personas: composite FK + built-in update guard

Two persona-infrastructure correctness fixes that go together because
the second one (FK with CASCADE) requires the first (preventing the
persona row from being mutated out from under its facts).

1. update_persona handler refuses name/systemPrompt edits to built-ins
   (409). includeAllMemories stays editable — that's a per-user
   preference, not the persona's identity. Mirrors the existing
   delete_persona guard. The DAO is intentionally permissive so the
   guard sits at the HTTP layer; persona_dao test pins that contract.

2. Migration 2026-05-10 adds user_id to entity_facts and a composite
   FK (user_id, persona_id) -> personas(user_id, persona_id) ON DELETE
   CASCADE. This closes two issues at once:

   - Persona orphans: deleting a custom persona used to leave its
     facts dangling forever, readable only via PersonaFilter::All.
     CASCADE now wipes them with the persona row.

   - Multi-user fact leakage: PersonaFilter::Single("default") used
     to surface every user's default-scoped facts. PersonaFilter is
     now { user_id, persona_id } and all read paths
     (get_facts_for_entity, list_facts, get_recent_activity) filter
     on user_id first. upsert_fact's dedup key extends to user_id so
     identical claims under shared persona names from different
     users no longer corroborate-bump each other's confidence.

   - user_id threads from Claims.sub.parse::<i32>().unwrap_or(1) at
     the chat / insight handlers through ChatTurnRequest, the
     streaming agentic loop, execute_tool, and into the leaf tools
     (tool_store_fact, tool_recall_facts_for_photo). The ".unwrap_or(1)"
     accommodates Apollo's service token whose sub is non-numeric on
     legacy mints.

   - Backfill picks the smallest user_id matching each legacy fact's
     persona_id so the FK holds for already-stored rows.

Five new knowledge_dao tests with FK-on connection: persona scoping
isolation, All-variant union per-user, dedup not crossing users,
CASCADE delete, FK rejection of unknown personas. Plus
dao_update_does_not_block_built_ins documenting where the
HTTP-layer guard lives.

Apollo coordinates separately — the matching changes there add the
/api/personas proxy and start sending persona_id on photo-chat turns.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Cameron Cordes
2026-05-10 13:30:35 -04:00
parent 79a1168724
commit fbd769e475
12 changed files with 629 additions and 28 deletions

