pydantic / pydantic-ai

Platform metadata records Python as the primary language, with 34,234,401 Python bytes alongside much smaller Makefile, PLpgSQL and Shell counts. It is MIT licensed; the latest recorded release is v2.54.0, published 2026-10-03, when the repository had 20385 stargazers. GitHub topics include agent-framework, genai, harness, harness-engineering, llm, pydantic and python.

The repository description — 'How Python does AI. Agents, realtime voice, image generation, embeddings. Every model, every interface, typed end to end.' — and the homepage https://pydantic.dev/pydantic-ai are author-supplied metadata, not analyzer observations.

Structurally, the analyzer counted 4082 analyzed files.

Executable starts appear as Python __main__ guards across the pydantic_ai_slim and clai packages, .github/scripts, examples and tests. Two packages have __main__ guards: pydantic_ai_slim/pydantic_ai/__main__.py calls cli_exit() and clai/clai/__main__.py calls _cli.cli_exit('clai'). Under .github/scripts/ there are many more verified guards, including agent_provider_health.py, agent_spend_report.py and check_api_compatibility.py.

Examples show the usage pattern: pydantic_model.py runs agent.run_sync('The windy city in the US of A.') and prints result.output and result.usage; data_analyst.py calls analyst_agent.run_sync with a prompt and a deps object. chat_app.py starts a uvicorn server for pydantic_ai_examples.chat_app:app, and a module under tests/ calls mcp.run().

The one bounded static flow traced in detail starts at .claude/skills/testing-skill/parse_cassette.py, passes through main() and parse_cassette(), and terminates at unresolved external boundaries such as argparse.ArgumentParser and args.cassette.exists; the analyzer marks them unresolved_boundary, so nothing beyond them is established. The summary's 'detected starting point', examples/pydantic_ai_examples/slack_lead_qualifier/app.py, rests on a filename heuristic and is not established as the primary entry.

The analyzer designates docs, src and tests as top-level subsystems. Dependency records cite pydantic_evals/pyproject.toml and pydantic_graph/pyproject.toml as source paths, alongside the root pyproject.toml.

Static call-degree roles (inferred by the analyzer) sketch the library's shape: messages.py draws 403 incoming calls and sends 70, while exceptions.py draws 421 with none outgoing — both look like widely shared layers. agent/__init__.py shows the opposite profile, 4 incoming and 203 outgoing calls, and is labeled a service/core candidate. models/__init__.py (113 in, 52 out) and models/openai.py (8 in, 151 out) sit between.

In total the analyzer recorded 150000 relations — 124623 calls and 25377 imports — of which 125702 remain external or unresolved and 24298 resolved to local targets.

The root pyproject.toml declares a large dev group: pytest, anyio, trio, coverage, dirty-equals, inline-snapshot, hypothesis, pytest-mock, pytest-xdist, pytest-split, mypy, pyright and ruff. Subpackages carry their own manifests: pydantic_graph lists anyio, logfire-api, pydantic and typing-inspection as runtime dependencies, and pydantic_evals lists logfire as optional. Manifest coverage is partial: the parser reads requirements files and pyproject.toml (PEP 621/735, Poetry tables) but not setup.py, setup.cfg, Pipfile or lock files, and version constraints are recorded as written.

The repository holds 2514 test-like files and 8 build/dependency manifests. tests/conftest.py defines pytest_configure (lines 101-113) and an anyio_backend function (lines 367-370).

Two coverage gaps: bounded analysis leaves files over 400 kB, minified bundles and vendored directories unparsed, and the packet reports five files in that state without establishing what their absence costs this analysis.

• README.md first — the analyzer classifies it as a project-level important file.
• Then pyproject.toml and LICENSE, also project-level important files.
• Then the library core: pydantic_ai_slim/pydantic_ai/agent/__init__.py, where the extracted Agent class spans lines 476-4290; follow with messages.py and exceptions.py (for example SystemPromptPart, ModelRetry).
• Then one runnable example, examples/pydantic_ai_examples/pydantic_model.py, to see run_sync usage and the result object in context.
• The analyzer's suggested path agrees: start from manifests and important symbols, trace the bounded execution path, and inspect unresolved boundaries before changing code.

This overview rests on static analysis only: reflection, runtime dependency injection, dynamic imports, monkey-patching, generated code, framework runtime wiring and dynamic dispatch are not resolved. Nothing here establishes runtime behavior — install/run/build inference is partial and installation commands are not verified. README-derived statements would remain author claims even if quoted, since extraction is line-based and skips headings, fenced code, HTML, badges, tables and shell prompts.

Relation coverage is incomplete: 125702 of the 150000 recorded relations remain external or unresolved, and the recorded total equals the documented 150000-per-repository cap. The packet also reports three entrypoint paths with no statically detected call steps; what those entrypoints do is not established.

The analysis resolves no wiring for how plugins or extensions load; dynamic imports and framework runtime wiring fall under the same static limits.

• models/__init__.py defines known_model_names (lines 120-127), the Model class (lines 454-1149), StreamedResponse (lines 1153-1417) and parse_model_id (lines 1645-1660).
• pydantic_ai_slim/pydantic_ai/exceptions.py defines ModelRetry (59-99), ToolFailed (102-149), CallDeferred (152-167), ApprovalRequired (170-185) and RunCancelled (270-449).
• pydantic_ai_slim/pydantic_ai/messages.py defines SystemPromptPart, FileUrl, VideoUrl, AudioUrl, ImageUrl, DocumentUrl and BinaryContent.