crates/core-api/src/provider.rs

@@ -183,6 +183,15 @@ pub trait ApiProvider: Send + Sync {
         None
     }
 
+    /// The dynamic-tool-loading (DTL) serialization format this provider's models
+    /// speak, e.g. `"anthropic_tool_reference"` or `"kimi_system_tools"`. Applied
+    /// only to a model that opts in via the `tool_search` capability. `None` = no
+    /// DTL (activated tools ride in the top-level `tools`). Returned as a string so
+    /// core-api needs no dependency on the engine's `DtlMode` — the caller parses it.
+    fn dtl_format(&self) -> Option<&str> {
+        None
+    }
+
     async fn llm_model_info(
         &self,
         _record:   &LlmProviderRecord,

crates/llm-client/src/anthropic.rs

@@ -66,19 +66,36 @@ impl AnthropicClient {
     /// Converts OpenAI-format tool definitions to Anthropic format.
     /// OpenAI: { "type": "function", "function": { "name", "description", "parameters" } }
     /// Anthropic: { "name", "description", "input_schema" }
+    ///
+    /// DTL (tool search): a top-level `defer_loading: true` on the OpenAI tool
+    /// object is carried through to Anthropic's native `defer_loading` field. When
+    /// any tool is deferred, the cache breakpoint is placed on the last
+    /// **non-deferred** tool — a deferred tool cannot also carry `cache_control`
+    /// (the API 400s), and at least one tool must stay non-deferred anyway.
     fn convert_tools(tools: &[Value]) -> Vec<Value> {
-        tools
+        let has_deferred = tools.iter().any(|t| t["defer_loading"].as_bool() == Some(true));
+        let mut out: Vec<Value> = tools
             .iter()
             .filter_map(|t| {
                 let func = &t["function"];
                 let name = func["name"].as_str()?;
-                Some(json!({
+                let mut tool = json!({
                     "name":         name,
                     "description":  func["description"].as_str().unwrap_or(""),
                     "input_schema": func["parameters"],
-                }))
+                });
+                if t["defer_loading"].as_bool() == Some(true) {
+                    tool["defer_loading"] = json!(true);
+                }
+                Some(tool)
             })
-            .collect()
+            .collect();
+        if has_deferred {
+            if let Some(t) = out.iter_mut().rev().find(|t| t["defer_loading"].as_bool() != Some(true)) {
+                t["cache_control"] = json!({ "type": "ephemeral" });
+            }
+        }
+        out
     }
 
     /// Converts OpenAI-format message array to Anthropic format.
@@ -145,10 +162,25 @@ impl AnthropicClient {
                     let mut results: Vec<Value> = Vec::new();
                     while i < messages.len() && messages[i]["role"].as_str() == Some("tool") {
                         let tm = &messages[i];
+                        // DTL (custom tool search): a tool result carrying
+                        // `_tool_references` (set by the message builder on an
+                        // `activate_tools` result in AnthropicToolReference mode) becomes a
+                        // `content` array of `tool_reference` blocks, which the API expands
+                        // into the deferred tools' full definitions. Empty/absent → the
+                        // normal text result.
+                        let content: Value = match tm["_tool_references"].as_array() {
+                            Some(refs) if !refs.is_empty() => Value::Array(
+                                refs.iter()
+                                    .filter_map(|r| r.as_str())
+                                    .map(|name| json!({ "type": "tool_reference", "tool_name": name }))
+                                    .collect(),
+                            ),
+                            _ => Value::String(tm["content"].as_str().unwrap_or("").to_string()),
+                        };
                         results.push(json!({
                             "type":        "tool_result",
                             "tool_use_id": tm["tool_call_id"].as_str().unwrap_or(""),
-                            "content":     tm["content"].as_str().unwrap_or(""),
+                            "content":     content,
                         }));
                         i += 1;
                     }
@@ -164,7 +196,7 @@ impl AnthropicClient {
 
     /// Assembles the `/v1/messages` request body shared by the buffered and the
     /// streaming path (the caller adds `stream` on top).
-    fn tools_body(&self, system: Option<String>, messages: Vec<Value>, tools: Vec<Value>, options: &ChatOptions) -> Value {
+    fn tools_body(&self, system: Option<Value>, messages: Vec<Value>, tools: Vec<Value>, options: &ChatOptions) -> Value {
         let max_tokens = options.max_tokens.unwrap_or(4096);
         let mut body = json!({
             "model":      options.model,
@@ -173,22 +205,47 @@ impl AnthropicClient {
             "tools":      tools,
         });
 
-        if let Some(sys) = system              { body["system"]      = sys.into(); }
+        if let Some(sys) = system              { body["system"]      = sys; }
         if let Some(t)   = options.temperature { body["temperature"] = t.into(); }
         self.apply_extra(&mut body);
         body
     }
 
-    /// Collects ALL system-role messages (main prompt, mid-conversation
+    /// Collects ALL system-role messages (main prompt, mid-conversation summary,
-    /// summary, tail_reminder) into a single `system:` string. The Anthropic
+    /// tail_reminder) into the single `system` parameter the Anthropic API accepts.
-    /// API only accepts a single system parameter.
+    ///
-    fn merged_system(messages: &[Value]) -> Option<String> {
+    /// Returns a plain string in the common case. When any system message carries
-        let parts: Vec<&str> = messages
+    /// **structured** content (a text-block array, e.g. the static prompt tagged
+    /// with `cache_control` when prompt caching is on), it returns the array form
+    /// instead so the cache breakpoint survives into `system`. String-content
+    /// messages become plain text blocks (no cache_control).
+    fn merged_system(messages: &[Value]) -> Option<Value> {
+        let sys: Vec<&Value> = messages
             .iter()
             .filter(|m| m["role"].as_str() == Some("system"))
-            .filter_map(|m| m["content"].as_str())
             .collect();
-        if parts.is_empty() { None } else { Some(parts.join("\n\n---\n\n")) }
+        if sys.is_empty() { return None; }
+
+        if !sys.iter().any(|m| m["content"].is_array()) {
+            let parts: Vec<&str> = sys.iter().filter_map(|m| m["content"].as_str()).collect();
+            return if parts.is_empty() { None } else { Some(Value::String(parts.join("\n\n---\n\n"))) };
+        }
+
+        let mut blocks: Vec<Value> = Vec::new();
+        for m in &sys {
+            match &m["content"] {
+                Value::String(s) if !s.is_empty() => blocks.push(json!({ "type": "text", "text": s })),
+                Value::Array(arr) => {
+                    for b in arr {
+                        if b["type"].as_str() == Some("text") {
+                            blocks.push(b.clone());
+                        }
+                    }
+                }
+                _ => {}
+            }
+        }
+        if blocks.is_empty() { None } else { Some(Value::Array(blocks)) }
     }
 
     fn url(&self) -> String {

crates/skald-core/src/chat_hub/mod.rs

@@ -452,7 +452,7 @@ impl ChatHub {
     /// The next LLM turn will start with no MCP servers activated.
     pub async fn reset_mcp(&self, source_id: &str) -> anyhow::Result<()> {
         let session_id = self.get_or_create_session(source_id, &self.default_agent).await?;
-        crate::db::session_mcp_grants::revoke_all(&self.db, session_id).await?;
+        crate::db::activated_tools::revoke_all_session(&self.db, session_id).await?;
         info!(source_id, session_id, "ChatHub: MCP grants reset");
         Ok(())
     }

crates/skald-core/src/compactor.rs

@@ -338,6 +338,17 @@ impl ContextCompactor {
 
         let summary_id = chat_summaries::save(pool, stack_id, &summary_text, last_covered_id).await?;
 
