime import. Brings 14 packages that nothing else needs.	Move out of runtime into docs tooling.
twine>=6.2.0,<8.0.0	7.0.0	Release tooling; no runtime import. Brings 14 packages that nothing else needs. 7.x reads metadata 2.5 wheels (#431).	Move out of runtime into release tooling.
jsonschema[format-nongpl]>=4.18.0,<5.0.0	4.26.0	lib/workflow/edge.py, native backend output validation.	Retain format-nongpl extra and >=4.18 registry semantics.
referencing>=0.36.0,<1.0.0	0.36.2	lib/workflow/edge.py.	Retain explicit direct dependency.
lark>=1.1.0,<2.0.0	1.3.1	lib/workflow/feel.py.	Retain explicit direct dependency.
semantic-kernel>=1.39.4,<2.0.0	1.47.0	Two importing files: lib/vector_store.py (record contracts, 11 dynamically imported connectors) and lib/text_chunker.py. Also telemetry helper names, logging namespaces and 10 test files. Brings 26 packages that nothing else needs (aiortc, av, azure-ai-agents, azure-ai-projects, scipy, openapi-core, prance, pybars4 and others). Its metadata caps mcp<2.0, pydantic<2.14 and pinecone~=7.0.	Remove. Replace with HoloDeck-owned adapters on native store clients and an in-house tiktoken splitter; see the selection evidence.
markitdown[all]>=0.1.4,<0.2.0	0.1.9	lib/file_processor.py. Now a stable release (was 0.1.5b1). [all] brings 26 packages that nothing else needs, including audio/speech, YouTube, Azure Document Intelligence and azure-ai-contentunderstanding>=1.2.0b1, which has only prerelease versions.	Retain MarkItDown; replace [all] with [pdf,docx,pptx,xlsx] (add xls/outlook only on evidence). Base MarkItDown still requires magika, which pulls onnxruntime. Declare directly imported openpyxl, pdfminer.six and python-pptx. Verify with format fixtures (FR-021).
pypdf>=6.21.4	7.0.0	lib/pdf_processor/page_extractor.py. Now on 7.x.	Retain; preserve security floor.
azure-ai-evaluation>=1.13.0,<2.0.0	1.18.11	lib/evaluators/azure_ai.py. Own closure is 61 packages; only 3 are exclusive (msrest, oauthlib, requests-oauthlib).	Remove completely (FR-030–031, AC-13). Unchanged decision.
evaluate>=0.4.6,<0.5.0	0.4.6	lib/evaluators/nlp_metrics.py: only ROUGE and METEOR use evaluate.load, which fetches metric scripts from the Hugging Face Hub at first use. No release since 2025-09. Brings datasets, pyarrow, dill, multiprocess, xxhash.	Remove. Call rouge_score and NLTK METEOR directly; freeze fixtures first to prove equal scores (FR-025). Drops the datasets and pyarrow constraints if no other package needs them.
aiofiles>=25.1.0,<26.0.0	25.1.0	No import in src/ or tests/; only HoloDeck requires it.	Remove.
rouge-score>=0.1.2,<0.2.0	0.1.2	ROUGE backend, reached through evaluate today. Last release 2022, sdist only; requires absl-py, nltk, numpy, six.	Retain and call directly. No maintained drop-in found.
absl-py>=2.3.1,<3.0.0	2.5.1	No direct import; rouge-score already requires it.	Remove the direct declaration.
sacrebleu>=2.5.1,<3.0.0	2.6.0	lib/evaluators/nlp_metrics.py, called directly for BLEU.	Retain.
anthropic>=0.72.0,<2.7.0	2.6.0	No import in src/ or tests/; only HoloDeck requires it. DeepEval does not declare it. The <2.7.0 cap excludes 2.7.0, which moves to httpx2.	Remove after a Claude backend and DeepEval-on-Anthropic test run without it.
ollama>=0.4,<1.0	0.6.3	No import in src/ or tests/; only HoloDeck requires it.	Remove after Ollama runtime and embedding tests run without it.
deepeval>=4.3.3,<5	4.3.3	Eight evaluator modules under lib/evaluators/deepeval/. Now 4.x. Still requires pytest, pytest-asyncio, pytest-xdist, pytest-repeat, pytest-rerunfailures and posthog at runtime; no slim extra exists. No longer declares sentry-sdk.	Retain (FR-025, AC-09). Move to an evaluation profile so the base host does not install pytest and posthog.
