Goal

• Read one or more Python files passed by the user and generate a markdown file explain_code.<descr>.md that documents their classes, functions, interfaces, and relationships

Workflow

Read the Files

• Read every Python file <FILE> passed by the user or referenced in the request

Generate the Document

• Generate the file in the directory where you are running, named explain_code.<descr>.md
  • <descr> is a short slug for what was read, typically the source file's basename without extension (e.g., batch_call_auction for batch_call_auction.py); pick a combined slug when documenting several files together
• Include the five sections below, in this exact order
• Under each section, print one level-2 header (## ) per file read, using the same path that was used to read the file in Step 1 (this disambiguates files that share a basename in different directories)
• Omit a section, or a file's subsection within it, entirely if there is nothing to report; never print an empty header or a placeholder bullet like -
• Do not print result to screen but only to file

Verify

• Follow the # Verification checklist below before returning the result

File Description

• Print a nested bullet list for each file describing what are the abstractions and functionalities the file contains

• Keep the files ordered in terms of use relationship

  • E.g., if file X is imported by Y, then X should come before Y in the list

• E.g.,

  - `batch_call_auction.py`
    - In-memory order book for a batch call auction
    - Queues buy/sell orders, buckets by capability tier
    - Clears each tier at a uniform price once per round
  - `passthrough_proxy.py`
    - LLM provider gateway
    - Routes a prompt to a registered provider
    - Times the call, estimates cost, and logs every request/response pair for
      later query
  

Class Description

• For each file, print one bullet per Python class with a description under 20 words

• E.g.,

  # Class Description
  
  ## `research/Noesis/batch_call_auction.py`
  
  - `Bid`
    - Buyer order: (tasks, min tier, max latency, min reliability, max price)
  - `Ask`
    - Seller order: (tasks, tier, typical latency/reliability, min price)
  - `Fill`
    - One matched buyer/seller trade at a tier's uniform clearing price
  - `TierClearResult`
    - Outcome of clearing one tier (price, fills, unfilled quantities)
  - `OrderBookStore`
    - Abstract pluggable storage backend for pending bids/asks
  
  ## <file.py>
  ...
  

Class Interface

• For each file, print every class with the class it descends from (or (dataclass) if it has no explicit base) and a nested bullet per method with its full signature and a description under 20 words

• If a class defines no custom methods (e.g., a plain dataclass with only an auto-generated constructor), say so instead of listing a signature

• E.g.,

  # Class Interface
  
  ## `research/Noesis/batch_call_auction.py`
  
  - `Bid(dataclass)`
    - `__init__(self, buyer_id: str, n_tasks: int, c_level_min: str, l_max: float, r_min: float, p_max: float) -> None`
      - Validates and stores one buy order
  
  - `Fill(dataclass)`
    - Plain data holder, no custom methods
  
  - `OrderBookStore(abc.ABC)`
    - `add_bid(self, bid: Bid) -> None`
      - Queue a bid
    - `clear(self) -> None`
      - Abstract, drop every stored bid/ask
  
  ## <file.py>
  ...
  

Function Interface

• For each file, list its module-level functions only, i.e., functions defined outside any class (methods are already covered under Class Interface)

• Include underscore-prefixed module-level functions when they hold core logic; skip trivial one-line helpers

• One line per function: full signature and a description under 20 words

• E.g.,

  # Function Interface
  
  ## `research/Noesis/batch_call_auction.py`
  
  - `_match_orders_in_tier(c_level: str, bids: List[Bid], asks: List[Ask]) -> TierClearResult`
    - Matches one tier's bids/asks and computes uniform clearing price
  
  ## <file.py>
  ...
  

Function Relationship

• For each module-level function from # Function Interface, list who calls it and what it calls, scoped to the files read

• Skip trivial one-line method delegations already implied by # Class Relationships (e.g., a method that only forwards to a stored collaborator)

• E.g.,

  # Function Relationship
  
  ## `research/Noesis/batch_call_auction.py`
  
  - `_match_orders_in_tier()`
    - Called by: `OrderBook.clear_round()`, once per capability tier
    - Calls: no other functions in this file
  
  ## <file.py>
  ...
  

Class Relationships

• Print the relationships among all classes across all files read, grouped under level-2 headers

• Typical groups are ## Inheritance, ## Composition, ## Uses, and ## Mirrors (schema correspondence with no code dependency); only include a group if it applies, and only list classes for which the relationship holds

• E.g.,

  # Class Relationships
  
  ## Inheritance
  
  - `OrderBookStore`
    - `_InMemoryOrderBookStore` from `batch_call_auction.py`
    - `PostgresOrderBookStore` from `postgres_store.py`
  
  ## Composition
  
  - `OrderBook`
    - Holds one `OrderBookStore`
  
  ## Uses
  
  - `OrderBook.clear_round()` calls `_match_orders_in_tier()` per capability
    tier and returns `Dict[str, TierClearResult]`
  - `TierClearResult` aggregates `List[Fill]`
  
  ## Mirrors
  
  - `BidRequest` (API layer) <-> `Bid` (matching engine): same fields, no
    shared code
  

Conventions

• Follow .claude/skills/markdown.rules.md and .claude/skills/text.rules.md for markdown and text formatting
• Keep every bullet description under 20 words
• Enclose class names, function names, and file paths in backticks

Constraints

• Do not modify the Python files being read
• Base every bullet strictly on what the code shows; do not invent classes, functions, or relationships that are not present
• Never print an empty section header or a placeholder bullet (- with no text)

Verification

• Every class and function name mentioned exists in a file that was read
• Every file that was read appears as a level-2 header in each section where it has content, using the same path used to read it
• No section or subsection is present with zero bullets under it
• Every description is under 20 words