+        // DTL: activations pinned to a message that was just compacted away would
+        // otherwise lose their render position (the Kimi `system`+`tools` block).
+        // Re-anchor them onto the first surviving message. Best-effort — a failure
+        // only means the model may re-activate a tool after compaction.
+        let first_surviving_id = messages[split].id;
+        if let Err(e) = crate::db::activated_tools::reanchor_compacted(
+            pool, stack_id, last_covered_id, first_surviving_id,
+        ).await {
+            warn!(stack_id, error = %e, "compactor: failed to re-anchor DTL activations");
+        }
+
         info!(
             stack_id,
             summary_id,

crates/skald-core/src/db/activated_tools.rs

@@ -0,0 +1,171 @@
+//! Persisted tool-group activations (the effect of `activate_tools`).
+//!
+//! One row per activated group, anchored at the assistant `message_id` that
+//! triggered it. Replaces the old `session_mcp_grants` / `stack_mcp_grants`
+//! pair: `stack_id IS NULL` is a session-scoped activation (root agent), a
+//! non-NULL `stack_id` is a sub-agent-frame activation (deleted on frame exit).
+//!
+//! The activation is the durable **effect**; the tool *call* itself lives in
+//! `chat_llm_tools`. Keeping them separate means "which groups are active" is a
+//! direct query, not a parse of `activate_tools` call arguments.
+//!
+//! `kind`/`ref` normalise the activated group: `('builtin', 'config')` for the
+//! reserved built-in group, `('mcp', <server name>)` for an MCP server.
+
+use anyhow::Result;
+use sqlx::SqlitePool;
+
+/// One activation row, anchored at the message that triggered it.
+#[derive(Debug, Clone)]
+pub struct Activation {
+    /// The assistant `chat_history.id` whose tool call triggered this activation.
+    pub message_id: i64,
+    /// `'builtin'` (the reserved `config` group) or `'mcp'` (a server).
+    pub kind: String,
+    /// The group reference: `'config'`, or the MCP server name.
+    pub ref_: String,
+}
+
+/// Persist a tool-group activation. `stack_id = None` → session-scoped (root
+/// agent); `Some(id)` → stack-scoped (sub-agent frame). Idempotent via the
+/// `COALESCE(stack_id, -1)`-based unique index (INSERT OR IGNORE).
+pub async fn grant(
+    pool: &SqlitePool,
+    session_id: i64,
+    stack_id: Option<i64>,
+    message_id: i64,
+    kind: &str,
+    ref_: &str,
+) -> Result<()> {
+    sqlx::query(
+        "INSERT OR IGNORE INTO activated_tools (session_id, stack_id, message_id, kind, ref)
+         VALUES (?, ?, ?, ?, ?)",
+    )
+    .bind(session_id)
+    .bind(stack_id)
+    .bind(message_id)
+    .bind(kind)
+    .bind(ref_)
+    .execute(pool)
+    .await?;
+    Ok(())
+}
+
+/// Session-scoped activated group refs (root agent). The in-memory grant set is
+/// seeded from this at config-build time. Replaces
+/// `session_mcp_grants::list_for_session`.
+pub async fn list_refs_session(pool: &SqlitePool, session_id: i64) -> Result<Vec<String>> {
+    let rows = sqlx::query_as::<_, (String,)>(
+        "SELECT ref FROM activated_tools
+         WHERE session_id = ? AND stack_id IS NULL
+         ORDER BY id",
+    )
+    .bind(session_id)
+    .fetch_all(pool)
+    .await?;
+    Ok(rows.into_iter().map(|(r,)| r).collect())
+}
+
+/// Stack-scoped activated group refs (sub-agent frame). Sub-agents do **not**
+/// inherit session-scoped grants — they start from their own frame only, exactly
+/// as with the old `stack_mcp_grants::list_for_stack`.
+pub async fn list_refs_stack(pool: &SqlitePool, stack_id: i64) -> Result<Vec<String>> {
+    let rows = sqlx::query_as::<_, (String,)>(
+        "SELECT ref FROM activated_tools
+         WHERE stack_id = ?
+         ORDER BY id",
+    )
+    .bind(stack_id)
+    .fetch_all(pool)
+    .await?;
+    Ok(rows.into_iter().map(|(r,)| r).collect())
+}
+
+/// Activations in effect for one scope, up to and including `upto_message_id`,
+/// ordered by the anchoring message. Used by the DTL serializer to place the
+/// injected tool block (Kimi) at the position where it was activated. The scope
+/// mirrors the in-memory grant set: root (`stack_id = None`) sees session-scoped
+/// activations; a sub-agent (`Some(id)`) sees only its own frame's.
+pub async fn list_active_at(
+    pool: &SqlitePool,
+    session_id: i64,
+    stack_id: Option<i64>,
+    upto_message_id: i64,
+) -> Result<Vec<Activation>> {
+    let rows = match stack_id {
+        None => {
+            sqlx::query_as::<_, (i64, String, String)>(
+                "SELECT message_id, kind, ref FROM activated_tools
+                 WHERE session_id = ? AND stack_id IS NULL AND message_id <= ?
+                 ORDER BY message_id, id",
+            )
+            .bind(session_id)
+            .bind(upto_message_id)
+            .fetch_all(pool)
+            .await?
+        }
+        Some(sid) => {
+            sqlx::query_as::<_, (i64, String, String)>(
+                "SELECT message_id, kind, ref FROM activated_tools
+                 WHERE stack_id = ? AND message_id <= ?
+                 ORDER BY message_id, id",
+            )
+            .bind(sid)
+            .bind(upto_message_id)
+            .fetch_all(pool)
+            .await?
+        }
+    };
+    Ok(rows
+        .into_iter()
+        .map(|(message_id, kind, ref_)| Activation { message_id, kind, ref_ })
+        .collect())
+}
+
+/// Clear all session-scoped activations for a session (the `/resettools` path).
+/// Stack-scoped rows are ephemeral (removed on frame exit) and there are none
+/// between turns, so only the session scope needs clearing.
+pub async fn revoke_all_session(pool: &SqlitePool, session_id: i64) -> Result<()> {
+    sqlx::query("DELETE FROM activated_tools WHERE session_id = ? AND stack_id IS NULL")
+        .bind(session_id)
+        .execute(pool)
+        .await?;
+    Ok(())
+}
+
+/// Remove a stack frame's activations. Called when the frame terminates.
+pub async fn delete_for_stack(pool: &SqlitePool, stack_id: i64) -> Result<()> {
+    sqlx::query("DELETE FROM activated_tools WHERE stack_id = ?")
+        .bind(stack_id)
+        .execute(pool)
+        .await?;
+    Ok(())
+}
+
+/// Re-anchor activations pinned to a message of `stack_id` that was just compacted
+/// (chat_history id ≤ `covers_up_to`) onto `new_anchor` (the first surviving
+/// message), so the DTL serializer still renders them after compaction instead of
+/// losing the injection point. Scoped to this stack's messages via a subquery —
+/// `message_id` is a global autoincrement, so a bare `<=` would also match other
+/// stacks' rows. No unique-index conflict: only `message_id` changes, and there is
+/// at most one row per `(session, stack, kind, ref)`.
+pub async fn reanchor_compacted(
+    pool: &SqlitePool,
+    stack_id: i64,
+    covers_up_to: i64,
+    new_anchor: i64,
+) -> Result<()> {
+    sqlx::query(
+        "UPDATE activated_tools SET message_id = ?
+         WHERE message_id IN (
+             SELECT id FROM chat_history
+             WHERE session_stack_id = ? AND id <= ?
+         )",
+    )
+    .bind(new_anchor)
+    .bind(stack_id)
+    .bind(covers_up_to)
+    .execute(pool)
+    .await?;
+    Ok(())
+}

