llm: restore request logging lost in the agent-loop migration

The LLM-requests page has been empty since 24ee5b8: deleting
`session/handler/llm_call.rs` dropped both halves of the request log.

The kernel builds the `ModelRequest` itself and sets `log: None`, so the
`LoggingModel` decorator — wrapped once per model by `LlmManager` — wrote
every metadata row with a NULL `user_id`, while the page (and the detail
endpoint) filter on it. Nothing wrote `llm_request_payloads` at all any
more, so the payload viewer had nothing to show either.

Correlation cannot come from `LlmManager`: it builds one shared client
per model and does not know whose traffic it serves. It now comes from
the `ModelSelector`, the one component that knows both the model and the
owner: `SkaldSelector::with_log(RequestLogTarget)` wraps the model it
hands out, so metadata lands in `llm_requests` attributed to the user and
the payload lands in that user's own encrypted DB, keyed by `request_id`.
Session and frame are read off the request's `conversation`/`frame`,
which makes kernel rounds, sub-agent frames and compaction summaries all
attributed with no extra plumbing (`ModelRequest::log` stays unused).

The compactor's summariser call is attributed too, which it never was:
`try_compact`/`force_compact` now take the owner (its selector is built
per compaction, so it can carry the target).

Three tests in `llm::logging` lock this down — the owner/session/frame
columns, the error row with the provider's rejected body, and the
metadata-only path when no owner pool is available.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

CLAUDE.md

@@ -92,7 +92,7 @@ Two rules keep the boundary real, and both are enforced by the compiler:
 | `crates/skald-core/src/clarification/` | `ClarificationManager`: background-session question/answer |
 | `crates/skald-core/src/elicitation/` | `ElicitationManager` + bridge: MCP server-initiated input (`elicitation/create`), surfaced in the Inbox; secrets never logged/persisted |
 | `crates/skald-core/src/inbox.rs` | `Inbox`: unified façade for pending approvals + clarifications + elicitations (wraps ApprovalManager, ClarificationManager, ElicitationManager). The managers already emit the `*Requested`/`*Resolved` lifecycle events on the per-user bus; `ws.rs` forwards them to every connected client of that user regardless of `source`, so the web UI updates live (see `sidebar.js` row) |
-| `crates/skald-core/src/llm/` | LLM client abstraction (OpenAI-compat, Anthropic, Ollama…). OpenAI-compatible provider *types* are runtime data, not code: `providers/declared.rs` loads `providers.yaml` at boot (see Config); only non-OpenAI-compatible or bespoke providers (anthropic, ollama, openai, openrouter) stay native. **Retriability** (`llm_call.rs::is_retriable_llm_error`) keys on the real HTTP status via `llm_client::http_status` (a structured `LlmError { status }` from the client, else a `reqwest::Error` in the chain), **not** a substring of the message — a model id/token count containing "404"/"401" no longer mis-classifies; 401/403/404/422 don't retry, 400/429/5xx/network do |
+| `crates/skald-core/src/llm/` | LLM client abstraction (OpenAI-compat, Anthropic, Ollama…). OpenAI-compatible provider *types* are runtime data, not code: `providers/declared.rs` loads `providers.yaml` at boot (see Config); only non-OpenAI-compatible or bespoke providers (anthropic, ollama, openai, openrouter) stay native. **Retriability** (`Model::is_retriable`, `agent-loop`) keys on the real HTTP status carried by `ModelError { status }`, **not** a substring of the message — a model id/token count containing "404"/"401" cannot mis-classify; 401/403/404/422 don't retry, 400/429/5xx/network do. **Request logging** is the `logging.rs::LoggingModel` decorator, attached by the *caller's* `ModelSelector` (`loop_adapters/selector.rs::SkaldSelector::with_log`) — never by `LlmManager`, which builds one shared client per model and cannot know whose traffic it serves. The decorator's `RequestLogTarget` carries the owner: metadata → `llm_requests` in the registry (`user_id`, the column the UI filters on), payload bodies/headers → `llm_request_payloads` in that user's own encrypted DB, keyed by `request_id`; session + frame come from the request's own `conversation`/`frame`, so kernel rounds, sub-agent frames and compaction summaries are all attributed with no extra plumbing (`ModelRequest::log` is unused here) |
 | `crates/skald-core/src/transcribe/` | Transcription providers |
 | `crates/skald-core/src/image_generate/` | Image generation providers |
 | `crates/skald-core/src/memory/` | Agent memory tools |

crates/skald-core/src/compactor.rs

@@ -34,7 +34,6 @@ use std::sync::Arc;
 use agent_loop::compaction::{CompactionMode, should_compact};
 use agent_loop::manager::LoopManager;
 use agent_loop::model::ModelHint;
-use serde_json::json;
 use sqlx::SqlitePool;
 use tracing::{info, warn};
 
