Architecture¶

src/clijson/
├── api.py            parse(), parse_session(), parse_file(): the pipeline
├── commands.py       command grammar, abbreviation-aware matching, pipe handling
├── registry.py       Parser base class, @register, plugin discovery
├── platforms.py      platform catalogue, aliases, prompts, fingerprints, detection
├── result.py         ParseResult (data + provenance)
├── models.py         vendor-neutral schemas ("intents") and helpers
├── textutils.py      building blocks for parser authors
├── engines/
│   ├── generic.py    heuristic engine: tables, key/value, indented sections
│   ├── structured.py Junos | display json / xml, generic XML/JSON
│   ├── config.py     running-config → tree (indent, curly, set)
│   └── external.py   optional ntc-templates / Genie adapters
├── parsers/          dedicated parsers, one package per platform
│   ├── iosxr/  junos/  vrp/
├── cli.py            `clijson` command
├── server.py         HTTP API (stdlib only)
└── live.py           scrapli / netmiko collection

The pipeline (clijson.parse)¶

1. Cleaning. Normalise CRLF, strip ANSI escapes, backspaces and pager prompts (--More--, ---- More ----), convert tabs and dedent pasted text.
2. Prompt and echo extraction. If the first line is a prompt plus a command (RP/0/RP0/CPU0:PE1#show bgp sum, lab@mx1> show route, <NE40E>display version), take the command and hostname from it. Trailing bare prompts and {master} lines are dropped. A bare echoed command (show bgp summary on the first line) also works. IOS XR timestamp lines go into metadata.timestamp.
3. Platform resolution. In order of precedence: the explicit argument, then the prompt, then detection (command verb, weighted fingerprints such as interface naming, Junos table names and Huawei legend lines). If detection is still ambiguous and a command is known, trial parsing runs every vendor's parser for that command and keeps the richest result.
4. Structured short-cut. | display json / | display xml (or output that looks like JSON/XML) goes straight to the structured engine.
5. Resolution. The command is split into a base command and | pipes. Filtering pipes add a warning. The base is scored against every pattern registered for the platform.
6. Engines. native (dedicated parser), then ntc, genie, and finally generic. If a dedicated parser raises, the failure is recorded as a warning and the next engine takes over (raise_on_error=True disables this).
7. Normalization. When asked, the parser's normalize() maps its native output to the vendor-neutral schema.

Design choices¶

• Native output keeps vendor vocabulary. Nothing is thrown away, and field names are snake_case versions of what the device prints. The normalized view sits alongside it rather than replacing it.
• Flat lists with context tags instead of deep vendor hierarchies. For example, BGP neighbors are one list, and each neighbor carries instance, vrf and address_family. This makes filtering and DataFrames easy.
• Values are typed (int, float, bool, None) where that is unambiguous. Identifiers such as asdot ASNs (65000.100), zero-padded IDs and interface names stay strings.
• Per-line matching (textutils.match_lines) instead of multi-line regexes, so a short line can never steal fields from the next one.
• Zero dependencies. Optional features (YAML, pretty tables, TextFSM, Genie, SSH) are extras and are detected at runtime.

Confidence¶