AbstractRuntime 0.2.0__py3-none-any.whl → 0.4.0__py3-none-any.whl

abstractruntime/core/runtime.py

@@ -20,8 +20,12 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from datetime import datetime, timezone
-from typing import Any, Callable, Dict, Optional
+from typing import Any, Callable, Dict, Optional, List
+import copy
 import inspect
+import json
+import os
+import re
 
 from .config import RuntimeConfig
 from .models import (
@@ -38,13 +42,164 @@ from .models import (
 )
 from .spec import WorkflowSpec
 from .policy import DefaultEffectPolicy, EffectPolicy
-from ..storage.base import LedgerStore, RunStore
+from ..storage.base import LedgerStore, RunStore, QueryableRunStore
+from .event_keys import build_event_wait_key
 
 
 def utc_now_iso() -> str:
     return datetime.now(timezone.utc).isoformat()
 
 
+_DEFAULT_GLOBAL_MEMORY_RUN_ID = "global_memory"
+_SAFE_RUN_ID_PATTERN = re.compile(r"^[a-zA-Z0-9_-]+$")
+
+
+def _ensure_runtime_namespace(vars: Dict[str, Any]) -> Dict[str, Any]:
+    runtime_ns = vars.get("_runtime")
+    if not isinstance(runtime_ns, dict):
+        runtime_ns = {}
+        vars["_runtime"] = runtime_ns
+    return runtime_ns
+
+
+def _ensure_control_namespace(vars: Dict[str, Any]) -> Dict[str, Any]:
+    runtime_ns = _ensure_runtime_namespace(vars)
+    control = runtime_ns.get("control")
+    if not isinstance(control, dict):
+        control = {}
+        runtime_ns["control"] = control
+    return control
+
+
+def _is_paused_run_vars(vars: Any) -> bool:
+    if not isinstance(vars, dict):
+        return False
+    runtime_ns = vars.get("_runtime")
+    if not isinstance(runtime_ns, dict):
+        return False
+    control = runtime_ns.get("control")
+    if not isinstance(control, dict):
+        return False
+    return bool(control.get("paused") is True)
+
+
+def _is_pause_wait(waiting: Any, *, run_id: str) -> bool:
+    if waiting is None:
+        return False
+    try:
+        reason = getattr(waiting, "reason", None)
+        reason_value = reason.value if hasattr(reason, "value") else str(reason) if reason else None
+    except Exception:
+        reason_value = None
+    if reason_value != WaitReason.USER.value:
+        return False
+    try:
+        wait_key = getattr(waiting, "wait_key", None)
+        if isinstance(wait_key, str) and wait_key == f"pause:{run_id}":
+            return True
+    except Exception:
+        pass
+    try:
+        details = getattr(waiting, "details", None)
+        if isinstance(details, dict) and details.get("kind") == "pause":
+            return True
+    except Exception:
+        pass
+    return False
+
+
+def _record_node_trace(
+    *,
+    run: RunState,
+    node_id: str,
+    effect: Effect,
+    outcome: "EffectOutcome",
+    idempotency_key: Optional[str],
+    reused_prior_result: bool,
+    duration_ms: Optional[float] = None,
+    max_entries_per_node: int = 100,
+) -> None:
+    """Record a JSON-safe per-node execution trace in run.vars["_runtime"].
+
+    This trace is runtime-owned and durable (stored in RunStore checkpoints).
+    It exists to support higher-level hosts (AbstractFlow, AbstractCode, etc.)
+    that need structured "scratchpad"/debug information without inventing
+    host-specific persistence formats.
+    """
+
+    runtime_ns = _ensure_runtime_namespace(run.vars)
+    traces = runtime_ns.get("node_traces")
+    if not isinstance(traces, dict):
+        traces = {}
+        runtime_ns["node_traces"] = traces
+
+    node_trace = traces.get(node_id)
+    if not isinstance(node_trace, dict):
+        node_trace = {"node_id": node_id, "steps": []}
+        traces[node_id] = node_trace
+
+    steps = node_trace.get("steps")
+    if not isinstance(steps, list):
+        steps = []
+        node_trace["steps"] = steps
+
+    wait_dict: Optional[Dict[str, Any]] = None
+    if outcome.status == "waiting" and outcome.wait is not None:
+        w = outcome.wait
+        wait_dict = {
+            "reason": w.reason.value if hasattr(w.reason, "value") else str(w.reason),
+            "wait_key": w.wait_key,
+            "until": w.until,
+            "resume_to_node": w.resume_to_node,
+            "result_key": w.result_key,
+            "prompt": w.prompt,
+            "choices": w.choices,
+            "allow_free_text": w.allow_free_text,
+            "details": w.details,
+        }
+
+    entry: Dict[str, Any] = {
+        "ts": utc_now_iso(),
+        "node_id": node_id,
+        "status": outcome.status,
+        "idempotency_key": idempotency_key,
+        "reused_prior_result": reused_prior_result,
+        "effect": {
+            "type": effect.type.value,
+            "payload": effect.payload,
+            "result_key": effect.result_key,
+        },
+    }
+    if isinstance(duration_ms, (int, float)) and duration_ms >= 0:
+        # UI/UX consumers use this for per-step timing badges (kept JSON-safe).
+        entry["duration_ms"] = float(duration_ms)
+    if outcome.status == "completed":
+        entry["result"] = outcome.result
+    elif outcome.status == "failed":
+        entry["error"] = outcome.error
+    elif wait_dict is not None:
+        entry["wait"] = wait_dict
+
+    # Ensure the trace remains JSON-safe even if a handler violates the contract.
+    try:
+        json.dumps(entry)
+    except TypeError:
+        entry = {
+            "ts": entry.get("ts"),
+            "node_id": node_id,
+            "status": outcome.status,
+            "idempotency_key": idempotency_key,
+            "reused_prior_result": reused_prior_result,
+            "effect": {"type": effect.type.value, "result_key": effect.result_key},
+            "error": "non_json_safe_trace_entry",
+        }
+
+    steps.append(entry)
+    if max_entries_per_node > 0 and len(steps) > max_entries_per_node:
+        del steps[: max(0, len(steps) - max_entries_per_node)]
+    node_trace["updated_at"] = utc_now_iso()
+
+
 @dataclass
 class DefaultRunContext:
     def now_iso(self) -> str:
@@ -95,6 +250,7 @@ class Runtime:
         artifact_store: Optional[Any] = None,
         effect_policy: Optional[EffectPolicy] = None,
         config: Optional[RuntimeConfig] = None,
+        chat_summarizer: Optional[Any] = None,
     ):
         self._run_store = run_store
         self._ledger_store = ledger_store
@@ -103,6 +259,7 @@ class Runtime:
         self._artifact_store = artifact_store
         self._effect_policy: EffectPolicy = effect_policy or DefaultEffectPolicy()
         self._config: RuntimeConfig = config or RuntimeConfig()
+        self._chat_summarizer = chat_summarizer
 
         self._handlers: Dict[EffectType, EffectHandler] = {}
         self._register_builtin_handlers()
@@ -169,6 +326,43 @@ class Runtime:
         if "_limits" not in vars:
             vars["_limits"] = self._config.to_limits_dict()
 
+        # Ensure a durable `_runtime` namespace exists and seed default provider/model metadata
+        # from the Runtime config (best-effort).
+        #
+        # Rationale:
+        # - The Runtime is the orchestration authority (ADR-0001/0014), and `start()` is the
+        #   choke point where durable run state is initialized.
+        # - Agents/workflows should not have to guess/duplicate routing metadata to make prompt
+        #   composition decisions (e.g. native-tools => omit Tools(session) prompt catalogs).
+        runtime_ns = vars.get("_runtime")
+        if not isinstance(runtime_ns, dict):
+            runtime_ns = {}
+            vars["_runtime"] = runtime_ns
+        try:
+            provider_id = getattr(self._config, "provider", None)
+            model_id = getattr(self._config, "model", None)
+            if isinstance(provider_id, str) and provider_id.strip():
+                runtime_ns.setdefault("provider", provider_id.strip())
+            if isinstance(model_id, str) and model_id.strip():
+                runtime_ns.setdefault("model", model_id.strip())
+        except Exception:
+            pass
+
+        # Seed tool-support metadata from model capabilities (best-effort).
+        #
+        # This makes the native-vs-prompted tools decision explicit and durable in run state,
+        # so adapters/UI helpers don't have to guess or re-run AbstractCore detection logic.
+        try:
+            caps = getattr(self._config, "model_capabilities", None)
+            if isinstance(caps, dict):
+                tool_support = caps.get("tool_support")
+                if isinstance(tool_support, str) and tool_support.strip():
+                    ts = tool_support.strip()
+                    runtime_ns.setdefault("tool_support", ts)
+                    runtime_ns.setdefault("supports_native_tools", ts == "native")
+        except Exception:
+            pass
+
         run = RunState.new(
             workflow_id=workflow.workflow_id,
             entry_node=workflow.entry_node,
@@ -205,6 +399,87 @@ class Runtime:
         run.status = RunStatus.CANCELLED
         run.error = reason or "Cancelled"
         run.waiting = None
+        try:
+            control = _ensure_control_namespace(run.vars)
+            control.pop("paused", None)
+        except Exception:
+            pass
+        run.updated_at = utc_now_iso()
+        self._run_store.save(run)
+        return run
+
+    def pause_run(self, run_id: str, *, reason: Optional[str] = None) -> RunState:
+        """Pause a run (durably) until it is explicitly resumed.
+
+        Semantics:
+        - Pausing a RUNNING run transitions it to WAITING with a synthetic USER wait.
+        - Pausing a WAITING run (non-USER waits such as UNTIL/EVENT/SUBWORKFLOW) sets a
+          runtime-owned `paused` flag so schedulers/event emitters can skip it.
+        - Pausing an ASK_USER wait is a no-op (already blocked by user input).
+        """
+        run = self.get_state(run_id)
+
+        if run.status in (RunStatus.COMPLETED, RunStatus.FAILED, RunStatus.CANCELLED):
+            return run
+
+        # If already paused, keep as-is.
+        if _is_paused_run_vars(run.vars):
+            return run
+
+        # Don't interfere with real user prompts (ASK_USER).
+        if run.status == RunStatus.WAITING and run.waiting is not None:
+            if getattr(run.waiting, "reason", None) == WaitReason.USER and not _is_pause_wait(run.waiting, run_id=run_id):
+                return run
+
+        control = _ensure_control_namespace(run.vars)
+        control["paused"] = True
+        control["paused_at"] = utc_now_iso()
+        if isinstance(reason, str) and reason.strip():
+            control["pause_reason"] = reason.strip()
+
+        if run.status == RunStatus.RUNNING:
+            run.status = RunStatus.WAITING
+            run.waiting = WaitState(
+                reason=WaitReason.USER,
+                wait_key=f"pause:{run.run_id}",
+                resume_to_node=run.current_node,
+                prompt="Paused",
+                choices=None,
+                allow_free_text=False,
+                details={"kind": "pause"},
+            )
+
+        run.updated_at = utc_now_iso()
+        self._run_store.save(run)
+        return run
+
+    def resume_run(self, run_id: str) -> RunState:
+        """Resume a previously paused run (durably).
+
+        If the run was paused while RUNNING, this clears the synthetic pause wait
+        and returns the run to RUNNING. If the run was paused while WAITING
+        (UNTIL/EVENT/SUBWORKFLOW), this only clears the paused flag.
+        """
+        run = self.get_state(run_id)
+
+        if run.status in (RunStatus.COMPLETED, RunStatus.FAILED, RunStatus.CANCELLED):
+            return run
+
+        if not _is_paused_run_vars(run.vars):
+            return run
+
+        try:
+            control = _ensure_control_namespace(run.vars)
+            control.pop("paused", None)
+            control.pop("pause_reason", None)
+            control["resumed_at"] = utc_now_iso()
+        except Exception:
+            pass
+
+        if run.status == RunStatus.WAITING and _is_pause_wait(run.waiting, run_id=run_id):
+            resume_to = getattr(run.waiting, "resume_to_node", None)
+            self._apply_resume_payload(run, payload={}, override_node=resume_to)
+
         run.updated_at = utc_now_iso()
         self._run_store.save(run)
         return run
@@ -218,6 +493,83 @@ class Runtime:
     def get_ledger(self, run_id: str) -> list[dict[str, Any]]:
         return self._ledger_store.list(run_id)
 