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

@@ -1,3 +1,4 @@
+pub mod activated_tools;
 pub mod approval_rules;
 pub mod project_members;
 pub mod projects;
@@ -26,10 +27,8 @@ pub mod role_capabilities;
 pub mod roles;
 pub mod scheduled_jobs;
 pub mod scratchpad;
-pub mod session_mcp_grants;
 pub mod shared_folders;
 pub mod sources;
-pub mod stack_mcp_grants;
 pub mod tool_permission_groups;
 pub mod users;
 
@@ -809,26 +808,40 @@ pub async fn create_owner_tables(pool: &SqlitePool) -> Result<()> {
     .execute(pool)
     .await?;
 
+    // Tool-group activations — the durable **effect** of `activate_tools`. One
+    // row per activated group, anchored at the assistant `message_id` that
+    // triggered it. `stack_id IS NULL` = session-scoped (root agent); non-NULL =
+    // sub-agent frame (removed on frame exit). `kind`/`ref`: ('builtin','config')
+    // or ('mcp', <server>). All FKs are owner→owner.
     sqlx::query(
-        "CREATE TABLE IF NOT EXISTS session_mcp_grants (
+        "CREATE TABLE IF NOT EXISTS activated_tools (
             id         INTEGER PRIMARY KEY AUTOINCREMENT,
-            session_id INTEGER NOT NULL,
+            session_id INTEGER NOT NULL REFERENCES chat_sessions(id),
-            mcp_name   TEXT    NOT NULL,
+            stack_id   INTEGER          REFERENCES chat_sessions_stack(id),
-            granted_at TEXT    NOT NULL DEFAULT (datetime('now')),
+            message_id INTEGER NOT NULL REFERENCES chat_history(id),
-            UNIQUE(session_id, mcp_name)
+            kind       TEXT    NOT NULL CHECK(kind IN ('builtin', 'mcp')),
+            ref        TEXT    NOT NULL,
+            created_at TEXT    NOT NULL DEFAULT (datetime('now'))
         )",
     )
     .execute(pool)
     .await?;
-
+    // Dedup per scope. A plain UNIQUE(session_id, stack_id, kind, ref) would NOT
+    // dedup session-scoped rows: SQLite treats two NULL `stack_id`s as distinct,
+    // so INSERT OR IGNORE would pile up duplicates. COALESCE folds NULL to -1.
     sqlx::query(
-        "CREATE TABLE IF NOT EXISTS stack_mcp_grants (
+        "CREATE UNIQUE INDEX IF NOT EXISTS ux_activated_tools
-            id         INTEGER PRIMARY KEY AUTOINCREMENT,
+         ON activated_tools(session_id, COALESCE(stack_id, -1), kind, ref)",
-            stack_id   INTEGER NOT NULL,
+    )
-            mcp_name   TEXT    NOT NULL,
+    .execute(pool)
-            granted_at TEXT    NOT NULL DEFAULT (datetime('now')),
+    .await?;
-            UNIQUE(stack_id, mcp_name)
+    sqlx::query(
-        )",
+        "CREATE INDEX IF NOT EXISTS idx_activated_tools_stack ON activated_tools(stack_id)",
+    )
+    .execute(pool)
+    .await?;
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_activated_tools_msg ON activated_tools(message_id)",
     )
     .execute(pool)
     .await?;
@@ -1080,8 +1093,9 @@ mod tests {
         one("INSERT INTO chat_summaries (stack_id, content, covers_up_to_message_id) VALUES (1, 's', 1)")
             .await.unwrap();
         one("INSERT INTO session_scratchpad (session_id, key, value) VALUES (1, 'k', 'v')").await.unwrap();
-        one("INSERT INTO session_mcp_grants (session_id, mcp_name) VALUES (1, 'm')").await.unwrap();
+        // Both activation scopes: session-scoped (stack_id NULL) + stack-scoped.
-        one("INSERT INTO stack_mcp_grants (stack_id, mcp_name) VALUES (1, 'm')").await.unwrap();
+        one("INSERT INTO activated_tools (session_id, stack_id, message_id, kind, ref) VALUES (1, NULL, 1, 'mcp', 'm')").await.unwrap();
+        one("INSERT INTO activated_tools (session_id, stack_id, message_id, kind, ref) VALUES (1, 1, 1, 'mcp', 'm')").await.unwrap();
         one("INSERT INTO scheduled_jobs (id, title, cron, prompt, session_id) VALUES (1, 't', '* * * * *', 'p', 1)")
             .await.unwrap();
         one("INSERT INTO job_runs (job_id, started_at, status) VALUES (1, 'now', 'completed')").await.unwrap();

crates/skald-core/src/db/session_mcp_grants.rs

@@ -1,36 +0,0 @@
-use anyhow::Result;
-use sqlx::SqlitePool;
-
-/// Grant access to an MCP server for a session.
-/// Uses INSERT OR IGNORE so calling it multiple times is safe.
-pub async fn grant(pool: &SqlitePool, session_id: i64, mcp_name: &str) -> Result<()> {
-    sqlx::query(
-        "INSERT OR IGNORE INTO session_mcp_grants (session_id, mcp_name)
-         VALUES (?, ?)"
-    )
-    .bind(session_id)
-    .bind(mcp_name)
-    .execute(pool)
-    .await?;
-    Ok(())
-}
-
-/// Revoke all MCP grants for a session.
-pub async fn revoke_all(pool: &SqlitePool, session_id: i64) -> Result<()> {
-    sqlx::query("DELETE FROM session_mcp_grants WHERE session_id = ?")
-        .bind(session_id)
-        .execute(pool)
-        .await?;
-    Ok(())
-}
-
-/// Returns the names of all MCP servers granted for this session.
-pub async fn list_for_session(pool: &SqlitePool, session_id: i64) -> Result<Vec<String>> {
-    let rows = sqlx::query_as::<_, (String,)>(
-        "SELECT mcp_name FROM session_mcp_grants WHERE session_id = ? ORDER BY granted_at"
-    )
-    .bind(session_id)
-    .fetch_all(pool)
-    .await?;
-    Ok(rows.into_iter().map(|(name,)| name).collect())
-}

