Skip to content
Octopus Octopus Octopus

digest-source

digest-source is how the consigliere workspace gets filled. You drop an input and describe it in one line; the skill turns it into grounded, routed memory. It does capture only — querying is context-status, and the heuristics loop is playbook-review.

One command

/octopus:digest-source <source> "natural-language description"
# <source> = pasted text | path/to/file.pdf | JIRA-123 | confluence-url

The description is the routing. Describe the input as you would to a colleague — “the alignment meeting with the payments team about the new checkout flow” — and the skill figures out where it belongs, then shows you what it understood before writing anything.

What it does, in six steps

  1. Resolve + guard. Reads consigliere.workspace; every write stays inside it (the write-guard from consigliere-bootstrap). Nothing ever lands in a team repo.
  2. Snapshot. Materializes the input under sources/YYYY/MM/ with origin / kind / fetched_at frontmatter. The snapshot is immutable — it is the grounding base everything traces back to.
  3. Route by inference, confirm before writing. Infers the target context or cross-cutting project from your description and the existing tree, and asks you to confirm. If the node doesn’t exist, it offers to create it. Ambiguous? It asks — it never guesses.
  4. Extract six grounded fields. Status, blockers + owner, decisions, the map of systems/areas, actions + owners, and political risk — the signals that never reach Jira. Every line cites its source line (src: …). It asserts only what is in the snapshot; if it isn’t there, it isn’t in the digest (this reuses audit-grounding).
  5. Preview. Shows the journal block, the state diff, and the pointer lines before touching disk — and surfaces any relevant heuristic (“owner tends to delay → FUP?”) as a suggestion, never as fact.
  6. Write. Appends a dated block to journal.md, rewrites the materialized state.md, and for a transversal project fans out a one-line pointer into each context it crosses — so each context stays self-sufficient when you later ask “how’s payments?”.

Sources today

Pasted text, local PDFs/files, and Jira (via MCP) work end-to-end now. Confluence needs the Atlassian MCP; until that lands, the skill asks you to export the page as PDF or paste it rather than failing silently.

Where it sits

digest-source is the second member of the consigliere bundle, built on the consigliere-bootstrap scaffold. It feeds the memory that context-status reads and the heuristics that playbook-review promotes.