+    def subscribe_ledger(
+        self,
+        callback: Callable[[Dict[str, Any]], None],
+        *,
+        run_id: Optional[str] = None,
+    ) -> Callable[[], None]:
+        """Subscribe to ledger append events (in-process only).
+
+        This is an optional capability: not all LedgerStore implementations
+        support subscriptions. When unavailable, wrap the configured store with
+        `abstractruntime.storage.observable.ObservableLedgerStore`.
+        """
+        subscribe = getattr(self._ledger_store, "subscribe", None)
+        if not callable(subscribe):
+            raise RuntimeError(
+                "Configured LedgerStore does not support subscriptions. "
+                "Wrap it with ObservableLedgerStore to enable `subscribe_ledger()`."
+            )
+        return subscribe(callback, run_id=run_id)
+
+    # ---------------------------------------------------------------------
+    # Trace Helpers (Runtime-Owned)
+    # ---------------------------------------------------------------------
+
+    def get_node_traces(self, run_id: str) -> Dict[str, Any]:
+        """Return runtime-owned per-node traces for a run.
+
+        Traces are stored in `RunState.vars["_runtime"]["node_traces"]`.
+        Returns a deep copy so callers can safely inspect without mutating the run.
+        """
+        run = self.get_state(run_id)
+        runtime_ns = run.vars.get("_runtime")
+        traces = runtime_ns.get("node_traces") if isinstance(runtime_ns, dict) else None
+        return copy.deepcopy(traces) if isinstance(traces, dict) else {}
+
+    def get_node_trace(self, run_id: str, node_id: str) -> Dict[str, Any]:
+        """Return a single node trace object for a run.
+
+        Returns an empty `{node_id, steps: []}` object when missing.
+        """
+        traces = self.get_node_traces(run_id)
+        trace = traces.get(node_id)
+        if isinstance(trace, dict):
+            return trace
+        return {"node_id": node_id, "steps": []}
+
+    # ---------------------------------------------------------------------
+    # Evidence Helpers (Runtime-Owned)
+    # ---------------------------------------------------------------------
+
+    def list_evidence(self, run_id: str) -> list[dict[str, Any]]:
+        """List evidence records for a run (index entries only).
+
+        Evidence is indexed as `kind="evidence"` items inside `vars["_runtime"]["memory_spans"]`.
+        """
+        run = self.get_state(run_id)
+        runtime_ns = run.vars.get("_runtime")
+        spans = runtime_ns.get("memory_spans") if isinstance(runtime_ns, dict) else None
+        if not isinstance(spans, list):
+            return []
+        out: list[dict[str, Any]] = []
+        for s in spans:
+            if not isinstance(s, dict):
+                continue
+            if s.get("kind") != "evidence":
+                continue
+            out.append(copy.deepcopy(s))
+        return out
+
+    def load_evidence(self, evidence_id: str) -> Optional[dict[str, Any]]:
+        """Load an evidence record payload from ArtifactStore by id."""
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            raise RuntimeError("Evidence requires an ArtifactStore; configure runtime.set_artifact_store(...)")
+        payload = artifact_store.load_json(str(evidence_id))
+        return payload if isinstance(payload, dict) else None
+
     # ---------------------------------------------------------------------
     # Limit Management
     # ---------------------------------------------------------------------
@@ -340,7 +692,10 @@ class Runtime:
 
     def tick(self, *, workflow: WorkflowSpec, run_id: str, max_steps: int = 100) -> RunState:
         run = self.get_state(run_id)
-        if run.status in (RunStatus.COMPLETED, RunStatus.FAILED):
+        # Terminal runs never progress.
+        if run.status in (RunStatus.COMPLETED, RunStatus.FAILED, RunStatus.CANCELLED):
+            return run
+        if _is_paused_run_vars(run.vars):
             return run
         if run.status == RunStatus.WAITING:
             # For WAIT_UNTIL we can auto-unblock if time passed
@@ -352,15 +707,40 @@ class Runtime:
             else:
                 return run
 
+        # IMPORTANT (Web hosts / concurrency):
+        # A run may be paused/cancelled by an external control plane (e.g. AbstractFlow Web UI)
+        # while we're blocked inside a long-running effect (LLM/tool execution).
+        #
+        # We make `tick()` resilient to that by re-loading the persisted RunState before
+        # committing any updates. If an external pause/cancel is observed, we stop without
+        # overwriting it.
+        def _abort_if_externally_controlled() -> Optional[RunState]:
+            try:
+                latest = self.get_state(run_id)
+            except Exception:
+                return None
+            if latest.status == RunStatus.CANCELLED:
+                return latest
+            if _is_paused_run_vars(latest.vars):
+                return latest
+            return None
+
         steps = 0
         while steps < max_steps:
             steps += 1
 
+            controlled = _abort_if_externally_controlled()
+            if controlled is not None:
+                return controlled
+
             handler = workflow.get_node(run.current_node)
             plan = handler(run, self._ctx)
 
             # Completion
             if plan.complete_output is not None:
+                controlled = _abort_if_externally_controlled()
+                if controlled is not None:
+                    return controlled
                 run.status = RunStatus.COMPLETED
                 run.output = plan.complete_output
                 run.updated_at = utc_now_iso()
@@ -377,6 +757,9 @@ class Runtime:
             if plan.effect is None:
                 if not plan.next_node:
                     raise ValueError(f"Node '{plan.node_id}' returned no effect and no next_node")
+                controlled = _abort_if_externally_controlled()
+                if controlled is not None:
+                    return controlled
                 run.current_node = plan.next_node
                 run.updated_at = utc_now_iso()
                 self._run_store.save(run)
@@ -387,6 +770,13 @@ class Runtime:
                 run=run, node_id=plan.node_id, effect=plan.effect
             )
             prior_result = self._find_prior_completed_result(run.run_id, idempotency_key)
+            reused_prior_result = prior_result is not None
+
+            # Measure effect execution duration (wall-clock). This is used for
+            # host-side UX (badges, throughput estimates) and is stored in the
+            # runtime-owned node trace (JSON-safe).
+            import time
+            t0 = time.perf_counter()
 
             if prior_result is not None:
                 # Reuse prior result - skip re-execution
@@ -401,7 +791,42 @@ class Runtime:
                     default_next_node=plan.next_node,
                 )
 
+            duration_ms = float((time.perf_counter() - t0) * 1000.0)
+
+            # Evidence capture (runtime-owned, durable):
+            # After tool execution completes, record provenance-first evidence for a small set of
+            # external-boundary tools (web_search/fetch_url/execute_command). This must happen
+            # BEFORE we persist node traces / result_key outputs so run state remains bounded.
+            try:
+                if (
+                    not reused_prior_result
+                    and plan.effect.type == EffectType.TOOL_CALLS
+                    and outcome.status == "completed"
+                ):
+                    self._maybe_record_tool_evidence(
+                        run=run,
+                        node_id=plan.node_id,
+                        effect=plan.effect,
+                        tool_results=outcome.result,
+                    )
+            except Exception:
+                # Evidence capture should never crash the run; failures are recorded in run vars.
+                pass
+
+            _record_node_trace(
+                run=run,
+                node_id=plan.node_id,
+                effect=plan.effect,
+                outcome=outcome,
+                idempotency_key=idempotency_key,
+                reused_prior_result=reused_prior_result,
+                duration_ms=duration_ms,
+            )
+
             if outcome.status == "failed":
+                controlled = _abort_if_externally_controlled()
+                if controlled is not None:
+                    return controlled
                 run.status = RunStatus.FAILED
                 run.error = outcome.error or "unknown error"
                 run.updated_at = utc_now_iso()
@@ -410,6 +835,9 @@ class Runtime:
 
             if outcome.status == "waiting":
                 assert outcome.wait is not None
+                controlled = _abort_if_externally_controlled()
+                if controlled is not None:
+                    return controlled
                 run.status = RunStatus.WAITING
                 run.waiting = outcome.wait
                 run.updated_at = utc_now_iso()
@@ -420,16 +848,85 @@ class Runtime:
             if plan.effect.result_key and outcome.result is not None:
                 _set_nested(run.vars, plan.effect.result_key, outcome.result)
 
+            # Terminal effect node: treat missing next_node as completion.
+            #
+            # Rationale: StepPlan.complete_output is evaluated *before* effects
+            # execute, so an effectful node cannot both execute an effect and
+            # complete the run in a single StepPlan. Allowing next_node=None
+            # makes "end on an effect node" valid (Blueprint-style UX).
             if not plan.next_node:
-                raise ValueError(f"Node '{plan.node_id}' executed effect but did not specify next_node")
+                controlled = _abort_if_externally_controlled()
+                if controlled is not None:
+                    return controlled
+                run.status = RunStatus.COMPLETED
+                run.output = {"success": True, "result": outcome.result}
+                run.updated_at = utc_now_iso()
+                self._run_store.save(run)
+                return run
+            controlled = _abort_if_externally_controlled()
+            if controlled is not None:
+                return controlled
             run.current_node = plan.next_node
             run.updated_at = utc_now_iso()
             self._run_store.save(run)
 
         return run
 
-    def resume(self, *, workflow: WorkflowSpec, run_id: str, wait_key: Optional[str], payload: Dict[str, Any]) -> RunState:
+    def _maybe_record_tool_evidence(
+        self,
+        *,
+        run: RunState,
+        node_id: str,
+        effect: Effect,
+        tool_results: Optional[Dict[str, Any]],
+    ) -> None:
+        """Best-effort evidence capture for TOOL_CALLS.
+
+        This is intentionally non-fatal: evidence capture must not crash the run,
+        but failures should be visible in durable run state for debugging.
+        """
+        if effect.type != EffectType.TOOL_CALLS:
+            return
+        if not isinstance(tool_results, dict):
+            return
+        payload = effect.payload if isinstance(effect.payload, dict) else {}
+        tool_calls = payload.get("tool_calls")
+        if not isinstance(tool_calls, list) or not tool_calls:
+            return
+
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            return
+
+        try:
+            from ..evidence import EvidenceRecorder
+
+            EvidenceRecorder(artifact_store=artifact_store).record_tool_calls(
+                run=run,
+                node_id=str(node_id or ""),
+                tool_calls=list(tool_calls),
+                tool_results=tool_results,
+            )
+        except Exception as e:
+            runtime_ns = _ensure_runtime_namespace(run.vars)
+            warnings = runtime_ns.get("evidence_warnings")
+            if not isinstance(warnings, list):
+                warnings = []
+                runtime_ns["evidence_warnings"] = warnings
+            warnings.append({"ts": utc_now_iso(), "node_id": str(node_id or ""), "error": str(e)})
+
+    def resume(
+        self,
+        *,
+        workflow: WorkflowSpec,
+        run_id: str,
+        wait_key: Optional[str],
+        payload: Dict[str, Any],
+        max_steps: int = 100,
+    ) -> RunState:
         run = self.get_state(run_id)
+        if _is_paused_run_vars(run.vars):
+            raise ValueError("Run is paused")
         if run.status != RunStatus.WAITING or run.waiting is None:
             raise ValueError("Run is not waiting")
 
@@ -440,14 +937,101 @@ class Runtime:
         resume_to = run.waiting.resume_to_node
         result_key = run.waiting.result_key
 
