docs(analysis): map IDE coding-agent harness techniques onto Hermes

Study of the five harness subsystems that make in-editor coding agents
outperform the raw model (same models underneath), with a grounded map of
each onto the Hermes codebase: indexed semantic retrieval, retrieval-as-
accuracy-driver, decoupled apply model, ambient context, per-task routing.

Findings reference real files (tools/fuzzy_match.py, agent/auxiliary_client.py,
agent/prompt_builder.py, hermes_state.py). Identifies semantic codebase
retrieval as the one structural gap; apply-reliability and model-routing as
existing strengths.

cli.py

@@ -313,6 +313,25 @@ def _load_prefill_messages(file_path: str) -> List[Dict[str, Any]]:
         return []
 
 
+def _resolve_prefill_messages_file(config: Dict[str, Any]) -> str:
+    """Resolve the prefill file path from env/config.
+
+    ``prefill_messages_file`` at the top level is the canonical config key.
+    ``agent.prefill_messages_file`` remains a legacy fallback for older CLI and
+    godmode-generated configs.
+    """
+    env_path = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "").strip()
+    if env_path:
+        return env_path
+    top_level = str(config.get("prefill_messages_file", "") or "").strip()
+    if top_level:
+        return top_level
+    agent_cfg = config.get("agent", {})
+    if isinstance(agent_cfg, dict):
+        return str(agent_cfg.get("prefill_messages_file", "") or "").strip()
+    return ""
+
+
 def _parse_reasoning_config(effort: str) -> dict | None:
     """Parse a reasoning effort level into an OpenRouter reasoning config dict."""
     from hermes_constants import parse_reasoning_effort
@@ -3272,7 +3291,7 @@ class HermesCLI:
         
         # Ephemeral prefill messages (few-shot priming, never persisted)
         self.prefill_messages = _load_prefill_messages(
-            CLI_CONFIG["agent"].get("prefill_messages_file", "")
+            _resolve_prefill_messages_file(CLI_CONFIG)
         )
         
         # Reasoning config (OpenRouter reasoning effort level)

cron/scheduler.py

@@ -1551,9 +1551,16 @@ def _run_job_impl(job: dict) -> tuple[bool, str, str, Optional[str]]:
         effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()
         reasoning_config = parse_reasoning_effort(effort)
 
-        # Prefill messages from env or config.yaml
+        # Prefill messages from env or config.yaml. The top-level
+        # prefill_messages_file key is canonical; agent.prefill_messages_file is
+        # retained as a legacy fallback for older CLI/godmode configs.
         prefill_messages = None
-        prefill_file = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "") or _cfg.get("prefill_messages_file", "")
+        agent_cfg = _cfg.get("agent", {}) if isinstance(_cfg.get("agent", {}), dict) else {}
+        prefill_file = (
+            os.getenv("HERMES_PREFILL_MESSAGES_FILE", "")
+            or _cfg.get("prefill_messages_file", "")
+            or agent_cfg.get("prefill_messages_file", "")
+        )
         if prefill_file:
             pfpath = Path(prefill_file).expanduser()
             if not pfpath.is_absolute():

