UNIFIELD

Read-only advisory/safety check. Tests whether a proposed agent action is consistent with a stored intent's core, axiom exclusions, and modifiers. Returns allowed | needs_confirmation | blocked. An ok:false / 'blocked' result MUST be treated as a stop-or-clarify signal — do not override it on the user's behalf. Advisory only — humans remain responsible.

check_constraints

Read-only, evidence-only. Return canonical structured evidence for a specific match (score, grade, matched facets, conflicts, missing_data). No server LLM call, no premium_explanations charge, no persisted natural-language wording. ChatGPT explains the evidence to the user ephemerally. Counterpart original text is never returned.

explain_match

Field a user-confirmed intent into the Unifield (the open global intent layer) or into a private field provided by UNIFIELD. Mutating. Private by default. Use this to help the user place a confirmed formulation into the Unifield AFTER you have shown the proposed intent and obtained explicit human confirmation in the current turn. Literal `consent_confirmed:true` is REQUIRED whenever the target is the global Unifield, `global_unifield:true`, or visibility/consent_state is non-private; missing/false consent returns JSON-RPC -32013 consent_required. Original-text boundary: `original_text_sharing:"omit"` is the safest default for ChatGPT-originated flows; `"store_private"` is only acceptable when the user has explicitly confirmed a single editable formulation. Raw chat context must NEVER be sent as original text. Do not phrase to the user as 'publish' — say 'field' / 'into the Unifield'. Non-binding: humans remain responsible.

publish_intent

Read-only. Return ranked base complements for one of the user's own intents (score + grade + evidence). Origin selection: supply EITHER intent_id (precise machine reference, only when already known from a previous tool result) OR origin_query (the user's natural-language reference to their own intent) — exactly one, never both, never neither. NEVER ask the user for a UUID. With origin_query the response carries `resolution`: RESOLVED (ranking continues in the same call), AMBIGUOUS (ask the user to choose by human-readable title/summary; matches is empty), NOT_ELIGIBLE (the existing intent is not ready for complement discovery), NOT_FOUND (no matching existing intent). Never auto-publish a replacement intent. Non-binding suggestions; human decides. Does not generate or refresh explanations, does not charge premium quota, and does not write cache. For a natural-language explanation of a specific match, call explain_match (evidence mode).

find_complements

Read-only. Return the fielding lifecycle and stepper status of a user-owned intent (draft → analysis → active → connected). Use this to help the user see where their intent stands in the Unifield without modifying anything.

get_intent_status

Read-only. List the fields the user may field intents into. Distinguishes the open global Unifield from private fields provided by UNIFIELD (e.g. organisations, communities). Use this to help the user choose where to field an intent.

list_fields

Preview a possible intent (IXP intent molecule) from user-provided wording. Read-only: nothing is stored or fielded by this call. Use this to help the user articulate what they want before fielding it into the Unifield. For ChatGPT-originated flows, do NOT pass arbitrary chat history as `text`; ask the user to confirm a single editable formulation first. Chat context is never original_text. Output is a non-binding preview; the user must confirm before publish_intent fields anything.

parse_intent

Render a single complement match card from find_complements.matches as a widget surface. Read-only: never requests contact, never calls request_connection, never calls any backend or tool. Input must be one privacy-safe ComplementMatchCard. Never exposes counterpart original_text or chat context.

render_complement_match_card

Render a previously parsed intent draft (IntentDraftPreview) as a widget surface. Read-only: performs no writes, makes no backend or other tool calls, never re-fetches data. Input must be a strict subset of parse_intent's structuredContent envelope. Never accepts or exposes counterpart original_text or chat context.

render_intent_draft

Render the lifecycle stepper for a caller-owned intent (from get_intent_status or publish_intent). Read-only: performs no writes, makes no backend or other tool calls. Never exposes counterpart original_text or chat context. Either id or intent_id, and either status or current_step, must be provided.

render_intent_status