feat: add profile-scoped local Hindsight eval path

Closes #1010

agent/context_snapshots.py

@@ -0,0 +1,69 @@
+"""First-class context snapshot artifacts for live runtime memory evaluation."""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+from typing import Any
+
+from hermes_constants import get_hermes_home
+
+
+_SAFE_SEGMENT_RE = re.compile(r"[^A-Za-z0-9_.-]+")
+
+
+class ContextSnapshotRecorder:
+    """Write per-call prompt-composition artifacts for a Hermes session."""
+
+    def __init__(self, session_id: str, *, enabled: bool = False, base_dir: str | Path | None = None):
+        self.session_id = session_id or "session"
+        self.enabled = bool(enabled)
+        self.base_dir = Path(base_dir) if base_dir else get_hermes_home() / "reports" / "context_snapshots"
+
+    @property
+    def session_dir(self) -> Path:
+        safe_session = _SAFE_SEGMENT_RE.sub("_", self.session_id).strip("._") or "session"
+        return self.base_dir / safe_session
+
+    def record_call(
+        self,
+        api_call_count: int,
+        *,
+        system_prompt: str,
+        memory_provider_system_prompt: str = "",
+        memory_prefetch_raw: str = "",
+        memory_context_block: str = "",
+        api_user_message: str = "",
+        api_messages: list[dict[str, Any]] | None = None,
+        metadata: dict[str, Any] | None = None,
+    ) -> Path | None:
+        if not self.enabled:
+            return None
+
+        call_dir = self.session_dir / f"call_{api_call_count:03d}"
+        call_dir.mkdir(parents=True, exist_ok=True)
+
+        self._write_text(call_dir / "system_prompt.txt", system_prompt or "")
+        self._write_text(call_dir / "memory_provider_system_prompt.txt", memory_provider_system_prompt or "")
+        self._write_text(call_dir / "memory_prefetch_raw.txt", memory_prefetch_raw or "")
+        self._write_text(call_dir / "memory_context_block.txt", memory_context_block or "")
+        self._write_text(call_dir / "api_user_message.txt", api_user_message or "")
+        self._write_json(call_dir / "api_messages.json", api_messages or [])
+        self._write_json(
+            call_dir / "metadata.json",
+            {
+                "session_id": self.session_id,
+                "api_call_count": api_call_count,
+                **(metadata or {}),
+            },
+        )
+        return call_dir
+
+    @staticmethod
+    def _write_text(path: Path, content: str) -> None:
+        path.write_text(content, encoding="utf-8")
+
+    @staticmethod
+    def _write_json(path: Path, payload: Any) -> None:
+        path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")