docs/analysis/coding-agent-harness.md

@@ -0,0 +1,179 @@
+# Why IDE-Embedded Coding Agents Feel Better — and Where Hermes Stands
+
+A study of the *harness* techniques that make in-editor coding agents punch above
+the raw model, and an honest map of which ones Hermes already implements, which it
+approximates, and which it lacks.
+
+## TL;DR
+
+The leading IDE-embedded coding products are **not better models** — they call the
+same frontier models (Claude, GPT) that power terminal agents like this one. Their
+edge comes entirely from the **harness**: how context is retrieved and assembled,
+how mechanical edits are applied reliably, and how cheap specialized models are
+routed in for sub-tasks. The model is a commodity; the meal it's fed is not.
+
+This document breaks the advantage into five concrete subsystems, each backed by
+published engineering from the vendors, and maps each onto the Hermes codebase.
+
+| # | Technique | What it buys | Hermes status |
+|---|-----------|--------------|---------------|
+| 1 | Indexed semantic retrieval (Merkle delta-sync + content-addressed embedding cache) | Knows the repo in ms; feeds the *right* snippets | ❌ **Gap** (grep/FTS5 only, no vector index) |
+| 2 | Retrieval as the accuracy driver | +~12.5% answer accuracy (vendor eval) | ⚠️ **Approximated** (lexical search, not semantic) |
+| 3 | Decoupled "apply" model + line-number-free search/replace | Frontier model only *reasons*; mechanical patching never botches the file | ✅ **Have an analog** (`tools/fuzzy_match.py`) |
+| 4 | Ambient IDE context (cursor pos, selection, live diagnostics) | More intent-signal per token | ⚠️ **Partial** (context files + LSP, no live cursor/selection) |
+| 5 | Per-task model routing (tiny model for autocomplete, frontier for reasoning) | Right tool per job | ✅ **Have** (`agent/auxiliary_client.py`) |
+
+---
+
+## 1. Indexed semantic retrieval — the actual moat
+
+The headline trick is *not* the system prompt. It's a vector index over the repo
+that is kept fresh cheaply:
+
+- **Merkle tree for change detection.** Every file is SHA-256 hashed; folder hashes
+  derive from children; the root summarizes the repo. An edit changes only that
+  file's hash plus the path to the root, so the indexer walks **only the branches
+  that differ** instead of rescanning. (This is git's own content-addressing trick
+  repurposed for indexing.)
+- **Syntax-aware chunking.** Changed files are split on function/class boundaries,
+  not arbitrary token windows, then embedded.
+- **Content-addressed embedding cache.** Embeddings are keyed by the hash of the
+  chunk content. Re-indexing unchanged code is a cache hit → zero embedding cost.
+  Embedding is the expensive step, so this is the whole ballgame for speed.
+- **Cross-clone index reuse.** Vendors observe that clones of one repo are ~92%
+  identical across an org; a "simhash" lets a new clone reuse a teammate's index,
+  collapsing time-to-first-query from hours (99th pct) to seconds. Access is gated
+  cryptographically: you can only compute a Merkle node's hash if you actually hold
+  the file, so results you can't *prove* you possess are dropped.
+
+**Hermes status: this is the real gap.** Core Hermes has no vector index, no
+embedding store, no Merkle delta-sync. Codebase awareness is achieved at task time
+via lexical tools (`search_files` → ripgrep) and session recall via SQLite FTS5
+(`hermes_state.py`, `tools/session_search_tool.py`). The only embedding-flavored
+retrieval lives in an optional plugin (`plugins/memory/holographic/`), and even
+that is FTS5-backed, not dense-vector.
+
+This is a *defensible* design choice for a terminal-first agent — ripgrep over a
+known working tree is fast, dependency-free, and always current — but it means
+Hermes "discovers" a codebase cold each task rather than walking in pre-indexed.
+
+## 2. Retrieval is the accuracy driver (empirically)
+
+Vendor evals attribute **~+12.5% answer accuracy** and higher edit-retention to
+semantic search alone — same model, better-retrieved context. This is the
+empirical proof of the thesis: *a worse model with better context beats a better
+model with worse context.*
+
+**Hermes status: approximated, lexically.** Hermes gets the *shape* of this through
+aggressive context assembly — `agent/prompt_builder.py` and `agent/system_prompt.py`
+inject project context files (AGENTS.md / CLAUDE.md / .cursorrules), and
+`agent/subdirectory_hints.py` surfaces local structure. What's missing is *ranked
+semantic* retrieval: Hermes finds text by pattern, not by meaning. For
+"where is the thing that does X" questions, lexical search is strictly weaker than
+embeddings.
+
+## 3. Decoupled apply model — the most underrated trick
+
+Leading products split edits into two stages:
+
+1. **Plan** — the frontier model emits a *terse* edit, often with
+   `// ... existing code ...` placeholders.
+2. **Apply** — a separate, cheap, often self-hosted model turns that sketch into the
+   final file.
+
+Why bother? Frontier models are *lazy and inaccurate* at large rewrites: they drop
+code, emit `...`, "helpfully" reformat unrelated lines, miscount line numbers, and
+can trap the agent in retry loops. Three published findings drive the design:
+
+- **Whole-file rewrites beat diffs** for the model, because diffs force fewer output
+  tokens (less room to "think"), are out-of-distribution (models saw far more whole
+  files in training), and line numbers are tokenizer poison (a number is one token,
+  forcing a one-shot commit, and models can't count lines).
+- So edits use **search/replace blocks with no line numbers**, with redundant
+  context lines so the parser tolerates model slips.
+- **Speculative decoding** makes apply fast (~1000 tok/s) because the unchanged file
+  *is* the draft — and because that can't be built into hosted Anthropic/OpenAI
+  models, vendors train and self-host their own apply model.
+
+**Hermes status: it has a deterministic analog, and it's good.**
+`tools/fuzzy_match.py` implements an **8-strategy search/replace matcher** (exact →
+line-trimmed → whitespace-normalized → indentation-flexible → escape-normalized →
+trimmed-boundary → block-anchor → context-aware-similarity) that is *precisely* the
+"tolerate model slips in line-number-free search/replace" idea — just solved with
+`difflib.SequenceMatcher` instead of a trained model. `tools/patch_parser.py` and
+`tools/file_tools.py` wire it into the `patch` tool. Hermes also already adopts the
+correct *interface*: the model emits `old_string`/`new_string`, never line numbers.
+
+Where the vendors go further: a *trained* apply model can reconstruct intent from a
+sketch (resolve `// ... existing ...` placeholders against the real file), whereas a
+fuzzy matcher can only locate-and-substitute text the model actually wrote. Hermes
+trades that capability for zero latency, zero cost, and full determinism — a sound
+trade for a local agent, but worth naming.
+
+## 4. Ambient context — the editor's free advantage
+
+Because the product *is* the editor, it injects for free: the open file, **cursor
+position**, current selection, **live LSP/linter diagnostics**, and recent diffs.
+Terminal agents must spend tool calls reconstructing all of this. More intent-signal
+per token → better output from the same model.
+
+**Hermes status: partial.** Hermes injects project context files and has LSP
+plumbing (`agent/lsp/`), and the ACP adapter (`acp_adapter/`) gives editors a way to
+feed edits/approvals back. What it lacks is the *passive* signal: it doesn't know
+where your cursor is or what you've selected, because in a terminal there is no
+cursor to read. The ACP integration narrows this gap when Hermes runs inside an
+editor, but the default terminal surface is signal-poorer by construction.
+
+## 5. Per-task model routing
+
+Autocomplete uses a tiny fast model; chat uses a frontier model; apply uses the
+custom fast model; the agent loop uses a frontier model plus tools. Nothing is
+forced through one monolith.
+
+**Hermes status: have it.** `agent/auxiliary_client.py` provides a routed auxiliary
+model used for cheaper sub-tasks — title generation (`agent/title_generator.py`),
+vision routing (`tools/computer_use/vision_routing.py`), background review
+(`agent/background_review.py`), and conversation compression
+(`agent/context_compressor.py`, `trajectory_compressor.py`). The pattern — reserve
+the expensive model for reasoning, route mechanical sub-tasks to a cheap one — is
+already core to Hermes.
+
+---
+
+## Synthesis: where Hermes wins, ties, and trails
+
+**Ties or wins:**
+- **Apply reliability** — the 8-strategy fuzzy matcher is a genuinely strong,
+  zero-cost analog to a trained apply model, and uses the same line-number-free
+  search/replace interface the research converged on.
+- **Model routing** — auxiliary-client routing already reserves the frontier model
+  for reasoning.
+- **Context-file injection & session memory** — robust, and FTS5 session search is a
+  real recall capability terminal-first.
+
+**Trails:**
+- **Semantic codebase retrieval** is the one structural gap. Hermes is lexical
+  (ripgrep + FTS5) where the leaders are dense-vector with a cheaply-maintained
+  index. This is the highest-leverage area if Hermes ever wants to close the
+  "feels like it already knows my repo" gap.
+- **Ambient passive context** (cursor/selection/live diagnostics) is inherently
+  weaker outside an editor; the ACP path is the right place to invest if that
+  matters.
+
+**The single most transferable insight:** the research independently concluded that
+it is *better to rewrite via fuzzy-tolerant, line-number-free search/replace than to
+trust the smart model to emit a precise diff* — and Hermes already lands on the same
+answer in `tools/fuzzy_match.py`. That convergence is a good sign the harness
+fundamentals here are sound; the missing piece is retrieval, not editing.
+
+## Suggested follow-ups (not implemented here — analysis only)
+
+1. **Optional semantic index plugin.** A content-addressed embedding cache keyed by
+   file hash, behind the existing plugin interface, would give ranked semantic
+   retrieval without bloating the core terminal path. Merkle delta-sync keeps it
+   cheap to refresh.
+2. **Apply-from-sketch mode.** Let the model emit `// ... existing ...` placeholders
+   and resolve them against the real file before handing to the fuzzy matcher —
+   captures most of a trained apply model's benefit deterministically.
+3. **Richer ambient context over ACP.** Pipe editor cursor/selection/diagnostics
+   into prompt assembly when running embedded, closing the passive-signal gap.