crates/skald-core/src/db/stack_mcp_grants.rs

@@ -1,36 +0,0 @@
-use anyhow::Result;
-use sqlx::SqlitePool;
-
-/// Persist an MCP grant scoped to a specific stack frame (sub-agent).
-/// Uses INSERT OR IGNORE so calling it multiple times is safe.
-pub async fn grant(pool: &SqlitePool, stack_id: i64, mcp_name: &str) -> Result<()> {
-    sqlx::query(
-        "INSERT OR IGNORE INTO stack_mcp_grants (stack_id, mcp_name)
-         VALUES (?, ?)",
-    )
-    .bind(stack_id)
-    .bind(mcp_name)
-    .execute(pool)
-    .await?;
-    Ok(())
-}
-
-/// Returns the names of all MCP servers granted for this stack frame.
-pub async fn list_for_stack(pool: &SqlitePool, stack_id: i64) -> Result<Vec<String>> {
-    let rows = sqlx::query_as::<_, (String,)>(
-        "SELECT mcp_name FROM stack_mcp_grants WHERE stack_id = ? ORDER BY granted_at",
-    )
-    .bind(stack_id)
-    .fetch_all(pool)
-    .await?;
-    Ok(rows.into_iter().map(|(name,)| name).collect())
-}
-
-/// Removes all MCP grants for a stack frame. Called when the frame terminates.
-pub async fn delete_for_stack(pool: &SqlitePool, stack_id: i64) -> Result<()> {
-    sqlx::query("DELETE FROM stack_mcp_grants WHERE stack_id = ?")
-        .bind(stack_id)
-        .execute(pool)
-        .await?;
-    Ok(())
-}

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

@@ -531,6 +531,16 @@ fn build_entry(
         context_length: model.context_length,
         prompt_cache,
         capabilities:   model.capabilities.clone(),
+        // DTL is opt-in per model (the `tool_search` capability); the wire *format*
+        // comes from the model's provider (native, or `providers.yaml`) — no
+        // hardcoded model list.
+        dtl:            if model.capabilities.iter().any(|c| c == "tool_search") {
+            registry.get(&provider.provider)
+                .and_then(|p| p.dtl_format().map(crate::llm::dtl_mode_from_format))
+                .unwrap_or(crate::llm::DtlMode::None)
+        } else {
+            crate::llm::DtlMode::None
+        },
     })
 }
 

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

@@ -26,6 +26,50 @@ pub struct LlmEntry {
     /// Input capabilities of the resolved model (`vision`, `video`, …), from
     /// `llm_models.capabilities`. Drives multimodal attachment inlining.
     pub capabilities:    Vec<String>,
+    /// Dynamic-tool-loading serialization mode for this model (resolved from
+    /// `capabilities` + provider type). Selects how a session's *activated* tools
+    /// are put on the wire so that activating one does not invalidate the
+    /// provider's prompt-cache prefix.
+    pub dtl:             DtlMode,
+}
+
+/// Per-model dynamic-tool-loading (DTL) serialization mode. Resolved in
+/// `build_entry` from the model's provider (via [`dtl_mode_from_format`]) gated by
+/// the `tool_search` capability. It selects how a session's activated tools are serialized so
+/// that an `activate_tools` call does not break the provider's prompt-cache
+/// prefix. The persistence layer (`activated_tools`) is model-agnostic; this is
+/// the model-aware half that renders that state per provider.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum DtlMode {
+    /// Today's behaviour: activated tools ride in the top-level `tools` array.
+    /// Correct, but every activation invalidates the cache from that point on.
+    /// The fallback for Ollama / LM Studio / generic OpenAI-compat providers.
+    #[default]
+    None,
+    /// Anthropic Messages API: candidate tools are declared `defer_loading:true`
+    /// and loaded via a custom client-side `tool_reference` expansion emitted at
+    /// the `activate_tools` result (preserves the cache; no 5-result cap).
+    AnthropicToolReference,
+    /// Kimi K3 (OpenAI-compatible): activated tools are injected as `system`
+    /// messages carrying a `tools` field, appended at the activation position so
+    /// the prefix stays byte-identical (append-only).
+    KimiSystemTools,
+}
+
+/// Parses a provider-declared DTL format name — from a native provider
+/// (`AnthropicProvider::dtl_format`) or from `providers.yaml` (`dtl:` on a declared
+/// provider) — into a [`DtlMode`]. Unknown names → [`DtlMode::None`].
+///
+/// The *format* is a property of the provider (which wire its client speaks);
+/// whether a given model *uses* it is gated separately by the `tool_search`
+/// capability (see `build_entry`). So there is no hardcoded model list — enabling
+/// a new Kimi-compatible provider is a `providers.yaml` edit.
+pub fn dtl_mode_from_format(fmt: &str) -> DtlMode {
+    match fmt {
+        "anthropic_tool_reference" => DtlMode::AnthropicToolReference,
+        "kimi_system_tools"        => DtlMode::KimiSystemTools,
+        _                          => DtlMode::None,
+    }
 }
 
 // ── Provider ──────────────────────────────────────────────────────────────────

crates/skald-core/src/llm/providers/anthropic.rs

@@ -25,6 +25,12 @@ impl ApiProvider for AnthropicProvider {
         &[ServiceType::Llm]
     }
 
+    fn dtl_format(&self) -> Option<&str> {
+        // Every Anthropic model that opts in (via the `tool_search` capability)
+        // uses the custom client-side tool_reference format.
+        Some("anthropic_tool_reference")
+    }
+
     async fn list_llm_models(&self, _record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
         Ok(None)
     }
@@ -97,9 +103,16 @@ impl ApiProvider for AnthropicProvider {
                 .with_context(|| format!("provider '{}': api_key required for anthropic", record.name))?;
             // Merge model extra_params + reasoning (thinking) into the request body.
             let extra = extra_with_reasoning(self, model);
+            // Prompt caching is enabled exactly when this model runs dynamic tool
+            // loading (the `tool_search` capability → custom tool_reference): the
+            // deferred toolset keeps the tools prefix stable and the message builder
+            // tags the static system block with cache_control, which the client
+            // renders into the `system` array. Without DTL the native Anthropic path
+            // stays uncached, as before.
+            let prompt_cache = model.capabilities.iter().any(|c| c == "tool_search");
             Ok(BuiltLlmClient {
                 client: Arc::new(AnthropicClient::with_extra_body(key, extra)),
-                prompt_cache: false,
+                prompt_cache,
             })
         })())
     }

crates/skald-core/src/llm/providers/declared.rs

@@ -59,6 +59,11 @@ struct ProviderSpec {
     fields:  Vec<FieldSpec>,
     models:  Option<ModelsSpec>,
     reasoning: Option<ReasoningSpec>,
+    /// Dynamic-tool-loading wire format for this provider's models (e.g.
+    /// `kimi_system_tools`). Applied only to a model with the `tool_search`
+    /// capability. Absent → no DTL for this provider.
+    #[serde(default)]
+    dtl: Option<String>,
 }
 
 #[derive(Debug, Default, serde::Deserialize)]
@@ -506,6 +511,10 @@ impl ApiProvider for DeclaredProvider {
         LLM_ONLY
     }
 
+    fn dtl_format(&self) -> Option<&str> {
+        self.spec.dtl.as_deref()
+    }
+
     async fn list_llm_models(
         &self,
         record: &LlmProviderRecord,

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

@@ -6,7 +6,7 @@ use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 use tracing::info;
 
