A skill for batch LLM Wiki ingestion that turns the manual "ingest → continue → continue" cycle into a semi-autonomous loop. Built on Loop Engineering principles: a scheduler reads the queue, the Maker runs the existing ingest pipeline to generate Wiki pages, and an independent Checker performs deterministic verification — stop immediately on failure, proceed only on success.
- Batch-ingests multiple sources from
raw/directories - Reuses the existing Maker pipeline to create
wiki/sources,wiki/concepts,wiki/entitiespages - Tracks state in a local append-only journal instead of relying on chat context
- Uses an independent Checker to limit auto-continuation to mechanical errors — semantic issues always escalate to a human
Maker-Checker separation is the core:
- Maker: runs the standard ingest pipeline — converts a raw source into Wiki pages, updates
wiki/index.mdandwiki/log.md - Checker: runs independently of the Maker (separate process / deterministic script), referencing
references/checker.py— never the same LLM invocation - Journal: writes every processing step to
~/.hermes/ingestion/run_journal.jsonl— used to skip completed items and detect updated sources - Fail fast: whitelisted mechanical errors get one fix attempt; if the re-check still fails, or any non-whitelisted issue appears, the loop stops immediately and asks for human intervention
This is the Loop Engineering boundary: scheduling, generation, verification, and state are separated — the model never gets to both generate and sign off on its own output.
The following local state must exist:
~/.hermes/ingestion/config.json- Must contain at least
vault_root - Typically also defines
raw_dirsto scan
- Must contain at least
~/.hermes/ingestion/run_journal.jsonl- Append-only run log
- Create with
touchif it doesn't exist yet
- Wiki directory structure and base files
wiki/SCHEMA.md— tag taxonomywiki/index.md— navigable indexwiki/log.md— ingestion log- Directories:
wiki/sources,wiki/concepts,wiki/entities
Say this in a Hermes / agent conversation:
消化队列
This triggers the skill to:
- Read config and journal
- Scan for unprocessed raw sources
- Process each file through Maker → Checker
- On success, write journal and continue; on failure, stop and report immediately
For single articles, use the single-file ingest directly — it's lighter and faster.