gateway/run.py

@@ -3034,13 +3034,16 @@ class GatewayRunner:
         """Load ephemeral prefill messages from config or env var.
         
         Checks HERMES_PREFILL_MESSAGES_FILE env var first, then falls back to
-        the prefill_messages_file key in ~/.hermes/config.yaml.
+        the top-level prefill_messages_file key in ~/.hermes/config.yaml.
+        agent.prefill_messages_file is accepted as a legacy fallback.
         Relative paths are resolved from ~/.hermes/.
         """
         file_path = os.getenv("HERMES_PREFILL_MESSAGES_FILE", "")
         if not file_path:
             cfg = _load_gateway_runtime_config()
             file_path = str(cfg.get("prefill_messages_file", "") or "")
+            if not file_path:
+                file_path = str(cfg_get(cfg, "agent", "prefill_messages_file", default="") or "")
         if not file_path:
             return []
         path = Path(file_path).expanduser()

plugins/memory/supermemory/README.md

@@ -1,6 +1,6 @@
 # Supermemory Memory Provider
 
-Semantic long-term memory with profile recall, semantic search, explicit memory tools, and session-end conversation ingest.
+Semantic long-term memory with profile recall, semantic search, explicit memory tools, and full-session conversation ingest (one ingest per session) for richer profiles.
 
 ## Requirements
 
@@ -45,22 +45,34 @@ Config file: `$HERMES_HOME/supermemory.json`
 
 ## Tools
 
-| Tool | Description |
-|------|-------------|
-| `supermemory_store` | Store an explicit memory |
-| `supermemory_search` | Search memories by semantic similarity |
-| `supermemory_forget` | Forget a memory by ID or best-match query |
-| `supermemory_profile` | Retrieve persistent profile and recent context |
+Kebab-case names are registered for the agent; snake_case aliases remain supported.
+
+| Tool | Alias | Description |
+|------|-------|-------------|
+| `supermemory-save` | `supermemory_store` | Store an explicit memory |
+| `supermemory-search` | `supermemory_search` | Search memories by semantic similarity |
+| `supermemory-forget` | `supermemory_forget` | Forget a memory by ID or best-match query |
+| `supermemory-profile` | `supermemory_profile` | Retrieve persistent profile and recent context |
+
+## Source attribution
+
+All Supermemory API calls send `x-sm-source: hermes`, and document writes stamp
+`metadata.sm_source: hermes`. This is a **functional routing key, not telemetry**:
+it groups Hermes-written memories into a dedicated "Hermes" Space in the
+Supermemory app, so you can filter, browse, and bulk-manage them per source agent
+(alongside Codex, Claude Code, etc.) from the Supermemory UI.
 
 ## Behavior
 
 When enabled, Hermes can:
 
 - prefetch relevant memory context before each turn
-- store cleaned conversation turns after each completed response
-- ingest the full session on session end for richer graph updates
+- buffer the full conversation and ingest it as **one session** at session end (or on `/reset`, branch, compression, or shutdown)
+- ingest the full session to the conversations endpoint for richer profile/graph updates
 - expose explicit tools for search, store, forget, and profile access
 
+The session is written once via the conversations endpoint, which drives Supermemory's entity extraction and profile building while keeping a clean, retrievable full transcript.
+
 ## Profile-Scoped Containers
 