@@ -45,6 +44,7 @@ use crate::config::CompactionConfig;
 use crate::config_store::GlobalConfigManager;
 use crate::db::chat_history;
 use crate::llm::LlmManager;
+use crate::llm::logging::RequestLogTarget;
 use crate::loop_adapters::history::SqliteHistory;
 use crate::loop_adapters::selector::SkaldSelector;
 
@@ -112,7 +112,8 @@ impl ContextCompactor {
     pub async fn try_compact(
         &self,
         manager:           &Arc<LoopManager>,
-        pool:              &SqlitePool,
+        pool:              &Arc<SqlitePool>,
+        user_id:           &str,
         session_id:        i64,
         stack_id:          i64,
         last_input_tokens: u32,
@@ -136,7 +137,7 @@ impl ContextCompactor {
             "compactor: threshold exceeded, starting compaction"
         );
 
-        self.do_compact(manager, session_id, stack_id, effective_tokens).await
+        self.do_compact(manager, pool, user_id, session_id, stack_id, effective_tokens).await
     }
 
     /// Force compaction regardless of the token threshold.
@@ -146,7 +147,8 @@ impl ContextCompactor {
     pub async fn force_compact(
         &self,
         manager:      &Arc<LoopManager>,
-        pool:         &SqlitePool,
+        pool:         &Arc<SqlitePool>,
+        user_id:      &str,
         session_id:   i64,
         stack_id:     i64,
         is_ephemeral: bool,
@@ -162,7 +164,7 @@ impl ContextCompactor {
             "compactor: manual compaction triggered"
         );
 
-        self.do_compact(manager, session_id, stack_id, effective_tokens).await
+        self.do_compact(manager, pool, user_id, session_id, stack_id, effective_tokens).await
     }
 
     /// Runs the library's compaction on the frame with Skald's model policy,
@@ -171,9 +173,12 @@ impl ContextCompactor {
     /// Model: the instance-wide Settings pick (`compaction_model`) wins; empty,
     /// unset, or naming a model that no longer exists all degrade to AUTO
     /// selection by `compaction.strength` from config.yml.
+    #[allow(clippy::too_many_arguments)]
     async fn do_compact(
         &self,
         manager:          &Arc<LoopManager>,
+        pool:             &Arc<SqlitePool>,
+        user_id:          &str,
         session_id:       i64,
         stack_id:         i64,
         effective_tokens: u32,
@@ -184,13 +189,14 @@ impl ContextCompactor {
         let outcome = manager
             .new_compaction(conv, agent_loop::ids::FrameId(stack_id))
             .mode(CompactionMode::Auto { keep_tail: self.config.keep_recent })
-            // Strength is Skald's, captured here (D14): a pin bypasses it.
+            // Strength is Skald's, captured here (D14): a pin bypasses it. The
-            .selector(Arc::new(SkaldSelector::new(
+            // owner rides along so the summariser's call shows up in the
-                Arc::clone(&self.llm_manager),
+            // requests log like any other (session/frame come from the request).
-                self.config.strength,
+            .selector(Arc::new(
-            )))
+                SkaldSelector::new(Arc::clone(&self.llm_manager), self.config.strength)
+                    .with_log(RequestLogTarget::user(user_id, Arc::clone(pool))),
+            ))
             .model(hint)
-            .log(json!({ "session_id": session_id, "stack_id": stack_id }))
             .run()
             .await?;
 

crates/skald-core/src/db/llm_requests/mod.rs

@@ -1,7 +1,9 @@
 //! DB operations for the `llm_requests` table (metadata only).
 //!
-//! Every `chat_with_tools` call is logged here by the
+//! Every model call is logged here by the
-//! [`crate::llm::logging::LoggingModel`] decorator.
+//! [`crate::llm::logging::LoggingModel`] decorator, which the caller's
+//! `ModelSelector` attaches to the model it hands out (that is where the owner
+//! of the traffic is known — `user_id` is what the UI filters on).
 //! Payloads (request/response bodies + headers) live in `llm_request_payloads`
 //! in the owner bucket (`{userid}.db`), correlated by `request_id`.
 //! Rows are retained for `llm.request_log.retention_days` days (default 14).