docs/hindsight-local-eval.md

@@ -0,0 +1,132 @@
+# Hindsight local eval homes for live Hermes runtime testing
+
+Issue: #1010
+Parent: #985
+
+This document defines a reproducible, profile-scoped evaluation layout for baseline / MemPalace / Hindsight comparisons without requiring Hindsight Cloud.
+
+## Eval home layout
+
+Use three separate `HERMES_HOME` directories so each run has isolated config, memory, sessions, and artifacts.
+
+```text
+~/.hermes/profiles/atlas-baseline/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  reports/context_snapshots/
+
+~/.hermes/profiles/atlas-mempalace/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  reports/context_snapshots/
+  plugins/           # if a local MemPalace plugin is installed for this eval lane
+
+~/.hermes/profiles/atlas-hindsight/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  hindsight/config.json
+  reports/context_snapshots/
+```
+
+## Hindsight local config
+
+The Hindsight provider already loads config from `$HERMES_HOME/hindsight/config.json` first. For the local eval lane, prefer `local_embedded` so Hermes can bring up a local Hindsight daemon without cloud signup.
+
+Example `~/.hermes/profiles/atlas-hindsight/hindsight/config.json`:
+
+```json
+{
+  "mode": "local_embedded",
+  "memory_mode": "context",
+  "recall_prefetch_method": "recall",
+  "llm_provider": "ollama",
+  "llm_model": "gemma3:12b",
+  "api_url": "http://localhost:8888"
+}
+```
+
+Notes:
+- `local_embedded` avoids any Hindsight Cloud dependency.
+- If `profile` is omitted, Hermes now derives a stable local Hindsight profile name from the active profile identity / `HERMES_HOME` instead of collapsing all local runs into the shared legacy `hermes` profile.
+- `local_external` remains valid if you already run a local Hindsight server yourself.
+
+## Runtime switching procedure
+
+Switch by exporting `HERMES_HOME` before launching Hermes.
+
+### 1. Baseline
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-baseline"
+unset HERMES_CONTEXT_SNAPSHOTS
+hermes chat
+```
+
+### 2. MemPalace lane
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-mempalace"
+export HERMES_CONTEXT_SNAPSHOTS=1
+hermes chat
+```
+
+### 3. Hindsight lane
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-hindsight"
+export HERMES_CONTEXT_SNAPSHOTS=1
+hermes chat
+```
+
+## Raw artifact capture
+
+When `HERMES_CONTEXT_SNAPSHOTS=1` is enabled, Hermes writes first-class prompt-composition artifacts under the active home by default.
+
+Artifact tree:
+
+```text
+$HERMES_HOME/reports/context_snapshots/<session-id>/call_001/
+  system_prompt.txt
+  memory_provider_system_prompt.txt
+  memory_prefetch_raw.txt
+  memory_context_block.txt
+  api_user_message.txt
+  api_messages.json
+  metadata.json
+```
+
+Minimum files a benchmark should inspect:
+- `system_prompt.txt`
+- `memory_prefetch_raw.txt`
+- `memory_context_block.txt`
+- `api_user_message.txt`
+- `api_messages.json`
+
+These prove:
+- what the system prompt was
+- what the provider prefetched
+- what entered `<memory-context>`
+- what the final API user message looked like
+- what full payload reached the model
+
+## Follow-on benchmark workflow
+
+A benchmark issue can now consume this path without redoing integration work:
+1. pick one eval home (`atlas-baseline`, `atlas-mempalace`, `atlas-hindsight`)
+2. export the corresponding `HERMES_HOME`
+3. run Hermes on the same prompt set
+4. compare the snapshot artifacts in `reports/context_snapshots/`
+5. score recall quality and answer quality separately
+
+## Why this is sovereign
+
+- no hosted Hindsight Cloud dependency is required
+- the Hindsight config is profile-scoped under `hindsight/config.json`
+- the runtime artifacts stay under the active `HERMES_HOME`
+- switching between baseline / MemPalace / Hindsight is just a `HERMES_HOME` swap

plugins/memory/hindsight/__init__.py

@@ -178,6 +178,25 @@ def _load_config() -> dict:
     }
 
 
+def _derive_local_profile_name(agent_identity: str = "", hermes_home: str = "") -> str:
+    """Return a stable profile name for local embedded Hindsight storage.
+
+    Prefer the active Hermes profile identity when available, otherwise fall back
+    to the basename of the active HERMES_HOME path. This prevents all local
+    Hindsight eval homes from sharing the legacy default profile name "hermes".
+    """
+    from pathlib import Path
+    import re
+
+    raw = (agent_identity or "").strip()
+    if not raw and hermes_home:
+        raw = Path(hermes_home).name.strip()
+    if not raw:
+        raw = "hermes"
+    safe = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip(".-_")
+    return safe or "hermes"
+
+
 # ---------------------------------------------------------------------------
 # MemoryProvider implementation
 # ---------------------------------------------------------------------------
@@ -468,6 +487,8 @@ class HindsightMemoryProvider(MemoryProvider):
 
     def initialize(self, session_id: str, **kwargs) -> None:
         self._session_id = session_id
+        hermes_home = str(kwargs.get("hermes_home") or "")
+        agent_identity = str(kwargs.get("agent_identity") or "")
 
         # Check client version and auto-upgrade if needed
         try:
@@ -500,6 +521,11 @@ class HindsightMemoryProvider(MemoryProvider):
         # "local" is a legacy alias for "local_embedded"
         if self._mode == "local":
             self._mode = "local_embedded"
+        if self._mode == "local_embedded" and not self._config.get("profile"):
+            self._config["profile"] = _derive_local_profile_name(
+                agent_identity=agent_identity,
+                hermes_home=hermes_home,
+            )
         self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "")
         default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL
         self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url)

run_agent.py

@@ -20,7 +20,6 @@ Usage:
     response = agent.run_conversation("Tell me about the latest Python updates")
 """
 
