::documents (rag)

Ingest PDFs, text, markdown, HTML, CSV, and Word (.docx) files into a local knowledge base the agent can search with search_knowledge_base. Retrieval is cited like any other source. Embeddings come from the tier's embedder model (qwen3-embedding:8b) and the vector store is cached under database/cache/.

commands

/docs                       list documents (and workspace files)
/docs add <path>            ingest a file
/docs remove <name>
/docs sync [--force]        re-scan the documents folder

Adding a document no longer runs a model call to summarize it — manifests carry the file's own first line, so ingest is fast and untrusted document text is never fed through a model at ingest time. Retrieved chunks pass through the injection quarantine. Tuning knobs (rag.chunk_size, rag.chunk_overlap, rag.k) live in config.yaml; changing the chunking forces a re-embed on the next sync.

::memory

remember stores a durable fact; recall reads facts back. Memory persists across sessions in a plain markdown file (database/memory/memory.md) and is fully inspectable and editable with /memory.

commands

/memory                 list facts
/memory add <fact>
/memory forget <n>

::workspace instructions

/init surveys your workspace and drafts SATURDAY.md — standing instructions loaded every turn, like a per-project system prompt. Edit it by hand any time; /init --force redrafts it.

::@file mentions

@path tokens in a prompt expand to the file's contents (clamped), with fuzzy path completion; @"path with spaces" works. Attachments and piped stdin are scanned for instruction-shaped content and warn when they carry it, but never block.

::undo

Every write takes a pre-write snapshot. /undo reverts the file changes of the last turn that wrote anything; /undo list shows what can be reverted.

::sessions

Conversations autosave crash-safely; /resume continues the last one, /resume save [name] names one, /resume list lists them, /resume <name> restores. Sessions are plain .json files under database/sessions/ — manage them there. Long histories auto-compact once the context window fills past runtime.compact_threshold.