crates/skald-core/src/llm/logging.rs

@@ -1,15 +1,26 @@
 //! Transparent logging decorator for any [`agent_loop::model::Model`].
 //!
-//! [`LoggingModel`] intercepts every `complete` call, measures the duration,
+//! [`LoggingModel`] intercepts every `complete` call, measures the duration and
-//! and persists a **metadata-only** row to `llm_requests` in `system.db`
+//! persists, fire-and-forget:
-//! (fire-and-forget). Per-request correlation (session/stack/user id) travels
-//! in [`ModelRequest::log`], set by the caller; the payload (request/response
-//! bodies) is returned to the caller inside [`ModelResponse::raw`] /
-//! [`ModelError::raw`] so it can be written to the user's own database.
 //!
-//! The split keeps conversation content (payloads) behind the user key while
+//! * a **metadata-only** row in `llm_requests` (`system.db`) — cost, tokens,
-//! metadata (cost, tokens, timing) stays in the admin-readable registry.
+//!   timing, plus the correlation the UI filters on (`user_id`, `session_id`,
-//! (Successor of `chatbot::logging::LoggingChatbotClient`, blueprint D13.)
+//!   `stack_id`);
+//! * the **payload** (request/response bodies + headers) in
+//!   `llm_request_payloads` in the caller's own database, keyed by the same
+//!   `request_id`.
+//!
+//! The split keeps conversation content behind the user key while metadata
+//! stays in the admin-readable registry (§5.1).
+//!
+//! **Correlation is the decorator's, not the request's.** The owner
+//! ([`RequestLogTarget`]) is captured when the model is handed out — the
+//! `ModelSelector` builds one decorator per selection, and it is the only place
+//! in the process that knows *whose* traffic this is. Session and frame come
+//! from the request itself (`conversation` / `frame`), so every caller — a
+//! round of the kernel loop, a sub-agent frame, a compaction summary — is
+//! attributed with no extra plumbing. `ModelRequest::log` is therefore unused
+//! by this host.
 
 use std::sync::Arc;
 use std::time::Instant;
@@ -19,19 +30,63 @@ use sqlx::SqlitePool;
 use tokio::sync::mpsc;
 use tracing::warn;
 
-use agent_loop::model::{Model, ModelError, ModelRequest, ModelResponse, StreamDelta};
+use agent_loop::model::{Model, ModelError, ModelRequest, ModelResponse, RawMeta, StreamDelta};
 
-use crate::db::llm_requests;
+use crate::db::{llm_request_payloads, llm_requests};
+use crate::loop_adapters::history::SqliteHistory;
+
+/// Who the logged traffic belongs to.
+#[derive(Clone, Default)]
+pub struct RequestLogTarget {
+    /// Owner of the call — the `user_id` column of the metadata row, and what
+    /// the LLM-requests page filters on.
+    pub user_id:  Option<String>,
+    /// The owner's own (SQLCipher) pool: destination of the payload rows.
+    /// `None` disables payload logging, keeping metadata only.
+    pub payloads: Option<Arc<SqlitePool>>,
+}
+
+impl RequestLogTarget {
+    /// A user's traffic: metadata attributed to them, payloads in their pool.
+    pub fn user(user_id: impl Into<String>, pool: Arc<SqlitePool>) -> Self {
+        Self { user_id: Some(user_id.into()), payloads: Some(pool) }
+    }
+}
 
 pub struct LoggingModel {
     inner:      Arc<dyn Model>,
-    pool:       Arc<SqlitePool>,
+    /// `system.db` — the registry the metadata row lands in.
+    registry:   Arc<SqlitePool>,
     model_name: String,
+    target:     RequestLogTarget,
 }
 