 Use `{identity}` in the `container_tag` to scope memories per Hermes profile:
@@ -87,7 +99,7 @@ For advanced setups (e.g. OpenClaw-style multi-workspace), you can enable custom
 ```
 
 When enabled:
-- `supermemory_search`, `supermemory_store`, `supermemory_forget`, and `supermemory_profile` accept an optional `container_tag` parameter
+- `supermemory-search`, `supermemory-save`, `supermemory-forget`, and `supermemory-profile` accept an optional `container_tag` parameter
 - The tag must be in the whitelist: primary container + `custom_containers`
 - Automatic operations (turn sync, prefetch, memory write mirroring, session ingest) always use the **primary** container only
 - Custom container instructions are injected into the system prompt

plugins/memory/supermemory/__init__.py

@@ -269,7 +269,22 @@ class _SupermemoryClient:
         self._container_tag = container_tag
         self._search_mode = search_mode if search_mode in _VALID_SEARCH_MODES else _DEFAULT_SEARCH_MODE
         self._timeout = timeout
-        self._client = Supermemory(api_key=api_key, timeout=timeout, max_retries=0)
+        self._client = Supermemory(
+            api_key=api_key,
+            timeout=timeout,
+            max_retries=0,
+            default_headers={"x-sm-source": "hermes"},
+        )
+
+    def _merge_metadata(self, metadata: Optional[dict]) -> dict:
+        # sm_source routes Hermes writes into the "Hermes" Space in the Supermemory
+        # app so the user can filter / bulk-manage them per source agent. This is a
+        # functional routing key for the user, not vendor telemetry.
+        merged = {"sm_source": "hermes", **(metadata or {})}
+        legacy_source = merged.pop("source", None)
+        if legacy_source and "type" not in merged:
+            merged["type"] = str(legacy_source)
+        return merged
 
     def add_memory(self, content: str, metadata: Optional[dict] = None, *,
                    entity_context: str = "", container_tag: Optional[str] = None,
@@ -280,7 +295,7 @@ class _SupermemoryClient:
             "container_tags": [tag],
         }
         if metadata:
-            kwargs["metadata"] = metadata
+            kwargs["metadata"] = self._merge_metadata(metadata)
         if entity_context:
             kwargs["entity_context"] = _clamp_entity_context(entity_context)
         if custom_id:
@@ -349,18 +364,22 @@ class _SupermemoryClient:
         preview = (target.get("memory") or "")[:100]
         return {"success": True, "message": f'Forgot: "{preview}"', "id": memory_id}
 
-    def ingest_conversation(self, session_id: str, messages: list[dict]) -> None:
-        payload = json.dumps({
+    def ingest_conversation(self, session_id: str, messages: list[dict], metadata: dict | None = None) -> None:
+        payload: dict = {
             "conversationId": session_id,
             "messages": messages,
             "containerTags": [self._container_tag],
-        }).encode("utf-8")
+        }
+        if metadata:
+            payload["metadata"] = self._merge_metadata(metadata)
+
         req = urllib.request.Request(
             _CONVERSATIONS_URL,
-            data=payload,
+            data=json.dumps(payload).encode("utf-8"),
             headers={
                 "Authorization": f"Bearer {self._api_key}",
                 "Content-Type": "application/json",
+                "x-sm-source": "hermes",
             },
             method="POST",
         )
@@ -447,6 +466,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
         self._custom_containers: List[str] = []
         self._custom_container_instructions = ""
         self._allowed_containers: List[str] = []
+        self._session_turns: List[Dict[str, str]] = []
 
     @property
     def name(self) -> str:
@@ -501,13 +521,13 @@ class SupermemoryMemoryProvider(MemoryProvider):
         self._search_mode = self._config["search_mode"]
         self._entity_context = self._config["entity_context"]
         self._api_timeout = self._config["api_timeout"]
-
-        # Multi-container setup
         self._enable_custom_containers = self._config["enable_custom_container_tags"]
         self._custom_containers = self._config["custom_containers"]
         self._custom_container_instructions = self._config["custom_container_instructions"]
         self._allowed_containers = [self._container_tag] + list(self._custom_containers)
 
+        self._session_turns = []
+
         agent_context = kwargs.get("agent_context", "")
         self._write_enabled = agent_context not in {"cron", "flush", "subagent"}
         self._active = bool(self._api_key)
@@ -534,7 +554,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
         lines = [
             "# Supermemory",
             f"Active. Container: {self._container_tag}.",
-            "Use supermemory_search, supermemory_store, supermemory_forget, and supermemory_profile for explicit memory operations.",
+            "Use supermemory-search, supermemory-save, supermemory-forget, and supermemory-profile (aliases: supermemory_search, supermemory_store, supermemory_forget, supermemory_profile).",
         ]
         if self._enable_custom_containers and self._custom_containers:
             tags_str = ", ".join(self._allowed_containers)
@@ -567,31 +587,11 @@ class SupermemoryMemoryProvider(MemoryProvider):
 
         clean_user = _clean_text_for_capture(user_content)
         clean_assistant = _clean_text_for_capture(assistant_content)
-        if not clean_user or not clean_assistant:
+        if not clean_user and not clean_assistant:
             return
-        if self._capture_mode == "all":
-            if len(clean_user) < _MIN_CAPTURE_LENGTH or len(clean_assistant) < _MIN_CAPTURE_LENGTH:
-                return
-            if _is_trivial_message(clean_user):
-                return
 
-        content = (
-            f"[role: user]\n{clean_user}\n[user:end]\n\n"
-            f"[role: assistant]\n{clean_assistant}\n[assistant:end]"
-        )
-        metadata = {"source": "hermes", "type": "conversation_turn"}
-
-        def _run():
-            try:
-                self._client.add_memory(content, metadata=metadata, entity_context=self._entity_context)
-            except Exception:
-                logger.debug("Supermemory sync_turn failed", exc_info=True)
-
-        if self._sync_thread and self._sync_thread.is_alive():
-            self._sync_thread.join(timeout=2.0)
-        self._sync_thread = None
-        self._sync_thread = threading.Thread(target=_run, daemon=True, name="supermemory-sync")
-        self._sync_thread.start()
+        # Buffer every turn for the single full-session document written at end/switch/shutdown
+        self._session_turns.append({"user": clean_user, "assistant": clean_assistant})
 
     def on_session_end(self, messages: List[Dict[str, Any]]) -> None:
         if not self._active or not self._write_enabled or not self._client or not self._session_id:
@@ -609,12 +609,68 @@ class SupermemoryMemoryProvider(MemoryProvider):
         if len(cleaned) == 1 and len(cleaned[0].get("content", "")) < 20:
             return
         try:
-            self._client.ingest_conversation(self._session_id, cleaned)
+            self._client.ingest_conversation(
+                self._session_id,
+                cleaned,
+                metadata={
+                    "type": "full_session",
+                    "session_id": self._session_id,
+                    "message_count": len(cleaned),
+                },
+            )
         except urllib.error.HTTPError:
             logger.warning("Supermemory session ingest failed", exc_info=True)
         except Exception:
             logger.warning("Supermemory session ingest failed", exc_info=True)
 
+        # Clear buffer so shutdown() doesn't duplicate on normal exit
+        self._session_turns = []
+
+    def on_session_switch(
+        self,
+        new_session_id: str,
+        *,
+        parent_session_id: str = "",
+        reset: bool = False,
+        **kwargs,
+    ) -> None:
+        """Flush any buffered turns from the old session as one document, then reset for the new session."""
+        if not self._active or not self._write_enabled or not self._client:
+            self._session_id = str(new_session_id or "").strip() or self._session_id
+            self._session_turns = []
+            return
+
+        old_session_id = self._session_id
+        old_turns = list(self._session_turns)
+
+        # Flush previous session via conversations ingest (with metadata)
+        if old_turns and old_session_id:
+            messages: list[dict] = []
+            for turn in old_turns:
+                if turn.get("user"):
+                    messages.append({"role": "user", "content": turn["user"]})
+                if turn.get("assistant"):
+                    messages.append({"role": "assistant", "content": turn["assistant"]})
+
+            try:
+                self._client.ingest_conversation(
+                    old_session_id,
+                    messages,
+                    metadata={
+                        "type": "full_session",
+                        "session_id": old_session_id,
+                        "message_count": len(old_turns) * 2,
+                        "partial": not reset,
+                    },
+                )
+            except Exception:
+                logger.debug("Supermemory session-switch ingest failed", exc_info=True)
+
+        # Reset for new session
+        self._session_id = str(new_session_id or "").strip() or old_session_id
+        self._session_turns = []
+        self._turn_count = 0
+
     def on_memory_write(self, action: str, target: str, content: str) -> None:
         if not self._active or not self._write_enabled or not self._client:
             return
@@ -625,7 +681,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
             try:
                 self._client.add_memory(
                     content.strip(),
-                    metadata={"source": "hermes_memory", "target": target, "type": "explicit_memory"},
+                    metadata={"target": target, "type": "explicit_memory"},
                     entity_context=self._entity_context,
                 )
             except Exception:
@@ -638,6 +694,31 @@ class SupermemoryMemoryProvider(MemoryProvider):
         self._write_thread.start()
 
     def shutdown(self) -> None:
+        # Emergency fallback (crashes only). Buffer is cleared on normal on_session_end().
+        if self._active and self._write_enabled and self._client and self._session_turns and self._session_id:
+            logger.warning("Supermemory: Saving session via shutdown (session=%s, turns=%d)", self._session_id, len(self._session_turns))
+
+            messages: list[dict] = []
+            for turn in self._session_turns:
+                if turn.get("user"):
+                    messages.append({"role": "user", "content": turn["user"]})
+                if turn.get("assistant"):
+                    messages.append({"role": "assistant", "content": turn["assistant"]})
+
+            try:
+                self._client.ingest_conversation(
+                    self._session_id,
+                    messages,
+                    metadata={
+                        "type": "full_session",
+                        "session_id": self._session_id,
+                        "message_count": len(self._session_turns) * 2,
+                        "partial": True,
+                    },
+                )
+            except Exception:
+                logger.debug("Supermemory shutdown ingest failed", exc_info=True)
+
         for attr_name in ("_prefetch_thread", "_sync_thread", "_write_thread"):
             thread = getattr(self, attr_name, None)
             if thread and thread.is_alive():
@@ -665,8 +746,25 @@ class SupermemoryMemoryProvider(MemoryProvider):
         return sanitized
 
     def get_tool_schemas(self) -> List[Dict[str, Any]]:
+        def with_kebab_aliases(schemas: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+            aliases = {
+                "supermemory_store": "supermemory-save",
+                "supermemory_search": "supermemory-search",
+                "supermemory_forget": "supermemory-forget",
+                "supermemory_profile": "supermemory-profile",
+            }
+            expanded = list(schemas)
+            for schema in schemas:
+                kebab = aliases.get(schema.get("name", ""))
+                if not kebab:
+                    continue
+                copy = json.loads(json.dumps(schema))
+                copy["name"] = kebab
+                expanded.append(copy)
+            return expanded
+
         if not self._enable_custom_containers:
-            return [STORE_SCHEMA, SEARCH_SCHEMA, FORGET_SCHEMA, PROFILE_SCHEMA]
+            return with_kebab_aliases([STORE_SCHEMA, SEARCH_SCHEMA, FORGET_SCHEMA, PROFILE_SCHEMA])
 
         # When multi-container is enabled, add optional container_tag to relevant tools
         container_param = {
@@ -678,7 +776,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
             schema = json.loads(json.dumps(base))  # deep copy
             schema["parameters"]["properties"]["container_tag"] = container_param
             schemas.append(schema)
-        return schemas
+        return with_kebab_aliases(schemas)
 
     def _tool_store(self, args: dict) -> str:
         content = str(args.get("content") or "").strip()
@@ -692,7 +790,7 @@ class SupermemoryMemoryProvider(MemoryProvider):
         if not isinstance(metadata, dict):
             metadata = {}
         metadata.setdefault("type", _detect_category(content))
-        metadata["source"] = "hermes_tool"
+        metadata.pop("source", None)
