people-context

Add a normalized-deduplicated alias to an existing person. `kind` is one of `nickname`, `native_script`, `transliteration`, `handle`, `former_name`, or `other`, and defaults to `other`. The published schema carries the enum, so an unlisted value is refused before the alias is built rather than dropped.

Record that an existing person held a role in an existing group. Record only what was stated or confirmed. Leave dates out when they are unknown — never guess a school year or fill a January 1 date. `temporal_basis` is `unknown` (no dates), `period` (at least one known bound; a missing bound means unknown, not open-ended), or `ongoing` (asserted current, no end date); it is inferred from the dates when omitted, except that `ongoing` must be stated. `role` defaults to `member`. `sensitivity` defaults to `personal`. Concurrent and historical memberships are all kept.

Correct one staged candidate before it is committed, returning the revised batch. The patch replaces the named fields and leaves the rest alone. Nothing is committed: confirming an amendment is not acceptance of the batch.

Record the last day a historically true membership held. Use this when someone left a group; the membership's earlier facts are preserved. It refuses a membership that already records an end (`already_ended`) or an end before its start (`ends_before_start`). Use `correct_record` with entity type `group_membership` when a stored membership was simply wrong.

Commit accepted people and resolvable interactions idempotently. Accepting a withdrawn candidate refuses the whole commit. Supplying the `expected_batch_digest` from the review the user approved refuses it if the batch moved after they saw it.

Transition one active reminder to completed.

Correct whitelisted assertion fields in place with before/after audit.

Create one identified group: a class, cohort, team, department, club, household, or community. Every call creates a new group; an equal name is never treated as the same group. Use `find_groups` first and reuse the returned `id` when the user confirms it is the same one. `organization_id` must name an existing organization and only places the group under it — it creates no affiliation. `sensitivity` defaults to `personal`; a `sensitive` or `restricted` group is withheld from ordinary reads together with every membership in it.

Explain how two people are connected through groups they were both recorded in, and when. Use for shared-background, introduction, and "how do they know each other" questions. Pass each person as an id (`person_a_id`, `person_b_id`) or a name (`person_a`, `person_b`). Each connection is one membership of each person in the same group, cited in full. `label` is `classmates` (both `student` in a `class`) or `teammates` (both `participant` in a `team`) only when the recorded dates prove a common day (`temporal: overlap`, with the `overlap` period); otherwise the connection is `shared_context` only. `temporal` is `disjoint` when the dates exclude a common time and `unknown` when they cannot tell. Sharing a club, household, or community is not friendship, kinship, or acquaintance; different groups under one organization, and people linked through a third person, share nothing here. `direct_relationships` are recorded relationships between the two, kept apart from derived results. No result means no shared group was found, not that they do not know each other. Only ordinary groups and memberships are used; `truncated` and `memberships_truncated` report a partial answer.

Return one shortest relationship path, or a structured not-connected result.

List groups whose normalized name contains `name`, optionally of one `kind`. Results are candidates, not identity: two groups with the same name stay distinct, so ask the user which one is meant when several match. Sensitive and restricted groups are never returned by this ordinary tool. Ordered by normalized name then id; `truncated` reports more.

Hard-delete a person or record and redact identifying audit history.

Return sensitivity-gated signal for client-composed communication advice. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline; an ambiguous name returns candidates instead of guidance.

Return one person's stored facts, traits, observations, and affiliations and how they relate. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline. Use this before proposing maintenance. `signals` names pairs of records that share a predicate or category and says how they stand — `duplicate_fact`, `restated_fact`, `contradictory_fact`, `succeeding_fact`, `duplicate_trait`, `divergent_trait` — comparing normalized values and inclusive validity periods only. It decides nothing: reading the evidence and proposing a `correct_record`, a `supersede_fact`, or a `merge_people` for the user to approve is your job, and several observations supporting one trait are separate evidence rather than duplicates. `affiliations` carries the person's stored roles at organizations — employment, education, membership — with their ids, dates, provenance, and any import receipt, so an incoming CV claim can be checked against what is already recorded. It has its own page and its own `affiliations_truncated` flag, and no signal is computed over it: two roles at one organization may be a promotion, a rehire, or two concurrent posts, and only you can read which. A `valid_to` of null means the stored assertion set no end, not that the role is confirmed current, and `created_at` is when the row was written rather than when the role began. A page that truncates is not evidence that a role is absent. This read never writes. Sensitive and restricted records are never returned, and a trait names only evidence that is itself ordinary. Affiliations carry no disclosure level at all, so this collection adds no sensitivity control and is no route to a protected fact — background that needs one belongs in a fact. An unknown or removed person returns `found: false` rather than an error.