-use crate::db::{chat_history, chat_llm_tools, chat_sessions_stack, scratchpad, stack_mcp_grants};
+use crate::db::{activated_tools, chat_history, chat_llm_tools, chat_sessions_stack, scratchpad};
 use crate::events::ServerEvent;
 
 use super::{ChatSessionHandler, MAX_AGENT_DEPTH, TurnOutcome};
@@ -108,8 +108,8 @@ impl ChatSessionHandler {
         // Sub-agents never inject live user input.
         let outcome = self.run_agent_turn(child.id, &child_config, token, tx, None).await;
 
-        if let Err(e) = stack_mcp_grants::delete_for_stack(pool, child.id).await {
+        if let Err(e) = activated_tools::delete_for_stack(pool, child.id).await {
-            tracing::warn!(stack_id = child.id, error = %e, "dispatch_sub_agent: failed to delete stack MCP grants");
+            tracing::warn!(stack_id = child.id, error = %e, "dispatch_sub_agent: failed to delete stack activations");
         }
 
         let parent_agent_id = parent_config.agent_id.clone();
@@ -165,7 +165,7 @@ impl ChatSessionHandler {
         stack_id:      i64,
         depth:         i64,
     ) -> anyhow::Result<AgentRunConfig> {
-        let persisted_grants = stack_mcp_grants::list_for_stack(&self.db, stack_id)
+        let persisted_grants = activated_tools::list_refs_stack(&self.db, stack_id)
             .await
             .unwrap_or_default();
         let active_mcp_grants: Arc<RwLock<HashSet<String>>> =
@@ -205,8 +205,6 @@ impl ChatSessionHandler {
 
         {
             let activate_tool = crate::tools::activate_tools::ActivateTools {
-                pool:              Arc::clone(&self.db),
-                session_id:        self.session_id,
                 stack_id:          Some(stack_id),
                 mcp:               Arc::clone(&self.mcp),
                 active_mcp_grants: Arc::clone(&active_mcp_grants),

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

@@ -140,7 +140,7 @@ impl ChatSessionHandler {
         // Load persisted session grants from DB (MCP server names and/or the reserved
         // `config` keyword), then inject `activate_tools` so the LLM can activate
         // additional groups on demand.
-        let persisted = crate::db::session_mcp_grants::list_for_session(
+        let persisted = crate::db::activated_tools::list_refs_session(
             &self.db, self.session_id,
         ).await.unwrap_or_default();
 
@@ -148,14 +148,10 @@ impl ChatSessionHandler {
             Arc::new(RwLock::new(persisted.into_iter().collect()));
 
         {
-            let pool_clone   = Arc::clone(&self.db);
-            let session_id   = self.session_id;
             let mcp_clone    = Arc::clone(&self.mcp);
             let grants_clone = Arc::clone(&active_mcp_grants);
 
             let activate_tool = crate::tools::activate_tools::ActivateTools {
-                pool:              pool_clone,
-                session_id,
                 stack_id:          None,
                 mcp:               mcp_clone,
                 active_mcp_grants: grants_clone,

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

@@ -3,6 +3,7 @@ use std::sync::{Arc, RwLock};
 
 use serde_json::Value;
 
+use crate::llm::DtlMode;
 use crate::mcp::McpProvider;
 use crate::tools::Tool;
 use crate::tools::tool_names as tn;
@@ -56,10 +57,10 @@ pub struct AgentRunConfig {
     pub mcp: Arc<dyn McpProvider>,
     /// Set of MCP server names currently granted (activated) for this agent run.
     ///
-    /// - Root agents: pre-populated from `session_mcp_grants` DB at config-build time;
+    /// - Root agents: pre-populated from the `activated_tools` table (session-scoped
-    ///   updated in-place by `activate_tools`.
+    ///   rows) at config-build time; updated in-place by `activate_tools`.
     /// - Sub-agents: starts empty; populated by `activate_tools` (stack-scoped, no
-    ///   session leak); deleted from DB when the stack frame terminates.
+    ///   session leak); the frame's rows are deleted when the stack frame terminates.
     ///
     /// May also contain the reserved keyword `"config"`, which unlocks the built-in
     /// `Config`-category tools (`config_tool_defs`) rather than an MCP server.
@@ -79,31 +80,39 @@ impl AgentRunConfig {
     ///
     /// Dynamic groups are re-queried every call so that an `activate_tools` call in
     /// round N makes the tools visible in round N+1 without rebuilding the whole config.
-    pub fn all_tool_defs(&self) -> Vec<Value> {
+    pub fn all_tool_defs(&self, dtl: DtlMode) -> Vec<Value> {
         let mut defs = self.base_tool_defs.clone();
 
-        // Dynamic groups: read the currently-granted set (MCP server names + `config`).
+        match dtl {
-        let granted: HashSet<String> = self.active_mcp_grants
+            // Anthropic custom tool_reference: declare EVERY accessible MCP tool +
-            .read()
+            // the config group as `defer_loading:true`, on every turn. The toolset
-            .map(|g| g.clone())
+            // is stable (cache-safe) and `activate_tools` loads the needed ones via
-            .unwrap_or_default();
+            // tool_reference. Deferred defs are excluded from the prompt prefix by
-
+            // the API and cost nothing until referenced.
-        // MCP servers: include tools for the granted server names.
+            DtlMode::AnthropicToolReference => {
-        let servers: Vec<String> = granted.iter()
+                defs.extend(self.mcp.tools().iter().map(|t| deferred(t.to_openai_definition())));
-            .filter(|n| n.as_str() != crate::tools::tool_names::CONFIG_GROUP)
+                defs.extend(self.config_tool_defs.iter().cloned().map(deferred));
-            .cloned()
+            }
-            .collect();
+            // Kimi K3: activated MCP/config tools are injected as `system` messages
-        if !servers.is_empty() {
+            // by the message builder, so they are NOT in the top-level tools here.
-            defs.extend(
+            DtlMode::KimiSystemTools => {}
-                self.mcp.tools_for(&servers)
+            // Today's behaviour: only the currently-granted MCP servers + config.
-                    .iter()
+            DtlMode::None => {
-                    .map(|t| t.to_openai_definition()),
+                let granted: HashSet<String> = self.active_mcp_grants
-            );
+                    .read()
-        }
+                    .map(|g| g.clone())
-
+                    .unwrap_or_default();
-        // `config` group: include the built-in Config-category tools on demand.
+                let servers: Vec<String> = granted.iter()
-        if granted.contains(crate::tools::tool_names::CONFIG_GROUP) {
+                    .filter(|n| n.as_str() != crate::tools::tool_names::CONFIG_GROUP)
-            defs.extend(self.config_tool_defs.iter().cloned());
+                    .cloned()
+                    .collect();
+                if !servers.is_empty() {
+                    defs.extend(self.mcp.tools_for(&servers).iter().map(|t| t.to_openai_definition()));
+                }
+                if granted.contains(crate::tools::tool_names::CONFIG_GROUP) {
+                    defs.extend(self.config_tool_defs.iter().cloned());
+                }
+            }
         }
 
         defs.extend(self.memory_tools.iter().map(|t| t.openai_definition()));
@@ -167,3 +176,12 @@ impl AgentRunConfig {
         }
     }
 }
+
+/// Tags an OpenAI tool definition as deferred (Anthropic tool search): the API
+/// keeps it out of the prompt prefix until `activate_tools` references it. The
+/// flag rides on the top-level tool object; `AnthropicClient::convert_tools`
+/// maps it to Anthropic's native `defer_loading` field.
+fn deferred(mut def: Value) -> Value {
+    def["defer_loading"] = Value::Bool(true);
+    def
+}