View File

@@ -57,12 +57,25 @@ pub struct FactFilter {
/// Persona scoping for fact reads. `Single` filters to one persona's
/// view; `All` is the hive-mind read used when a persona has
/// `include_all_memories=true` in the personas table. Entities and
/// photo-links are always shared and don't take a persona filter.
/// `include_all_memories=true` in the personas table. Both variants
/// carry `user_id` because facts are user-isolated — two users with
/// the same 'default' persona must not see each other's facts (this
/// is enforced at the schema level by the composite FK in migration
/// 2026-05-10). Entities and photo-links are always shared and don't
/// take a persona filter.
#[derive(Debug, Clone)]
pub enum PersonaFilter {
Single(String),
All,
Single { user_id: i32, persona_id: String },
All { user_id: i32 },
}
impl PersonaFilter {
pub fn user_id(&self) -> i32 {
match self {
Self::Single { user_id, .. } => *user_id,
Self::All { user_id } => *user_id,
}
}
}
pub struct EntityPatch {
@@ -598,12 +611,14 @@ impl KnowledgeDao for SqliteKnowledgeDao {
let mut conn = self.connection.lock().expect("KnowledgeDao lock");
// Look for an identical active fact AUTHORED BY THE SAME
// PERSONA. The same claim from a different persona is a
// (USER, PERSONA). The same claim from a different persona
// or from a different user with the same persona name — is a
// separate fact (each persona's voice/confidence is its own),
// not a confidence bump on someone else's row.
let mut dup_query = entity_facts
.filter(subject_entity_id.eq(fact.subject_entity_id))
.filter(predicate.eq(&fact.predicate))
.filter(user_id.eq(fact.user_id))
.filter(persona_id.eq(&fact.persona_id))
.filter(status.ne("rejected"))
.into_boxed();
@@ -665,8 +680,9 @@ impl KnowledgeDao for SqliteKnowledgeDao {
let mut q = entity_facts
.filter(subject_entity_id.eq(entity_id))
.filter(status.ne("rejected"))
.filter(user_id.eq(persona.user_id()))
.into_boxed();
if let PersonaFilter::Single(pid) = persona {
if let PersonaFilter::Single { persona_id: pid, .. } = persona {
q = q.filter(persona_id.eq(pid.clone()));
}
q.load::<EntityFact>(conn.deref_mut())
@@ -688,6 +704,11 @@ impl KnowledgeDao for SqliteKnowledgeDao {
let mut query = entity_facts.into_boxed();
let mut count_query = entity_facts.into_boxed();
// user_id always applies — facts are user-isolated.
let uid = filter.persona.user_id();
query = query.filter(user_id.eq(uid));
count_query = count_query.filter(user_id.eq(uid));
if let Some(eid) = filter.entity_id {
query = query.filter(subject_entity_id.eq(eid));
count_query = count_query.filter(subject_entity_id.eq(eid));
@@ -701,7 +722,7 @@ impl KnowledgeDao for SqliteKnowledgeDao {
query = query.filter(predicate.eq(pred));
count_query = count_query.filter(predicate.eq(pred));
}
if let PersonaFilter::Single(ref pid) = filter.persona {
if let PersonaFilter::Single { persona_id: ref pid, .. } = filter.persona {
query = query.filter(persona_id.eq(pid.clone()));
count_query = count_query.filter(persona_id.eq(pid.clone()));
}
@@ -901,8 +922,9 @@ impl KnowledgeDao for SqliteKnowledgeDao {
let mut facts_q = ef::entity_facts
.filter(ef::created_at.gt(since))
.filter(ef::user_id.eq(persona.user_id()))
.into_boxed();
if let PersonaFilter::Single(pid) = persona {
if let PersonaFilter::Single { persona_id: pid, .. } = persona {
facts_q = facts_q.filter(ef::persona_id.eq(pid.clone()));
}
let recent_facts = facts_q
@@ -919,3 +941,339 @@ impl KnowledgeDao for SqliteKnowledgeDao {
.map_err(|_| DbError::new(DbErrorKind::QueryError))
}
}
#[cfg(test)]
mod tests {
//! Persona scoping + composite-FK invariants for entity_facts.
//!
//! These tests pin three contracts that are silently regressable:
//!
//! 1. PersonaFilter::Single isolates per (user_id, persona_id). Two
//! users with the same 'default' persona must not see each
//! other's facts (multi-user leakage was a latent bug before
//! migration 2026-05-10 added user_id + composite FK).
//!
//! 2. PersonaFilter::All scopes to a single user but unions across
//! that user's personas. Hive-mind for human browsing of
//! /knowledge/*; never crosses users.
//!
//! 3. Deleting a persona CASCADEs to the user's facts under that
//! persona — and ONLY that user's, ONLY that persona's. Other
//! users sharing the persona_id name keep their facts.
//!