python-ulid>=3.1.0,<5.0.0	4.1.0	Claude backend and serve sessions/protocols/models. Now 4.x.	Retain.
ml-dtypes>=0.4.0,<0.7.0	0.6.0	No direct import; only HoloDeck requires it. Cap widened to <0.7 in #431.	Remove the direct declaration unless a reverse dependency appears after SK removal.
mcp>=1.30.0,<2	1.30.0	Consumed by native SDKs and tool bridges. Latest upstream is 2.3.0. SK requires mcp<2.0. claude-agent-sdk 0.2.179 and openai-agents 0.24.0 both allow mcp<3.	Retain. Lift the <2 cap after SK removal, and only after the Claude backend passes on MCP 2: the manifest comment records a past Server.list_tools failure.
inquirerpy>=0.3.4,<0.4.0	0.3.4	cli/utils/wizard.py.	Retain.
fastapi>=0.115.0,<1.0.0	0.162.2	serve/server.py, REST protocol, middleware, tool-init routes.	Retain; consider explicit serving profile.
starlette>=1.3.1	1.10.0	serve/middleware.py; also FastAPI/MCP dependency.	Retain with security floor.
uvicorn[standard]>=0.34.0,<1.0.0	0.54.0	cli/commands/serve.py.	Retain.
ag-ui-protocol>=0.1.18,<1.0.0	0.1.22	Chat executor, Claude backend, AG-UI server/protocol. Upstream 1.0 needs a multimodal migration (Dependabot #413).	Retain; 1.0 migration is separate work.
werkzeug>=3.1.9	3.1.9	No direct import; security floor. Required by dash and flask (dashboard) and by openapi-core (SK).	Move the floor to the dashboard profile/constraints after SK removal.
opentelemetry-sdk>=1.20.0,<2.0.0	1.45.2	Observability providers, exporters, serving, optimizer, SDK tracing.	Retain.
opentelemetry-exporter-otlp-proto-grpc>=1.20.0,<2.0.0	1.45.2	lib/observability/exporters/otlp.py.	Retain.
opentelemetry-exporter-otlp-proto-http>=1.20.0,<2.0.0	1.45.2	lib/observability/exporters/otlp.py.	Retain.
rank-bm25>=0.2.2,<0.3.0	0.2.2	lib/keyword_search.py sparse fallback. Last release 2022; numpy only; works on Python 3.14.	Retain. bm25s 0.3.16 is active but has no mode that reproduces BM25Okapi's idf floor, so rankings would change. Fallback if maintenance fails: vendor BM25Okapi (about 60 lines).
opensearch-py>=2.0.0,<4.0.0	3.2.0	lib/keyword_search.py production keyword index. Now 3.x, verified against OpenSearch 2.19 (#431).	Optionalize as a provider capability.
tiktoken>=0.12.0	0.14.0	lib/structured_chunker.py, contextual generation paths.	Retain; also the token counter for the replacement plain splitter.
aiohttp>=3.14.4	3.14.5	No direct import; LiteLLM/DeepEval transport; security floor.	Keep the floor; review direct vs constraint scope.
httpx>=0.27	0.28.1	lib/source_resolver.py.	Retain; becomes the single HTTP client.
authlib>=1.9.1	1.9.5	No direct import; only HoloDeck requires it.	Review: either a security floor for a dependency that has since left the graph, or dead. Remove if no reverse dependency remains.
azure-core>=1.38.0	1.41.0	deploy/deployers/azure_containerapps.py, lib/source_resolver.py.	Move with the Azure deploy/source capabilities.
claude-agent-sdk==0.2.179	0.2.179	Claude backend, hooks, tool/MCP adapters, lib/claude_context_generator.py.	Retain native agent execution and exact pin.
python-frontmatter>=1.1,<2.0	1.3.0	lib/prompt_version.py.	Retain.
optuna>=4.8.0	5.0.0	optimizer/proposers/numeric.py. Now 5.x; brings sqlalchemy 2.1 and alembic.	Optionalize with the optimizer.
litellm>=1.108.3,<1.109	1.108.3	lib/litellm_support.py, lib/llm_context_generator.py, telemetry instrumentation. Pinned to one minor (<1.109).	Retain as the embedding/context inference boundary.
bkflow-feel==1.2.0	1.2.0	lib/workflow/feel.py.	Retain exact pin.

