feat(workspace): layered memory with sensitivity-based privacy redirect (#1112)

* feat(workspace): layered memory with sensitivity-based privacy redirect

Introduce MemoryLayer type for named memory layers with sensitivity
levels and write permissions. Layers map to synthetic user_id values
in workspace tables, enabling shared/private memory isolation.

- Add MemoryLayer, LayerSensitivity types with default_for_user()
- Add layer-aware write methods (write_to_layer, append_to_layer)
- Add PatternPrivacyClassifier to guard shared layer writes
- Add optional 'layer' parameter to memory_write tool and HTTP API
- Add 'redirected' and 'actual_layer' fields to write response
- Add MEMORY_LAYERS env var (JSON) for layer configuration
- Workspace user_id now derived from GATEWAY_USER_ID (was hardcoded "default")
- 10 integration tests for layered memory operations

Addresses prerequisite for Issue #59 (multi-tenancy).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: add explicit default to memory_write layer schema

Add "default": "private" to the layer parameter's JSON schema so
LLM tool consumers can see the default without reading code.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* refactor: extract resolve_layer_target to deduplicate layer writes

Consolidate shared layer-lookup, writable check, and privacy
classification logic from write_to_layer and append_to_layer into a
single resolve_layer_target helper.

Flagged on #349 review — the duplication originates in this PR.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address review feedback on layered memory PR

- Fix email regex pipe bug in TLD character class (privacy.rs)
- Add append support to web memory_write handler via `append` field
- Validate MemoryLayer name/scope: reject empty, check duplicates
- Remove hardcoded 'private' default from tool schema; omit layer
  fields from output when no layer specified
- Document scope isolation risk for multi-tenant (Issue #59)

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address adversarial review findings

- CRITICAL: fix identity file protection bypass via trailing slash
  (normalize target path before protection checks)
- HIGH: check private layer is writable before privacy redirect
- HIGH: map LayerNotFound/ReadOnly to proper 4xx HTTP status codes
- HIGH: honor `append` field in non-layer HTTP write path
- MEDIUM: remove redundant DB fetch in append_to_layer (narrower
  TOCTOU window)
- MEDIUM: remove dead memory_write_handler from handlers/memory.rs

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* feat: opt-in privacy classifier, force override, confidence scoring

Address review feedback from @zmanian:

- Privacy classifier is now opt-in via with_privacy_classifier() instead
  of always-on. Default hardcoded patterns (doctor, therapy, email, phone)
  had unacceptable false positive rates in household contexts. LLM chooses
  the correct layer via system prompt; regex can't improve on that.
- Add ConfigurablePrivacyClassifier for operator-supplied patterns.
- PatternPrivacyClassifier defaults narrowed to hard PII only (SSN,
  credit card, credentials).
- Add force param to write_to_layer/append_to_layer to skip classifier.
- PrivacyClassifier trait returns SensitivityResult { is_sensitive,
  confidence } instead of bool, ready for probabilistic classifiers.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: remove redundant heartbeat match arm in memory_write

The heartbeat arm was identical to the catch-all — resolved_path
already points to paths::HEARTBEAT when target is "heartbeat".

Addresses review feedback from gemini-code-assist on #1112.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: return Result from PatternPrivacyClassifier::new()

Replace .expect() with proper error propagation per project
no-panics policy. Remove Default impl (unused in production).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* refactor: move memory_layers from GatewayConfig to WorkspaceConfig

Resolve merge conflicts between HEAD (transcription, search, env helpers)
and the workspace config branch. GatewayConfig no longer owns memory_layers;
WorkspaceConfig::resolve() handles parsing, validation (name length >64,
character set, empty scope, duplicates), and fallback defaults.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* test: strengthen privacy classifier and layer isolation coverage

Add 8 privacy classifier edge case tests (format variants, keywords,
longer documents, empty/partial inputs) and 5 layer write isolation
integration tests (cross-scope invisibility, overwrite, empty path,
sensitive-to-private no-redirect).

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* fix: tautological test assertion and add WorkspaceConfig validation tests

Replace always-true `is_ok() || is_err()` in write_empty_path_to_layer
with actual behavior assertion (write succeeds with normalized empty path).

Add 8 unit tests for WorkspaceConfig::resolve() covering valid JSON parsing,
invalid JSON, empty/long/invalid-char layer names, empty scopes, duplicates,
and default fallback behavior.