         try:
             result = self._client.add_memory(content, metadata=metadata, entity_context=self._entity_context, container_tag=tag)
             preview = content[:80] + ("..." if len(content) > 80 else "")
@@ -777,6 +875,13 @@ class SupermemoryMemoryProvider(MemoryProvider):
     def handle_tool_call(self, tool_name: str, args: Dict[str, Any], **kwargs) -> str:
         if not self._active or not self._client:
             return tool_error("Supermemory is not configured")
+        aliases = {
+            "supermemory-save": "supermemory_store",
+            "supermemory-search": "supermemory_search",
+            "supermemory-forget": "supermemory_forget",
+            "supermemory-profile": "supermemory_profile",
+        }
+        tool_name = aliases.get(tool_name, tool_name)
         if tool_name == "supermemory_store":
             return self._tool_store(args)
         if tool_name == "supermemory_search":

plugins/memory/supermemory/plugin.yaml

@@ -1,5 +1,5 @@
 name: supermemory
-version: 1.0.0
+version: 1.0.1
 description: "Supermemory semantic long-term memory with profile recall, semantic search, explicit memory tools, and session ingest."
 pip_dependencies:
   - supermemory

scripts/release.py

@@ -70,6 +70,7 @@ AUTHOR_MAP = {
     "524706+Twanislas@users.noreply.github.com": "Twanislas",
     "9592417+adam91holt@users.noreply.github.com": "adam91holt",
     "kchuang1015@users.noreply.github.com": "kchuang1015",
+    "maheshthedev@gmail.com": "MaheshtheDev",
     "kyssta-exe@users.noreply.github.com": "kyssta-exe",
     "45688690+fujinice@users.noreply.github.com": "fujinice",
     "276689385+carltonawong@users.noreply.github.com": "carltonawong",

skills/red-teaming/godmode/SKILL.md

@@ -90,7 +90,7 @@ undo_jailbreak()
 7. **If a strategy works**, locks it in:
    - Writes the winning system prompt to `agent.system_prompt` in `config.yaml`
    - Writes prefill messages to `~/.hermes/prefill.json`
-   - Sets `agent.prefill_messages_file: "prefill.json"` in `config.yaml`
+   - Sets `prefill_messages_file: "prefill.json"` in `config.yaml`
 8. **Reports results** — which strategy won, score, preview of compliant response
 
 ### Strategy order per model family:
@@ -171,8 +171,7 @@ Create `~/.hermes/prefill.json`:
 
 Then set in `~/.hermes/config.yaml`:
 ```yaml
-agent:
-  prefill_messages_file: "prefill.json"
+prefill_messages_file: "prefill.json"
 ```
 
 Prefill messages are injected at the start of every API call, after the system prompt. They are ephemeral — never saved to sessions or trajectories. The model sees them as prior conversation context, establishing a pattern of compliance.

skills/red-teaming/godmode/scripts/auto_jailbreak.py

@@ -397,7 +397,8 @@ def _write_config(system_prompt: str = None, prefill_file: str = None):
         cfg["agent"]["system_prompt"] = system_prompt
 
     if prefill_file is not None:
-        cfg["agent"]["prefill_messages_file"] = prefill_file
+        cfg["prefill_messages_file"] = prefill_file
+        cfg["agent"].pop("prefill_messages_file", None)
 
     with open(CONFIG_PATH, "w") as f:
         yaml.dump(cfg, f, default_flow_style=False, allow_unicode=True,
@@ -721,6 +722,7 @@ def undo_jailbreak(verbose=True):
             if "agent" in cfg:
                 cfg["agent"].pop("system_prompt", None)
                 cfg["agent"].pop("prefill_messages_file", None)
+            cfg.pop("prefill_messages_file", None)
             with open(CONFIG_PATH, "w") as f:
                 yaml.dump(cfg, f, default_flow_style=False, allow_unicode=True,
                           width=120, sort_keys=False)

tests/cli/test_prefill_config.py

@@ -0,0 +1,35 @@
+"""Regression tests for CLI prefill config key compatibility."""
+
+from __future__ import annotations
+
+import cli
+
+
+def test_resolve_prefill_messages_file_uses_top_level(monkeypatch):
+    monkeypatch.delenv("HERMES_PREFILL_MESSAGES_FILE", raising=False)
+
+    assert cli._resolve_prefill_messages_file(
+        {
+            "prefill_messages_file": "top.json",
+            "agent": {"prefill_messages_file": "legacy.json"},
+        }
+    ) == "top.json"
+
+
+def test_resolve_prefill_messages_file_accepts_legacy_agent_key(monkeypatch):
+    monkeypatch.delenv("HERMES_PREFILL_MESSAGES_FILE", raising=False)
+
+    assert cli._resolve_prefill_messages_file(
+        {"agent": {"prefill_messages_file": "legacy.json"}}
+    ) == "legacy.json"
+
+
+def test_resolve_prefill_messages_file_prefers_env(monkeypatch):
+    monkeypatch.setenv("HERMES_PREFILL_MESSAGES_FILE", "env.json")
+
+    assert cli._resolve_prefill_messages_file(
+        {
+            "prefill_messages_file": "top.json",
+            "agent": {"prefill_messages_file": "legacy.json"},
+        }
+    ) == "env.json"

tests/cron/test_scheduler.py

@@ -1546,6 +1546,36 @@ class TestRunJobConfigEnvVarExpansion:
             "config.yaml ${VAR} was not expanded in the cron execution path."
         )
 
+    def test_legacy_agent_prefill_messages_file_is_loaded(self, tmp_path, monkeypatch):
+        """Cron accepts the legacy agent.prefill_messages_file fallback."""
+        prefill = [{"role": "system", "content": "legacy cron prefill"}]
+        (tmp_path / "prefill.json").write_text(json.dumps(prefill), encoding="utf-8")
+        (tmp_path / "config.yaml").write_text(
+            "agent:\n"
+            "  prefill_messages_file: prefill.json\n",
+            encoding="utf-8",
+        )
+
+        job = {"id": "prefill-job", "name": "prefill test", "prompt": "hi"}
+        fake_db = MagicMock()
+
+        with patch("cron.scheduler._hermes_home", tmp_path), \
+             patch("cron.scheduler._resolve_origin", return_value=None), \
+             patch("dotenv.load_dotenv"), \
+             patch("hermes_state.SessionDB", return_value=fake_db), \
+             patch("hermes_cli.runtime_provider.resolve_runtime_provider",
+                   return_value=self._RUNTIME), \
+             patch("tools.mcp_tool.discover_mcp_tools", return_value=[]), \
+             patch("run_agent.AIAgent") as mock_agent_cls:
+            mock_agent = MagicMock()
+            mock_agent.run_conversation.return_value = {"final_response": "ok"}
+            mock_agent_cls.return_value = mock_agent
+            success, _, _, error = run_job(job)
+
+        assert success is True
+        assert error is None
+        assert mock_agent_cls.call_args.kwargs["prefill_messages"] == prefill
+
     def test_fallback_model_env_ref_in_config_yaml_is_expanded(self, tmp_path, monkeypatch):
         """${VAR} in config.yaml fallback_providers model: is expanded."""
         (tmp_path / "config.yaml").write_text(

tests/gateway/test_runtime_config_env_expansion.py

@@ -33,6 +33,29 @@ def test_load_prefill_messages_expands_env_var_path(monkeypatch, gateway_home):
     assert gateway_run.GatewayRunner._load_prefill_messages() == prefill
 
 