Optional capabilities

Extra	Declared packages (locked)	Evidence and disposition
dev	pytest>=9.0.3 (9.2.0), pytest-cov>=4.1.0 (7.1.0), pytest-asyncio>=0.21.0 (1.4.1.post1), pytest-mock>=3.11.0 (3.17.0), pytest-xdist>=3.8.0 (3.8.0), black>=26.10.1 (26.10.1), ruff>=0.17.4 (0.17.4), mypy>=2.4.1 (2.4.1), pre-commit>=3.3.0 (4.7.0), tox>=4.0.0 (4.67.13), bandit[toml]>=1.8.0 (1.9.4), pip-audit>=2.7.0 (2.10.1), detect-secrets>=1.4.0 (1.5.0), types-PyYAML>=6.0.12 (6.0.12.20261103), mkdocstrings[python]>=0.25.0 (1.0.8), mkdocs-llmstxt>=0.4 (0.5.1)	Tests, linters, type checking, security and docs tooling; keep separate from production profiles.
pinecone	pinecone[asyncio,grpc]~=7.0 (7.0.2.dev1)	Held at 7.x only by SK (7.0.2.dev1, a dev build, is what the ~=7.0 pin resolves to; 7.1+ needs packaging<25). Latest client is 10.3.0. SK's connector fails to import with 8+. Dependabot ignores Pinecone majors until SK is removed.
postgres	psycopg[binary,pool]~=3.2 (3.3.8)	SK connector plus native filtered reads in lib/collection_filter.py; preserve record layout and filters.
qdrant	qdrant-client>=1.18,<2.0 (1.20.0)	Native async client in four files as well as the SK connector; live verify coverage. Preserve full-text hybrid/RRF and payload indexes.
chromadb	chromadb>=0.5,<1.6 (1.5.9)	SK connector plus native client setup. Now 1.5.x with a live verify case (#425). SK caps chromadb<1.6.
vectorstores	pinecone[asyncio,grpc]~=7.0 (7.0.2.dev1), psycopg[binary,pool]~=3.2 (3.3.8), qdrant-client>=1.18,<2.0 (1.20.0), chromadb>=0.5,<1.6 (1.5.9)	Convenience union of four provider extras; eleven connector names exist.
otel-prometheus	opentelemetry-exporter-prometheus>=0.60b0,<1.0.0 (0.66b2), prometheus-client>=0.17.0,<1.0.0 (0.26.0)	Declared/configured; no exporter implementation module found. Must not claim installation yields a working exporter.
otel-azmon	azure-monitor-opentelemetry-exporter>=1.0.0b24,<2.0.0 (1.0.0b59)	Declared/configured; no exporter implementation module found. Resolve supported status before packaging promise.
otel-all	opentelemetry-exporter-prometheus>=0.60b0,<1.0.0 (0.66b2), prometheus-client>=0.17.0,<1.0.0 (0.26.0), azure-monitor-opentelemetry-exporter>=1.0.0b24,<2.0.0 (1.0.0b59)	Convenience group for the two exporters above; same implementation gap.
claude-otel	otel-instrumentation-claude-agent-sdk>=0.0.6,<0.1.0 (0.0.6)	Lazily imported by the Claude backend; retain tracing/redaction behaviour.
openai-agents	openai-agents==0.24.0 (0.24.0), openai>=3,<4 (3.48.0)	Native OpenAI/Azure backend. Now on the openai 3 stack (#424); openai is declared here, which closes the earlier hidden-import gap.
dashboard	dash>=3.0,<5.0 (4.5.0), plotly>=5.20,<8.0 (7.1.0), pandas>=2.0 (2.3.3)	Dash 4 (#431), Plotly 7, pandas. Retain as a separate profile.
deploy	docker>=7.0.0 (7.2.0)	Docker SDK used by deploy/builder.py.
deploy-aws	boto3>=1.42.0 (1.43.126)	AWS deployer is unimplemented; boto3 is actually used by the S3 source resolver.
s3	boto3>=1.42.0 (1.43.126)	boto3 source resolution in lib/source_resolver.py.
azure-blob	azure-storage-blob>=12.19 (12.32.0b1)	Azure Blob source resolution in lib/source_resolver.py. The lock resolves a beta (12.32.0b1) because prereleases are allowed.
all-sources	boto3>=1.42.0 (1.43.126), azure-storage-blob>=12.19 (12.32.0b1)	Convenience union of implemented source resolvers.
deploy-gcp	google-cloud-run>=0.13.0 (0.16.1)	GCP deployer is unimplemented; no google-cloud-run import found.
deploy-azure	azure-mgmt-appcontainers>=5.1.0,<6 (5.1.0), azure-identity>=1.15.0 (1.27.0)	Implemented deploy/deployers/azure_containerapps.py; azure-mgmt-appcontainers moved to 5.x (#426).
deploy-all	docker>=7.0.0 (7.2.0), boto3>=1.42.0 (1.43.126), google-cloud-run>=0.13.0 (0.16.1), azure-mgmt-appcontainers>=5.1.0,<6 (5.1.0), azure-identity>=1.15.0 (1.27.0)	Includes implemented Docker/Azure and unimplemented AWS/GCP declarations; resolve advertised support.
temporal	temporalio==1.35.0 (1.35.0)	Durable activity/worker implementation; exact pin, moved to 1.35 with live tests (#431).

