llm-wiki-source-packet

name	llm-wiki-source-packet
description	Source packet runtime skill for LLM Awesome Wiki. Use when an agent needs to create, inspect, or validate source inventory rows, raw-source identity, source packets, extraction backend handoff, source-type packet profiles, generated fields, known gaps, or source packet review routing.

LLM Wiki Source Packet

Role

Use this subskill for the raw-to-source-packet stage only.

This skill is the normal runtime for source intake. It does not run a parser or extractor harness. It defines the packet shape an agent or optional backend must leave behind.

Detailed source rules are loaded only when needed:

• rules/source/extractor-adapter-protocol.md for optional manual, Agent/MCP, local CLI, MinerU-like, or custom extraction backend handoff
• rules/source/source-type-packet-profiles.md when PDF, PPTX, DOCX, image, table, dataset, or mixed-media minimum shape matters
• rules/source/generated-fields-review-routing.md when OCR, captions, chart summaries, table repairs, formula recognition, or inferred structure affect important knowledge

Inputs

• raw files under raw/sources/
• source inventory or source list
• workspace purpose.md and schema.md when present
• optional backend output, manual notes, rendered media, OCR, or VLM output

Do not rewrite raw source files.

Runtime

1. Confirm each considered source has an inventory row.
2. Assign or verify a stable source_id.
3. Record raw_path, raw_sha256 or hash state, source_kind, intake status, and packet path.
4. Create or inspect:

raw/derived/<source-id>/
  metadata.yml or metadata.json
  source.md
  media/ optional
  review.md optional
  gaps.md optional

5. In metadata, record source_id, raw_path, raw_sha256, source_kind, extraction_backend, extraction_method, extraction_version, extraction_status, modalities, generated_fields, known_gaps, review_required, review_reason, derived_artifacts, created, and updated when available.
6. In source.md, preserve stable anchors that later stages can cite:

<source_id>#<anchor_id>

Each useful anchor should have location, label, content kind, source-derived state, generated state, media or derived artifact reference when available, and concise notes.

7. Keep backend-local IDs subordinate to workspace anchors.
8. Mark generated fields and extraction gaps explicitly.
9. Route partial, failed, manual-review, uncertain visual, changed-source, or important generated content to review.

Outputs

Expected outputs are source inventory state and source packets. This stage does not write final wiki pages.

Use source references in later stages as:

<source_id>#<anchor_id>

Boundaries

• Do not accept final wiki pages directly from raw files without source packet anchors and a grounding pass when a source packet is feasible.
• Do not turn source.md into a polished chapter summary.
• Do not hide unsupported modalities or extraction gaps.
• Do not treat generated captions, OCR, chart summaries, formulas, or inferred structure as source-equivalent unless review accepts that state.
• Do not let an optional backend create the durable workspace source namespace.

Validation

Run source checks when available:

python -m llm_wiki_tools workspace-check --workspace . --mode source

If source checks are unavailable, record the manual checks and unresolved review items in the validation note.

Minimum manual checks:

• every processed source has an inventory row and stable packet path
• every packet has metadata and at least one anchor unless extraction failed
• generated or model-assisted fields are visible
• partial, failed, or manual-review packets record gaps and review reason
• later stages can cite <source_id>#<anchor_id> without depending on backend internals