+        # Keep track of what we actually persisted for this resume (tool resumes may
+        # merge blocked-by-allowlist entries back into the payload).
+        stored_payload: Dict[str, Any] = payload
+
         if result_key:
-            _set_nested(run.vars, result_key, payload)
+            # Tool waits may carry blocked-by-allowlist metadata. External hosts typically only execute
+            # the filtered subset of tool calls and resume with results for those calls. To keep agent
+            # semantics correct (and evidence indices aligned), merge blocked entries back into the
+            # resumed payload deterministically.
+            merged_payload: Dict[str, Any] = payload
+            try:
+                details = run.waiting.details if run.waiting is not None else None
+                if isinstance(details, dict):
+                    blocked = details.get("blocked_by_index")
+                    original_count = details.get("original_call_count")
+                    results = payload.get("results") if isinstance(payload, dict) else None
+                    if (
+                        isinstance(blocked, dict)
+                        and isinstance(original_count, int)
+                        and original_count > 0
+                        and isinstance(results, list)
+                        and len(results) != original_count
+                    ):
+                        merged_results: list[Any] = []
+                        executed_iter = iter(results)
+
+                        for idx in range(original_count):
+                            blocked_entry = blocked.get(str(idx))
+                            if isinstance(blocked_entry, dict):
+                                merged_results.append(blocked_entry)
+                                continue
+                            try:
+                                merged_results.append(next(executed_iter))
+                            except StopIteration:
+                                merged_results.append(
+                                    {
+                                        "call_id": "",
+                                        "name": "",
+                                        "success": False,
+                                        "output": None,
+                                        "error": "Missing tool result",
+                                    }
+                                )
+
+                        merged_payload = dict(payload)
+                        merged_payload["results"] = merged_results
+                        merged_payload.setdefault("mode", "executed")
+            except Exception:
+                merged_payload = payload
+
+            _set_nested(run.vars, result_key, merged_payload)
+            stored_payload = merged_payload
+            # Passthrough tool execution: the host resumes with tool results. We still want
+            # evidence capture and payload-bounding (store large parts as artifacts) before
+            # the run continues.
+            try:
+                details = run.waiting.details if run.waiting is not None else None
+                tool_calls_for_evidence = None
+                if isinstance(details, dict):
+                    tool_calls_for_evidence = details.get("tool_calls_for_evidence")
+                    if not isinstance(tool_calls_for_evidence, list):
+                        tool_calls_for_evidence = details.get("tool_calls")
+
+                if isinstance(tool_calls_for_evidence, list):
+                    from ..evidence import EvidenceRecorder
+
+                    artifact_store = self._artifact_store
+                    if artifact_store is not None and isinstance(payload, dict):
+                        EvidenceRecorder(artifact_store=artifact_store).record_tool_calls(
+                            run=run,
+                            node_id=str(run.current_node or ""),
+                            tool_calls=list(tool_calls_for_evidence or []),
+                            tool_results=merged_payload,
+                        )
+            except Exception:
+                pass
+
+        # Terminal waiting node: if there is no resume target, treat the resume payload as
+        # the final output instead of re-executing the waiting node again (which would
+        # otherwise create an infinite wait/resume loop).
+        if resume_to is None:
+            run.status = RunStatus.COMPLETED
+            run.waiting = None
+            run.output = {"success": True, "result": stored_payload}
+            run.updated_at = utc_now_iso()
+            self._run_store.save(run)
+            return run
 
         self._apply_resume_payload(run, payload=payload, override_node=resume_to)
         run.updated_at = utc_now_iso()
         self._run_store.save(run)
 
-        return self.tick(workflow=workflow, run_id=run_id)
+        if max_steps <= 0:
+            return run
+        return self.tick(workflow=workflow, run_id=run_id, max_steps=max_steps)
 
     # ---------------------------------------------------------------------
     # Internals
@@ -457,8 +1041,88 @@ class Runtime:
         self._handlers[EffectType.WAIT_EVENT] = self._handle_wait_event
         self._handlers[EffectType.WAIT_UNTIL] = self._handle_wait_until
         self._handlers[EffectType.ASK_USER] = self._handle_ask_user
+        self._handlers[EffectType.ANSWER_USER] = self._handle_answer_user
+        self._handlers[EffectType.EMIT_EVENT] = self._handle_emit_event
+        self._handlers[EffectType.MEMORY_QUERY] = self._handle_memory_query
+        self._handlers[EffectType.MEMORY_TAG] = self._handle_memory_tag
+        self._handlers[EffectType.MEMORY_COMPACT] = self._handle_memory_compact
+        self._handlers[EffectType.MEMORY_NOTE] = self._handle_memory_note
+        self._handlers[EffectType.MEMORY_REHYDRATE] = self._handle_memory_rehydrate
+        self._handlers[EffectType.VARS_QUERY] = self._handle_vars_query
         self._handlers[EffectType.START_SUBWORKFLOW] = self._handle_start_subworkflow
 
+    # Built-in memory helpers ------------------------------------------------
+
+    def _global_memory_run_id(self) -> str:
+        """Return the global memory run id (stable).
+
+        Hosts can override via `ABSTRACTRUNTIME_GLOBAL_MEMORY_RUN_ID`.
+        """
+        rid = os.environ.get("ABSTRACTRUNTIME_GLOBAL_MEMORY_RUN_ID")
+        rid = str(rid or "").strip()
+        if rid and _SAFE_RUN_ID_PATTERN.match(rid):
+            return rid
+        return _DEFAULT_GLOBAL_MEMORY_RUN_ID
+
+    def _ensure_global_memory_run(self) -> RunState:
+        """Load or create the global memory run used as the owner for `scope="global"` spans."""
+        rid = self._global_memory_run_id()
+        existing = self._run_store.load(rid)
+        if existing is not None:
+            return existing
+
+        run = RunState(
+            run_id=rid,
+            workflow_id="__global_memory__",
+            status=RunStatus.COMPLETED,
+            current_node="done",
+            vars={
+                "context": {"task": "", "messages": []},
+                "scratchpad": {},
+                "_runtime": {"memory_spans": []},
+                "_temp": {},
+                "_limits": {},
+            },
+            waiting=None,
+            output={"messages": []},
+            error=None,
+            created_at=utc_now_iso(),
+            updated_at=utc_now_iso(),
+            actor_id=None,
+            session_id=None,
+            parent_run_id=None,
+        )
+        self._run_store.save(run)
+        return run
+
+    def _resolve_session_root_run(self, run: RunState) -> RunState:
+        """Resolve the root run of the current run-tree (walk `parent_run_id`)."""
+        cur = run
+        seen: set[str] = set()
+        while True:
+            parent_id = getattr(cur, "parent_run_id", None)
+            if not isinstance(parent_id, str) or not parent_id.strip():
+                return cur
+            pid = parent_id.strip()
+            if pid in seen:
+                # Defensive: break cycles.
+                return cur
+            seen.add(pid)
+            parent = self._run_store.load(pid)
+            if parent is None:
+                return cur
+            cur = parent
+
+    def _resolve_scope_owner_run(self, base_run: RunState, *, scope: str) -> RunState:
+        s = str(scope or "").strip().lower() or "run"
+        if s == "run":
+            return base_run
+        if s == "session":
+            return self._resolve_session_root_run(base_run)
+        if s == "global":
+            return self._ensure_global_memory_run()
+        raise ValueError(f"Unknown memory scope: {scope}")
+
     def _find_prior_completed_result(
         self, run_id: str, idempotency_key: str
     ) -> Optional[Dict[str, Any]]:
@@ -579,16 +1243,226 @@ class Runtime:
     def _handle_wait_event(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
         wait_key = effect.payload.get("wait_key")
         if not wait_key:
-            return EffectOutcome.failed("wait_event requires payload.wait_key")
+            # Allow structured payloads (scope/name) so hosts can compute stable keys.
+            scope = effect.payload.get("scope", "session")
+            name = effect.payload.get("name") or effect.payload.get("event_name")
+            if not isinstance(name, str) or not name.strip():
+                return EffectOutcome.failed("wait_event requires payload.wait_key or payload.name")
+
+            session_id = effect.payload.get("session_id") or run.session_id or run.run_id
+            try:
+                wait_key = build_event_wait_key(
+                    scope=str(scope or "session"),
+                    name=str(name),
+                    session_id=str(session_id) if session_id is not None else None,
+                    workflow_id=run.workflow_id,
+                    run_id=run.run_id,
+                )
+            except Exception as e:
+                return EffectOutcome.failed(f"wait_event invalid payload: {e}")
         resume_to = effect.payload.get("resume_to_node") or default_next_node
+        # Optional UX metadata for hosts:
+        # - "prompt"/"choices"/"allow_free_text" enable durable human-in-the-loop
+        #   waits using EVENT as the wakeup mechanism (useful for thin clients).
+        prompt: Optional[str] = None
+        try:
+            p = effect.payload.get("prompt")
+            if isinstance(p, str) and p.strip():
+                prompt = p
+        except Exception:
+            prompt = None
+
+        choices: Optional[List[str]] = None
+        try:
+            raw_choices = effect.payload.get("choices")
+            if isinstance(raw_choices, list):
+                normalized: List[str] = []
+                for c in raw_choices:
+                    if isinstance(c, str) and c.strip():
+                        normalized.append(c.strip())
+                choices = normalized
+        except Exception:
+            choices = None
+
+        allow_free_text = True
+        try:
+            aft = effect.payload.get("allow_free_text")
+            if aft is None:
+                aft = effect.payload.get("allowFreeText")
+            if aft is not None:
+                allow_free_text = bool(aft)
+        except Exception:
+            allow_free_text = True
+
+        details = None
+        try:
+            d = effect.payload.get("details")
+            if isinstance(d, dict):
+                details = dict(d)
+        except Exception:
+            details = None
         wait = WaitState(
             reason=WaitReason.EVENT,
             wait_key=str(wait_key),
             resume_to_node=resume_to,
             result_key=effect.result_key,
+            prompt=prompt,
+            choices=choices,
+            allow_free_text=allow_free_text,
+            details=details,
         )
         return EffectOutcome.waiting(wait)
 
+    def _handle_emit_event(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Emit a durable event and resume matching WAIT_EVENT runs.
+
+        Payload:
+          - name: str (required)  event name
+          - scope: str (optional, default "session")  "session" | "workflow" | "run" | "global"
+          - session_id: str (optional)  target session id (for cross-workflow targeted delivery)
+          - payload: dict (optional)  event payload delivered to listeners
+          - max_steps: int (optional, default 100)  tick budget per resumed run
+
+        Notes:
+        - This is durable because it resumes WAIT_EVENT runs via Runtime.resume(), which
+          checkpoints run state and appends ledger records for subsequent steps.
+        - Delivery is best-effort and at-least-once; listeners should be idempotent if needed.
+        """
+        name = effect.payload.get("name") or effect.payload.get("event_name")
+        if not isinstance(name, str) or not name.strip():
+            return EffectOutcome.failed("emit_event requires payload.name")
+
+        scope = effect.payload.get("scope", "session")
+        target_session_id = effect.payload.get("session_id")
+        payload = effect.payload.get("payload") or {}
+        if not isinstance(payload, dict):
+            payload = {"value": payload}
+
+        # NOTE: we intentionally resume listeners with max_steps=0 (no execution).
+        # Hosts (web backend, workers, schedulers) should drive RUNNING runs and
+        # stream their StepRecords deterministically (better observability and UX).
+        try:
+            max_steps = int(effect.payload.get("max_steps", 0) or 0)
+        except Exception:
+            max_steps = 0
+        if max_steps < 0:
+            max_steps = 0
+
+        # Determine target scope id (default: current session/run).
+        session_id = target_session_id
+        if session_id is None and str(scope or "session").strip().lower() == "session":
+            session_id = run.session_id or run.run_id
+
+        try:
+            wait_key = build_event_wait_key(
+                scope=str(scope or "session"),
+                name=str(name),
+                session_id=str(session_id) if session_id is not None else None,
+                workflow_id=run.workflow_id,
+                run_id=run.run_id,
+            )
+        except Exception as e:
+            return EffectOutcome.failed(f"emit_event invalid payload: {e}")
+
+        # Wildcard listeners ("*") receive all events within the same scope_id.
+        wildcard_wait_key: Optional[str] = None
+        try:
+            wildcard_wait_key = build_event_wait_key(
+                scope=str(scope or "session"),
+                name="*",
+                session_id=str(session_id) if session_id is not None else None,
+                workflow_id=run.workflow_id,
+                run_id=run.run_id,
+            )
+        except Exception:
+            wildcard_wait_key = None
+
+        if self._workflow_registry is None:
+            return EffectOutcome.failed(
+                "emit_event requires a workflow_registry to resume target runs. "
+                "Set it via Runtime(workflow_registry=...) or runtime.set_workflow_registry(...)."
+            )
+
+        if not isinstance(self._run_store, QueryableRunStore):
+            return EffectOutcome.failed(
+                "emit_event requires a QueryableRunStore to find waiting runs. "
+                "Use InMemoryRunStore/JsonFileRunStore or provide a queryable store."
+            )
+
+        # Find all runs waiting for this event key.
+        candidates = self._run_store.list_runs(
+            status=RunStatus.WAITING,
+            wait_reason=WaitReason.EVENT,
+            limit=10_000,
+        )
+
+        delivered_to: list[str] = []
+        resumed: list[Dict[str, Any]] = []
+        envelope = {
+            "event_id": effect.payload.get("event_id") or None,
+            "name": str(name),
+            "scope": str(scope or "session"),
+            "session_id": str(session_id) if session_id is not None else None,
+            "payload": dict(payload),
+            "emitted_at": utc_now_iso(),
+            "emitter": {
+                "run_id": run.run_id,
+                "workflow_id": run.workflow_id,
+                "node_id": run.current_node,
+            },
+        }
+
+        available_in_session: list[str] = []
+        prefix = f"evt:session:{session_id}:"
+
+        for r in candidates:
+            if _is_paused_run_vars(getattr(r, "vars", None)):
+                continue
+            w = getattr(r, "waiting", None)
+            if w is None:
+                continue
+            wk = getattr(w, "wait_key", None)
+            if isinstance(wk, str) and wk.startswith(prefix):
+                # Help users debug name mismatches (best-effort).
+                suffix = wk[len(prefix) :]
+                if suffix and suffix not in available_in_session and len(available_in_session) < 15:
+                    available_in_session.append(suffix)
+            if wk != wait_key and (wildcard_wait_key is None or wk != wildcard_wait_key):
+                continue
+
+            wf = self._workflow_registry.get(r.workflow_id)
+            if wf is None:
+                # Can't resume without the spec; skip but include diagnostic in result.
+                resumed.append({"run_id": r.run_id, "status": "skipped", "error": "workflow_not_registered"})
+                continue
+
+            try:
+                # Resume using the run's own wait_key (supports wildcard listeners).
+                resume_key = wk if isinstance(wk, str) and wk else None
+                new_state = self.resume(
+                    workflow=wf,
+                    run_id=r.run_id,
+                    wait_key=resume_key,
+                    payload=envelope,
+                    max_steps=max_steps,
+                )
+                delivered_to.append(r.run_id)
+                resumed.append({"run_id": r.run_id, "status": new_state.status.value})
+            except Exception as e:
+                resumed.append({"run_id": r.run_id, "status": "error", "error": str(e)})
+
+        out: Dict[str, Any] = {
+            "wait_key": wait_key,
+            "name": str(name),
+            "scope": str(scope or "session"),
+            "delivered": len(delivered_to),
+            "delivered_to": delivered_to,
+            "resumed": resumed,
+        }
+        if not delivered_to and available_in_session:
+            out["available_listeners_in_session"] = available_in_session
+        return EffectOutcome.completed(out)
+
     def _handle_wait_until(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
         until = effect.payload.get("until")
         if not until:
@@ -628,6 +1502,22 @@ class Runtime:
         )
         return EffectOutcome.waiting(wait)
 