-import ast
 import asyncio
 import base64
 import concurrent.futures
@@ -605,6 +604,8 @@ class AIAgent:
         checkpoint_max_snapshots: int = 50,
         pass_session_id: bool = False,
         persist_session: bool = True,
+        context_snapshots_enabled: bool | None = None,
+        context_snapshots_dir: str | None = None,
     ):
         """
         Initialize the AI Agent.
@@ -1130,6 +1131,43 @@ class AIAgent:
         except Exception:
             _agent_cfg = {}
 
+        def _is_enabled(value):
+            if isinstance(value, bool):
+                return value
+            return str(value).strip().lower() in {"1", "true", "yes", "on"}
+
+        _debug_cfg = _agent_cfg.get("debug", {}) if isinstance(_agent_cfg, dict) else {}
+        if not isinstance(_debug_cfg, dict):
+            _debug_cfg = {}
+        _snapshot_cfg = _debug_cfg.get("context_snapshots", {})
+        if not isinstance(_snapshot_cfg, dict):
+            _snapshot_cfg = {}
+        _snapshots_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS")
+        _snapshots_dir_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS_DIR")
+        if context_snapshots_enabled is None:
+            if _snapshots_env is not None:
+                self._context_snapshots_enabled = _is_enabled(_snapshots_env)
+            else:
+                self._context_snapshots_enabled = _is_enabled(_snapshot_cfg.get("enabled", False))
+        else:
+            self._context_snapshots_enabled = bool(context_snapshots_enabled)
+        self._context_snapshots_dir = (
+            context_snapshots_dir
+            or _snapshots_dir_env
+            or _snapshot_cfg.get("dir")
+            or None
+        )
+        try:
+            from agent.context_snapshots import ContextSnapshotRecorder
+            self._context_snapshot_recorder = ContextSnapshotRecorder(
+                session_id=self.session_id,
+                enabled=self._context_snapshots_enabled,
+                base_dir=self._context_snapshots_dir,
+            )
+        except Exception as _snapshot_err:
+            logger.debug("Context snapshot recorder init failed: %s", _snapshot_err)
+            self._context_snapshot_recorder = None
+
         # Persistent memory (MEMORY.md + USER.md) -- loaded from disk
         self._memory_store = None
         self._memory_enabled = False
@@ -3329,119 +3367,6 @@ class AIAgent:
 
     _VALID_API_ROLES = frozenset({"system", "user", "assistant", "tool", "function", "developer"})
 