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

@@ -45,7 +45,6 @@ impl ChatSessionHandler {
         stack_id:      i64,
         config:        &AgentRunConfig,
         active_grants: &HashSet<String>,
-        tool_defs:     &[Value],
         req_scope:     Option<&str>,
         req_strength:  Option<LlmStrength>,
         cur_name:      &mut String,
@@ -57,6 +56,9 @@ impl ChatSessionHandler {
         let mut tried_this_round: Vec<String> = vec![cur_name.clone()];
 
         loop {
+            // Re-derive the tool defs for the model actually serving this attempt:
+            // a fallback across DTL modes must re-shape (deferred candidates or not).
+            let cur_tool_defs = config.all_tool_defs(cur_llm.dtl);
             let request_id = uuid::Uuid::new_v4().to_string();
             let options = ChatOptions {
                 model:       cur_llm.model.clone(),
@@ -72,8 +74,8 @@ impl ChatSessionHandler {
             // open directly — keyed on the model actually serving this attempt, so a
             // fallback to a text-only model drops the claim. `None` (no media
             // capability) leaves the shared defs untouched, avoiding a clone.
-            let annotated = media_annotated_tools(tool_defs, &cur_llm.capabilities);
+            let annotated = media_annotated_tools(&cur_tool_defs, &cur_llm.capabilities);
-            let defs: &[Value] = annotated.as_deref().unwrap_or(tool_defs);
+            let defs: &[Value] = annotated.as_deref().unwrap_or(&cur_tool_defs);
 
             // Clone the Arc so the in-flight future does not borrow `cur_llm` across
             // the fallback reassignment below. On cancel we drop the future
@@ -164,11 +166,13 @@ impl ChatSessionHandler {
                     // settings (e.g. switching from OpenRouter/Anthropic to DeepSeek)
                     // or different input capabilities (a non-vision fallback drops
                     // inline media back to the textual path block).
+                    let activation_stack = if config.depth == 0 { None } else { Some(stack_id) };
                     match self.build_openai_messages(
                         &self.db, stack_id, &config.agent_id,
                         config.extra_system.as_deref(), config.extra_system_dynamic.as_deref(),
                         config.tail_reminder.as_deref(), active_grants,
                         &config.system_substitutions, cur_llm.prompt_cache, &cur_llm.capabilities,
+                        cur_llm.dtl, &config.config_tool_defs, activation_stack,
                     ).await {
                         Ok(m)  => *messages = m,
                         Err(e) => return RoundLlm::Failed(e),

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

@@ -117,8 +117,11 @@ impl ChatSessionHandler {
             // Messages are (re)built with the current model's prompt_cache flag.
             // On fallback within the same round `call_llm_round` rebuilds them again
             // if the replacement model has a different prompt_cache setting.
-            let mut messages = self.build_openai_messages(pool, stack_id, &config.agent_id, config.extra_system.as_deref(), config.extra_system_dynamic.as_deref(), config.tail_reminder.as_deref(), &active_grants_snapshot, &config.system_substitutions, cur_llm.prompt_cache, &cur_llm.capabilities).await?;
+            // Activation scope for the DTL serializer: session-scoped for the root
-            let tool_defs    = config.all_tool_defs();
+            // agent (stack_id NULL), the frame itself for a sub-agent.
+            let activation_stack = if config.depth == 0 { None } else { Some(stack_id) };
+            let mut messages = self.build_openai_messages(pool, stack_id, &config.agent_id, config.extra_system.as_deref(), config.extra_system_dynamic.as_deref(), config.tail_reminder.as_deref(), &active_grants_snapshot, &config.system_substitutions, cur_llm.prompt_cache, &cur_llm.capabilities, cur_llm.dtl, &config.config_tool_defs, activation_stack).await?;
+            let tool_defs    = config.all_tool_defs(cur_llm.dtl);
 
             // Record every tool actually offered to the LLM so the Security-groups
             // UI can list/gate dynamically-injected tools. Cheap no-op once each
@@ -128,7 +131,7 @@ impl ChatSessionHandler {
             // One LLM call for this round, with automatic model fallback on
             // retriable errors. `cur_name`/`cur_llm`/`messages` are updated in place.
             let turn_result = match self.call_llm_round(
-                stack_id, config, &active_grants_snapshot, &tool_defs,
+                stack_id, config, &active_grants_snapshot,
                 req_scope.as_deref(), req_strength,
                 &mut cur_name, &mut cur_llm, &mut messages, token, &em,
             ).await {
@@ -274,6 +277,25 @@ impl ChatSessionHandler {
             DispatchResult::AbortPending => return Ok(CallFlow::End(TurnOutcome::Cancelled)),
         };
 
+        // Persist the durable effect of `activate_tools`, anchored at the assistant
+        // `message_id` that triggered it (the anchor the DTL serializer positions
+        // injected tool blocks against). The in-memory grant set was already updated
+        // inside the tool; this records it across turns and restarts.
+        if call.name == crate::tools::tool_names::ACTIVATE_TOOLS {
+            if let Some(groups) = call.arguments.get("groups").and_then(|g| g.as_array()) {
+                // Root (depth 0) → session-scoped (stack_id NULL); sub-agent → its frame.
+                let anchor_stack = if config.depth == 0 { None } else { Some(stack_id) };
+                for g in groups.iter().filter_map(|v| v.as_str()) {
+                    let kind = if g == crate::tools::tool_names::CONFIG_GROUP { "builtin" } else { "mcp" };
+                    if let Err(e) = crate::db::activated_tools::grant(
+                        pool, self.session_id, anchor_stack, message_id, kind, g,
+                    ).await {
+                        tracing::warn!(session_id = self.session_id, group = g, error = %e, "activate_tools: failed to persist activation");
+                    }
+                }
+            }
+        }
+
         match self.record_tool_outcome(
             tool_call_id, &call.name, &call.arguments, outcome, preview, em, Some(all_tool_calls),
         ).await? {

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

@@ -10,6 +10,7 @@ use core_api::user_fs::UserFs;
 use crate::compactor::{ContextCompactor, SUMMARY_PREFIX};
 use crate::config::DatetimeConfig;
 use crate::db::{chat_history, chat_llm_tools, chat_summaries};
+use crate::llm::DtlMode;
 use crate::mcp::McpProvider;
 use crate::tools::tool_names as tn;
 
@@ -108,6 +109,12 @@ impl MessageBuilder {
         // Input capabilities of the resolved model (`vision`, `video`, …) —
         // drives inline media for current-turn attachments.
         capabilities:         &[String],
+        // Dynamic-tool-loading mode for the resolved model, plus the config-group
+        // tool defs (needed to resolve activated tools). `activation_stack` scopes
+        // the activation read: `None` = session-scoped (root), `Some(id)` = frame.
+        dtl:                  DtlMode,
+        config_tool_defs:     &[Value],
+        activation_stack:     Option<i64>,
     ) -> anyhow::Result<Vec<Value>> {
         let pool = &*self.pool;
 
@@ -254,6 +261,15 @@ impl MessageBuilder {
             media_turn_start -= 1;
         }
 