+    def _handle_answer_user(
+        self, run: RunState, effect: Effect, default_next_node: Optional[str]
+    ) -> EffectOutcome:
+        """Handle ANSWER_USER effect.
+
+        This effect is intentionally non-blocking: it completes immediately and
+        returns the message payload so the host UI can render it.
+        """
+        message = effect.payload.get("message")
+        if message is None:
+            # Backward/compat convenience aliases.
+            message = effect.payload.get("text") or effect.payload.get("content")
+        if message is None:
+            return EffectOutcome.failed("answer_user requires payload.message")
+        return EffectOutcome.completed({"message": str(message)})
+
     def _handle_start_subworkflow(
         self, run: RunState, effect: Effect, default_next_node: Optional[str]
     ) -> EffectOutcome:
@@ -664,6 +1554,8 @@ class Runtime:
 
         sub_vars = effect.payload.get("vars") or {}
         is_async = bool(effect.payload.get("async", False))
+        wait_for_completion = bool(effect.payload.get("wait", False))
+        include_traces = bool(effect.payload.get("include_traces", False))
         resume_to = effect.payload.get("resume_to_node") or default_next_node
 
         # Start the subworkflow with parent tracking
@@ -676,8 +1568,29 @@ class Runtime:
         )
 
         if is_async:
-            # Async mode: return immediately with sub_run_id
-            # The child is started but not ticked - caller is responsible for driving it
+            # Async mode: start the child and return immediately.
+            #
+            # If `wait=True`, we *also* transition the parent into a durable WAITING state
+            # so a host (e.g. AbstractFlow WebSocket runner loop) can:
+            # - tick the child run incrementally (and stream node traces in real time)
+            # - resume the parent once the child completes (by calling runtime.resume(...))
+            #
+            # Without `wait=True`, this remains fire-and-forget.
+            if wait_for_completion:
+                wait = WaitState(
+                    reason=WaitReason.SUBWORKFLOW,
+                    wait_key=f"subworkflow:{sub_run_id}",
+                    resume_to_node=resume_to,
+                    result_key=effect.result_key,
+                    details={
+                        "sub_run_id": sub_run_id,
+                        "sub_workflow_id": workflow_id,
+                        "async": True,
+                    },
+                )
+                return EffectOutcome.waiting(wait)
+
+            # Fire-and-forget: caller is responsible for driving/observing the child.
             return EffectOutcome.completed({"sub_run_id": sub_run_id, "async": True})
 
         # Sync mode: run the subworkflow until completion or waiting
@@ -689,10 +1602,10 @@ class Runtime:
 
         if sub_state.status == RunStatus.COMPLETED:
             # Subworkflow completed - return its output
-            return EffectOutcome.completed({
-                "sub_run_id": sub_run_id,
-                "output": sub_state.output,
-            })
+            result: Dict[str, Any] = {"sub_run_id": sub_run_id, "output": sub_state.output}
+            if include_traces:
+                result["node_traces"] = self.get_node_traces(sub_run_id)
+            return EffectOutcome.completed(result)
 
         if sub_state.status == RunStatus.FAILED:
             # Subworkflow failed - propagate error
@@ -721,6 +1634,1523 @@ class Runtime:
         # Unexpected status
         return EffectOutcome.failed(f"Unexpected subworkflow status: {sub_state.status.value}")
 
