supragnosis

Define ontology types (T-Box) for a workspace: give an entity type or relation type a name and a natural-language definition of what it means. Use this to record the vocabulary of the ontology so a type is not just a bare label but has a stated meaning (e.g. entity type 'Driver' = 'a kernel module that ...'; relation type 'depends_on' = 'X requires Y at runtime'). A description is required. Scoped to the workspace. Read them back via the supragnosis://workspace/{ws}/types resource.

Look up an entity and its relations/provenance by entity id.

Get one proposal's folded state (kind/targets/into/rationale/state/verdicts/workspace) by id, with its belief diff and checks, read in the workspace it lives in; a merged claim_promotion also carries granted_tier, the tier it granted. A tbox_change or a recall carries no_effect_yet: its kind has no commit effect yet, so it cannot be merged.

List proposals in a workspace (newest first) with their folded state (open/blocked/merged/rejected/withdrawn), kind, targets, verdict count and the workspace each lives in; a merged claim_promotion also carries granted_tier, the tier it granted. A verdict counts only in its proposal's workspace. A tbox_change or a recall carries no_effect_yet: its kind has no commit effect yet, so it cannot be merged.

Ingest a knowledge fragment - how anything new enters the graph. Stores it as an immutable observation (the source of truth: superseded later, never edited or deleted) and links the entities and relations you extracted alongside it into the ontology. Use it as facts, decisions and the reasons behind them come up; to reshape knowledge that is already stored - merging duplicates, changing trust, retracting - use `propose` instead. Returns the observation id, the linked entity/relation ids and the workspace written to; a write that makes a new workspace adds a note naming the ones that hold knowledge.

Open a proposal to change the canon - the one path by which anything enters shared knowledge. Use it to reshape what is already there rather than to add: entity_merge and entity_split for identity, claim_promotion and claim_demotion for trust, tbox_change for the type vocabulary, recall for retracting a contaminated derivation tree - those two have no commit effect yet, so neither can be merged until it has one. To state something new, use `observe` instead. A proposal is itself an observation and changes nothing until `review` accepts it; `get_proposal` shows where it stands, and `kind` describes what each one does and which arguments it needs. Returns the proposal id, the workspace, its state and found - true when an identical proposal was already there, since a proposal is its content - with a note saying what is left.

Cast a verdict on a proposal. decision is merge (accept), reject, comment, or withdraw. In a single-user workspace the verdict is self-attested. The state is a deterministic fold of the events; a merge is the absorbing outcome. A claim_promotion merged through this agent surface grants at most host_signed - promotion to human_confirmed is the person's own act in the Supragnosis app: never cast it yourself, by this tool or any other way. A merge on a recall is refused here for the same reason, and a merge on a tbox_change or a recall is refused on every surface while the kind has no commit effect; comment, reject and withdraw are accepted. Returns the recorded observation id and the workspace it was filed in, the proposal's own.

Search knowledge (entities/observations). The response mode reports the surface actually used: hybrid (semantic+keyword) or keyword (keyword-only degrade - an empty result in this mode may be a recall failure). score is for rank comparison only - its scale differs per mode, so the absolute value is not a confidence.

Pull the workspace's missing knowledge from the configured sync servers and apply it (verify signatures -> dedup/absorb -> re-materialize the projection). Returns per-server accepted/rejected counts.

Push the workspace's local knowledge to the configured sync servers (backfill-stamp unstamped attestations, then send what each server lacks per its advertised version vector). Only share-whitelisted workspaces leave this node.

Federation sync status of this node: node id, public key, share whitelist, configured sync servers, the workspace's version vector (what this node holds per origin), `config_notes` (federation configuration this build worked around rather than obeyed - read these, they are the reason something is not syncing), and `negotiated` - what each host last said this node may reach, as a three-way difference against the share whitelist (`both` / `local_only` = listed here but not admitted there, a setup error / `peer_only` = admitted there but not shared here). A host with state `unknown` has not answered yet; that is not an empty grant set. Read-only.

Traverse the graph from an entity, following relation direction (from->to). Returns the entities reachable within the maximum hops (max_depth), along with their shortest distance. Reachability runs THROUGH a relation endpoint that no entity row describes yet (partial ingest), but such an endpoint is not listed - so a hit can sit deeper than any listed neighbor. That gap is unknown, not a missing edge.

Survey a workspace's main co-occurrence contexts (hyperedges - the set of entities asserted together in a single observation). For cold-start orientation: before searching, grasp by name 'what is here and in what clusters'. Clusters are sorted by size (number of co-occurring entities) and sources is the number of supporting observations. This is a directional signal, not an asserted relation - confirm actual relations/details with search_knowledge/get_entity. To solidify a recurring context into first-class structure, observe a group entity (e.g. type Context) with member_of relations from each member and derived_from listing the co-asserting observation ids - the assertion then carries provenance and trust like any other edge.