+        // DTL: resolve the tools activated at each assistant message (empty unless a
+        // DTL mode is active). Drives the Kimi `system`+`tools` injection and the
+        // Anthropic `tool_reference` markers emitted in the history loop below.
+        let activation_defs: HashMap<i64, Vec<Value>> = if matches!(dtl, DtlMode::None) {
+            HashMap::new()
+        } else {
+            self.resolve_activation_defs(activation_stack, config_tool_defs).await
+        };
+
         for (idx, entry) in history.iter().enumerate() {
             let is_previous_turn = current_turn_boundary.is_some_and(|b| idx < b);
 
@@ -378,11 +394,29 @@ impl MessageBuilder {
                                 tc.arguments.as_deref(),
                             );
 
-                            out.push(json!({
+                            let mut tool_msg = json!({
                                 "role":         "tool",
                                 "tool_call_id": format!("tc_{}", tc.id),
                                 "content":      result_content,
-                            }));
+                            });
+                            // Anthropic DTL: an `activate_tools` result becomes a set of
+                            // `tool_reference`s (the activated groups' tool names) that the
+                            // client renders as tool_reference blocks and the API expands
+                            // into the deferred tools. `_tool_references` is a neutral
+                            // marker other clients ignore.
+                            if matches!(dtl, DtlMode::AnthropicToolReference)
+                                && tc.name == tn::ACTIVATE_TOOLS
+                                && let Some(adefs) = activation_defs.get(&entry.id)
+                            {
+                                let names: Vec<Value> = adefs.iter()
+                                    .filter_map(|d| d["function"]["name"].as_str())
+                                    .map(|n| Value::String(n.to_string()))
+                                    .collect();
+                                if !names.is_empty() {
+                                    tool_msg["_tool_references"] = Value::Array(names);
+                                }
+                            }
+                            out.push(tool_msg);
                         }
 
                         // Media a tool produced this turn (e.g. read_file on an
@@ -411,6 +445,16 @@ impl MessageBuilder {
                                 }
                             }
                         }
+
+                        // Kimi K3 DTL: inject the tools activated at this assistant
+                        // message as a `system` message carrying a `tools` field, right
+                        // after its tool-result group (append-only → cache-safe).
+                        if matches!(dtl, DtlMode::KimiSystemTools)
+                            && let Some(adefs) = activation_defs.get(&entry.id)
+                            && !adefs.is_empty()
+                        {
+                            out.push(json!({ "role": "system", "tools": adefs }));
+                        }
                     }
                 }
             }
@@ -585,7 +629,49 @@ impl MessageBuilder {
         ))
     }
 
-    fn render_mcp_list(&self, active_mcp_grants: &HashSet<String>) -> String {
+    /// Resolves the tools activated at each assistant message, for DTL
+    /// serialization. Returns `message_id → deduped OpenAI tool defs`. MCP groups
+    /// resolve to the server's live tool defs; the `config` group resolves to the
+    /// passed-in config-category defs. Scope mirrors the in-memory grant set:
+    /// `None` = session-scoped (root), `Some(id)` = the sub-agent frame.
+    async fn resolve_activation_defs(
+        &self,
+        activation_stack: Option<i64>,
+        config_tool_defs: &[Value],
+    ) -> HashMap<i64, Vec<Value>> {
+        let activations = crate::db::activated_tools::list_active_at(
+            &self.pool, self.session_id, activation_stack, i64::MAX,
+        ).await.unwrap_or_default();
+
+        let mut map: HashMap<i64, Vec<Value>> = HashMap::new();
+        for a in activations {
+            let resolved: Vec<Value> = match a.kind.as_str() {
+                "mcp" => self.mcp.tools_for(&[a.ref_.clone()])
+                    .iter().map(|t| t.to_openai_definition()).collect(),
+                "builtin" if a.ref_ == tn::CONFIG_GROUP => config_tool_defs.to_vec(),
+                _ => Vec::new(),
+            };
+            let entry = map.entry(a.message_id).or_default();
+            for d in resolved {
+                if let Some(name) = d["function"]["name"].as_str().map(str::to_string) {
+                    if !entry.iter().any(|e| e["function"]["name"].as_str() == Some(name.as_str())) {
+                        entry.push(d);
+                    }
+                }
+            }
+        }
+        map
+    }
+
+    /// A **static** catalogue of the MCP servers this user can load — identical
+    /// regardless of which are currently active. Static is deliberate: this text
+    /// sits inside the `cache_control: ephemeral` system block, so the old
+    /// Available/Active split (which moved a server between tables on activation)
+    /// invalidated the prompt-cache prefix on every `activate_tools` call. Which
+    /// servers are active is already visible to the model through the injected
+    /// tools. The grant set is retained in the signature only for call-site
+    /// stability (the DTL serializer may consume it later).
+    fn render_mcp_list(&self, _active_mcp_grants: &HashSet<String>) -> String {
         let all_servers: std::collections::BTreeSet<String> = self.mcp.tools()
             .into_iter()
             .map(|t| t.server_name)
@@ -597,37 +683,17 @@ impl MessageBuilder {
 
         let descriptions = self.mcp.server_descriptions();
 
-        let hidden: Vec<&String> = all_servers.iter()
+        let mut out = String::from(
-            .filter(|n| !active_mcp_grants.contains(*n))
+            "## MCP servers\n\nConnectors you can load with `activate_tools([\"name\"])`. \
-            .collect();
+             Once loaded, a server's tools are callable as `mcp__<name>__<tool>`:\n\n",
-        let active: Vec<&String> = all_servers.iter()
+        );
-            .filter(|n| active_mcp_grants.contains(*n))
+        out.push_str("| Server | Description |\n|--------|-------------|\n");
-            .collect();
+        for name in &all_servers {
-
+            let desc = descriptions.get(name)
-        let mut out = String::from("## MCP servers\n");
+                .and_then(|d| d.as_deref())
-
+                .unwrap_or("—");
-        if !hidden.is_empty() {
+            out.push_str(&format!("| `{name}` | {desc} |\n"));
-            out.push_str("\n**Available** — call `activate_tools([\"name\"])` to load tools:\n\n");
-            out.push_str("| Server | Description |\n|--------|-------------|\n");
-            for name in &hidden {
-                let desc = descriptions.get(*name)
-                    .and_then(|d| d.as_deref())
-                    .unwrap_or("—");
-                out.push_str(&format!("| `{name}` | {desc} |\n"));
-            }
         }
-
-        if !active.is_empty() {
-            out.push_str("\n**Active** — tools callable as `mcp__<name>__<tool>`:\n\n");
-            out.push_str("| Server | Description |\n|--------|-------------|\n");
-            for name in &active {
-                let desc = descriptions.get(*name)
-                    .and_then(|d| d.as_deref())
-                    .unwrap_or("—");
-                out.push_str(&format!("| `{name}` | {desc} |\n"));
-            }
-        }
-
         out
     }
 }

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

@@ -3,6 +3,7 @@ use std::sync::Arc;
 
 use serde_json::Value;
 
+use crate::llm::DtlMode;
 use super::ChatSessionHandler;
 use super::message_builder::MessageBuilder;
 