Return one group and a bounded page of its recorded memberships. These are recorded memberships only: sharing a group is context, not proof that members know each other. Sensitive and restricted groups read as `found: false`, and sensitive or restricted memberships are omitted without affecting `truncated`.

Assemble a minimal-disclosure context bundle for one person. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline: an ambiguous name returns the candidates instead of context. Returns narrow identity fields, active relationships and affiliations, and one ranked facts/interactions slice capped by `max_items`. Set `include_communication=true` (or a `purpose` mentioning communication) to include communication traits. Sensitive and restricted records are never returned by this ordinary tool, and leave no trace that they exist. `truncated` says the item budget cut the list.

Return one person's recent history, newest first, as a bounded chronology. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline. Entries project durable records — interactions, observations, facts, affiliations, relationships, and traits — with the stored timestamp each was placed by and which field that was. Sensitive and restricted records are never returned by this ordinary tool, and a trait names only evidence that is itself ordinary. An unknown or removed person returns `found: false` rather than an error.

Return active relationship structure around one person, capped for bounded disclosure. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline.

Return people you have not interacted with recently, using ordinary interactions only.

Extract and atomically stage header-only candidates from a supported source without bodies. Accepted `source_type` values are `email`, `mbox`, `vcard`, `ics`, `linkedin`, `outlook`, and `whatsapp`. `self_sender` is an optional chat-export label for the user, such as a display name or a bare phone number, used to omit the user's own messages. A `path` import records a receipt for the file it read, so importing that exact file again reports the existing batch instead of staging a second copy. `forced` says the repeat is intentional: it stages the same content as a distinct processing session and never weakens the duplicate rule for later calls. It is also the only way past `source_previously_redacted` after a hard forget — and for `mbox`, which is read from a path and cannot be resubmitted as inline content, the only way at all.

Return a bounded page of one person's recorded group memberships with each group. Pass `person_id` from `resolve_person`, or `person` (a name or alias) to resolve inline. Only ordinary memberships in ordinary groups are returned. An unknown or removed person returns `found: false`.

List pull-based reminders, due-dated first and communication notes last. Filter by `person_id`, or by `person` (a name or alias) resolved inline; omit both for every person's reminders.

Merge a duplicate person into a primary person atomically.

Record a time-aware fact about an existing person. `sensitivity` defaults to `personal`; `sensitive` and `restricted` records are withheld from ordinary reads. Prefer `remember` for a single statement named by a person's name.

Record a concise interaction summary after validating all participants.

Record a subjective observation, separate from disclosed context.

Record a derived trait with validated category and provenance. `category` is one of `communication_style`, `temperament`, `values`, `preference`, `topics_to_avoid`, or `other`. Cite the observation or interaction it rests on in `evidence_note` or `evidence_ids` where you can.

Record one thing the user stated about one person, in a single call. Use this when the user directly tells you something durable — "Alice from Acme prefers short emails", "Bob is my manager", "I had coffee with Dana today". `person` is the name as the user said it: it is resolved first, and a new person is created only when nobody matches. `org`/`role` record an affiliation, `relationship` records how the user relates to them (e.g. `manager_of`, `friend_of`, from the user's point of view), and `note` records the statement as a `fact`, `trait`, or `interaction`. Leave `kind` as `auto` to classify the note by a fixed keyword rule, or set it explicitly when you know. Pass `occurred_at` when the statement says an interaction happened earlier ("met Dana last week"); without it such a note is refused rather than dated today, because that date is what the staleness report reads. Identity is never guessed for a write: `status: ambiguous` or `unconfirmed` returns candidates and records nothing — ask the user, then call again with the exact canonical name or a unique alias. Everything recorded commits in one transaction and is audited like the individual tools. For material you extracted or inferred from a transcript rather than a direct statement, use `stage_candidates` so the user reviews it first.

Create a new person or update the existing one matching `name`. Use this to durably record someone the user talks about. `aliases` is a list of `{value, kind?, lang?, script?}` objects (kinds: nickname, native_script, transliteration, handle, former_name, other); new aliases are merged into an existing record. Set `summary` to describe who they are, and `is_self=True` only for the user themselves. Returns the person and whether it was created.

Resolve a name, nickname, or partial reference to candidate people. Call this first whenever the user mentions someone, before asking who they mean. Returns ranked candidates with a score and match reason. If two or more candidates are close, the result is flagged `ambiguous` and all are returned so you can disambiguate with extra context or a clarifying question. An empty candidate list means no confident match — use `remember` or `remember_person` to create a new record. Put distinguishing context in `hints` (`org`, `role`, `relationship`) rather than in `query`.

