| name | knowledge-synthesize |
| model | haiku |
| description | Surface cross-node connections in a knowledge root. The deterministic core (octopus synthesize, over the octopus kr registry) ranks candidates — shared-target, co-mention, and --node relevance (forgotten-but-relevant); this skill judges which are real, flags contradictions, and seeds missing links with --fix.
|
| triggers | {"paths":["docs/**","knowledge/**","CONTEXT.md"],"keywords":["synthesize","connection","related","forgotten","contradiction"],"tools":[]} |
/octopus:knowledge-synthesize
Purpose
A knowledge base is worth more than the sum of its notes only if something
traverses it and surfaces links manual recall misses. Every root is a silo by
default. This skill surfaces cross-node connections and the forgotten-but-relevant
past note, and flags where a node contradicts an authority it cites.
The mechanical candidate-finding is deterministic in the octopus synthesize
core, which reads nodes and links from octopus kr. This skill adds the
judgment the core can't make.
Invocation
/octopus:knowledge-synthesize [--root <id>] [--node <path>] [--fix]
--root <id> — one root (e.g. docs, memory, consigliere); default: every resolved root.--node <path> — forgotten-but-relevant lookup for that node (ranked by shared entities).--fix — seed a missing link where a mention resolves to exactly one node.
Run the core directly with octopus synthesize [--root <id>] [--node <path>] [--fix].
Signals
Each candidate is one line: kind|root|a|b|signal|score.
- shared-target — node pair
a,b that both link the same third node (signal); score = intersection size. They relate via that target.
- co-mention — an entity (
a) appearing across score nodes with no node of its own. A recurring topic with no home.
- relevant — (with
--node) other nodes ranked by shared-entity overlap with the focus node.
Contradiction
The core does not decide contradictions — it surfaces a node together with the authorities it links (e.g. ADRs). Read the node and its linked authority; judge whether the node asserts something the authority negates, and report it with the citing line. This is the judgment that justifies the skill over the bare core.
Fix Mode
--fix seeds a relative link only when a mention's entity resolves to exactly one node title (unambiguous) and the link is not already present. Multi-target or fuzzy matches stay report-only. The edit is a plain append the user can git revert.
Entity Extraction (language-neutral core + LLM free-text pass)
The deterministic core's ks_entities extracts only structural, language-neutral entities — [[mentions]] and `code` spans. It does not guess free-text entities: a hardcoded English regex + stopword list would miss accented entities ([[Política Fiscal]], Gestão de Estoque) and silo every non-English root.
Free-text / multilingual entity detection is yours (the LLM). When a root is not wikilinked (e.g. docs/ with relative links), read the nodes and identify the proper nouns / domain terms in whatever language they are written in — pt-br, en, etc. This linguistic step is exactly what the model layer is for.
Run this free-text pass on the cheapest model tier. It is light NLP, not reasoning — the costly structural work already happened for free in bash. When dispatching it through Octopus, route to the cheapest model (Claude: octopus ask <role> --skill knowledge-synthesize --model haiku "…"; other assistants: their fastest/cheapest model). Do not spend a frontier model on entity tokenization.
