Output Formats

astwire supports three output formats, selected with -f/--format. All three formatters implement the same BaseFormatter.format_partition() interface and receive the same underlying FileRecord list — they differ only in how they render it.

#LLM-Optimized XML (-f llm, default)

astwire src/ -f llm -o prompt.xml
# omitting -f / -o writes context.xml with the same format, since llm is the default

A low-overhead XML schema tuned for prompt injection — no directory tree, no Markdown fencing overhead:

<context>
<file path="config.py" language="python">
# Module contents
</file>
</context>

With -i --graph, a <graph> block of <edge from="..." to="..."/> tags is inserted before <files>:

<context>
<graph>
<edge from="cli.py" to="core/analyzer.py"/>
</graph>
<files>
config.py
</files>
...
</context>

Best for: programmatic prompt construction where every extra token in formatting overhead is a token not spent on actual code content.

#Markdown (-f markdown)

astwire src/ -f markdown -o context.md

Features an ASCII directory tree (when show_tree is enabled, the default), structured file headers, and language-tagged fenced code blocks:

# Context Bundle

## Project Structure
```text
├── cli.py
└── core/
    ├── analyzer.py
    └── ast_crawler.py

#Files

#cli.py

...file content...

With `-i --graph`, a `## Dependency Graph` section is inserted between the tree and `## Files`, one bullet per file listing its outgoing local imports:

```text
## Dependency Graph

- `cli.py` → `core/analyzer.py`, `core/ast_crawler.py`

Without --graph (or without -i), the section is omitted entirely.

Best for: human review, pasting into a chat-style LLM interface, or any workflow where readability matters as much as token efficiency.

#JSON (-f json)

astwire src/ -f json -o context.json

Structured schema containing partition metadata and individual file content records:

{
  "part_index": 1,
  "total_parts": 1,
  "graph": [
    { "from": "cli.py", "to": "config.py" }
  ],
  "files": [
    {
      "path": "config.py",
      "language": "python",
      "content": "...",
      "skipped": false,
      "skip_reason": null,
      "redactions": 0
    }
  ]
}

graph is always present — an empty array [] unless run with -i --graph.

Best for: feeding astwire's output into another tool or pipeline programmatically (e.g. a custom prompt builder, a CI step that inspects redaction counts, or a dashboard).

#Choosing a format

#Partition-aware formatting

All three formatters render one partition at a time — when --max-tokens produces multiple partitions, each formatter is called once per partition, and each gets part_index/total_parts context so it can label itself accordingly (Markdown adds a [Part N of M] header; the LLM formatter adds part="N" total_parts="M" attributes to <context>; JSON includes both as top-level fields).

See ../architecture/diagrams.md#5-token-bounded-partitioning---max-tokens for how partitioning interacts with formatting.