-    @staticmethod
-    def _normalize_tool_call_arguments(arguments: Any) -> tuple[str, bool]:
-        """Return ``(normalized_text, is_complete)`` for tool-call arguments.
-
-        Conservative by design: repairs harmless formatting quirks common in
-        Gemma 4 / Ollama output (whitespace, trailing commas, Python-style
-        single-quoted dicts, bare key/value pairs) but does NOT auto-close
-        truncated JSON objects. Truly incomplete fragments must remain marked
-        incomplete so the agent can retry instead of silently dropping fields.
-        """
-        if isinstance(arguments, (dict, list)):
-            return json.dumps(arguments, ensure_ascii=False, separators=(",", ":")), True
-        if arguments is None:
-            return "{}", True
-        if not isinstance(arguments, str):
-            arguments = str(arguments)
-
-        text = arguments.strip()
-        if not text:
-            return "{}", True
-
-        def _parse_candidate(candidate: str):
-            try:
-                return json.loads(candidate)
-            except (json.JSONDecodeError, TypeError, ValueError):
-                pass
-            try:
-                return ast.literal_eval(candidate)
-            except (SyntaxError, ValueError):
-                return None
-
-        candidates: list[str] = [text]
-
-        trimmed_trailing_commas = re.sub(r",\s*([}\]])", r"\1", text)
-        if trimmed_trailing_commas != text:
-            candidates.append(trimmed_trailing_commas)
-
-        if ":" in text and not text.startswith(("{", "[")):
-            wrapped = "{" + text + "}"
-            candidates.append(wrapped)
-            quoted_keys = re.sub(
-                r'([\{,]\s*)([A-Za-z_][A-Za-z0-9_\-]*)(\s*:)',
-                r'\1"\2"\3',
-                wrapped,
-            )
-            if quoted_keys != wrapped:
-                candidates.append(quoted_keys)
-                trimmed_quoted_keys = re.sub(r",\s*([}\]])", r"\1", quoted_keys)
-                if trimmed_quoted_keys != quoted_keys:
-                    candidates.append(trimmed_quoted_keys)
-
-        seen: set[str] = set()
-        for candidate in candidates:
-            if candidate in seen:
-                continue
-            seen.add(candidate)
-            parsed = _parse_candidate(candidate)
-            if isinstance(parsed, (dict, list)):
-                return json.dumps(parsed, ensure_ascii=False, separators=(",", ":")), True
-
-        return text, False
-
-    @staticmethod
-    def _merge_consecutive_assistant_tool_call_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Merge adjacent assistant messages that each carry tool_calls.
-
-        Some providers emit parallel tool calls as multiple consecutive assistant
-        messages instead of a single assistant message with multiple tool calls.
-        Merge only adjacent assistant/tool-call messages; any non-assistant
-        boundary flushes the current batch.
-        """
-        merged: List[Dict[str, Any]] = []
-        pending: Optional[Dict[str, Any]] = None
-
-        def _flush_pending() -> None:
-            nonlocal pending
-            if pending is not None:
-                merged.append(pending)
-                pending = None
-
-        for msg in messages:
-            if not isinstance(msg, dict):
-                _flush_pending()
-                merged.append(msg)
-                continue
-
-            role = msg.get("role")
-            tool_calls = msg.get("tool_calls")
-            if role == "assistant" and isinstance(tool_calls, list) and tool_calls:
-                if pending is None:
-                    pending = copy.deepcopy(msg)
-                    continue
-
-                pending_tool_calls = pending.get("tool_calls")
-                if not isinstance(pending_tool_calls, list):
-                    pending_tool_calls = []
-                    pending["tool_calls"] = pending_tool_calls
-                pending_tool_calls.extend(copy.deepcopy(tool_calls))
-
-                pending_content = pending.get("content") or ""
-                current_content = msg.get("content") or ""
-                if pending_content and current_content:
-                    pending["content"] = pending_content + "\n" + current_content
-                elif current_content:
-                    pending["content"] = current_content
-                continue
-
-            _flush_pending()
-            merged.append(msg)
-
-        _flush_pending()
-        return merged
-
     @staticmethod
     def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Fix orphaned tool_call / tool_result pairs before every LLM call.
@@ -3461,7 +3386,7 @@ class AIAgent:
                 )
                 continue
             filtered.append(msg)
-        messages = AIAgent._merge_consecutive_assistant_tool_call_messages(filtered)
+        messages = filtered
 
         surviving_call_ids: set = set()
         for msg in messages:
@@ -5368,9 +5293,12 @@ class AIAgent:
                 mock_tool_calls = []
                 for idx in sorted(tool_calls_acc):
                     tc = tool_calls_acc[idx]
-                    arguments, is_complete = self._normalize_tool_call_arguments(tc["function"]["arguments"])
-                    if not is_complete:
-                        has_truncated_tool_args = True
+                    arguments = tc["function"]["arguments"]
+                    if arguments and arguments.strip():
+                        try:
+                            json.loads(arguments)
+                        except json.JSONDecodeError:
+                            has_truncated_tool_args = True
                     mock_tool_calls.append(SimpleNamespace(
                         id=tc["id"],
                         type=tc["type"],
@@ -6674,7 +6602,6 @@ class AIAgent:
                     response_item_id if isinstance(response_item_id, str) else None,
                 )
 