 impl LoggingModel {
-    pub fn new(inner: Arc<dyn Model>, pool: Arc<SqlitePool>, model_name: impl Into<String>) -> Self {
+    pub fn new(
-        Self { inner, pool, model_name: model_name.into() }
+        inner:      Arc<dyn Model>,
+        registry:   Arc<SqlitePool>,
+        model_name: impl Into<String>,
+        target:     RequestLogTarget,
+    ) -> Self {
+        Self { inner, registry, model_name: model_name.into(), target }
+    }
+
+    /// Persists the request/response bodies in the owner's own database.
+    /// Fire-and-forget: a failed write must never break the turn.
+    fn spawn_payload(&self, request_id: &str, raw: &RawMeta) {
+        let Some(pool) = self.target.payloads.clone() else { return };
+        let row = llm_request_payloads::PayloadRow {
+            request_id:       request_id.to_string(),
+            request_json:     raw.request_body.as_ref().map(|v| v.to_string()).unwrap_or_default(),
+            request_headers:  raw.request_headers.as_ref().map(|v| v.to_string()),
+            response_json:    raw.response_body.as_ref().map(|v| v.to_string()),
+            response_headers: raw.response_headers.as_ref().map(|v| v.to_string()),
+        };
+        tokio::spawn(async move {
+            if let Err(e) = llm_request_payloads::insert(&pool, row).await {
+                warn!(error = %e, "llm_request_payloads: failed to insert");
+            }
+        });
     }
 }
 
@@ -46,14 +101,13 @@ impl Model for LoggingModel {
         let result = self.inner.complete(req, deltas).await;
         let duration_ms = start.elapsed().as_millis() as i64;
 
-        // Per-request correlation set by the caller (llm_call / compactor).
+        // Correlation: the owner is ours, the conversation/frame are the call's.
-        let log        = req.log.clone().unwrap_or_default();
+        let session_id = SqliteHistory::session_id(&req.conversation).ok();
-        let session_id = log["session_id"].as_i64();
+        let stack_id   = Some(req.frame.0);
-        let stack_id   = log["stack_id"].as_i64();
+        let user_id    = self.target.user_id.clone();
-        let user_id    = log["user_id"].as_str().map(str::to_string);
         let request_id = Some(req.request_id.clone());
         let model_name = self.model_name.clone();
-        let pool       = Arc::clone(&self.pool);
+        let pool       = Arc::clone(&self.registry);
 