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

* style: cargo fmt after staging merge

Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>
Co-authored-by: [email protected] <[email protected]>
This commit is contained in:
standardtoaster
2026-03-20 22:15:29 -07:00
committed by GitHub
co-authored by Claude Opus 4.6 [email protected] <[email protected]>
parent 0d1a5c210b
commit 212d661e20
13 changed files with 1363 additions and 78 deletions
+178 -3
View File
@@ -45,6 +45,8 @@ mod document;
mod embedding_cache;
mod embeddings;
pub mod hygiene;
pub mod layer;
pub mod privacy;
#[cfg(feature = "postgres")]
mod repository;
mod search;
@@ -61,6 +63,17 @@ pub use search::{
FusionStrategy, RankedResult, SearchConfig, SearchResult, fuse_results, reciprocal_rank_fusion,
};
/// Result of a layer-aware write operation.
///
/// Contains the written document plus metadata about whether the write
/// was redirected to a different layer (e.g., sensitive content redirected
/// from shared to private).
pub struct WriteResult {
pub document: MemoryDocument,
pub redirected: bool,
pub actual_layer: String,
}
use std::sync::Arc;
use chrono::{NaiveDate, Utc};
@@ -344,20 +357,29 @@ pub struct Workspace {
bootstrap_completed: std::sync::atomic::AtomicBool,
/// Default search configuration applied to all queries.
search_defaults: SearchConfig,
/// Memory layers this workspace has access to.
memory_layers: Vec<crate::workspace::layer::MemoryLayer>,
/// Optional privacy classifier for shared layer writes.
/// When None, writes go exactly where requested — no silent redirect.
privacy_classifier: Option<Arc<dyn crate::workspace::privacy::PrivacyClassifier>>,
}
impl Workspace {
/// Create a new workspace backed by a PostgreSQL connection pool.
#[cfg(feature = "postgres")]
pub fn new(user_id: impl Into<String>, pool: Pool) -> Self {
let user_id_str = user_id.into();
let memory_layers = crate::workspace::layer::MemoryLayer::default_for_user(&user_id_str);
Self {
user_id: user_id.into(),
user_id: user_id_str,
agent_id: None,
storage: WorkspaceStorage::Repo(Repository::new(pool)),
embeddings: None,
bootstrap_pending: std::sync::atomic::AtomicBool::new(false),
bootstrap_completed: std::sync::atomic::AtomicBool::new(false),
search_defaults: SearchConfig::default(),
memory_layers,
privacy_classifier: None,
}
}
@@ -365,14 +387,18 @@ impl Workspace {
///
/// Use this for libSQL or any other backend that implements the Database trait.
pub fn new_with_db(user_id: impl Into<String>, db: Arc<dyn crate::db::Database>) -> Self {
let user_id_str = user_id.into();
let memory_layers = crate::workspace::layer::MemoryLayer::default_for_user(&user_id_str);
Self {
user_id: user_id.into(),
user_id: user_id_str,
agent_id: None,
storage: WorkspaceStorage::Db(db),
embeddings: None,
bootstrap_pending: std::sync::atomic::AtomicBool::new(false),
bootstrap_completed: std::sync::atomic::AtomicBool::new(false),
search_defaults: SearchConfig::default(),
memory_layers,
privacy_classifier: None,
}
}
@@ -444,6 +470,32 @@ impl Workspace {
self
}
/// Configure memory layers for this workspace.
///
/// Also updates read_user_ids to include all layer scopes.
pub fn with_memory_layers(mut self, layers: Vec<crate::workspace::layer::MemoryLayer>) -> Self {
self.memory_layers = layers;
self
}
/// Set a privacy classifier for shared layer writes.
///
/// When set, writes to shared layers are checked against the classifier
/// and redirected to the private layer if sensitive content is detected.
/// When unset (the default), writes go exactly where requested.
pub fn with_privacy_classifier(
mut self,
classifier: Arc<dyn crate::workspace::privacy::PrivacyClassifier>,
) -> Self {
self.privacy_classifier = Some(classifier);
self
}
/// Get the configured memory layers.
pub fn memory_layers(&self) -> &[crate::workspace::layer::MemoryLayer] {
&self.memory_layers
}
/// Get the user ID.
pub fn user_id(&self) -> &str {
&self.user_id
@@ -501,7 +553,9 @@ impl Workspace {
/// Append content to a file.
///
/// Creates the file if it doesn't exist.
/// Adds a newline separator between existing and new content.
/// Uses a single `\n` separator (suitable for log-style entries).