@@ -23,6 +24,9 @@ impl ChatSessionHandler {
         system_substitutions: &HashMap<String, String>,
         cache_hints:          bool,
         capabilities:         &[String],
+        dtl:                  DtlMode,
+        config_tool_defs:     &[Value],
+        activation_stack:     Option<i64>,
     ) -> anyhow::Result<Vec<Value>> {
         let project_root = self.run_context.read().await
             .as_ref()
@@ -45,6 +49,6 @@ impl ChatSessionHandler {
         // `pool` is passed in from the caller (always `&self.db`) but we take
         // ownership via Arc::clone above so the signature stays backward-compatible.
         let _ = pool; // suppress unused-variable warning; MessageBuilder uses its own Arc
-        builder.build(stack_id, agent_id, extra_system_static, extra_system_dynamic, tail_reminder, active_mcp_grants, system_substitutions, cache_hints, capabilities).await
+        builder.build(stack_id, agent_id, extra_system_static, extra_system_dynamic, tail_reminder, active_mcp_grants, system_substitutions, cache_hints, capabilities, dtl, config_tool_defs, activation_stack).await
     }
 }

crates/skald-core/src/tools/activate_tools.rs

@@ -3,7 +3,6 @@ use std::sync::{Arc, RwLock};
 
 use anyhow::Result;
 use serde_json::{Value, json};
-use sqlx::SqlitePool;
 
 use crate::mcp::McpProvider;
 use crate::tools::tool_names::CONFIG_GROUP;
@@ -16,26 +15,23 @@ use crate::tools::{Tool, ToolDescriptionLength, truncate_label, MAX_LABEL_SHORT}
 /// - the reserved keyword `"config"` — loads all built-in `Config`-category
 ///   tools (system configuration: MCP/plugin/cron management, secrets).
 ///
-/// When the LLM calls `activate_tools(["gmail", "config"])`:
+/// When the LLM calls `activate_tools(["gmail", "config"])`, this tool updates
-/// - The in-memory grant set is updated immediately, so the group's tools appear
+/// the in-memory grant set **immediately**, so the group's tools appear in the
-///   in the *next LLM round* of the current turn (via `all_tool_defs()`).
+/// *next LLM round* of the current turn (via `all_tool_defs()`).
-/// - If `stack_id` is `None` (root agent): grants are persisted to
-///   `session_mcp_grants` — they survive across turns and restarts.
-/// - If `stack_id` is `Some(id)` (sub-agent): grants are persisted to
-///   `stack_mcp_grants` for that stack frame — they survive restarts but are
-///   deleted when the frame terminates (`dispatch_call_agent` calls
-///   `stack_mcp_grants::delete_for_stack` on cleanup).
 ///
-/// The `session_mcp_grants` / `stack_mcp_grants` tables store the group string
+/// The **durable** record of the activation is written by the round loop
-/// verbatim, so `"config"` is persisted just like an MCP server name.
+/// (`handle_tool_call`), not here: it anchors the activation to the assistant
+/// `message_id` that triggered it, in the owner table `activated_tools`
+/// (`stack_id NULL` = session-scoped root grant; `Some(id)` = sub-agent frame
+/// grant, deleted on frame exit). Splitting the write this way lets the loop
+/// supply the `message_id` — the anchor the DTL serializer positions injected
+/// tool blocks against — which this tool does not have.
 ///
 /// Not in the global `ToolRegistry` — injected as an `InterfaceTool` in
-/// `build_agent_config` (root) and `dispatch_call_agent` (sub-agents).
+/// `build_agent_config` (root) and `build_sub_agent_config` (sub-agents).
 pub struct ActivateTools {
-    pub pool:               Arc<SqlitePool>,
+    /// `None` for root agents, `Some(stack_id)` for sub-agents. Used only to
-    pub session_id:         i64,
+    /// label the confirmation message; the durable scope is decided by the loop.
-    /// `None` for root agents (session-scoped grants).
-    /// `Some(stack_id)` for sub-agents (stack-scoped grants, deleted on frame exit).
     pub stack_id:           Option<i64>,
     pub mcp:                Arc<dyn McpProvider>,
     /// Shared in-memory grant set. Updated in-place on every call so subsequent
@@ -97,32 +93,11 @@ impl Tool for ActivateTools {
             .map(|t| t.server_name.clone())
             .collect();
 
-        let pool       = Arc::clone(&self.pool);
+        // Update the in-memory set so the next LLM round sees the new grants.
-        let session_id = self.session_id;
+        // The durable record (with the triggering `message_id`) is written by the
-        let stack_id   = self.stack_id;
+        // round loop after this tool returns.
-        let grants_set = Arc::clone(&self.active_mcp_grants);
-
-        // Persist to DB (session-scoped or stack-scoped) and update in-memory set.
-        // The reserved `config` group is stored verbatim, exactly like a server name.
-        tokio::task::block_in_place(|| {
-            tokio::runtime::Handle::current().block_on(async {
-                for name in &names {
-                    match stack_id {
-                        None => {
-                            crate::db::session_mcp_grants::grant(&pool, session_id, name).await?;
-                        }
-                        Some(sid) => {
-                            crate::db::stack_mcp_grants::grant(&pool, sid, name).await?;
-                        }
-                    }
-                }
-                anyhow::Ok(())
-            })
-        })?;
-
-        // Update in-memory set so the next LLM round sees the new grants.
         {
-            let mut set = grants_set.write()
+            let mut set = self.active_mcp_grants.write()
                 .map_err(|_| anyhow::anyhow!("activate_tools: lock poisoned"))?;
             for name in &names {
                 set.insert(name.clone());
@@ -142,7 +117,7 @@ impl Tool for ActivateTools {
             })
             .collect();
 
-        let scope = match stack_id {
+        let scope = match self.stack_id {
             None    => "session".to_string(),
             Some(s) => format!("stack {s}"),
         };

providers.yaml

@@ -14,6 +14,8 @@
 #   base_url_overridable: the DB instance's base_url overrides this default
 #   api_key:              required | optional | none  (default: required)
 #   prompt_cache:         true | false                (default: false)
+#   dtl:                  DTL wire format for this provider's models that carry
+#                         the `tool_search` capability: kimi_system_tools (default: none)
 #   ui:                   { color, icon, description }   (color/icon required)
 #   fields:               UI form fields [{ key, label, required, secret }]
 #   models:               remote catalog config — omit to disable listing:
@@ -46,6 +48,7 @@ providers:
   - id: moonshot
     name: "Moonshot AI pay-as-you-go"
     base_url: "https://api.moonshot.ai/v1"
+    dtl: kimi_system_tools
     ui:
       color: "#2563eb"
       icon: "bi-moon-stars"
@@ -64,6 +67,7 @@ providers:
   - id: moonshot_code
     name: "Moonshot AI Kimi Code"
     base_url: "https://api.kimi.com/coding/v1"
+    dtl: kimi_system_tools
     ui:
       color: "#000000"
       icon: "bi-code-slash"
@@ -81,7 +85,7 @@ providers:
       # Fills the metadata the endpoint may omit (endpoint values always win):
       # k3 → 1M context + native vision and video input; kimi-for-coding → 256k.
       enrich:
-        - { match: "k3*", context_length: 1048576, vision: true, add_capabilities: [video] }
+        - { match: "k3*", context_length: 1048576, vision: true, add_capabilities: [video, tool_search] }
         - { match: "kimi-for-coding*", context_length: 262144 }
     reasoning:
       # k3 exposes a graded reasoning_effort ("disabled" routes to K2.6);