         match &result {
             Ok(resp) => {
@@ -64,6 +118,9 @@ impl Model for LoggingModel {
                     usage.cache_read.map(|n| n as i64),
                     usage.cache_write.map(|n| n as i64),
                 );
+                if let Some(raw) = resp.raw() {
+                    self.spawn_payload(&req.request_id, raw);
+                }
                 tokio::spawn(async move {
                     if let Err(e) = llm_requests::insert(&pool, llm_requests::LlmRequestRow {
                         request_id,
@@ -83,6 +140,11 @@ impl Model for LoggingModel {
                 });
             }
             Err(e) => {
+                // Only an HTTP failure carries a body (a provider 400 is exactly
+                // what the debug page is for); network/parse errors carry none.
+                if let Some(raw) = e.raw.as_ref() {
+                    self.spawn_payload(&req.request_id, raw);
+                }
                 let error_text = e.to_string();
                 tokio::spawn(async move {
                     if let Err(log_err) = llm_requests::insert(&pool, llm_requests::LlmRequestRow {
@@ -111,3 +173,166 @@ impl Model for LoggingModel {
         self.inner.is_retriable(err)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use agent_loop::ids::{ConversationId, FrameId};
+    use agent_loop::model::Usage;
+    use agent_loop::testing::{FakeModel, Step};
+    use serde_json::{Value, json};
+
+    fn temp_db_path(tag: &str) -> String {
+        let mut p = std::env::temp_dir();
+        let nanos = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH).unwrap().as_nanos();
+        p.push(format!("skald-test-{tag}-{}-{nanos}.db", std::process::id()));
+        p.to_string_lossy().into_owned()
+    }
+
+    fn cleanup(path: &str) {
+        for suffix in ["", "-wal", "-shm"] {
+            let _ = std::fs::remove_file(format!("{path}{suffix}"));
+        }
+    }
+
+    fn raw() -> RawMeta {
+        RawMeta {
+            request_headers:  Some(json!({ "authorization": "REDACTED" })),
+            request_body:     Some(json!({ "model": "m", "messages": [] })),
+            response_headers: Some(json!({ "content-type": "application/json" })),
+            response_body:    Some(json!({ "choices": [] })),
+        }
+    }
+
+    fn request(request_id: &str, session_id: i64, stack_id: i64) -> ModelRequest {
+        ModelRequest {
+            messages:     Vec::new(),
+            tools:        Vec::new(),
+            model:        "m".into(),
+            max_tokens:   None,
+            temperature:  None,
+            request_id:   request_id.into(),
+            conversation: ConversationId::new(format!("session:{session_id}")),
+            frame:        FrameId(stack_id),
+            extras:       Value::Null,
+            log:          None,
+        }
+    }
+
+    /// The rows are written fire-and-forget from a spawned task.
+    async fn wait_for(pool: &SqlitePool, sql: &'static str) -> i64 {
+        for _ in 0..100 {
+            let n = sqlx::query_scalar::<_, i64>(sql).fetch_one(pool).await.unwrap();
+            if n > 0 {
+                return n;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+        }
+        panic!("no row appeared for: {sql}");
+    }
+
+    /// The regression that made the LLM-requests page empty after the
+    /// agent-loop migration: a row written with no `user_id` (the page filters
+    /// on it) and no payload.
+    #[tokio::test]
+    async fn logs_metadata_with_owner_and_payload() {
+        let path = temp_db_path("llmlog-ok");
+        let pool = Arc::new(crate::db::init_system_pool(&path).await.unwrap());
+
+        let mut resp = ModelResponse::message("hi");
+        *resp.usage_mut() = Usage {
+            input_tokens:  Some(11),
+            output_tokens: Some(7),
+            ..Usage::default()
+        };
+        let ModelResponse::Message { content, reasoning, usage, .. } = resp else { unreachable!() };
+        let scripted = ModelResponse::Message { content, reasoning, usage, raw: Some(raw()) };
+
+        let inner = Arc::new(FakeModel::new("m", vec![Step { result: Ok(scripted), deltas: Vec::new(), pending: false }]));
+        let model = LoggingModel::new(
+            inner,
+            Arc::clone(&pool),
+            "gpt-test",
+            RequestLogTarget::user("u-1", Arc::clone(&pool)),
+        );
+
+        model.complete(&request("req-1", 42, 7), None).await.unwrap();
+
+        wait_for(&pool, "SELECT COUNT(*) FROM llm_requests").await;
+        let (user_id, session_id, stack_id, model_name, input, output): (Option<String>, Option<i64>, Option<i64>, String, Option<i64>, Option<i64>) =
+            sqlx::query_as("SELECT user_id, session_id, stack_id, model_name, input_tokens, output_tokens
+                            FROM llm_requests WHERE request_id = 'req-1'")
+                .fetch_one(&*pool).await.unwrap();
+        assert_eq!(user_id.as_deref(), Some("u-1"), "the page filters on user_id");
+        assert_eq!(session_id, Some(42));
+        assert_eq!(stack_id, Some(7));
+        assert_eq!(model_name, "gpt-test");
+        assert_eq!((input, output), (Some(11), Some(7)));
+
+        wait_for(&pool, "SELECT COUNT(*) FROM llm_request_payloads").await;
+        let body: String = sqlx::query_scalar(
+            "SELECT request_json FROM llm_request_payloads WHERE request_id = 'req-1'")
+            .fetch_one(&*pool).await.unwrap();
+        assert!(body.contains("\"messages\""), "payload not persisted: {body}");
+
+        pool.close().await;
+        cleanup(&path);
+    }
+
+    /// A failed call is logged too, with the provider's rejected body attached.
+    #[tokio::test]
+    async fn logs_error_row_and_error_payload() {
+        let path = temp_db_path("llmlog-err");
+        let pool = Arc::new(crate::db::init_system_pool(&path).await.unwrap());
+
+        let err = ModelError::new(Some(400), "bad request").with_raw(raw());
+        let inner = Arc::new(FakeModel::new("m", vec![Step { result: Err(err), deltas: Vec::new(), pending: false }]));
+        let model = LoggingModel::new(
+            inner,
+            Arc::clone(&pool),
+            "gpt-test",
+            RequestLogTarget::user("u-2", Arc::clone(&pool)),
+        );
+
+        assert!(model.complete(&request("req-2", 5, 9), None).await.is_err());
+
+        wait_for(&pool, "SELECT COUNT(*) FROM llm_requests").await;
+        let (user_id, error_text): (Option<String>, Option<String>) =
+            sqlx::query_as("SELECT user_id, error_text FROM llm_requests WHERE request_id = 'req-2'")
+                .fetch_one(&*pool).await.unwrap();
+        assert_eq!(user_id.as_deref(), Some("u-2"));
+        assert!(error_text.unwrap_or_default().contains("bad request"));
+
+        wait_for(&pool, "SELECT COUNT(*) FROM llm_request_payloads").await;
+
+        pool.close().await;
+        cleanup(&path);
+    }
+
+    /// No target pool ⇒ metadata only (payloads are the owner's, and an owner
+    /// with a locked database must not silently lose the metadata row).
+    #[tokio::test]
+    async fn without_payload_pool_only_metadata_is_written() {
+        let path = temp_db_path("llmlog-meta");
+        let pool = Arc::new(crate::db::init_system_pool(&path).await.unwrap());
+
+        let inner = Arc::new(FakeModel::new("m", vec![Step::message("hi")]));
+        let model = LoggingModel::new(
+            inner,
+            Arc::clone(&pool),
+            "gpt-test",
+            RequestLogTarget { user_id: Some("u-3".into()), payloads: None },
+        );
+
+        model.complete(&request("req-3", 1, 2), None).await.unwrap();
+
+        wait_for(&pool, "SELECT COUNT(*) FROM llm_requests").await;
+        let payloads: i64 = sqlx::query_scalar("SELECT COUNT(*) FROM llm_request_payloads")
+            .fetch_one(&*pool).await.unwrap();
+        assert_eq!(payloads, 0);
+
+        pool.close().await;
+        cleanup(&path);
+    }
+}