-                normalized_args, _ = self._normalize_tool_call_arguments(tool_call.function.arguments)
                 tc_dict = {
                     "id": call_id,
                     "call_id": call_id,
@@ -6682,7 +6609,7 @@ class AIAgent:
                     "type": tool_call.type,
                     "function": {
                         "name": tool_call.function.name,
-                        "arguments": normalized_args,
+                        "arguments": tool_call.function.arguments
                     },
                 }
                 # Preserve extra_content (e.g. Gemini thought_signature) so it
@@ -8256,12 +8183,17 @@ class AIAgent:
         # Use original_user_message (clean input) — user_message may contain
         # injected skill content that bloats / breaks provider queries.
         _ext_prefetch_cache = ""
+        _memory_provider_prompt_cache = ""
         if self._memory_manager:
             try:
                 _query = original_user_message if isinstance(original_user_message, str) else ""
                 _ext_prefetch_cache = self._memory_manager.prefetch_all(_query) or ""
             except Exception:
                 pass
+            try:
+                _memory_provider_prompt_cache = self._memory_manager.build_system_prompt() or ""
+            except Exception:
+                pass
 
         while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call:
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
@@ -8329,6 +8261,8 @@ class AIAgent:
             # However, providers like Moonshot AI require a separate 'reasoning_content' field
             # on assistant messages with tool_calls. We handle both cases here.
             api_messages = []
+            _current_api_user_message = ""
+            _current_memory_context_block = ""
             for idx, msg in enumerate(messages):
                 api_msg = msg.copy()
 
@@ -8343,12 +8277,15 @@ class AIAgent:
                         _fenced = build_memory_context_block(_ext_prefetch_cache)
                         if _fenced:
                             _injections.append(_fenced)
+                            _current_memory_context_block = _fenced
                     if _plugin_user_context:
                         _injections.append(_plugin_user_context)
                     if _injections:
                         _base = api_msg.get("content", "")
                         if isinstance(_base, str):
                             api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections)
+                    if isinstance(api_msg.get("content"), str):
+                        _current_api_user_message = api_msg["content"]
 
                 # For ALL assistant messages, pass reasoning back to the API
                 # This ensures multi-turn reasoning context is preserved
@@ -8383,7 +8320,13 @@ class AIAgent:
                 from agent.privacy_filter import PrivacyFilter
                 pf = PrivacyFilter()
                 # Sanitize messages before they reach the provider
-                api_messages = pf.sanitize_messages(api_messages)
+                _pf_result = pf.sanitize_messages(api_messages)
+                if isinstance(_pf_result, tuple):
+                    api_messages, _pf_report = _pf_result
+                    if getattr(pf, "last_report", None) is None:
+                        pf.last_report = _pf_report
+                else:
+                    api_messages = _pf_result
                 if pf.last_report and pf.last_report.had_redactions:
                     logger.info(f"Privacy Filter: Redacted sensitive data from turn payload. Details: {pf.last_report.summary()}")
             except Exception as e:
@@ -8454,6 +8397,27 @@ class AIAgent:
                     new_tcs.append(tc)
                 am["tool_calls"] = new_tcs
 
+            if self._context_snapshot_recorder:
+                try:
+                    self._context_snapshot_recorder.record_call(
+                        api_call_count,
+                        system_prompt=effective_system,
+                        memory_provider_system_prompt=_memory_provider_prompt_cache,
+                        memory_prefetch_raw=_ext_prefetch_cache,
+                        memory_context_block=_current_memory_context_block,
+                        api_user_message=_current_api_user_message,
+                        api_messages=api_messages,
+                        metadata={
+                            "model": self.model,
+                            "provider": self.provider,
+                            "platform": self.platform or "",
+                            "api_mode": self.api_mode,
+                            "memory_providers": [p.name for p in getattr(self._memory_manager, "providers", [])],
+                        },
+                    )
+                except Exception as _snapshot_err:
+                    logger.debug("Context snapshot capture failed: %s", _snapshot_err)
+
             # Calculate approximate request size for logging
             total_chars = sum(len(str(msg)) for msg in api_messages)
             approx_tokens = estimate_messages_tokens_rough(api_messages)
@@ -10143,15 +10107,21 @@ class AIAgent:
                     # Handle empty strings as empty objects (common model quirk)
                     invalid_json_args = []
                     for tc in assistant_message.tool_calls:
-                        normalized_args, is_complete = self._normalize_tool_call_arguments(tc.function.arguments)
-                        tc.function.arguments = normalized_args
-                        if not is_complete:
-                            try:
-                                json.loads(normalized_args)
-                            except json.JSONDecodeError as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
-                            except Exception as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
+                        args = tc.function.arguments
+                        if isinstance(args, (dict, list)):
+                            tc.function.arguments = json.dumps(args)
+                            continue
+                        if args is not None and not isinstance(args, str):
+                            tc.function.arguments = str(args)
+                            args = tc.function.arguments
+                        # Treat empty/whitespace strings as empty object
+                        if not args or not args.strip():
+                            tc.function.arguments = "{}"
+                            continue
+                        try:
+                            json.loads(args)
+                        except json.JSONDecodeError as e:
+                            invalid_json_args.append((tc.function.name, str(e)))
                     