Vector store providers: what works today

Checked in the full development environment (all extras installed) at 00d1ed4 by importing each SK connector module. "Repo use" counts files under sample/, docs/examples, docs/guides and templates that set the provider.

Provider	SK connector imports	HoloDeck extra for the client	Repo use	Live verify
qdrant	Yes	qdrant	10	Yes
chromadb	Yes	chromadb	22	Yes
postgres	Yes	postgres	8	No
pinecone	Yes (7.x only)	pinecone	4	No
in-memory	Yes	none needed	1	No
sql-server	Yes (module only; needs pyodbc, not installed)	none	0	No
azure-ai-search	No: azure.search missing	none	0	No
weaviate	No: weaviate missing	none	0	No
faiss	No: faiss missing	none	0	No
azure-cosmos-nosql	No: azure.cosmos missing	none	0	No
azure-cosmos-mongo	No: azure.cosmos missing	none	0	No

Six of the eleven provider names cannot run from any documented HoloDeck install. A user must install the client by hand. This is evidence for the provider-scope decision in the spec (FR-014); it is not a decision to retire them.

Build and dependency groups

• Build: hatchling, hatch-vcs; retain VCS version/wheel behaviour.
• [dependency-groups].dev: pyasn1>=0.6.5, types-requests>=2.32.4.20250913, virtualenv>=21.14.9; distinct from the published dev extra. Reconcile contributor/CI install commands.
• tool.uv.prerelease = "allow": the manifest comment names SK's azure-ai-agents. MarkItDown [all] also needs it (azure-ai-contentunderstanding has only prereleases). The setting lets other packages resolve prereleases: the lock holds pinecone 7.0.2.dev1, azure-storage-blob 12.32.0b1, lxml 7.0.0b1 and wrapt 1.17.4rc1. Removing SK and MarkItDown [all] should allow prerelease = "if-necessary-or-explicit"; OTel 0.x beta packages are explicit pins.
• tool.uv.override-dependencies: pytz>=2024 overrides bkflow-feel's old cap. Preserve FEEL conformance evidence.
• The published wheel carries no readme, so the PyPI page has no description (twine warns). Fix with the packaging changes.

Hidden direct imports and transitive ownership

• Imported but not declared in core: openpyxl, pdfminer, pptx (all through MarkItDown [all]), grpc (through the OTLP gRPC exporter), pandas (dashboard extra; in the base install it arrives only through azure-ai-evaluation and evaluate), weaviate (no extra). Each needs an owning declaration before its current carrier is removed.
• openai is now declared in the openai-agents extra.
• semantic_kernel appears as imports in two source files, and as connector path strings, telemetry helper names (enable_semantic_kernel_telemetry), logging namespaces and default filter namespaces in six more. Import counting alone understates removal scope.