+    # Built-in memory handlers ---------------------------------------------
+
+    def _handle_memory_query(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle MEMORY_QUERY.
+
+        This effect supports provenance-first recall over archived memory spans stored in ArtifactStore.
+        It is intentionally metadata-first and embedding-free (semantic retrieval belongs in AbstractMemory).
+
+        Payload (all optional unless otherwise stated):
+          - span_id: str | int | list[str|int]  (artifact_id or 1-based index into _runtime.memory_spans)
+          - query: str                          (keyword substring match)
+          - since: str                          (ISO8601, span intersection filter)
+          - until: str                          (ISO8601, span intersection filter)
+          - tags: dict[str, str|list[str]]      (span tag filter; values can be multi-valued)
+          - tags_mode: "all"|"any"              (default "all"; AND/OR across tag keys)
+          - authors: list[str]                  (alias: usernames; matches span.created_by case-insensitively)
+          - locations: list[str]                (matches span.location case-insensitively)
+          - limit_spans: int                    (default 5)
+          - deep: bool                          (default True when query is set; scans archived messages)
+          - deep_limit_spans: int               (default 50)
+          - deep_limit_messages_per_span: int   (default 400)
+          - connected: bool                     (include connected spans via time adjacency + shared tags)
+          - neighbor_hops: int                  (default 1 when connected=True)
+          - connect_by: list[str]               (default ["topic","person"])
+          - max_messages: int                   (default 80; total messages rendered across all spans)
+          - tool_name: str                      (default "recall_memory"; for formatting)
+          - call_id: str                        (tool-call id passthrough)
+        """
+        from .vars import ensure_namespaces, parse_vars_path, resolve_vars_path
+
+        ensure_namespaces(run.vars)
+        runtime_ns = run.vars.get("_runtime")
+        if not isinstance(runtime_ns, dict):
+            runtime_ns = {}
+            run.vars["_runtime"] = runtime_ns
+
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            return EffectOutcome.failed(
+                "MEMORY_QUERY requires an ArtifactStore; configure runtime.set_artifact_store(...)"
+            )
+
+        payload = dict(effect.payload or {})
+        tool_name = str(payload.get("tool_name") or "recall_memory")
+        call_id = str(payload.get("call_id") or "memory")
+
+        # Scope routing (run-tree/global). Scope affects which run owns the span index queried.
+        scope = str(payload.get("scope") or "run").strip().lower() or "run"
+        if scope not in {"run", "session", "global", "all"}:
+            return EffectOutcome.failed(f"Unknown memory_query scope: {scope}")
+
+        # Return mode controls whether we include structured meta in the tool result.
+        return_mode = str(payload.get("return") or payload.get("return_mode") or "rendered").strip().lower() or "rendered"
+        if return_mode not in {"rendered", "meta", "both"}:
+            return EffectOutcome.failed(f"Unknown memory_query return mode: {return_mode}")
+
+        query = payload.get("query")
+        query_text = str(query or "").strip()
+        since = payload.get("since")
+        until = payload.get("until")
+        tags = payload.get("tags")
+        tags_dict: Optional[Dict[str, Any]] = None
+        if isinstance(tags, dict):
+            # Accept str or list[str] values. Ignore reserved key "kind".
+            out_tags: Dict[str, Any] = {}
+            for k, v in tags.items():
+                if not isinstance(k, str) or not k.strip():
+                    continue
+                if k == "kind":
+                    continue
+                if isinstance(v, str) and v.strip():
+                    out_tags[k.strip()] = v.strip()
+                elif isinstance(v, (list, tuple)):
+                    vals = [str(x).strip() for x in v if isinstance(x, str) and str(x).strip()]
+                    if vals:
+                        out_tags[k.strip()] = vals
+            tags_dict = out_tags or None
+
+        tags_mode_raw = payload.get("tags_mode")
+        if tags_mode_raw is None:
+            tags_mode_raw = payload.get("tagsMode")
+        if tags_mode_raw is None:
+            tags_mode_raw = payload.get("tag_mode")
+        tags_mode = str(tags_mode_raw or "all").strip().lower() or "all"
+        if tags_mode in {"and"}:
+            tags_mode = "all"
+        if tags_mode in {"or"}:
+            tags_mode = "any"
+        if tags_mode not in {"all", "any"}:
+            tags_mode = "all"
+
+        def _norm_str_list(value: Any) -> list[str]:
+            if value is None:
+                return []
+            if isinstance(value, str):
+                v = value.strip()
+                return [v] if v else []
+            if not isinstance(value, list):
+                return []
+            out: list[str] = []
+            for x in value:
+                if isinstance(x, str) and x.strip():
+                    out.append(x.strip())
+            # preserve order but dedup (case-insensitive)
+            seen: set[str] = set()
+            deduped: list[str] = []
+            for s in out:
+                key = s.lower()
+                if key in seen:
+                    continue
+                seen.add(key)
+                deduped.append(s)
+            return deduped
+
+        authors = _norm_str_list(payload.get("authors") if "authors" in payload else payload.get("usernames"))
+        if not authors:
+            authors = _norm_str_list(payload.get("users"))
+        locations = _norm_str_list(payload.get("locations") if "locations" in payload else payload.get("location"))
+
+        try:
+            limit_spans = int(payload.get("limit_spans", 5) or 5)
+        except Exception:
+            limit_spans = 5
+        if limit_spans < 1:
+            limit_spans = 1
+
+        deep = payload.get("deep")
+        if deep is None:
+            deep_enabled = bool(query_text)
+        else:
+            deep_enabled = bool(deep)
+
+        try:
+            deep_limit_spans = int(payload.get("deep_limit_spans", 50) or 50)
+        except Exception:
+            deep_limit_spans = 50
+        if deep_limit_spans < 1:
+            deep_limit_spans = 1
+
+        try:
+            deep_limit_messages_per_span = int(payload.get("deep_limit_messages_per_span", 400) or 400)
+        except Exception:
+            deep_limit_messages_per_span = 400
+        if deep_limit_messages_per_span < 1:
+            deep_limit_messages_per_span = 1
+
+        connected = bool(payload.get("connected", False))
+        try:
+            neighbor_hops = int(payload.get("neighbor_hops", 1) or 1)
+        except Exception:
+            neighbor_hops = 1
+        if neighbor_hops < 0:
+            neighbor_hops = 0
+
+        connect_by = payload.get("connect_by")
+        if isinstance(connect_by, list):
+            connect_keys = [str(x) for x in connect_by if isinstance(x, (str, int, float)) and str(x).strip()]
+        else:
+            connect_keys = ["topic", "person"]
+
+        try:
+            max_messages = int(payload.get("max_messages", -1) or -1)
+        except Exception:
+            max_messages = -1
+        # `-1` means "no truncation" for rendered messages.
+        if max_messages < -1:
+            max_messages = -1
+        if max_messages != -1 and max_messages < 1:
+            max_messages = 1
+
+        from ..memory.active_context import ActiveContextPolicy, TimeRange
+
+        # Select run(s) to query.
+        runs_to_query: list[RunState] = []
+        if scope == "run":
+            runs_to_query = [run]
+        elif scope == "session":
+            runs_to_query = [self._resolve_scope_owner_run(run, scope="session")]
+        elif scope == "global":
+            runs_to_query = [self._resolve_scope_owner_run(run, scope="global")]
+        else:  # all
+            # Deterministic order; dedup by run_id.
+            root = self._resolve_scope_owner_run(run, scope="session")
+            global_run = self._resolve_scope_owner_run(run, scope="global")
+            seen_ids: set[str] = set()
+            for r in (run, root, global_run):
+                if r.run_id in seen_ids:
+                    continue
+                seen_ids.add(r.run_id)
+                runs_to_query.append(r)
+
+        # Collect per-run span indexes (metadata) and summary maps for rendering.
+        spans_by_run_id: dict[str, list[dict[str, Any]]] = {}
+        all_spans: list[dict[str, Any]] = []
+        all_summary_by_artifact: dict[str, str] = {}
+        for target in runs_to_query:
+            spans = ActiveContextPolicy.list_memory_spans_from_run(target)
+            # `memory_spans` is a general span-like index (conversation spans, notes, evidence, etc).
+            # MEMORY_QUERY is specifically for provenance-first *memory recall*, not evidence retrieval.
+            spans = [s for s in spans if not (isinstance(s, dict) and str(s.get("kind") or "") == "evidence")]
+            spans_by_run_id[target.run_id] = spans
+            all_spans.extend([dict(s) for s in spans if isinstance(s, dict)])
+            all_summary_by_artifact.update(ActiveContextPolicy.summary_text_by_artifact_id_from_run(target))
+
+        # Resolve explicit span ids if provided.
+        span_id_payload = payload.get("span_id")
+        span_ids_payload = payload.get("span_ids")
+        explicit_ids = span_ids_payload if isinstance(span_ids_payload, list) else span_id_payload
+
+        all_selected: list[str] = []
+
+        if explicit_ids is not None:
+            explicit_list = list(explicit_ids) if isinstance(explicit_ids, list) else [explicit_ids]
+
+            # Indices are inherently scoped to a single run's span list; for `scope="all"`,
+            # require stable artifact ids to avoid ambiguity.
+            if scope == "all":
+                for x in explicit_list:
+                    if isinstance(x, int):
+                        return EffectOutcome.failed("memory_query scope='all' requires explicit span_ids as artifact ids (no indices)")
+                    if isinstance(x, str) and x.strip().isdigit():
+                        return EffectOutcome.failed("memory_query scope='all' requires explicit span_ids as artifact ids (no indices)")
+                # Treat as artifact ids.
+                all_selected = _dedup_preserve_order([str(x).strip() for x in explicit_list if str(x).strip()])
+            else:
+                # Single-run resolution for indices.
+                target = runs_to_query[0]
+                spans = spans_by_run_id.get(target.run_id, [])
+                all_selected = ActiveContextPolicy.resolve_span_ids_from_spans(explicit_list, spans)
+        else:
+            # Filter spans per target and union.
+            time_range = None
+            if since or until:
+                time_range = TimeRange(
+                    start=str(since) if since else None,
+                    end=str(until) if until else None,
+                )
+
+            for target in runs_to_query:
+                spans = spans_by_run_id.get(target.run_id, [])
+                matches = ActiveContextPolicy.filter_spans_from_run(
+                    target,
+                    artifact_store=artifact_store,
+                    time_range=time_range,
+                    tags=tags_dict,
+                    tags_mode=tags_mode,
+                    authors=authors or None,
+                    locations=locations or None,
+                    query=query_text or None,
+                    limit=limit_spans,
+                )
+                selected = [str(s.get("artifact_id") or "") for s in matches if isinstance(s, dict) and s.get("artifact_id")]
+
+                if deep_enabled and query_text:
+                    # Deep scan is bounded and should respect metadata filters (tags/authors/locations/time).
+                    deep_candidates = ActiveContextPolicy.filter_spans_from_run(
+                        target,
+                        artifact_store=artifact_store,
+                        time_range=time_range,
+                        tags=tags_dict,
+                        tags_mode=tags_mode,
+                        authors=authors or None,
+                        locations=locations or None,
+                        query=None,
+                        limit=deep_limit_spans,
+                    )
+                    selected = _dedup_preserve_order(
+                        selected
+                        + _deep_scan_span_ids(
+                            spans=deep_candidates,
+                            artifact_store=artifact_store,
+                            query=query_text,
+                            limit_spans=deep_limit_spans,
+                            limit_messages_per_span=deep_limit_messages_per_span,
+                        )
+                    )
+
+                if connected and selected:
+                    connect_candidates = ActiveContextPolicy.filter_spans_from_run(
+                        target,
+                        artifact_store=artifact_store,
+                        time_range=time_range,
+                        tags=tags_dict,
+                        tags_mode=tags_mode,
+                        authors=authors or None,
+                        locations=locations or None,
+                        query=None,
+                        limit=max(1000, len(spans)),
+                    )
+                    selected = _dedup_preserve_order(
+                        _expand_connected_span_ids(
+                            spans=connect_candidates,
+                            seed_artifact_ids=selected,
+                            connect_keys=connect_keys,
+                            neighbor_hops=neighbor_hops,
+                            limit=max(limit_spans, len(selected)),
+                        )
+                    )
+
+                all_selected = _dedup_preserve_order(all_selected + selected)
+
+        rendered_text = ""
+        if return_mode in {"rendered", "both"}:
+            # Render output (provenance + messages). Note: this may load artifacts.
+            rendered_text = _render_memory_query_output(
+                spans=all_spans,
+                artifact_store=artifact_store,
+                selected_artifact_ids=all_selected,
+                summary_by_artifact=all_summary_by_artifact,
+                max_messages=max_messages,
+            )
+
+        # Structured meta output (for deterministic workflows).
+        meta: dict[str, Any] = {}
+        if return_mode in {"meta", "both"}:
+            # Index span record by artifact id (first match wins deterministically).
+            by_artifact: dict[str, dict[str, Any]] = {}
+            for s in all_spans:
+                try:
+                    aid = str(s.get("artifact_id") or "").strip()
+                except Exception:
+                    aid = ""
+                if not aid or aid in by_artifact:
+                    continue
+                by_artifact[aid] = s
+
+            matches: list[dict[str, Any]] = []
+            for aid in all_selected:
+                span = by_artifact.get(aid)
+                if not isinstance(span, dict):
+                    continue
+                m: dict[str, Any] = {
+                    "span_id": aid,
+                    "kind": span.get("kind"),
+                    "created_at": span.get("created_at"),
+                    "from_timestamp": span.get("from_timestamp"),
+                    "to_timestamp": span.get("to_timestamp"),
+                    "tags": span.get("tags") if isinstance(span.get("tags"), dict) else {},
+                }
+                for k in ("created_by", "location"):
+                    if k in span:
+                        m[k] = span.get(k)
+                # Include known preview fields without enforcing a global schema.
+                for k in ("note_preview", "message_count", "summary_message_id"):
+                    if k in span:
+                        m[k] = span.get(k)
+                matches.append(m)
+
+            meta = {"matches": matches, "span_ids": list(all_selected)}
+
+        result = {
+            "mode": "executed",
+            "results": [
+                {
+                    "call_id": call_id,
+                    "name": tool_name,
+                    "success": True,
+                    "output": rendered_text if return_mode in {"rendered", "both"} else "",
+                    "error": None,
+                    "meta": meta if meta else None,
+                }
+            ],
+        }
+        return EffectOutcome.completed(result=result)
+
+    def _handle_vars_query(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle VARS_QUERY.
+
+        This is a JSON-safe, runtime-owned introspection primitive intended for:
+        - progressive recall/debugging (e.g., inspect `scratchpad`)
+        - host tooling parity (schema-only tools that map to runtime effects)
+
+        Payload (all optional unless stated):
+          - path: str                 (default "scratchpad"; supports dot path or JSON pointer "/a/b/0")
+          - keys_only: bool           (default False; when True, return keys/length instead of full value)
+          - target_run_id: str        (optional; inspect another run state)
+          - tool_name: str            (default "inspect_vars"; for tool-style output)
+          - call_id: str              (tool-call id passthrough)
+        """
+        import json
+
+        from .vars import ensure_namespaces, parse_vars_path, resolve_vars_path
+
+        payload = dict(effect.payload or {})
+        tool_name = str(payload.get("tool_name") or "inspect_vars")
+        call_id = str(payload.get("call_id") or "vars")
+
+        target_run_id = payload.get("target_run_id")
+        if target_run_id is not None:
+            target_run_id = str(target_run_id).strip() or None
+
+        path = payload.get("path")
+        if path is None:
+            path = payload.get("var_path")
+        path_text = str(path or "").strip() or "scratchpad"
+
+        keys_only = bool(payload.get("keys_only", False))
+
+        target_run = run
+        if target_run_id and target_run_id != run.run_id:
+            loaded = self._run_store.load(target_run_id)
+            if loaded is None:
+                return EffectOutcome.completed(
+                    result={
+                        "mode": "executed",
+                        "results": [
+                            {
+                                "call_id": call_id,
+                                "name": tool_name,
+                                "success": False,
+                                "output": None,
+                                "error": f"Unknown target_run_id: {target_run_id}",
+                            }
+                        ],
+                    }
+                )
+            target_run = loaded
+
+        ensure_namespaces(target_run.vars)
+
+        try:
+            tokens = parse_vars_path(path_text)
+            value = resolve_vars_path(target_run.vars, tokens)
+        except Exception as e:
+            return EffectOutcome.completed(
+                result={
+                    "mode": "executed",
+                    "results": [
+                        {
+                            "call_id": call_id,
+                            "name": tool_name,
+                            "success": False,
+                            "output": None,
+                            "error": str(e),
+                        }
+                    ],
+                }
+            )
+
+        out: Dict[str, Any] = {"path": path_text, "type": type(value).__name__}
+        if keys_only:
+            if isinstance(value, dict):
+                out["keys"] = sorted([str(k) for k in value.keys()])
+            elif isinstance(value, list):
+                out["length"] = len(value)
+            else:
+                out["value"] = value
+        else:
+            out["value"] = value
+
+        text = json.dumps(out, ensure_ascii=False, indent=2, sort_keys=True, default=str)
+
+        return EffectOutcome.completed(
+            result={
+                "mode": "executed",
+                "results": [
+                    {
+                        "call_id": call_id,
+                        "name": tool_name,
+                        "success": True,
+                        "output": text,
+                        "error": None,
+                    }
+                ],
+            }
+        )
+
+    def _handle_memory_tag(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle MEMORY_TAG.
+
+        Payload (required unless stated):
+          - span_id: str | int   (artifact_id or 1-based index into `_runtime.memory_spans`)
+          - tags: dict[str,str]  (merged into span["tags"] by default)
+          - merge: bool          (optional, default True; when False, replaces span["tags"])
+          - tool_name: str       (optional; for tool-style output, default "remember")
+          - call_id: str         (optional; passthrough for tool-style output)
+
+        Notes:
+        - This mutates the in-run span index (`_runtime.memory_spans`) only; it does not change artifacts.
+        - Tagging is intentionally JSON-safe (string->string).
+        """
+        import json
+
+        from .vars import ensure_namespaces
+
+        ensure_namespaces(run.vars)
+        runtime_ns = run.vars.get("_runtime")
+        if not isinstance(runtime_ns, dict):
+            runtime_ns = {}
+            run.vars["_runtime"] = runtime_ns
+
+        spans = runtime_ns.get("memory_spans")
+        if not isinstance(spans, list):
+            return EffectOutcome.failed("MEMORY_TAG requires _runtime.memory_spans to be a list")
+
+        payload = dict(effect.payload or {})
+        tool_name = str(payload.get("tool_name") or "remember")
+        call_id = str(payload.get("call_id") or "memory")
+
+        span_id = payload.get("span_id")
+        tags = payload.get("tags")
+        if span_id is None:
+            return EffectOutcome.failed("MEMORY_TAG requires payload.span_id")
+        if not isinstance(tags, dict) or not tags:
+            return EffectOutcome.failed("MEMORY_TAG requires payload.tags as a non-empty dict[str,str]")
+
+        merge = bool(payload.get("merge", True))
+
+        clean_tags: Dict[str, str] = {}
+        for k, v in tags.items():
+            if isinstance(k, str) and isinstance(v, str) and k and v:
+                clean_tags[k] = v
+        if not clean_tags:
+            return EffectOutcome.failed("MEMORY_TAG requires at least one non-empty string tag")
+
+        artifact_id: Optional[str] = None
+        target_index: Optional[int] = None
+
+        if isinstance(span_id, int):
+            idx = span_id - 1
+            if idx < 0 or idx >= len(spans):
+                return EffectOutcome.failed(f"Unknown span index: {span_id}")
+            span = spans[idx]
+            if not isinstance(span, dict):
+                return EffectOutcome.failed(f"Invalid span record at index {span_id}")
+            artifact_id = str(span.get("artifact_id") or "").strip() or None
+            target_index = idx
+        elif isinstance(span_id, str):
+            s = span_id.strip()
+            if not s:
+                return EffectOutcome.failed("MEMORY_TAG requires a non-empty span_id")
+            if s.isdigit():
+                idx = int(s) - 1
+                if idx < 0 or idx >= len(spans):
+                    return EffectOutcome.failed(f"Unknown span index: {s}")
+                span = spans[idx]
+                if not isinstance(span, dict):
+                    return EffectOutcome.failed(f"Invalid span record at index {s}")
+                artifact_id = str(span.get("artifact_id") or "").strip() or None
+                target_index = idx
+            else:
+                artifact_id = s
+        else:
+            return EffectOutcome.failed("MEMORY_TAG requires span_id as str or int")
+
+        if not artifact_id:
+            return EffectOutcome.failed("Could not resolve span_id to an artifact_id")
+
+        if target_index is None:
+            for i, span in enumerate(spans):
+                if not isinstance(span, dict):
+                    continue
+                if str(span.get("artifact_id") or "") == artifact_id:
+                    target_index = i
+                    break
+
+        if target_index is None:
+            return EffectOutcome.failed(f"Unknown span_id: {artifact_id}")
+
+        target = spans[target_index]
+        if not isinstance(target, dict):
+            return EffectOutcome.failed(f"Invalid span record at index {target_index + 1}")
+
+        existing_tags = target.get("tags")
+        if not isinstance(existing_tags, dict):
+            existing_tags = {}
+
+        if merge:
+            merged_tags = dict(existing_tags)
+            merged_tags.update(clean_tags)
+        else:
+            merged_tags = dict(clean_tags)
+
+        target["tags"] = merged_tags
+        target["tagged_at"] = utc_now_iso()
+        if run.actor_id:
+            target["tagged_by"] = str(run.actor_id)
+
+        rendered_tags = json.dumps(merged_tags, ensure_ascii=False, sort_keys=True)
+        text = f"Tagged span_id={artifact_id} tags={rendered_tags}"
+
+        result = {
+            "mode": "executed",
+            "results": [
+                {
+                    "call_id": call_id,
+                    "name": tool_name,
+                    "success": True,
+                    "output": text,
+                    "error": None,
+                }
+            ],
+        }
+        return EffectOutcome.completed(result=result)
+
+    def _handle_memory_compact(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle MEMORY_COMPACT.
+
+        This is a runtime-owned compaction of a run's active context:
+        - archives the compacted messages to ArtifactStore (provenance preserved)
+        - inserts a system summary message that includes `span_id=...` (LLM-visible handle)
+        - updates `_runtime.memory_spans` index with metadata/tags
+
+        Payload (optional unless stated):
+          - preserve_recent: int        (default 6; preserves N most recent non-system messages)
+          - compression_mode: str       ("light"|"standard"|"heavy", default "standard")
+          - focus: str                  (optional; topic to prioritize)
+          - target_run_id: str          (optional; defaults to current run)
+          - tool_name: str              (optional; for tool-style output, default "compact_memory")
+          - call_id: str                (optional)
+        """
+        import json
+        from uuid import uuid4
+
+        from .vars import ensure_namespaces
+        from ..memory.compaction import normalize_messages, split_for_compaction, span_metadata_from_messages
+
+        ensure_namespaces(run.vars)
+
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            return EffectOutcome.failed(
+                "MEMORY_COMPACT requires an ArtifactStore; configure runtime.set_artifact_store(...)"
+            )
+
+        payload = dict(effect.payload or {})
+        tool_name = str(payload.get("tool_name") or "compact_memory")
+        call_id = str(payload.get("call_id") or "memory")
+
+        target_run_id = payload.get("target_run_id")
+        if target_run_id is not None:
+            target_run_id = str(target_run_id).strip() or None
+
+        try:
+            preserve_recent = int(payload.get("preserve_recent", 6) or 6)
+        except Exception:
+            preserve_recent = 6
+        if preserve_recent < 0:
+            preserve_recent = 0
+
+        compression_mode = str(payload.get("compression_mode") or "standard").strip().lower()
+        if compression_mode not in ("light", "standard", "heavy"):
+            compression_mode = "standard"
+
+        focus = payload.get("focus")
+        focus_text = str(focus).strip() if isinstance(focus, str) else ""
+        focus_text = focus_text or None
+
+        # Resolve which run is being compacted.
+        target_run = run
+        if target_run_id and target_run_id != run.run_id:
+            loaded = self._run_store.load(target_run_id)
+            if loaded is None:
+                return EffectOutcome.failed(f"Unknown target_run_id: {target_run_id}")
+            target_run = loaded
+        ensure_namespaces(target_run.vars)
+
+        ctx = target_run.vars.get("context")
+        if not isinstance(ctx, dict):
+            return EffectOutcome.failed("MEMORY_COMPACT requires vars.context to be a dict")
+        messages_raw = ctx.get("messages")
+        if not isinstance(messages_raw, list) or not messages_raw:
+            return EffectOutcome.completed(
+                result={
+                    "mode": "executed",
+                    "results": [
+                        {
+                            "call_id": call_id,
+                            "name": tool_name,
+                            "success": True,
+                            "output": "No messages to compact.",
+                            "error": None,
+                        }
+                    ],
+                }
+            )
+
+        now_iso = utc_now_iso
+        messages = normalize_messages(messages_raw, now_iso=now_iso)
+        split = split_for_compaction(messages, preserve_recent=preserve_recent)
+
+        if not split.older_messages:
+            return EffectOutcome.completed(
+                result={
+                    "mode": "executed",
+                    "results": [
+                        {
+                            "call_id": call_id,
+                            "name": tool_name,
+                            "success": True,
+                            "output": f"Nothing to compact (non-system messages <= preserve_recent={preserve_recent}).",
+                            "error": None,
+                        }
+                    ],
+                }
+            )
+
+        # ------------------------------------------------------------------
+        # 1) LLM summary - use integration layer summarizer if available
+        # ------------------------------------------------------------------
+        #
+        # When chat_summarizer is injected (from AbstractCore integration layer),
+        # use it for adaptive chunking based on max_tokens. This handles cases
+        # where the environment can't use the model's full context window
+        # (e.g., GPU memory constraints).
+        #
+        # When max_tokens == -1 (AUTO): Uses model's full capability
+        # When max_tokens > 0: Chunks messages if they exceed the limit
+
+        sub_run_id: Optional[str] = None  # Track for provenance if using fallback
+
+        if self._chat_summarizer is not None:
+            # Use AbstractCore's BasicSummarizer with adaptive chunking
+            try:
+                summarizer_result = self._chat_summarizer.summarize_chat_history(
+                    messages=split.older_messages,
+                    preserve_recent=0,  # Already split; don't preserve again
+                    focus=focus_text,
+                    compression_mode=compression_mode,
+                )
+                summary_text_out = summarizer_result.get("summary", "(summary unavailable)")
+                key_points = list(summarizer_result.get("key_points") or [])
+                confidence = summarizer_result.get("confidence")
+            except Exception as e:
+                return EffectOutcome.failed(f"Summarizer failed: {e}")
+        else:
+            # Fallback: Original prompt-based approach (for non-AbstractCore runtimes)
+            older_text = "\n".join([f"{m.get('role')}: {m.get('content')}" for m in split.older_messages])
+            focus_line = f"Focus: {focus_text}\n" if focus_text else ""
+            mode_line = f"Compression mode: {compression_mode}\n"
+
+            prompt = (
+                "You are compressing older conversation context for an agent runtime.\n"
+                "Write a faithful, compact summary that preserves decisions, constraints, names, file paths, commands, and open questions.\n"
+                "Do NOT invent details. If something is unknown, say so.\n"
