[codex] Hosted review phase 4 metadata and provenance

docs/configuration.md

@@ -350,7 +350,12 @@ AGENTS.md files may begin with optional YAML frontmatter to control loading:
 ---
 memory_type: project
 priority: 10
-scope: project
+scope: repo
+trust: maintainer_approved
+visibility: public_review
+source: github_pr
+source_ref: OpenCoven/coven-code#123
+expires_at: 2099-12-31
 ---
 
 # My Project Notes
@@ -362,9 +367,22 @@ Frontmatter fields:
 
 | Field | Description |
 |-------|-------------|
-| `memory_type` | Informal label (currently informational only). |
+| `id` | Stable memory id used for hosted review citation, for example `mem_auth_policy`. If omitted, Coven Code derives a stable id from path and content. |
+| `memory_type` | Memory category label such as `project`, `user`, `reference`, or `feedback`. |
 | `priority` | Integer sort priority (lower numbers are prepended first within the same scope). |
-| `scope` | Informational label for documentation purposes. |
+| `scope` | Intended scope, such as `user`, `tenant`, `installation`, `repo`, `branch`, or `pr`. |
+| `trust` | Source trust. Hosted review enforces this against `hostedReview.memoryTrustThreshold`. Supported values include `system_policy`, `maintainer_approved`, `default_branch_code`, `model_inferred`, `contributor_input`, `fork_input`, and `unknown`. |
+| `visibility` | Intended review visibility: `public_review`, `private_review`, or `security_private`. Hosted public reviews exclude `security_private` memory by default. |
+| `source` | Provenance source kind, for example `manual`, `github_pr`, `github_pr_review`, or `session_memory_extraction`. |
+| `source_ref` | Source reference such as `owner/repo#123`, a commit SHA, or another non-secret audit handle. |
+| `expires_at` | Optional expiry date in `YYYY-MM-DD` format. Expired hosted memory is ignored. |
+| `created_at`, `created_by`, `session_id`, `transcript_ref`, `confidence` | Optional provenance fields for audit and review artifacts. |
+
+Local mode tolerates missing metadata for backward compatibility. Hosted review
+mode treats missing trust as `unknown`, ignores expired memory, and excludes
+memory below the configured trust threshold. Tagged hosted memory is injected
+with memory ids and provenance metadata; findings that rely on memory should
+include those ids in `memory_refs`.
 
 ### @include directives
 

docs/shared/hosted-review-phase-4-pr.md

@@ -0,0 +1,34 @@
+# Hosted Review Phase 4 PR Notes
+
+## Linked issues
+
+Fixes #105.
+Fixes #106.
+Fixes #112.
+
+## Summary
+
+- Extends AGENTS.md frontmatter with enforceable hosted metadata: stable id, trust, visibility, source, source_ref, expiry, created_at, created_by, session_id, transcript_ref, and confidence.
+- Enforces hosted memory metadata during loading: missing trust is lowest trust, expired memory is ignored, security-private memory is excluded from public hosted review, and trust must meet the configured threshold.
+- Renders hosted memory as tagged entries with stable ids, trust, visibility, source, source_ref, and session metadata so review output can cite memory refs.
+- Threads session-scoped provenance into hosted auto-extracted memory candidates.
+- Adds structured review-output parsing and validation helpers for `memory_refs` on memory-dependent findings.
+- Documents the hosted frontmatter contract and citation behavior.
+
+## Test evidence
+
+- `cargo fmt --all -- --check`
+- `cargo check --workspace`
+- `cargo clippy --workspace --all-targets -- -D warnings`
+- `cargo test -p claurst-core --lib claudemd --quiet`
+- `cargo test -p claurst-core --lib system_prompt --quiet`
+- `cargo test -p claurst-core --lib --quiet`
+- `cargo test -p claurst-query session_memory --quiet`
+- `cargo test -p claurst-commands structured_review_output --quiet`
+- `cargo test --workspace --quiet`
+
+## Risk notes
+
+- The slash `/review` command remains markdown-first; the structured review parser is available for hosted integrations that request JSON review artifacts.
+- Hosted memory without trust metadata is intentionally ignored unless policy lowers the trust threshold.
+- Provenance stores references and ids, not secret values.

src-rust/crates/commands/src/lib.rs

@@ -3382,6 +3382,39 @@ impl SlashCommand for LearnCommand {
 
 // ---- /review -------------------------------------------------------------
 
+#[derive(Debug, Clone, serde::Deserialize, PartialEq, Eq)]
+pub struct StructuredReviewOutput {
+    #[serde(default)]
+    pub findings: Vec<StructuredReviewFinding>,
+}
+
+#[derive(Debug, Clone, serde::Deserialize, PartialEq, Eq)]
+pub struct StructuredReviewFinding {
+    pub title: String,
+    #[serde(default)]
+    pub memory_refs: Vec<String>,
+    #[serde(default)]
+    pub memory_dependent: bool,
+}
+
+pub fn parse_structured_review_output(text: &str) -> Option<StructuredReviewOutput> {
+    serde_json::from_str::<StructuredReviewOutput>(text).ok()
+}
+
+pub fn validate_structured_review_memory_refs(review: &StructuredReviewOutput) -> Vec<String> {
+    review
+        .findings
+        .iter()
+        .filter(|finding| finding.memory_dependent && finding.memory_refs.is_empty())
+        .map(|finding| {
+            format!(
+                "finding '{}' is marked memory-dependent but has no memory_refs",
+                finding.title
+            )
+        })
+        .collect()
+}
+
 #[async_trait]
 impl SlashCommand for ReviewCommand {
     fn name(&self) -> &str {
@@ -3527,7 +3560,8 @@ impl SlashCommand for ReviewCommand {
              (1-3 sentences describing what changed)\n\n\
              ## Issues\n\
              (bulleted list: [CRITICAL|MAJOR|MINOR] file:line — description; \
-             omit section if none)\n\n\
+             omit section if none; if the issue depends on a loaded memory entry, \
+             include memory_refs: [\"mem_...\"] on that bullet)\n\n\
              ## Suggestions\n\
              (bulleted list of optional improvements; omit section if none)\n\n\
              ## Verdict\n\
@@ -10488,6 +10522,30 @@ mod tests {
         }
     }
 
+    #[test]
+    fn structured_review_output_parses_memory_refs() {
+        let review = parse_structured_review_output(
+            r#"{"findings":[{"title":"Auth check","memory_refs":["mem_auth"],"memory_dependent":true}]}"#,
+        )
+        .unwrap();
+
+        assert_eq!(review.findings[0].memory_refs, vec!["mem_auth"]);
+        assert!(validate_structured_review_memory_refs(&review).is_empty());
+    }
+
+    #[test]
+    fn structured_review_output_warns_when_memory_refs_missing() {
+        let review = parse_structured_review_output(
+            r#"{"findings":[{"title":"Auth check","memory_dependent":true}]}"#,
+        )
+        .unwrap();
+
+        let warnings = validate_structured_review_memory_refs(&review);
+        assert_eq!(warnings.len(), 1);
+        assert!(warnings[0].contains("Auth check"));
+        assert!(warnings[0].contains("memory_refs"));
+    }
+
     #[tokio::test]
     async fn test_learn_resolves_and_emits_skill_prompt() {
         // Resolvable by name and by alias.