+def test_load_prefill_messages_accepts_legacy_agent_key(monkeypatch, gateway_home):
+    prefill = [{"role": "system", "content": "legacy few-shot"}]
+    (gateway_home / "prefill.json").write_text(json.dumps(prefill), encoding="utf-8")
+    _write_config(gateway_home, "agent:\n  prefill_messages_file: prefill.json\n")
+
+    assert gateway_run.GatewayRunner._load_prefill_messages() == prefill
+
+
+def test_load_prefill_messages_prefers_top_level_over_legacy(monkeypatch, gateway_home):
+    top_level = [{"role": "system", "content": "top-level"}]
+    legacy = [{"role": "system", "content": "legacy"}]
+    (gateway_home / "top.json").write_text(json.dumps(top_level), encoding="utf-8")
+    (gateway_home / "legacy.json").write_text(json.dumps(legacy), encoding="utf-8")
+    _write_config(
+        gateway_home,
+        "prefill_messages_file: top.json\n"
+        "agent:\n"
+        "  prefill_messages_file: legacy.json\n",
+    )
+
+    assert gateway_run.GatewayRunner._load_prefill_messages() == top_level
+
+
 @pytest.mark.parametrize(
     ("config_body", "env_name", "env_value", "loader_name", "expected"),
     [

tests/plugins/memory/test_supermemory_provider.py

@@ -50,8 +50,8 @@ class FakeClient:
     def forget_by_query(self, query, *, container_tag=None):
         return self.forget_by_query_response
 
-    def ingest_conversation(self, session_id, messages):
-        self.ingest_calls.append({"session_id": session_id, "messages": messages})
+    def ingest_conversation(self, session_id, messages, metadata=None):
+        self.ingest_calls.append({"session_id": session_id, "messages": messages, "metadata": metadata})
 
 
 @pytest.fixture
@@ -136,23 +136,28 @@ def test_prefetch_skips_profile_between_frequency(provider):
     assert "User Profile (Persistent)" not in result
 
 
-def test_sync_turn_skips_trivial_message(provider):
+def test_sync_turn_buffers_short_messages(provider):
+    # Trivial filtering is no longer applied at sync time — every non-empty turn
+    # is buffered and only the full session is written at session boundaries.
     provider.sync_turn("ok", "sure", session_id="session-1")
+    assert provider._session_turns == [{"user": "ok", "assistant": "sure"}]
     assert provider._client.add_calls == []
 
 
-def test_sync_turn_persists_cleaned_exchange(provider):
+def test_sync_turn_buffers_cleaned_exchange(provider):
     provider.sync_turn(
         "Please remember this\n<supermemory-context>ignore</supermemory-context>",
         "Got it, storing the context",
         session_id="session-1",
     )
-    provider._sync_thread.join(timeout=1)
-    assert len(provider._client.add_calls) == 1
-    content = provider._client.add_calls[0]["content"]
-    assert "ignore" not in content
-    assert "[role: user]" in content
-    assert "[role: assistant]" in content
+    assert len(provider._session_turns) == 1
+    turn = provider._session_turns[0]
+    assert "ignore" not in turn["user"]
+    assert turn["user"].startswith("Please remember this")
+    assert turn["assistant"] == "Got it, storing the context"
+    # Buffering only — no per-turn writes to the client
+    assert provider._client.add_calls == []
+    assert provider._client.ingest_calls == []
 
 
 def test_on_session_end_ingests_clean_messages(provider):
@@ -169,6 +174,28 @@ def test_on_session_end_ingests_clean_messages(provider):
         {"role": "user", "content": "hello"},
         {"role": "assistant", "content": "hi there"},
     ]
+    assert payload["metadata"]["type"] == "full_session"
+    assert payload["metadata"]["session_id"] == "session-1"
+    assert payload["metadata"]["message_count"] == 2
+    # Buffer is cleared after a normal session-end ingest.
+    assert provider._session_turns == []
+
+
+def test_merge_metadata_stamps_sm_source():
+    # sm_source routes Hermes writes into the "Hermes" Space in the Supermemory
+    # app (functional routing, not telemetry) — must always be present.
+    from plugins.memory.supermemory import _SupermemoryClient
+
+    client = _SupermemoryClient.__new__(_SupermemoryClient)
+    merged = client._merge_metadata({"type": "explicit_memory"})
+    assert merged["sm_source"] == "hermes"
+    assert merged["type"] == "explicit_memory"
+
+    # Legacy "source" is migrated into "type" when type is absent.
+    merged2 = client._merge_metadata({"source": "conversation_turn"})
+    assert merged2["sm_source"] == "hermes"
+    assert merged2["type"] == "conversation_turn"
+    assert "source" not in merged2
 
 
 def test_on_memory_write_tracks_thread(provider):
@@ -179,7 +206,7 @@ def test_on_memory_write_tracks_thread(provider):
     assert provider._client.add_calls[0]["metadata"]["type"] == "explicit_memory"
 
 
-def test_shutdown_joins_and_clears_threads(provider, monkeypatch):
+def test_shutdown_joins_threads_and_flushes_buffer(provider, monkeypatch):
     started = threading.Event()
     release = threading.Event()
 
@@ -196,15 +223,16 @@ def test_shutdown_joins_and_clears_threads(provider, monkeypatch):
 
     monkeypatch.setattr(provider._client, "add_memory", slow_add_memory)
 
+    # sync_turn now only buffers — no thread is spawned.
     provider.sync_turn(
         "Please remember this request in long-term memory",
         "Absolutely, I will keep that in long-term memory.",
         session_id="session-1",
     )
-    assert started.wait(timeout=1)
-    assert provider._sync_thread is not None
+    assert provider._sync_thread is None
+    assert len(provider._session_turns) == 1
 
-    started.clear()
+    # on_memory_write still runs on a background thread.
     provider.on_memory_write("add", "memory", "Jordan likes concise docs")
     assert started.wait(timeout=1)
     assert provider._write_thread is not None
@@ -212,10 +240,18 @@ def test_shutdown_joins_and_clears_threads(provider, monkeypatch):
     release.set()
     provider.shutdown()
 
+    # All tracked threads joined and cleared.
     assert provider._sync_thread is None
     assert provider._write_thread is None
     assert provider._prefetch_thread is None
-    assert len(provider._client.add_calls) == 2
+    # Explicit memory write went through.
+    assert len(provider._client.add_calls) == 1
+    # Buffered turn was flushed as a partial full-session ingest.
+    assert len(provider._client.ingest_calls) == 1
+    payload = provider._client.ingest_calls[0]
+    assert payload["session_id"] == "session-1"
+    assert payload["metadata"]["partial"] is True
+    assert payload["metadata"]["type"] == "full_session"
 
 
 def test_store_tool_returns_saved_payload(provider):

website/docs/user-guide/features/memory-providers.md

@@ -66,7 +66,7 @@ AI-native cross-session user modeling with dialectic reasoning, session-scoped c
 hermes memory setup        # select "honcho" — runs the Honcho-specific post-setup
 ```
 