+                f"{mode_line}"
+                f"{focus_line}"
+                "Return STRICT JSON with keys: summary (string), key_points (array of strings), confidence (number 0..1).\n\n"
+                "OLDER MESSAGES (to be archived):\n"
+                f"{older_text}\n"
+            )
+
+            # Best-effort output budget for the summary itself.
+            limits = target_run.vars.get("_limits") if isinstance(target_run.vars.get("_limits"), dict) else {}
+            max_out = limits.get("max_output_tokens")
+            try:
+                max_out_tokens = int(max_out) if max_out is not None else None
+            except Exception:
+                max_out_tokens = None
+
+            llm_payload: Dict[str, Any] = {"prompt": prompt}
+            if max_out_tokens is not None:
+                llm_payload["params"] = {"max_tokens": max_out_tokens}
+
+            def llm_node(sub_run: RunState, sub_ctx) -> StepPlan:
+                return StepPlan(
+                    node_id="llm",
+                    effect=Effect(type=EffectType.LLM_CALL, payload=llm_payload, result_key="_temp.llm"),
+                    next_node="done",
+                )
+
+            def done_node(sub_run: RunState, sub_ctx) -> StepPlan:
+                temp = sub_run.vars.get("_temp") if isinstance(sub_run.vars.get("_temp"), dict) else {}
+                return StepPlan(node_id="done", complete_output={"response": temp.get("llm")})
+
+            wf = WorkflowSpec(workflow_id="wf_memory_compact_llm", entry_node="llm", nodes={"llm": llm_node, "done": done_node})
+
+            sub_run_id = self.start(
+                workflow=wf,
+                vars={"context": {"prompt": prompt}, "scratchpad": {}, "_runtime": {}, "_temp": {}, "_limits": dict(limits)},
+                actor_id=run.actor_id,
+                session_id=getattr(run, "session_id", None),
+                parent_run_id=run.run_id,
+            )
+
+            sub_state = self.tick(workflow=wf, run_id=sub_run_id)
+            if sub_state.status == RunStatus.WAITING:
+                return EffectOutcome.failed("MEMORY_COMPACT does not support waiting subworkflows yet")
+            if sub_state.status == RunStatus.FAILED:
+                return EffectOutcome.failed(sub_state.error or "Compaction LLM subworkflow failed")
+            response = (sub_state.output or {}).get("response")
+            if not isinstance(response, dict):
+                response = {}
+
+            content = response.get("content")
+            content_text = "" if content is None else str(content).strip()
+            lowered = content_text.lower()
+            if any(
+                keyword in lowered
+                for keyword in (
+                    "operation not permitted",
+                    "failed to connect",
+                    "connection refused",
+                    "timed out",
+                    "timeout",
+                    "not running",
+                    "model not found",
+                )
+            ):
+                return EffectOutcome.failed(f"Compaction LLM unavailable: {content_text}")
+
+            summary_text_out = content_text
+            key_points: list[str] = []
+            confidence: Optional[float] = None
+
+            # Parse JSON if present (support fenced output).
+            if content_text:
+                candidate = content_text
+                if "```" in candidate:
+                    # extract first JSON-ish block
+                    start = candidate.find("{")
+                    end = candidate.rfind("}")
+                    if 0 <= start < end:
+                        candidate = candidate[start : end + 1]
+                try:
+                    parsed = json.loads(candidate)
+                    if isinstance(parsed, dict):
+                        if parsed.get("summary") is not None:
+                            summary_text_out = str(parsed.get("summary") or "").strip() or summary_text_out
+                        kp = parsed.get("key_points")
+                        if isinstance(kp, list):
+                            key_points = [str(x) for x in kp if isinstance(x, (str, int, float))][:20]
+                        conf = parsed.get("confidence")
+                        if isinstance(conf, (int, float)):
+                            confidence = float(conf)
+                except Exception:
+                    pass
+
+            summary_text_out = summary_text_out.strip()
+            if not summary_text_out:
+                summary_text_out = "(summary unavailable)"
+
+        # ------------------------------------------------------------------
+        # 2) Archive older messages + update run state with summary
+        # ------------------------------------------------------------------
+
+        span_meta = span_metadata_from_messages(split.older_messages)
+        artifact_payload = {
+            "messages": split.older_messages,
+            "span": span_meta,
+            "created_at": now_iso(),
+        }
+        artifact_tags: Dict[str, str] = {
+            "kind": "conversation_span",
+            "compression_mode": compression_mode,
+            "preserve_recent": str(preserve_recent),
+        }
+        if focus_text:
+            artifact_tags["focus"] = focus_text
+
+        meta = artifact_store.store_json(artifact_payload, run_id=target_run.run_id, tags=artifact_tags)
+        archived_ref = meta.artifact_id
+
+        summary_message_id = f"msg_{uuid4().hex}"
+        summary_prefix = f"[CONVERSATION HISTORY SUMMARY span_id={archived_ref}]"
+        summary_metadata: Dict[str, Any] = {
+            "message_id": summary_message_id,
+            "kind": "memory_summary",
+            "compression_mode": compression_mode,
+            "preserve_recent": preserve_recent,
+            "source_artifact_id": archived_ref,
+            "source_message_count": int(span_meta.get("message_count") or 0),
+            "source_from_timestamp": span_meta.get("from_timestamp"),
+            "source_to_timestamp": span_meta.get("to_timestamp"),
+            "source_from_message_id": span_meta.get("from_message_id"),
+            "source_to_message_id": span_meta.get("to_message_id"),
+        }
+        if focus_text:
+            summary_metadata["focus"] = focus_text
+
+        summary_message = {
+            "role": "system",
+            "content": f"{summary_prefix}: {summary_text_out}",
+            "timestamp": now_iso(),
+            "metadata": summary_metadata,
+        }
+
+        new_messages = list(split.system_messages) + [summary_message] + list(split.recent_messages)
+        ctx["messages"] = new_messages
+        if isinstance(getattr(target_run, "output", None), dict):
+            target_run.output["messages"] = new_messages
+
+        runtime_ns = target_run.vars.get("_runtime")
+        if not isinstance(runtime_ns, dict):
+            runtime_ns = {}
+            target_run.vars["_runtime"] = runtime_ns
+        spans = runtime_ns.get("memory_spans")
+        if not isinstance(spans, list):
+            spans = []
+            runtime_ns["memory_spans"] = spans
+        span_record: Dict[str, Any] = {
+            "kind": "conversation_span",
+            "artifact_id": archived_ref,
+            "created_at": now_iso(),
+            "summary_message_id": summary_message_id,
+            "from_timestamp": span_meta.get("from_timestamp"),
+            "to_timestamp": span_meta.get("to_timestamp"),
+            "from_message_id": span_meta.get("from_message_id"),
+            "to_message_id": span_meta.get("to_message_id"),
+            "message_count": int(span_meta.get("message_count") or 0),
+            "compression_mode": compression_mode,
+            "focus": focus_text,
+        }
+        if run.actor_id:
+            span_record["created_by"] = str(run.actor_id)
+        spans.append(span_record)
+
+        if target_run is not run:
+            target_run.updated_at = now_iso()
+            self._run_store.save(target_run)
+
+        out = {
+            "llm_run_id": sub_run_id,
+            "span_id": archived_ref,
+            "summary_message_id": summary_message_id,
+            "preserve_recent": preserve_recent,
+            "compression_mode": compression_mode,
+            "focus": focus_text,
+            "key_points": key_points,
+            "confidence": confidence,
+        }
+        text = f"Compacted {len(split.older_messages)} messages into span_id={archived_ref}."
+        result = {
+            "mode": "executed",
+            "results": [
+                {
+                    "call_id": call_id,
+                    "name": tool_name,
+                    "success": True,
+                    "output": text,
+                    "error": None,
+                    "meta": out,
+                }
+            ],
+        }
+        return EffectOutcome.completed(result=result)
+
+    def _handle_memory_note(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle MEMORY_NOTE.
+
+        Store a small, durable memory note (key insight/decision) with tags and provenance sources.
+
+        Payload:
+          - note: str                (required)
+          - tags: dict[str,str]      (optional)
+          - sources: dict            (optional)
+              - run_id: str          (optional; defaults to current run_id)
+              - span_ids: list[str]  (optional; referenced span ids)
+              - message_ids: list[str] (optional; referenced message ids)
+          - target_run_id: str       (optional; defaults to current run_id)
+          - tool_name: str           (optional; default "remember_note")
+          - call_id: str             (optional; passthrough)
+        """
+        import json
+
+        from .vars import ensure_namespaces
+
+        ensure_namespaces(run.vars)
+        runtime_ns = run.vars.get("_runtime")
+        if not isinstance(runtime_ns, dict):
+            runtime_ns = {}
+            run.vars["_runtime"] = runtime_ns
+
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            return EffectOutcome.failed(
+                "MEMORY_NOTE requires an ArtifactStore; configure runtime.set_artifact_store(...)"
+            )
+
+        payload = dict(effect.payload or {})
+        tool_name = str(payload.get("tool_name") or "remember_note")
+        call_id = str(payload.get("call_id") or "memory")
+
+        base_run_id = str(payload.get("target_run_id") or run.run_id).strip() or run.run_id
+        base_run = run
+        if base_run_id != run.run_id:
+            loaded = self._run_store.load(base_run_id)
+            if loaded is None:
+                return EffectOutcome.failed(f"Unknown target_run_id: {base_run_id}")
+            base_run = loaded
+            ensure_namespaces(base_run.vars)
+
+        scope = str(payload.get("scope") or "run").strip().lower() or "run"
+        try:
+            target_run = self._resolve_scope_owner_run(base_run, scope=scope)
+        except Exception as e:
+            return EffectOutcome.failed(str(e))
+        ensure_namespaces(target_run.vars)
+
+        target_runtime_ns = target_run.vars.get("_runtime")
+        if not isinstance(target_runtime_ns, dict):
+            target_runtime_ns = {}
+            target_run.vars["_runtime"] = target_runtime_ns
+        spans = target_runtime_ns.get("memory_spans")
+        if not isinstance(spans, list):
+            spans = []
+            target_runtime_ns["memory_spans"] = spans
+
+        note = payload.get("note")
+        note_text = str(note or "").strip()
+        if not note_text:
+            return EffectOutcome.failed("MEMORY_NOTE requires payload.note (non-empty string)")
+
+        location_raw = payload.get("location")
+        location = str(location_raw).strip() if isinstance(location_raw, str) else ""
+
+        tags = payload.get("tags")
+        clean_tags: Dict[str, str] = {}
+        if isinstance(tags, dict):
+            for k, v in tags.items():
+                if isinstance(k, str) and isinstance(v, str) and k and v:
+                    if k == "kind":
+                        continue
+                    clean_tags[k] = v
+
+        sources = payload.get("sources")
+        sources_dict = dict(sources) if isinstance(sources, dict) else {}
+
+        def _norm_list(value: Any) -> list[str]:
+            if not isinstance(value, list):
+                return []
+            out: list[str] = []
+            for item in value:
+                if isinstance(item, str):
+                    s = item.strip()
+                    if s:
+                        out.append(s)
+                elif isinstance(item, int):
+                    out.append(str(item))
+            # preserve order but dedup
+            seen: set[str] = set()
+            deduped: list[str] = []
+            for s in out:
+                if s in seen:
+                    continue
+                seen.add(s)
+                deduped.append(s)
+            return deduped
+
+        # Provenance default: the run that emitted this effect (not the scope owner).
+        source_run_id = str(sources_dict.get("run_id") or run.run_id).strip() or run.run_id
+        span_ids = _norm_list(sources_dict.get("span_ids"))
+        message_ids = _norm_list(sources_dict.get("message_ids"))
+
+        created_at = utc_now_iso()
+        artifact_payload: Dict[str, Any] = {
+            "note": note_text,
+            "sources": {"run_id": source_run_id, "span_ids": span_ids, "message_ids": message_ids},
+            "created_at": created_at,
+        }
+        if location:
+            artifact_payload["location"] = location
+        if run.actor_id:
+            artifact_payload["actor_id"] = str(run.actor_id)
+        session_id = getattr(target_run, "session_id", None) or getattr(run, "session_id", None)
+        if session_id:
+            artifact_payload["session_id"] = str(session_id)
+
+        artifact_tags: Dict[str, str] = {"kind": "memory_note"}
+        artifact_tags.update(clean_tags)
+        meta = artifact_store.store_json(artifact_payload, run_id=target_run.run_id, tags=artifact_tags)
+        artifact_id = meta.artifact_id
+
+        preview = note_text
+        if len(preview) > 160:
+            preview = preview[:157] + "…"
+
+        span_record: Dict[str, Any] = {
+            "kind": "memory_note",
+            "artifact_id": artifact_id,
+            "created_at": created_at,
+            # Treat notes as point-in-time spans for time-range filtering.
+            "from_timestamp": created_at,
+            "to_timestamp": created_at,
+            "message_count": 0,
+            "note_preview": preview,
+        }
+        if location:
+            span_record["location"] = location
+        if clean_tags:
+            span_record["tags"] = dict(clean_tags)
+        if span_ids or message_ids:
+            span_record["sources"] = {"run_id": source_run_id, "span_ids": span_ids, "message_ids": message_ids}
+        if run.actor_id:
+            span_record["created_by"] = str(run.actor_id)
+
+        spans.append(span_record)
+
+        def _coerce_bool(value: Any) -> bool:
+            if isinstance(value, bool):
+                return bool(value)
+            if isinstance(value, (int, float)) and not isinstance(value, bool):
+                try:
+                    return float(value) != 0.0
+                except Exception:
+                    return False
+            if isinstance(value, str):
+                s = value.strip().lower()
+                if not s:
+                    return False
+                if s in {"false", "0", "no", "off"}:
+                    return False
+                if s in {"true", "1", "yes", "on"}:
+                    return True
+            return False
+
+        # Optional UX convenience: keep the stored note immediately visible to downstream LLM calls by
+        # rehydrating it into `base_run.context.messages` as a synthetic system message.
+        keep_raw = payload.get("keep_in_context")
+        if keep_raw is None:
+            keep_raw = payload.get("keepInContext")
+        keep_in_context = _coerce_bool(keep_raw)
+        kept: Optional[Dict[str, Any]] = None
+        if keep_in_context:
+            try:
+                from ..memory.active_context import ActiveContextPolicy
+