//! FKs aren't enabled by default on Diesel's SQLite connection;
//! `connection_with_fks_on()` flips the pragma so the cascade
//! actually fires in tests (mirroring runtime in production).
use super::*;
use crate::database::models::{InsertEntity, InsertEntityFact, InsertPersona};
use crate::database::test::in_memory_db_connection;
use diesel::connection::SimpleConnection;
fn connection_with_fks_on() -> Arc<Mutex<SqliteConnection>> {
let mut conn = in_memory_db_connection();
conn.batch_execute("PRAGMA foreign_keys = ON;")
.expect("enable foreign_keys pragma");
Arc::new(Mutex::new(conn))
}
fn create_user(conn: &Arc<Mutex<SqliteConnection>>, username: &str) -> i32 {
use crate::database::schema::users::dsl as u;
let mut c = conn.lock().unwrap();
diesel::insert_into(u::users)
.values((u::username.eq(username), u::password.eq("x")))
.execute(c.deref_mut())
.unwrap();
u::users
.filter(u::username.eq(username))
.select(u::id)
.first(c.deref_mut())
.unwrap()
}
fn create_persona_row(conn: &Arc<Mutex<SqliteConnection>>, uid: i32, pid: &str) {
use crate::database::schema::personas::dsl as p;
let mut c = conn.lock().unwrap();
diesel::insert_into(p::personas)
.values(InsertPersona {
user_id: uid,
persona_id: pid,
name: pid,
system_prompt: "test prompt",
is_built_in: false,
include_all_memories: false,
created_at: 0,
updated_at: 0,
})
.execute(c.deref_mut())
.unwrap();
}
fn make_entity(dao: &mut SqliteKnowledgeDao, name: &str) -> Entity {
let cx = opentelemetry::Context::new();
dao.upsert_entity(
&cx,
InsertEntity {
name: name.to_string(),
entity_type: "person".to_string(),
description: String::new(),
embedding: None,
confidence: 0.6,
status: "active".to_string(),
created_at: 0,
updated_at: 0,
},
)
.unwrap()
}
fn add_fact(
dao: &mut SqliteKnowledgeDao,
subject: i32,
predicate: &str,
value: &str,
user_id: i32,
persona_id: &str,
) -> EntityFact {
let cx = opentelemetry::Context::new();
let (fact, _) = dao
.upsert_fact(
&cx,
InsertEntityFact {
subject_entity_id: subject,
predicate: predicate.to_string(),
object_entity_id: None,
object_value: Some(value.to_string()),
source_photo: None,
source_insight_id: None,
confidence: 0.6,
status: "active".to_string(),
created_at: 0,
persona_id: persona_id.to_string(),
user_id,
},
)
.unwrap();
fact
}
#[test]
fn persona_filter_single_isolates_per_user() {
// Two users, same persona name. Each user's facts under that
// persona must NOT surface to the other user's reads — this is
// the multi-user leakage that motivated adding user_id.
let cx = opentelemetry::Context::new();
let conn = connection_with_fks_on();
let alice = create_user(&conn, "alice");
let bob = create_user(&conn, "bob");
create_persona_row(&conn, alice, "default");
create_persona_row(&conn, bob, "default");
let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
let entity = make_entity(&mut dao, "Cabin");
add_fact(&mut dao, entity.id, "located_in", "Vermont", alice, "default");
add_fact(&mut dao, entity.id, "color", "red", bob, "default");
let alice_view = dao
.get_facts_for_entity(
&cx,
entity.id,
&PersonaFilter::Single {
user_id: alice,
persona_id: "default".to_string(),
},
)
.unwrap();
assert_eq!(alice_view.len(), 1);
assert_eq!(alice_view[0].predicate, "located_in");
let bob_view = dao
.get_facts_for_entity(
&cx,
entity.id,
&PersonaFilter::Single {
user_id: bob,
persona_id: "default".to_string(),
},
)
.unwrap();
assert_eq!(bob_view.len(), 1);
assert_eq!(bob_view[0].predicate, "color");
}
#[test]
fn persona_filter_all_unions_across_personas_one_user() {
// include_all_memories=true → All variant: see this user's
// facts across all their personas. Must NOT include other
// users' facts even when they share a persona name.
let cx = opentelemetry::Context::new();
let conn = connection_with_fks_on();
let alice = create_user(&conn, "alice");
let bob = create_user(&conn, "bob");
create_persona_row(&conn, alice, "default");
create_persona_row(&conn, alice, "journal");
create_persona_row(&conn, bob, "default");
let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
let entity = make_entity(&mut dao, "Cabin");
add_fact(&mut dao, entity.id, "p1", "v1", alice, "default");
add_fact(&mut dao, entity.id, "p2", "v2", alice, "journal");
add_fact(&mut dao, entity.id, "p3", "v3", bob, "default");
let alice_all = dao
.get_facts_for_entity(&cx, entity.id, &PersonaFilter::All { user_id: alice })