High-impact locked transitive clusters

Root (locked)	Selected dependencies and implication
semantic-kernel 1.47.0	Own closure 98 packages; 26 exclusive: aiortc, aioice, av, pylibsrtp, azure-ai-agents, azure-ai-projects, scipy, openapi-core, prance, pybars4, cloudevents and others. Caps mcp<2.0, pydantic<2.14.
markitdown 0.1.9 with all	26 exclusive: magika, onnxruntime, pdfminer-six, pdfplumber, python-pptx, openpyxl, mammoth, speechrecognition, pydub, youtube-transcript-api, azure-ai-documentintelligence, azure-ai-contentunderstanding (prerelease).
azure-ai-evaluation 1.18.11	Own closure 61; only msrest, oauthlib, requests-oauthlib are exclusive. It is one of two carriers of pandas and NLTK into the base install.
evaluate 0.4.6	datasets, pyarrow, dill, multiprocess, xxhash exclusive. Downloads metric scripts at runtime.
deepeval 4.3.3	17 exclusive: pytest and four pytest plugins, posthog, questionary, pyfiglet, setuptools, wheel.
litellm 1.108.3	9 exclusive: boto3/botocore/s3transfer, tokenizers, h2 stack. Keep one inference-owner boundary.
optuna 5.0.0	sqlalchemy 2.1.8, alembic, mako, colorlog. sqlalchemy 2.1 conflicts with the LlamaIndex Postgres connector (sqlalchemy<2.1).
claude-agent-sdk 0.2.179 / openai-agents 0.24.0	Both allow mcp<3. Retaining both native runtimes preserves shared packages after SK removal.

Audit limits and acceptance implications

• This audit inspects declarations, AST import sites, dynamic connector paths, lock entries and connector imports in one developer environment. It does not certify that a provider is licensed, vulnerability-free or behaviour-compatible.
• Exclusive-package counts follow required dependencies and requested extras in uv.lock. Platform markers are ignored, so a count can differ by one or two per platform.
• Fresh isolated installs must exercise CLI, provider initialization, tool ingestion/retrieval, serving and evaluation profiles.

Non-Python hosting and build prerequisites

Sources: base image, local-wheel image, image generator, runtime validation, and Claude backend.

Dependency	Inspected behavior	Disposition / owner
Python and Linux base image	Both Dockerfiles now use python:3.13-slim; published base defaults to latest PyPI HoloDeck when no version argument is supplied.	Packaging: record supported Python/OS/architecture and immutable build inputs; test actual images.
Claude executable	Claude Agent SDK bundles a CLI binary; the backend runs it as a subprocess. Node is not required solely because the provider is Anthropic.	Backend: retain SDK executable provisioning, platform wheels, startup and lifecycle checks.
Node.js / NodeSource	Generator conditionally installs Node 22 through NodeSource; validator requires Node >=18 for MCP commands node, npx, yarn, or pnpm. Some generator comments still incorrectly attribute the need to Claude itself.	Backend/packaging: preserve conditional installation, verify supported versions and bootstrap provenance; correct comments during implementation.
MCP launchers and server packages	Selected stdio tools may require Node package managers, uv/uvx, Python or other configured executables and server packages. These are agent-specific dependencies, not all universal requirements.	Tool/deployment maintainers: inventory the selected agent's command/args and package versions, provide missing-launcher errors, do not install arbitrary user tool dependencies globally.
uv	Base images install via https://astral.sh/uv/install.sh; generated extras install uses --prerelease=allow.	Packaging: pin/review bootstrap and resolver inputs; record the actual installed graph.
curl, ca-certificates, bubblewrap, shell utilities	Both base images install the first three. Curl supplies bootstrap and health checks; Dockerfile documents bubblewrap for Claude subprocess environment scrubbing.	Deployment/security: verify runtime need and platform support; preserve TLS, health and credential isolation.
Docker daemon / build engine and registry	Docker SDK drives image builds; generated images default to ghcr.io/justinbarias/holodeck-base:latest.	Deployment: retain supported workflow, record base/agent digest, candidate version, extras and package/system manifest.
External services	Configured model APIs, vector stores, MCP servers, OTel collectors, optional OpenSearch/Temporal and cloud source/deployment services.	Respective adapter owners: compatibility and availability tests; distinguish service dependency from installed library.

