simonw / llm · Source walkthrough

This asset is a reading order over the analyzer's lessons, not a code listing or tutorial: four reading stops (orientation, execution story, core code, change guide) plus a closing limits section, each built from the analyzer's line-addressable blocks such as llm/__main__.py:1-4. Every line range below is written path:start-end and is exactly the lines value the analyzer recorded for that block in the analyzed revision's files; ranges are never computed or widened. A block is the analyzer's line-addressable unit of code, and its description is the analyzer's inference carrying a confidence score (0.97 for that module block), so treat block text as guidance, not established fact. Where a signature is quoted, it is the one-line signature from the analyzer's symbol records. The analyzer classifies README.md, pyproject.toml and LICENSE as project-level important files and suggests reading them early. There are no install, run or shell steps here, because installation commands are not verified. The repository's platform metadata describes it as a way to access large language models from the command-line; that description is author-claimed.

The analyzer's summary infers a Python project of 116 analyzed files, organized around docs, llm and tests, with llm/__main__.py as the apparent main starting place; the orientation lesson's two file-level module blocks, llm/__main__.py:1-4 and llm/cli.py:1-4407, are therefore the first reading targets - read them for layout before later stops descend into specific ranges. llm/__main__.py is a verified entrypoint whose recorded direct-execution excerpt invokes the cli function, and a second verified entrypoint record points at llm/cli.py, which the role table labels an entry/orchestration candidate with 105 incoming and 245 outgoing static calls.

• llm/__main__.py:1-4 - module block, symbol llm/__main__.py: described as the file-level container for the analyzed code in llm/__main__.py (confidence 0.97).
• llm/cli.py:1-4407 - module block, symbol llm/cli.py: described as the file-level container for the analyzed code in llm/cli.py (confidence 0.76); the role table labels this file an entry/orchestration candidate (105 incoming, 245 outgoing calls).

Read the execution-story lesson next: four blocks tracing how the analyzer reconstructs a run, plus the bounded paths listed in the packet. The analyzer reconstructed 13 bounded execution paths and the packet lists 5 of them: one runs from the entrypoint llm/__main__.py:__main__ to cli in llm/cli.py as a local call and terminates as a leaf; the other four start at llm/cli.py and end at unresolved boundaries - warnings.simplefilter, or load_plugins reaching hasattr, pm.load_setuptools_entrypoints, or LLM_LOAD_PLUGINS.split. These are static reconstructions, not runtime proof: the analyzer does not resolve reflection, dynamic imports or dynamic dispatch.

• llm/__main__.py:3-4 - main_guard, symbol __main__: gates code that should run when the Python file is executed directly rather than merely imported (confidence 0.99).
• llm/cli.py:541-565 - function, cli: encapsulates a named operation with no resolved downstream relation required by this block (confidence 0.97); recorded signature def cli():.
• llm/cli.py:1-4407 - module, llm/cli.py: the file-level container for the analyzed code in llm/cli.py (confidence 0.76).
• llm/plugins.py:23-52 - function, load_plugins: encapsulates an operation connected to 11 downstream call or import relations (confidence 0.76); recorded signature def load_plugins():.

The core-code lesson offers a compact reading set of eight blocks for the main implementation; the entrypoint guard, cli and load_plugins also appear at stop 2, and five blocks are new here. The set reaches beyond the llm/ package into docs/conf.py, where markdown_admonitions connects to 11 downstream relations and setup to 1 - read that file only if its behavior matters to you. The final llm/cli.py blocks cover an exception class, FragmentNotFound, and two helpers the analyzer says each connect to 4 downstream relations.

• llm/__main__.py:3-4 - main_guard, __main__: gates code that should run when the file is executed directly rather than merely imported (confidence 0.99).
• llm/cli.py:541-565 - function, cli: encapsulates a named operation with no resolved downstream relation required by the block (confidence 0.97).
• llm/plugins.py:23-52 - function, load_plugins: encapsulates an operation connected to 11 downstream call or import relations (confidence 0.76).
• docs/conf.py:6-21 - function, markdown_admonitions: encapsulates an operation connected to 11 downstream call or import relations (confidence 0.76).
• docs/conf.py:24-25 - function, setup: encapsulates an operation connected to 1 downstream call or import relation (confidence 0.76).
• llm/cli.py:99-100 - class, FragmentNotFound: groups related behavior under one class boundary (confidence 0.97); recorded signature class FragmentNotFound(Exception):.
• llm/cli.py:103-119 - function, display_stream_events: encapsulates an operation connected to 4 downstream call or import relations (confidence 0.76).
• llm/cli.py:122-133 - function, display_async_stream_events: encapsulates an operation connected to 4 downstream call or import relations (confidence 0.76).

Plan edits around the change-guide lesson's six blocks, which the analyzer presents for prioritization when planning a change; the two with the most recorded downstream relations are prompt (119) and logs_list (116). get_model appears three times in llm/__init__.py - function, straight_line and try blocks - and the role table labels llm/__init__.py a service/core candidate with 754 incoming calls and llm/parts.py one with 470.

• llm/parts.py:313-360 - class, StreamEvent: groups related behavior under one class boundary (confidence 0.97); recorded signature class StreamEvent:.
• llm/cli.py:685-1187 - function, prompt: encapsulates an operation connected to 119 downstream call or import relations (confidence 0.76).
• llm/__init__.py:345-363 - function, get_model: encapsulates an operation connected to 6 downstream call or import relations (confidence 0.97); recorded signature def get_model(name: str | None = None, _skip_async: bool = False) -> Model:.
• llm/cli.py:1878-2379 - function, logs_list: encapsulates an operation connected to 116 downstream call or import relations (confidence 0.76).
• llm/__init__.py:349-363 - try, get_model: groups an operation with explicit exception-handling control flow (confidence 0.97).
• llm/__init__.py:347-348 - straight_line, get_model: groups adjacent sequential statements so they can be explained as one bounded region (confidence 0.96).

The analysis is static only: reflection, runtime dependency injection, dynamic imports, monkey-patching, generated code, framework runtime wiring and dynamic dispatch are not resolved, so runtime behavior behind the listed paths is not established. Install/run/build inference is partial and installation commands are not verified, which is why this walkthrough includes no run steps. Dependency records come only from requirements-style manifests parsed by analyzer v0.10; pyproject.toml dependency tables are not parsed. README claim extraction is line-based and may capture code lines instead of prose claims, so treat README-derived text as author-claimed at best. The analyzer's recorded capabilities describe the analyzer itself, not this repository, and are excluded from repository evidence. Block annotations, finally, remain the analyzer's inferences with confidence scores, not verified descriptions of runtime behavior.