                     if invalid_json_args:
                         # Check if the invalid JSON is due to truncation rather

tests/agent/test_context_snapshots.py

@@ -0,0 +1,43 @@
+from pathlib import Path
+
+from agent.context_snapshots import ContextSnapshotRecorder
+
+
+def test_disabled_recorder_writes_nothing(tmp_path):
+    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=False, base_dir=tmp_path)
+
+    out = recorder.record_call(
+        1,
+        system_prompt="system",
+        api_messages=[{"role": "user", "content": "hello"}],
+    )
+
+    assert out is None
+    assert not (tmp_path / "session-1").exists()
+
+
+def test_enabled_recorder_writes_expected_artifacts(tmp_path):
+    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=True, base_dir=tmp_path)
+
+    out = recorder.record_call(
+        1,
+        system_prompt="system prompt",
+        memory_provider_system_prompt="# Hindsight Memory\nActive.",
+        memory_prefetch_raw="- remembered fact",
+        memory_context_block="<memory-context>\nremembered\n</memory-context>",
+        api_user_message="What do I prefer?\n\n<memory-context>\nremembered\n</memory-context>",
+        api_messages=[
+            {"role": "system", "content": "system prompt"},
+            {"role": "user", "content": "What do I prefer?"},
+        ],
+        metadata={"provider": "openai", "memory_providers": ["builtin", "hindsight"]},
+    )
+
+    assert out == tmp_path / "session-1" / "call_001"
+    assert (out / "system_prompt.txt").read_text(encoding="utf-8") == "system prompt"
+    assert (out / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
+    assert (out / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered fact"
+    assert "<memory-context>" in (out / "memory_context_block.txt").read_text(encoding="utf-8")
+    assert "What do I prefer?" in (out / "api_user_message.txt").read_text(encoding="utf-8")
+    assert (out / "api_messages.json").read_text(encoding="utf-8").startswith("[")
+    assert '"hindsight"' in (out / "metadata.json").read_text(encoding="utf-8")

tests/plugins/memory/test_hindsight_provider.py

@@ -596,3 +596,26 @@ class TestAvailability:
         monkeypatch.setenv("HINDSIGHT_MODE", "local")
         p = HindsightMemoryProvider()
         assert p.is_available()
+
+    def test_local_embedded_profile_defaults_to_agent_identity(self, tmp_path, monkeypatch):
+        config_path = tmp_path / "hindsight" / "config.json"
+        config_path.parent.mkdir(parents=True, exist_ok=True)
+        config_path.write_text(json.dumps({
+            "mode": "local_embedded",
+            "llm_provider": "ollama",
+            "llm_model": "gemma3:12b",
+        }))
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path,
+        )
+
+        p = HindsightMemoryProvider()
+        p.initialize(
+            session_id="test-session",
+            hermes_home=str(tmp_path / "profiles" / "atlas-hindsight"),
+            platform="cli",
+            agent_identity="atlas-hindsight",
+        )
+
+        assert p._config["profile"] == "atlas-hindsight"

tests/run_agent/test_run_agent.py

@@ -1037,138 +1037,6 @@ class TestBuildAssistantMessage:
         result = agent._build_assistant_message(msg, "tool_calls")
         assert "extra_content" not in result["tool_calls"][0]
 