The generator currently installs unversioned holodeck-ai[extras] independently of the audited lock. This can change the candidate graph even after a successful local wheel test. FR-029/AC-11 require reproducible base, agent, and worker image identity and resolved dependency evidence. There is no repository-owned Node package.json outside ignored samples in this baseline; sample frontend dependencies are not universal hosting dependencies.

Ownership and disposition gates

Owners apply to the rows in their domain; the spec acceptance IDs are the exit gates. Unresolved dispositions stay open until those gates provide evidence.

Domain	Owner	Gate
Config, CLI, native SDKs, MCP and shared utilities	Backend maintainers	AC-01–03; isolated imports and both native backend contracts
Retrieval, readers, tokenization and provider extras	Retrieval maintainers	AC-04–08; all-provider operations, stored data and format fixtures
HTTP, serving and deployment/source packages	Serving/deployment maintainers	AC-03/06/11; lifecycle, async I/O, actual deployer support
Evaluation and Optuna	Evaluation/optimizer maintainers	AC-09, AC-13; retain DeepEval and deterministic/code graders, remove Microsoft/legacy evaluation, verify migration and optional installs
OTel, exporter extras and telemetry dependencies	Observability maintainers	AC-10; actual exporter implementation, routing and redaction
Security floors, transitive-only declarations, docs/release/build/dev and profile resolution	Packaging/security maintainers	AC-01/02/11/12; reverse graph, clean wheels and reviewed constraints
FEEL/Temporal/schema primitives	Workflow maintainers	AC-09; conformance, replay and sandbox safety

Complete resolver constraints

These are current declarations (28; 20 at the first audit), not newly verified vulnerability claims. uv constraints do not constrain a downstream pip installation of a published wheel.

Constraint	Proposed disposition / gate
sentry-sdk>=2.0.0,<3.0.0	Stale: sentry-sdk is no longer in the lock (DeepEval 4 dropped it). Remove after confirming no profile resolves it.
filelock>=4.2.1	Retain until AC-11 reviews the candidate graph.
jaraco-context>=6.1.0	Retain until AC-11 reviews the candidate graph.
nltk>=3.10.4	Still needed: rouge-score requires NLTK, and METEOR uses it. The scoped security exception stays.
pyopenssl>=26.0.0	Retain until AC-11 reviews the candidate graph.
pillow>=13.0.0	Retain until AC-11 reviews the candidate graph.
protobuf>=5.29.6	Retain until AC-11 reviews the candidate graph.
python-multipart>=0.0.31	Retain until AC-11 reviews the candidate graph.
pyarrow>=23.0.1	Needed only while evaluate is installed; remove with it if no reverse dependency remains.
markdown>=3.10.2	Retain until AC-11 reviews the candidate graph.
pygments>=2.20.0	Retain until AC-11 reviews the candidate graph.
aiohttp>=3.14.4	Retain until AC-11 reviews the candidate graph.
pyjwt>=2.15.0	Retain until AC-11 reviews the candidate graph.
pip>=26.1.2	Retain until AC-11 reviews the candidate graph.
msgpack>=1.2.3	Retain until AC-11 reviews the candidate graph.
pydantic-settings>=2.14.2	Retain until AC-11 reviews the candidate graph.
joserfc>=1.7.6	Retain until AC-11 reviews the candidate graph.
soupsieve>=2.8.4	Retain until AC-11 reviews the candidate graph.
setuptools>=84.1.0	Retain until AC-11 reviews the candidate graph.
pymdown-extensions>=11.0.0	Retain until AC-11 reviews the candidate graph.
anyio>=4.14.2	Retain until AC-11 reviews the candidate graph.
certifi>=2026.10.18	Retain until AC-11 reviews the candidate graph.
dnspython>=2.9.0	Retain until AC-11 reviews the candidate graph.
fsspec>=2026.6.0	Retain until AC-11 reviews the candidate graph.
multidict>=6.9.1	Retain until AC-11 reviews the candidate graph.
oauthlib>=4.0.1	Retain until AC-11 reviews the candidate graph.
opentelemetry-api>=1.45.1	Retain until AC-11 reviews the candidate graph.
datasets>=5.0.1	Needed only while evaluate is installed; remove with it if no reverse dependency remains.

pyproject.toml and uv.lock remain the executable source of truth. Candidate changes must update this inventory with measured evidence.