crates/skald-core/src/llm/manager.rs

@@ -8,11 +8,9 @@ use sqlx::SqlitePool;
 use tokio::sync::RwLock;
 use tracing::{info, warn};
 
-use agent_loop::model::Model;
 use core_api::provider::LlmStrength;
 use crate::provider::{ApiProvider, ProviderRegistry, ReasoningMode};
 
-use super::logging::LoggingModel;
 use super::providers::RemoteLlmModelInfo;
 use super::{ClientStatus, LlmEntry, LlmModelInfo, LlmModelRecord, LlmProviderInfo, LlmProviderRecord};
 use super::db;
@@ -69,7 +67,9 @@ pub struct LlmManager {
     catalog:            RwLock<HashMap<i64, CachedCatalog>>,
     /// Per-model metadata cache, keyed by model display name. TTL = 1h.
     model_meta_cache:   RwLock<HashMap<String, CachedModelMeta>>,
-    /// When `true`, every LLM entry is wrapped with [`LoggingChatbotClient`].
+    /// `llm.requests_log.enabled` — when `true`, a selected model is wrapped
+    /// with [`crate::llm::logging::LoggingModel`] by the caller's selector
+    /// (which is what knows *whose* traffic it is). See [`Self::log_pool`].
     log_enabled: bool,
 }
 
@@ -172,6 +172,13 @@ impl LlmManager {
         provider.llm_model_info(&record, model_id).await.ok().flatten()
     }
 
+    /// The registry pool the request log lands in, or `None` when logging is
+    /// disabled (`llm.requests_log.enabled: false`). A selector wraps the model
+    /// it hands out only when this is `Some`.
+    pub fn log_pool(&self) -> Option<Arc<SqlitePool>> {
+        self.log_enabled.then(|| Arc::clone(&self.pool))
+    }
+
     pub async fn get(&self, name: &str) -> Option<Arc<LlmEntry>> {
         self.state.read().await.models.get(name).map(|s| s.entry.clone())
     }
@@ -457,9 +464,7 @@ impl LlmManager {
                 }
             };
 
-            let log_pool = self.log_enabled.then(|| Arc::clone(&self.pool));
+            let entry = match build_entry(&self.registry, &provider, &model, model.id) {
-
-            let entry = match build_entry(&self.registry, &provider, &model, model.id, log_pool) {
                 Ok(e)  => Arc::new(e),
                 Err(e) => {
                     warn!(model = %model.name, error = %e, "failed to build LLM entry, skipping");
@@ -501,22 +506,18 @@ fn build_entry(
     provider:   &LlmProviderRecord,
     model:      &LlmModelRecord,
     model_db_id: i64,
-    log_pool:   Option<Arc<SqlitePool>>,
 ) -> Result<LlmEntry> {
     let built = registry.get(&provider.provider)
         .ok_or_else(|| anyhow::anyhow!("unknown provider type '{}'", provider.provider))?
         .build_llm(provider, model)
         .ok_or_else(|| anyhow::anyhow!("provider '{}' does not support LLM", provider.provider))??;
 