-    def test_tool_call_arguments_normalized_from_gemma4_whitespace(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments='  \n  {"path": "README.md"}  \n  ',
-            call_id="c4",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-    def test_tool_call_arguments_normalized_from_single_quotes_and_trailing_comma(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments="{'path': 'README.md',}",
-            call_id="c5",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-
-class TestNormalizeToolCallArguments:
-    @pytest.mark.parametrize(
-        ("raw_args", "expected"),
-        [
-            ('{"q":"test"}', '{"q":"test"}'),
-            ('  \n  {"q": "test"}  \n  ', '{"q":"test"}'),
-            ('{"q": "test",}', '{"q":"test"}'),
-            ("{'q': 'test'}", '{"q":"test"}'),
-            ("{'path': 'README.md', 'mode': 'read'}", '{"path":"README.md","mode":"read"}'),
-            ('"path": "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md", mode: "read"', '{"path":"README.md","mode":"read"}'),
-            ({"path": "README.md"}, '{"path":"README.md"}'),
-            (["README.md", "docs.md"], '["README.md","docs.md"]'),
-            ('\t\n  ', '{}'),
-            ('{"nested": {"path": "README.md"}}', '{"nested":{"path":"README.md"}}'),
-        ],
-    )
-    def test_complete_args_are_normalized(self, raw_args, expected):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is True
-        assert normalized == expected
-
-    @pytest.mark.parametrize(
-        "raw_args",
-        [
-            '{"path": "README.md"',
-            '{"a": 1, "b"',
-            '{"path": [1, 2}',
-            "{'path': 'README.md'",
-            'path: "README.md", mode:',
-            '{"command": "echo hello",',
-        ],
-    )
-    def test_incomplete_args_are_not_marked_complete(self, raw_args):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is False
-        assert isinstance(normalized, str)
-        assert normalized == raw_args.strip()
-
-
-class TestSanitizeApiMessages:
-    def test_merges_consecutive_assistant_tool_call_messages(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "first",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "second",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "search_files", "arguments": '{"pattern":"TODO"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {"role": "tool", "tool_call_id": "c2", "content": "matches"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert len(sanitized) == 3
-        assert sanitized[0]["role"] == "assistant"
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2"]
-        assert sanitized[0]["content"] == "first\nsecond"
-
-    def test_does_not_merge_assistant_tool_call_messages_across_non_assistant_boundary(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c2", "content": "b.py"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assistant_msgs = [m for m in sanitized if m.get("role") == "assistant"]
-        assert len(assistant_msgs) == 2
-        assert assistant_msgs[0]["tool_calls"][0]["id"] == "c1"
-        assert assistant_msgs[1]["tool_calls"][0]["id"] == "c2"
-
-    def test_merge_preserves_tool_call_order(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c3", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"c.py"}'}}],
-            },
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2", "c3"]
-
 
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
@@ -3599,59 +3467,6 @@ class TestStreamingApiCall:
         assert tc[0].function.arguments == '{"path":"x.txt","content":"hel'
         assert resp.choices[0].finish_reason == "length"
 
-    @pytest.mark.parametrize(
-        ("raw_arguments", "expected"),
-        [
-            ('  \n  {"path": "x.txt"}  \n  ', '{"path":"x.txt"}'),
-            ("{'path': 'x.txt',}", '{"path":"x.txt"}'),
-            ('path: "x.txt", mode: "read"', '{"path":"x.txt","mode":"read"}'),
-        ],
-    )
-    def test_repairable_tool_call_args_do_not_upgrade_finish_reason_to_length(self, agent, raw_arguments, expected):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", raw_arguments)]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.name == "read_file"
-        assert tc[0].function.arguments == expected
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_tool_call_args_single_quotes_across_chunks_normalized(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", "{'path':")]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, " 'x.txt',}")]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_split_json_chunks_still_reassemble(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", '{"path":')]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, ' "x.txt"}')]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
     def test_ollama_reused_index_separate_tool_calls(self, agent):
         """Ollama sends every tool call at index 0 with different ids.
 