+                policy = ActiveContextPolicy(run_store=self._run_store, artifact_store=artifact_store)
+                out = policy.rehydrate_into_context_from_run(
+                    base_run,
+                    span_ids=[artifact_id],
+                    placement="end",
+                    dedup_by="message_id",
+                    max_messages=1,
+                )
+                kept = {"inserted": out.get("inserted", 0), "skipped": out.get("skipped", 0)}
+
+                # Persist when mutating a different run than the currently executing one.
+                if base_run is not run:
+                    base_run.updated_at = utc_now_iso()
+                    self._run_store.save(base_run)
+            except Exception as e:
+                kept = {"inserted": 0, "skipped": 0, "error": str(e)}
+
+        if target_run is not run:
+            target_run.updated_at = utc_now_iso()
+            self._run_store.save(target_run)
+
+        rendered_tags = json.dumps(clean_tags, ensure_ascii=False, sort_keys=True) if clean_tags else "{}"
+        text = f"Stored memory_note span_id={artifact_id} tags={rendered_tags}"
+        meta_out: Dict[str, Any] = {"span_id": artifact_id, "created_at": created_at, "note_preview": preview}
+        if isinstance(kept, dict):
+            meta_out["kept_in_context"] = kept
+
+        result = {
+            "mode": "executed",
+            "results": [
+                {
+                    "call_id": call_id,
+                    "name": tool_name,
+                    "success": True,
+                    "output": text,
+                    "error": None,
+                    "meta": meta_out,
+                }
+            ],
+        }
+        return EffectOutcome.completed(result=result)
+
+    def _handle_memory_rehydrate(self, run: RunState, effect: Effect, default_next_node: Optional[str]) -> EffectOutcome:
+        """Handle MEMORY_REHYDRATE.
+
+        This is a runtime-owned, deterministic mutation of `context.messages`:
+        - loads archived conversation span artifacts from ArtifactStore
+        - inserts them into `context.messages` with dedup
+        - persists the mutated run (RunStore checkpoint)
+
+        Payload (required unless stated):
+          - span_ids: list[str|int]  (required; artifact ids preferred; indices allowed)
+          - placement: str          ("after_summary"|"after_system"|"end", default "after_summary")
+          - dedup_by: str           (default "message_id")
+          - max_messages: int       (optional; max inserted messages)
+          - target_run_id: str      (optional; defaults to current run)
+        """
+        from .vars import ensure_namespaces
+
+        ensure_namespaces(run.vars)
+        artifact_store = self._artifact_store
+        if artifact_store is None:
+            return EffectOutcome.failed(
+                "MEMORY_REHYDRATE requires an ArtifactStore; configure runtime.set_artifact_store(...)"
+            )
+
+        payload = dict(effect.payload or {})
+        target_run_id = str(payload.get("target_run_id") or run.run_id).strip() or run.run_id
+
+        # Normalize span_ids (accept legacy `span_id` too).
+        raw_span_ids = payload.get("span_ids")
+        if raw_span_ids is None:
+            raw_span_ids = payload.get("span_id")
+        span_ids: list[Any] = []
+        if isinstance(raw_span_ids, list):
+            span_ids = list(raw_span_ids)
+        elif raw_span_ids is not None:
+            span_ids = [raw_span_ids]
+        if not span_ids:
+            return EffectOutcome.failed("MEMORY_REHYDRATE requires payload.span_ids (non-empty list)")
+
+        placement = str(payload.get("placement") or "after_summary").strip() or "after_summary"
+        dedup_by = str(payload.get("dedup_by") or "message_id").strip() or "message_id"
+        max_messages = payload.get("max_messages")
+
+        # Load the target run (may be different from current).
+        target_run = run
+        if target_run_id != run.run_id:
+            loaded = self._run_store.load(target_run_id)
+            if loaded is None:
+                return EffectOutcome.failed(f"Unknown target_run_id: {target_run_id}")
+            target_run = loaded
+            ensure_namespaces(target_run.vars)
+
+        # Best-effort: rehydrate only span kinds that are meaningful to inject into
+        # `context.messages` for downstream LLM calls.
+        #
+        # Rationale:
+        # - conversation_span: archived chat messages
+        # - memory_note: durable notes (rehydrated as a synthetic message by ActiveContextPolicy)
+        #
+        # Evidence and other span kinds are intentionally skipped by default.
+        from ..memory.active_context import ActiveContextPolicy
+
+        spans = ActiveContextPolicy.list_memory_spans_from_run(target_run)
+        resolved = ActiveContextPolicy.resolve_span_ids_from_spans(span_ids, spans)
+        if not resolved:
+            return EffectOutcome.completed(result={"inserted": 0, "skipped": 0, "artifacts": []})
+
+        kind_by_artifact: dict[str, str] = {}
+        for s in spans:
+            if not isinstance(s, dict):
+                continue
+            aid = str(s.get("artifact_id") or "").strip()
+            if not aid or aid in kind_by_artifact:
+                continue
+            kind_by_artifact[aid] = str(s.get("kind") or "").strip()
+
+        to_rehydrate: list[str] = []
+        skipped_artifacts: list[dict[str, Any]] = []
+        allowed_kinds = {"conversation_span", "memory_note"}
+        for aid in resolved:
+            kind = kind_by_artifact.get(aid, "")
+            if kind and kind not in allowed_kinds:
+                skipped_artifacts.append(
+                    {"span_id": aid, "inserted": 0, "skipped": 0, "error": None, "kind": kind}
+                )
+                continue
+            to_rehydrate.append(aid)
+
+        # Reuse the canonical policy implementation (no duplicated logic).
+        # Mutate the in-memory RunState to keep runtime tick semantics consistent.
+        policy = ActiveContextPolicy(run_store=self._run_store, artifact_store=artifact_store)
+        out = policy.rehydrate_into_context_from_run(
+            target_run,
+            span_ids=to_rehydrate,
+            placement=placement,
+            dedup_by=dedup_by,
+            max_messages=max_messages,
+        )
+
+        # Persist when mutating a different run than the currently executing one.
+        if target_run is not run:
+            target_run.updated_at = utc_now_iso()
+            self._run_store.save(target_run)
+
+        # Normalize output shape to match backlog expectations (`span_id` field, optional kind).
+        artifacts_out: list[dict[str, Any]] = []
+        artifacts = out.get("artifacts")
+        if isinstance(artifacts, list):
+            for a in artifacts:
+                if not isinstance(a, dict):
+                    continue
+                aid = str(a.get("artifact_id") or "").strip()
+                artifacts_out.append(
+                    {
+                        "span_id": aid,
+                        "inserted": a.get("inserted"),
+                        "skipped": a.get("skipped"),
+                        "error": a.get("error"),
+                        "kind": kind_by_artifact.get(aid) or None,
+                        "preview": a.get("preview"),
+                    }
+                )
+        artifacts_out.extend(skipped_artifacts)
+
+        return EffectOutcome.completed(
+            result={
+                "inserted": out.get("inserted", 0),
+                "skipped": out.get("skipped", 0),
+                "artifacts": artifacts_out,
+            }
+        )
+
+
+def _dedup_preserve_order(values: list[str]) -> list[str]:
+    seen: set[str] = set()
+    out: list[str] = []
+    for v in values:
+        s = str(v or "").strip()
+        if not s or s in seen:
+            continue
+        seen.add(s)
+        out.append(s)
+    return out
+
+
+def _span_sort_key(span: dict) -> tuple[str, str]:
+    """Sort key for span adjacency. Prefer from_timestamp, then created_at."""
+    from_ts = str(span.get("from_timestamp") or "")
+    created = str(span.get("created_at") or "")
+    return (from_ts or created, created)
+
+
+def _expand_connected_span_ids(
+    *,
+    spans: list[dict[str, Any]],
+    seed_artifact_ids: list[str],
+    connect_keys: list[str],
+    neighbor_hops: int,
+    limit: int,
+) -> list[str]:
+    """Expand seed spans to include deterministic neighbors (time + shared tags)."""
+    if not spans or not seed_artifact_ids:
+        return list(seed_artifact_ids)
+
+    ordered = [s for s in spans if isinstance(s, dict) and s.get("artifact_id")]
+    ordered.sort(key=_span_sort_key)
+    idx_by_artifact: dict[str, int] = {str(s["artifact_id"]): i for i, s in enumerate(ordered)}
+
+    # Build tag index for requested keys.
+    tag_index: dict[tuple[str, str], list[str]] = {}
+    for s in ordered:
+        tags = s.get("tags") if isinstance(s.get("tags"), dict) else {}
+        for k in connect_keys:
+            v = tags.get(k)
+            if isinstance(v, str) and v:
+                tag_index.setdefault((k, v), []).append(str(s["artifact_id"]))
+
+    out: list[str] = []
+    for aid in seed_artifact_ids:
+        if len(out) >= limit:
+            break
+        out.append(aid)
+
+        idx = idx_by_artifact.get(aid)
+        if idx is not None and neighbor_hops > 0:
+            for delta in range(1, neighbor_hops + 1):
+                for j in (idx - delta, idx + delta):
+                    if 0 <= j < len(ordered):
+                        out.append(str(ordered[j]["artifact_id"]))
+
+        if connect_keys:
+            s = ordered[idx] if idx is not None and 0 <= idx < len(ordered) else None
+            if isinstance(s, dict):
+                tags = s.get("tags") if isinstance(s.get("tags"), dict) else {}
+                for k in connect_keys:
+                    v = tags.get(k)
+                    if isinstance(v, str) and v:
+                        out.extend(tag_index.get((k, v), []))
+
+    return _dedup_preserve_order(out)[:limit]
+
+
+def _deep_scan_span_ids(
+    *,
+    spans: list[dict[str, Any]],
+    artifact_store: Any,
+    query: str,
+    limit_spans: int,
+    limit_messages_per_span: int,
+) -> list[str]:
+    """Fallback keyword scan over archived messages when metadata/summary is insufficient."""
+    q = str(query or "").strip().lower()
+    if not q:
+        return []
+
+    scanned = 0
+    matches: list[str] = []
+    for s in spans:
+        if scanned >= limit_spans:
+            break
+        if not isinstance(s, dict):
+            continue
+        artifact_id = s.get("artifact_id")
+        if not isinstance(artifact_id, str) or not artifact_id:
+            continue
+        scanned += 1
+
+        payload = artifact_store.load_json(artifact_id)
+        if not isinstance(payload, dict):
+            continue
+        messages = payload.get("messages")
+        if not isinstance(messages, list) or not messages:
+            continue
+
+        for m in messages[:limit_messages_per_span]:
+            if not isinstance(m, dict):
+                continue
+            content = m.get("content")
+            if not content:
+                continue
+            if q in str(content).lower():
+                matches.append(artifact_id)
+                break
+
+    return _dedup_preserve_order(matches)
+
+
+def _render_memory_query_output(
+    *,
+    spans: list[dict[str, Any]],
+    artifact_store: Any,
+    selected_artifact_ids: list[str],
+    summary_by_artifact: dict[str, str],
+    max_messages: int,
+) -> str:
+    if not selected_artifact_ids:
+        return "No matching memory spans."
+
+    span_by_id: dict[str, dict[str, Any]] = {
+        str(s.get("artifact_id")): s for s in spans if isinstance(s, dict) and s.get("artifact_id")
+    }
+
+    lines: list[str] = []
+    lines.append("Recalled memory spans (provenance-preserving):")
+
+    remaining: Optional[int] = None if int(max_messages) == -1 else int(max_messages)
+    for i, aid in enumerate(selected_artifact_ids, start=1):
+        span = span_by_id.get(aid, {})
+        kind = span.get("kind") or "span"
+        created = span.get("created_at") or ""
+        from_ts = span.get("from_timestamp") or ""
+        to_ts = span.get("to_timestamp") or ""
+        count = span.get("message_count") or ""
+        created_by = span.get("created_by") or ""
+        location = span.get("location") or ""
+        tags = span.get("tags") if isinstance(span.get("tags"), dict) else {}
+        tags_txt = ", ".join([f"{k}={v}" for k, v in sorted(tags.items()) if isinstance(v, str) and v])
+
+        lines.append("")
+        lines.append(f"[{i}] span_id={aid} kind={kind} msgs={count} created_at={created}")
+        if from_ts or to_ts:
+            lines.append(f"    time_range: {from_ts} .. {to_ts}")
+        if isinstance(created_by, str) and str(created_by).strip():
+            lines.append(f"    created_by: {str(created_by).strip()}")
+        if isinstance(location, str) and str(location).strip():
+            lines.append(f"    location: {str(location).strip()}")
+        if tags_txt:
+            lines.append(f"    tags: {tags_txt}")
+
+        summary = summary_by_artifact.get(aid)
+        if summary:
+            lines.append(f"    summary: {str(summary).strip()}")
+
+        if remaining is not None and remaining <= 0:
+            continue
+
+        payload = artifact_store.load_json(aid)
+        if not isinstance(payload, dict):
+            lines.append("    (artifact payload unavailable)")
+            continue
+        if kind == "memory_note" or "note" in payload:
+            note = str(payload.get("note") or "").strip()
+            if note:
+                lines.append("    note: " + note)
+            else:
+                lines.append("    (note payload missing note text)")
+
+            if not (isinstance(location, str) and location.strip()):
+                loc = payload.get("location")
+                if isinstance(loc, str) and loc.strip():
+                    lines.append(f"    location: {loc.strip()}")
+
+            sources = payload.get("sources")
+            if isinstance(sources, dict):
+                src_run = sources.get("run_id")
+                span_ids = sources.get("span_ids")
+                msg_ids = sources.get("message_ids")
+                if isinstance(src_run, str) and src_run:
+                    lines.append(f"    sources.run_id: {src_run}")
+                if isinstance(span_ids, list) and span_ids:
+                    cleaned = [str(x) for x in span_ids if isinstance(x, (str, int))]
+                    if cleaned:
+                        lines.append(f"    sources.span_ids: {', '.join(cleaned[:12])}")
+                if isinstance(msg_ids, list) and msg_ids:
+                    cleaned = [str(x) for x in msg_ids if isinstance(x, (str, int))]
+                    if cleaned:
+                        lines.append(f"    sources.message_ids: {', '.join(cleaned[:12])}")
+            continue
+
+        messages = payload.get("messages")
+        if not isinstance(messages, list):
+            lines.append("    (artifact missing messages)")
+            continue
+
+        # Render messages with a global cap.
+        rendered = 0
+        for m in messages:
+            if remaining is not None and remaining <= 0:
+                break
+            if not isinstance(m, dict):
+                continue
+            role = str(m.get("role") or "unknown")
+            content = str(m.get("content") or "")
+            ts = str(m.get("timestamp") or "")
+            prefix = f"    - {role}: "
+            if ts:
+                prefix = f"    - {ts} {role}: "
+            lines.append(prefix + content)
+            rendered += 1
+            if remaining is not None:
+                remaining -= 1
+
+        total = sum(1 for m in messages if isinstance(m, dict))
+        if remaining is not None and rendered < total:
+            lines.append(f"    (remaining {total - rendered} messages omitted by max_messages={int(max_messages)})")
+
+    return "\n".join(lines)
+
 
 def _set_nested(target: Dict[str, Any], dotted_key: str, value: Any) -> None:
     """Set nested dict value using dot notation."""