.unwrap();
let predicates: Vec<&str> = alice_all.iter().map(|f| f.predicate.as_str()).collect();
assert_eq!(predicates.len(), 2);
assert!(predicates.contains(&"p1"));
assert!(predicates.contains(&"p2"));
assert!(
!predicates.contains(&"p3"),
"All variant must not leak across users"
);
}
#[test]
fn upsert_fact_dedup_does_not_cross_users() {
// Two users insert the SAME claim (same subject + predicate +
// object_value) under the same persona name. Pre-fix, the
// dedup key was (subject, predicate, persona_id) and bob's
// insert would corroborate alice's row instead of creating a
// new one. Post-fix the key includes user_id, so each user
// gets their own row at confidence=0.6.
let conn = connection_with_fks_on();
let alice = create_user(&conn, "alice");
let bob = create_user(&conn, "bob");
create_persona_row(&conn, alice, "default");
create_persona_row(&conn, bob, "default");
let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
let entity = make_entity(&mut dao, "Cabin");
let alice_fact = add_fact(&mut dao, entity.id, "color", "red", alice, "default");
let bob_fact = add_fact(&mut dao, entity.id, "color", "red", bob, "default");
assert_ne!(alice_fact.id, bob_fact.id, "must be separate rows");
assert_eq!(alice_fact.confidence, 0.6);
assert_eq!(
bob_fact.confidence, 0.6,
"bob's row should not have been corroboration-bumped against alice's"
);
}
#[test]
fn deleting_persona_cascades_only_that_users_facts() {
// Composite FK + CASCADE: deleting alice's 'journal' persona
// wipes alice's journal facts but leaves alice's default
// facts AND bob's journal-named facts untouched.
let cx = opentelemetry::Context::new();
let conn = connection_with_fks_on();
let alice = create_user(&conn, "alice");
let bob = create_user(&conn, "bob");
create_persona_row(&conn, alice, "default");
create_persona_row(&conn, alice, "journal");
create_persona_row(&conn, bob, "journal");
let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
let entity = make_entity(&mut dao, "Cabin");
add_fact(&mut dao, entity.id, "p_alice_default", "x", alice, "default");
add_fact(&mut dao, entity.id, "p_alice_journal", "y", alice, "journal");
add_fact(&mut dao, entity.id, "p_bob_journal", "z", bob, "journal");
// Delete alice's journal persona — CASCADE should remove only
// alice's journal facts.
{
use crate::database::schema::personas::dsl as p;
let mut c = conn.lock().unwrap();
diesel::delete(
p::personas
.filter(p::user_id.eq(alice))
.filter(p::persona_id.eq("journal")),
)
.execute(c.deref_mut())
.unwrap();
}
// alice/default survives.
let alice_default = dao
.get_facts_for_entity(
&cx,
entity.id,
&PersonaFilter::Single {
user_id: alice,
persona_id: "default".to_string(),
},
)
.unwrap();
assert_eq!(alice_default.len(), 1);
assert_eq!(alice_default[0].predicate, "p_alice_default");
// alice/journal is gone.
let alice_journal = dao
.get_facts_for_entity(
&cx,
entity.id,
&PersonaFilter::Single {
user_id: alice,
persona_id: "journal".to_string(),
},
)
.unwrap();
assert!(
alice_journal.is_empty(),
"CASCADE should have removed alice's journal facts"
);
// bob/journal — same persona name, different user — untouched.
let bob_journal = dao
.get_facts_for_entity(
&cx,
entity.id,
&PersonaFilter::Single {
user_id: bob,
persona_id: "journal".to_string(),
},
)
.unwrap();
assert_eq!(bob_journal.len(), 1);
assert_eq!(bob_journal[0].predicate, "p_bob_journal");
}
#[test]
fn fact_insert_with_unknown_persona_is_rejected() {
// FK enforcement: inserting a fact whose (user_id, persona_id)
// pair has no matching personas row should fail. Protects
// against typo'd persona ids silently leaking into the table.
let cx = opentelemetry::Context::new();
let conn = connection_with_fks_on();
let alice = create_user(&conn, "alice");
// Note: NO persona row inserted for alice + 'ghost'.
let mut dao = SqliteKnowledgeDao::from_connection(conn.clone());
let entity = make_entity(&mut dao, "Cabin");
let result = dao.upsert_fact(
&cx,
InsertEntityFact {
subject_entity_id: entity.id,
predicate: "color".to_string(),
object_entity_id: None,
object_value: Some("red".to_string()),
source_photo: None,
source_insight_id: None,
confidence: 0.6,
status: "active".to_string(),
created_at: 0,
persona_id: "ghost".to_string(),
user_id: alice,
},
);
assert!(
result.is_err(),
"FK should reject fact whose persona doesn't exist"
);
}
}