tests/run_agent/test_run_agent_context_snapshots.py

@@ -0,0 +1,94 @@
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+import importlib
+import sys
+import types
+
+
+
+def _make_tool_defs(*names: str) -> list:
+    return [
+        {
+            "type": "function",
+            "function": {
+                "name": n,
+                "description": f"{n} tool",
+                "parameters": {"type": "object", "properties": {}},
+            },
+        }
+        for n in names
+    ]
+
+
+
+def _mock_response(content="Done", finish_reason="stop"):
+    msg = SimpleNamespace(content=content, tool_calls=None)
+    choice = SimpleNamespace(message=msg, finish_reason=finish_reason)
+    return SimpleNamespace(choices=[choice], usage=SimpleNamespace(prompt_tokens=1, completion_tokens=1, total_tokens=2))
+
+
+
+def _load_ai_agent():
+    sys.modules.setdefault("agent.auxiliary_client", types.SimpleNamespace(call_llm=lambda *a, **k: ""))
+    run_agent = importlib.import_module("run_agent")
+    return run_agent.AIAgent
+
+
+
+def test_run_conversation_writes_context_snapshot_artifacts(tmp_path):
+    AIAgent = _load_ai_agent()
+
+    class _FakePrivacyFilter:
+        def __init__(self):
+            self.last_report = None
+
+        def sanitize_messages(self, messages):
+            return list(messages)
+
+    with (
+        patch("run_agent.get_tool_definitions", return_value=_make_tool_defs("web_search")),
+        patch("run_agent.check_toolset_requirements", return_value={}),
+        patch("run_agent.OpenAI"),
+        patch("hermes_cli.plugins.invoke_hook", return_value=[]),
+        patch.dict(sys.modules, {"agent.privacy_filter": types.SimpleNamespace(PrivacyFilter=_FakePrivacyFilter)}),
+    ):
+        agent = AIAgent(
+            api_key="test-key-1234567890",
+            base_url="https://example.com/v1",
+            quiet_mode=True,
+            skip_context_files=True,
+            skip_memory=True,
+            context_snapshots_enabled=True,
+            context_snapshots_dir=str(tmp_path),
+        )
+
+    agent.client = MagicMock()
+    agent.client.chat.completions.create.return_value = _mock_response(content="Done")
+    agent._build_system_prompt = MagicMock(return_value="Core system prompt")
+    agent._memory_manager = MagicMock()
+    agent._memory_manager.prefetch_all.return_value = "- remembered preference"
+    agent._memory_manager.build_system_prompt.return_value = "# Hindsight Memory\nActive."
+    agent._memory_manager.providers = [
+        SimpleNamespace(name="builtin"),
+        SimpleNamespace(name="hindsight"),
+    ]
+
+    result = agent.run_conversation("What do I prefer?")
+
+    assert result["final_response"] == "Done"
+
+    call_dir = tmp_path / agent.session_id / "call_001"
+    assert call_dir.exists()
+    assert (call_dir / "system_prompt.txt").read_text(encoding="utf-8") == "Core system prompt"
+    assert (call_dir / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
+    assert (call_dir / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered preference"
+    assert "<memory-context>" in (call_dir / "memory_context_block.txt").read_text(encoding="utf-8")
+    api_user_message = (call_dir / "api_user_message.txt").read_text(encoding="utf-8")
+    assert "What do I prefer?" in api_user_message
+    assert "remembered preference" in api_user_message
+    api_messages = (call_dir / "api_messages.json").read_text(encoding="utf-8")
+    assert '"role": "system"' in api_messages
+    assert '"role": "user"' in api_messages
+    metadata = (call_dir / "metadata.json").read_text(encoding="utf-8")
+    assert '"hindsight"' in metadata

tests/test_hindsight_local_eval_doc.py

@@ -0,0 +1,22 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+DOC = ROOT / "docs" / "hindsight-local-eval.md"
+
+
+def test_hindsight_local_eval_doc_exists_and_covers_switching():
+    assert DOC.exists(), "missing Hindsight local eval doc"
+    text = DOC.read_text(encoding="utf-8")
+    for snippet in (
+        "atlas-baseline",
+        "atlas-mempalace",
+        "atlas-hindsight",
+        "HERMES_HOME",
+        "HERMES_CONTEXT_SNAPSHOTS",
+        "memory_prefetch_raw.txt",
+        "api_user_message.txt",
+        "local_embedded",
+        "hindsight/config.json",
+    ):
+        assert snippet in text