-    let inner        = built.client;
+    // The bare client: request logging is a per-caller decorator applied by the
+    // `ModelSelector` (it is the only place that knows the owner), not here.
+    let client       = built.client;
     let prompt_cache = built.prompt_cache;
     let extra        = model.extra_params.clone();
 
-    let client: Arc<dyn Model> = match log_pool {
-        Some(pool) => Arc::new(LoggingModel::new(inner, pool, &model.name)),
-        None       => inner,
-    };
-
     Ok(LlmEntry {
         client,
         model:          model.model_id.clone(),

crates/skald-core/src/loop_adapters/catalog.rs

@@ -24,6 +24,7 @@ use sqlx::SqlitePool;
 use crate::approval::ApprovalManager;
 use crate::clarification::ClarificationManager;
 use crate::llm::LlmManager;
+use crate::llm::logging::RequestLogTarget;
 use crate::loop_adapters::activation::SkaldToolActivator;
 use crate::loop_adapters::builtins::{SkaldAskUserTool, SkaldHumanChannel};
 use crate::loop_adapters::history::SqliteHistory;
@@ -109,8 +110,12 @@ impl AgentCatalog for SkaldAgentCatalog {
         let meta = crate::agents::load_task_meta(id).map_err(|e| anyhow::anyhow!("{e}"))?;
 
         // The child's own strength drives its selector (D14) — never the
-        // parent's resolved client.
+        // parent's resolved client. Its traffic is logged under the same owner
-        let selector = Arc::new(SkaldSelector::new(self.llm_manager.clone(), meta.strength));
+        // (the child's frame id already distinguishes it in the log).
+        let selector = Arc::new(
+            SkaldSelector::new(self.llm_manager.clone(), meta.strength)
+                .with_log(RequestLogTarget::user(self.user_id.clone(), self.pool.clone())),
+        );
         let model = meta.client.as_deref().map(ModelHint::name);
 
         // The child's system context: its own prompt, no per-turn extras.

crates/skald-core/src/loop_adapters/runtime.rs

@@ -30,6 +30,7 @@ use crate::approval::ApprovalManager;
 use crate::clarification::ClarificationManager;
 use crate::config::DatetimeConfig;
 use crate::llm::LlmManager;
+use crate::llm::logging::RequestLogTarget;
 use crate::loop_adapters::activation::{SkaldActivationSource, SkaldToolActivator};
 use crate::loop_adapters::async_task::CronExecutor;
 use crate::loop_adapters::builtins::{
@@ -132,9 +133,12 @@ impl UserLoopRuntime {
         }));
 
         // The default selector has no strength requirement; every turn overrides
-        // it with the agent's own (D14).
+        // it with the agent's own (D14). It carries the owner's log target, so a
-        let default_selector: Arc<dyn ModelSelector> =
+        // call served by it (recovery, compaction) is still attributed.
-            Arc::new(SkaldSelector::new(llm_manager.clone(), None));
+        let default_selector: Arc<dyn ModelSelector> = Arc::new(
+            SkaldSelector::new(llm_manager.clone(), None)
+                .with_log(RequestLogTarget::user(user_id.clone(), pool.clone())),
+        );
 
         let manager = Arc::new(
             LoopManager::builder()
@@ -208,6 +212,12 @@ impl UserLoopRuntime {
         &self.store
     }
 
+    /// Where this user's LLM traffic is logged: metadata in the registry
+    /// (attributed to them), payloads in their own encrypted pool.
+    pub fn log_target(&self) -> RequestLogTarget {
+        RequestLogTarget::user(self.user_id.clone(), self.pool.clone())
+    }
+
     /// The conversation id of a session — the store's encoding.
     pub fn conversation(session_id: i64) -> ConversationId {
         SqliteHistory::conversation(session_id)
@@ -259,10 +269,11 @@ impl UserLoopRuntime {
         extensions.insert(Arc::new(CallerUserId(self.user_id.clone())));
         extensions.insert(scope.clone());
 
-        // ── Selector: this agent's strength (D14) ──
+        // ── Selector: this agent's strength (D14) + the owner's request log ──
         let strength = crate::agents::load_meta(&frame_agent).ok().and_then(|m| m.strength);
-        let selector: Arc<dyn ModelSelector> =
+        let selector: Arc<dyn ModelSelector> = Arc::new(
-            Arc::new(SkaldSelector::new(self.llm_manager.clone(), strength));
+            SkaldSelector::new(self.llm_manager.clone(), strength).with_log(self.log_target()),
+        );
 