Return staged candidates and statuses for one batch. The response carries a `batch_digest`. Pass it back to `amend_candidate`, `withdraw_candidates`, or `commit_import` to have the action refused if somebody else changed the batch after this read.

Free-text search over stored people for browsing or lookup. Broader than `resolve_person`: use this to list who is known that matches a query, rather than to pin down a single identity. Returns ranked candidates.

Search active people and safe interaction summaries by multilingual semantic similarity. This optional local search requires an explicit `pctx reindex --semantic` first. It never downloads a model while serving a query and refuses to mix vectors from different model revisions.

Create an affiliation, resolving an org id or get/creating by name.

Store communication philosophy verbatim while auditing lengths only.

Create a directed relationship between two existing people.

Create a reminder for an existing person. `kind` is `follow_up` (dated, something to do), `occasion` (a date that recurs), or `communication_note` (undated guidance surfaced with the person's context).

Stage agent-extracted people, interactions, records, relationships, groups, and memberships. Candidate types are `person`, `interaction`, `affiliation`, `fact`, `observation`, `trait`, `relationship`, `group`, and `group_membership`. Use this after extracting concise candidates from user-provided notes, meeting transcripts, or other agent-visible text. Distinguish what was stated (`fact`), what happened in this source (`observation`), and what you inferred (`trait`, which requires an explicit `confidence` and a concise `evidence_note`). Relationship candidates carry batch-local `from_ref`/`to_ref` and are ordinary-disclosure only: omit a relationship the user would consider sensitive or restricted rather than staging it. References are batch-local; raw notes and source text must not be included in candidate fields. A request using `observation`, `trait`, or `relationship` is bounded to 500 candidates, a 128-character `source`, 1 MiB of candidate JSON, and 8 KiB per string. A `fact` or `affiliation` may record who asserted it in an optional `stated_by` of at most 256 characters: a person, a document, or a role. Attribution is not verification. A CV saying someone is analytical is that person's own claim about themselves, so stage it as a fact whose value says so and whose `stated_by` names them, never as an inferred `trait`. `stated_by` is separate from `source`, which names the process that wrote the row, and from `source_kind` below, which receipts the artifact read. Omit it when the attribution is unknown rather than guessing at a speaker. A trait may name the records it was drawn from. Give a supporting `observation` or `interaction` any short `evidence_ref` label of your own and list those labels in the trait's `evidence_refs`; use `evidence_ids` for records already stored. Evidence must be about the trait's own person, and one trait cites at most 32 references and ids combined, each at most 256 characters. A shared context is two candidates, never one. A `group` names an identified class, cohort, team, department, club, household, or community and carries a batch-local `ref`; a `group_membership` places one `person_ref` in one `group_ref` with a `role`, and requires a batch that passes `source_kind`, so a group committed in an earlier call can still be named. Group names never merge: staging looks nothing up by name, so committing a group candidate creates a new group unless you pass `group_id` — resolve an existing group with `find_groups` first and pass the id it returns. Record dates only where the source gave them. Absent dates stay absent and mean unknown, never "still going": `temporal_basis` is `unknown` without dates, `period` with them, and `ongoing` only when you set it because the source said so. Do not invent a year, a class number, or a yearly placement per grade; a confirmation that people stayed together records the people and the extent that were confirmed, and nothing more. "My classmate" with no identified group is a `relationship`, not an invented group. `source_kind` optionally records an import receipt for this batch. It is a machine category such as `meeting_transcript`, at most 128 characters of letters, digits, `.`, `_`, `-`, or `/` — never a person, a title, or a description; put any human wording in `label` instead. If you can compute a SHA-256 over the exact source artifact, pass it as `content_digest` (64 lowercase hex characters) so re-importing that same source can be detected; without one, no duplicate detection is promised. `extraction_fingerprint` is optional and should be omitted unless you have explicit, bounded configuration semantics for it. None of these fields may carry source text.

Close a fact that was true and open its replacement from an effective date. Use this when a stored value was historically correct and the real-world state changed; `correct_record` remains the tool for a value that was simply wrong. The old fact keeps its person, predicate, value, and provenance and is closed the day before `effective_from`; the replacement inherits the old assertion's original end date, so a bounded claim is never widened into an open-ended one. Person, predicate, and the replacement's end date cannot be changed here. Omitting `confidence` or `sensitivity` inherits the old fact's. Both rows commit together under one logical transaction, or neither commits.

Return ordinary birthdays and dated active reminders inside an inclusive upcoming window. Optionally narrow to one person by `person_id` or by `person` (a name or alias).

Drop staged candidates from consideration, returning the revised batch. A withdrawn candidate stays listed as `rejected` so the user can see what was dropped, and is never committed. Nothing else about the batch changes.