/// For semantic separation (e.g., memory entries), use `append_memory()`
/// which uses `\n\n`.
pub async fn append(&self, path: &str, content: &str) -> Result<(), WorkspaceError> {
let path = normalize_path(path);
let doc = self
@@ -526,6 +580,127 @@ impl Workspace {
Ok(())
}
/// Resolve the target scope for a layer write, optionally applying privacy guards.
///
/// Validates that the layer exists and is writable. When a privacy classifier
/// is configured on the workspace AND `force` is false, checks shared-layer
/// writes for sensitive content and redirects to the private layer.
///
/// By default no classifier is set — writes go exactly where requested.
/// This is intentional: the LLM chooses the correct layer via system prompt
/// guidance, and a regex classifier can't improve on that decision without
/// unacceptable false positive rates in household contexts (e.g., "doctor",
/// "therapy", phone numbers). Operators who want a safety net can configure
/// one via `with_privacy_classifier()`.
///
/// # Multi-tenant safety (Issue #59)
///
/// Layer scopes are currently used directly as `user_id` for DB operations.
/// In a multi-tenant deployment, an operator could configure a scope that
/// collides with another user's ID, granting write access to their data.
/// Future work should namespace or validate scopes to prevent this.
///
/// Returns `(scope, actual_layer_name, redirected)`.
fn resolve_layer_target(
&self,
layer_name: &str,
content: &str,
force: bool,
) -> Result<(String, String, bool), WorkspaceError> {
use crate::workspace::layer::{LayerSensitivity, MemoryLayer};
let layer = MemoryLayer::find(&self.memory_layers, layer_name).ok_or_else(|| {
WorkspaceError::LayerNotFound {
name: layer_name.to_string(),
}
})?;
if !layer.writable {
return Err(WorkspaceError::LayerReadOnly {
name: layer_name.to_string(),
});
}
if !force
&& layer.sensitivity == LayerSensitivity::Shared
&& let Some(ref classifier) = self.privacy_classifier
&& classifier.classify(content).is_sensitive
{
tracing::warn!(
layer = layer_name,
"Redirected sensitive content to private layer"
);
let private = MemoryLayer::private_layer(&self.memory_layers)
.ok_or(WorkspaceError::PrivacyRedirectFailed)?;
if !private.writable {
return Err(WorkspaceError::PrivacyRedirectFailed);
}
return Ok((private.scope.clone(), private.name.clone(), true));
}
Ok((layer.scope.clone(), layer_name.to_string(), false))
}
/// Write to a specific memory layer.
///
/// Checks that the layer exists and is writable. Uses the layer's scope
/// as the user_id for the database write. For shared layers, sensitive
/// content is automatically redirected to the private layer unless
/// `force` is set.
pub async fn write_to_layer(
&self,
layer_name: &str,
path: &str,
content: &str,
force: bool,
) -> Result<WriteResult, WorkspaceError> {
let (scope, actual_layer, redirected) =
self.resolve_layer_target(layer_name, content, force)?;
let path = normalize_path(path);
let doc = self
.storage
.get_or_create_document_by_path(&scope, self.agent_id, &path)
.await?;
self.storage.update_document(doc.id, content).await?;
self.reindex_document(doc.id).await?;
let document = self.storage.get_document_by_id(doc.id).await?;
Ok(WriteResult {
document,
redirected,
actual_layer,
})
}
/// Write to a layer, with append semantics.
pub async fn append_to_layer(
&self,
layer_name: &str,
path: &str,
content: &str,
force: bool,
) -> Result<WriteResult, WorkspaceError> {
let (scope, actual_layer, redirected) =
self.resolve_layer_target(layer_name, content, force)?;
let path = normalize_path(path);
let doc = self
.storage
.get_or_create_document_by_path(&scope, self.agent_id, &path)
.await?;
let new_content = if doc.content.is_empty() {
content.to_string()
} else {
format!("{}\n\n{}", doc.content, content)
};
self.storage.update_document(doc.id, &new_content).await?;
self.reindex_document(doc.id).await?;
let document = self.storage.get_document_by_id(doc.id).await?;
Ok(WriteResult {
document,
redirected,
actual_layer,
})
}
/// Check if a file exists.
pub async fn exists(&self, path: &str) -> Result<bool, WorkspaceError> {
let path = normalize_path(path);