         // The session's root frame; the store reuses the provisioned row.
         let frame = self

crates/skald-core/src/loop_adapters/selector.rs

@@ -8,10 +8,11 @@ use std::sync::Arc;
 
 use agent_loop::activation::ToolRendering;
 use agent_loop::async_trait;
-use agent_loop::model::{ModelHandle, ModelHint, ModelInfo, ModelSelector};
+use agent_loop::model::{Model, ModelHandle, ModelHint, ModelInfo, ModelSelector};
 use agent_loop::ids::ModelId;
 use serde_json::Value;
 
+use crate::llm::logging::{LoggingModel, RequestLogTarget};
 use crate::llm::{DtlMode, LlmEntry, LlmManager, LlmStrength};
 
 /// Maps Skald's per-model DTL mode to the crate's wire protocol (D15).
@@ -37,14 +38,42 @@ pub fn model_info_of(entry: &LlmEntry) -> ModelInfo {
 
 /// The selector handed to the loop manager for one turn: the manager's
 /// strength tiering + health + priority, behind the crate's seam.
+///
+/// It is also where **request logging** is attached: the selector is the only
+/// component that knows both the model and the owner of the call, so it wraps
+/// the model it hands out in a [`LoggingModel`] bound to that owner (see
+/// [`Self::with_log`]). Without it the metadata row would land with a NULL
+/// `user_id` and no payload — invisible in the LLM-requests page.
 pub struct SkaldSelector {
     manager:  Arc<LlmManager>,
     strength: Option<LlmStrength>,
+    log:      Option<RequestLogTarget>,
 }
 
 impl SkaldSelector {
     pub fn new(manager: Arc<LlmManager>, strength: Option<LlmStrength>) -> Self {
-        Self { manager, strength }
+        Self { manager, strength, log: None }
+    }
+
+    /// Attributes every call served by this selector to `target`. Honoured only
+    /// when instance-wide request logging is on (`llm.requests_log.enabled`).
+    pub fn with_log(mut self, target: RequestLogTarget) -> Self {
+        self.log = Some(target);
+        self
+    }
+
+    /// Wraps a resolved client in the logging decorator when both a target and
+    /// the registry pool are available; otherwise hands the bare client over.
+    fn instrument(&self, name: &str, entry: &LlmEntry) -> Arc<dyn Model> {
+        match (&self.log, self.manager.log_pool()) {
+            (Some(target), Some(registry)) => Arc::new(LoggingModel::new(
+                entry.client.clone(),
+                registry,
+                name,
+                target.clone(),
+            )),
+            _ => entry.client.clone(),
+        }
     }
 }
 
@@ -60,9 +89,10 @@ impl ModelSelector for SkaldSelector {
             let excluded: Vec<&str> = exclude.iter().map(String::as_str).collect();
             self.manager.select_excluding(&excluded, self.strength).await?
         };
+        let model = self.instrument(&name, &entry);
         Ok(ModelHandle {
             id:    name,
-            model: entry.client.clone(),
+            model,
             info:  model_info_of(&entry),
         })
     }

crates/skald-core/src/session/handler/mod.rs

@@ -452,7 +452,8 @@ impl ChatSessionHandler {
         match self.compactor {
             Some(ref compactor) => {
                 compactor.force_compact(
-                    self.loop_runtime.manager(), pool, self.session_id, stack.id, self.is_ephemeral,
+                    self.loop_runtime.manager(), pool, &self.user_id,
+                    self.session_id, stack.id, self.is_ephemeral,
                 ).await
             }
             None => Ok(false),
@@ -555,7 +556,8 @@ impl ChatSessionHandler {
         if let Some(ref compactor) = self.compactor {
             let last_tokens = self.last_input_tokens.load(Ordering::Relaxed);
             match compactor.try_compact(
-                self.loop_runtime.manager(), pool, self.session_id, stack.id, last_tokens, self.is_ephemeral,
+                self.loop_runtime.manager(), pool, &self.user_id,
+                self.session_id, stack.id, last_tokens, self.is_ephemeral,
             ).await {
                 Ok(true)  => info!(session_id = self.session_id, stack_id = stack.id, "handle_message: context compacted"),
                 Ok(false) => {}