@murphai/parsers

Local-first multimedia parsing for inbox attachments and other vault-backed media.

This package consumes attachment-level parse jobs from the inbox runtime, selects the best available local parser from a deterministic priority stack, and writes normalized derived artifacts under derived/inbox/... for later chat/model layers. The default ordering is text/native CLI first, with raw-PDF routing fallback handled in the model layer rather than by a second PDF OCR pass, and remote APIs nowhere in the default stack.

Current default stack

• text-like documents: built-in Node text reader
• native-text PDFs: pdftotext when available
• audio and extracted video audio: whisper.cpp
• media normalization: ffmpeg

Design rules

• original raw evidence stays under raw/inbox/...
• parse outputs are derived files, never canonical vault state
• provider discovery stays explicit and local-first
• adapters remain thin wrappers around mature open-source tools
• all modalities normalize into one parse result shape (text, markdown, chunks, tables, metadata)

Integration seams

• createInboxParserService(...) wraps scoped drain and requeue flows for an inbox runtime
• @murphai/inboxd now owns the inbox-plus-parser composition helpers: createParsedInboxPipeline(...) processes a capture and immediately drains any newly enqueued attachment jobs runInboxDaemonWithParsers(...) backfills parser jobs on startup and keeps future captures auto-drained

This keeps parsing additive to @murphai/inboxd: raw inbox evidence remains canonical, while parser outputs stay rebuildable under derived/inbox/**.

Toolchain config and discovery

• writeParserToolchainConfig(...) persists local command and model-path overrides under <vault>/.runtime/parsers/toolchain.json
• discoverParserToolchain(...) reports which local tools are currently available plus where each setting came from
• createConfiguredParserRegistry(...) builds a default registry from the discovered toolchain state