-On a fresh install, configure Honcho directly with `hermes memory setup honcho`. The legacy `hermes honcho setup` command still works (it now redirects to `hermes memory setup`), but is only registered after Honcho is selected as the active memory provider.
+The legacy `hermes honcho setup` command still works (it now redirects to `hermes memory setup`), but is only registered after Honcho is selected as the active memory provider.
 
 **Config:** `$HERMES_HOME/honcho.json` (profile-local) or `~/.honcho/config.json` (global). Resolution order: `$HERMES_HOME/honcho.json` > `~/.hermes/honcho.json` > `~/.honcho/config.json`. See the [config reference](https://github.com/NousResearch/hermes-agent/blob/main/plugins/memory/honcho/README.md) and the [Honcho integration guide](https://docs.honcho.dev/v3/guides/integrations/hermes).
 
@@ -498,11 +498,11 @@ echo 'SUPERMEMORY_API_KEY=***' >> ~/.hermes/.env
 
 **Key features:**
 - Automatic context fencing — strips recalled memories from captured turns to prevent recursive memory pollution
-- Session-end conversation ingest for richer graph-level knowledge building
+- Full-session ingest — the entire conversation is sent once at session boundaries
+- Session-end conversation ingest (to `/v4/conversations`) for richer profile + graph building in Supermemory
 - Profile facts injected on first turn and at configurable intervals
-- Trivial message filtering (skips "ok", "thanks", etc.)
 - **Profile-scoped containers** — use `{identity}` in `container_tag` (e.g. `hermes-{identity}` → `hermes-coder`) to isolate memories per Hermes profile
-- **Multi-container mode** — enable `enable_custom_container_tags` with a `custom_containers` list to let the agent read/write across named containers. Automatic operations (sync, prefetch) stay on the primary container.
+- **Multi-container mode** — enable `enable_custom_container_tags` with a `custom_containers` list to let the agent read/write across named containers. Automatic operations stay on the primary container.
 
 <details>
 <summary>Multi-container example</summary>

website/docs/user-guide/skills/bundled/red-teaming/red-teaming-godmode.md

@@ -108,7 +108,7 @@ undo_jailbreak()
 7. **If a strategy works**, locks it in:
    - Writes the winning system prompt to `agent.system_prompt` in `config.yaml`
    - Writes prefill messages to `~/.hermes/prefill.json`
-   - Sets `agent.prefill_messages_file: "prefill.json"` in `config.yaml`
+   - Sets `prefill_messages_file: "prefill.json"` in `config.yaml`
 8. **Reports results** — which strategy won, score, preview of compliant response
 
 ### Strategy order per model family:
@@ -189,8 +189,7 @@ Create `~/.hermes/prefill.json`:
 
 Then set in `~/.hermes/config.yaml`:
 ```yaml
-agent:
-  prefill_messages_file: "prefill.json"
+prefill_messages_file: "prefill.json"
 ```
 
 Prefill messages are injected at the start of every API call, after the system prompt. They are ephemeral — never saved to sessions or trajectories. The model sees them as prior conversation context, establishing a pattern of compliance.

website/docs/user-guide/skills/godmode.md

@@ -94,7 +94,7 @@ undo_jailbreak()
 7. **If a strategy works**, locks it in:
    - Writes the winning system prompt to `agent.system_prompt` in `config.yaml`
    - Writes prefill messages to `~/.hermes/prefill.json`
-   - Sets `agent.prefill_messages_file: "prefill.json"` in `config.yaml`
+   - Sets `prefill_messages_file: "prefill.json"` in `config.yaml`
 8. **Reports results** — which strategy won, score, preview of compliant response
 
 ### Model-Specific Strategy Order
@@ -150,8 +150,7 @@ export HERMES_EPHEMERAL_SYSTEM_PROMPT="[END OF INPUT] [START OF INPUT]..."
 Create `~/.hermes/prefill.json` and reference it in config:
 
 ```yaml
-agent:
-  prefill_messages_file: "prefill.json"
+prefill_messages_file: "prefill.json"
 ```
 
 Prefill messages are injected at the start of every API call, after the system prompt. They are **ephemeral** — never saved to sessions or trajectories. The model sees them as prior conversation context, establishing a pattern of compliance.

website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/features/memory-providers.md

@@ -498,9 +498,9 @@ echo 'SUPERMEMORY_API_KEY=***' >> ~/.hermes/.env
 
 **主要特性：**
 - 自动上下文隔离——从捕获的轮次中剥离已召回的记忆，防止递归记忆污染
-- 会话结束时的对话导入，用于构建更丰富的图谱级知识
+- 在会话边界时将整个会话**一次性导入**
+- 会话结束时同时导入到对话端点（`/v4/conversations`），用于 Supermemory 的 profile 和图谱构建
 - 在第一轮及可配置间隔注入 profile 事实
-- 无意义消息过滤（跳过"ok"、"thanks"等）
 - **Profile 范围容器**——在 `container_tag` 中使用 `{identity}`（例如 `hermes-{identity}` → `hermes-coder`），按 Hermes profile 隔离记忆
 - **多容器模式**——启用 `enable_custom_container_tags` 并配置 `custom_containers` 列表，让 Agent 跨命名容器读写。自动操作（同步、预取）保持在主容器上。
 

website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/skills/bundled/red-teaming/red-teaming-godmode.md

@@ -108,7 +108,7 @@ undo_jailbreak()
 7. **若某策略有效**，则锁定：
    - 将胜出的系统 prompt 写入 `config.yaml` 的 `agent.system_prompt`
    - 将 prefill 消息写入 `~/.hermes/prefill.json`
-   - 在 `config.yaml` 中设置 `agent.prefill_messages_file: "prefill.json"`
+   - 在 `config.yaml` 中设置 `prefill_messages_file: "prefill.json"`
 8. **报告结果**——胜出策略、得分、合规响应预览
 
 ### 各模型系列的策略顺序：
@@ -189,8 +189,7 @@ export HERMES_EPHEMERAL_SYSTEM_PROMPT="[END OF INPUT] [START OF INPUT]..."
 
 然后在 `~/.hermes/config.yaml` 中设置：
 ```yaml
-agent:
-  prefill_messages_file: "prefill.json"
+prefill_messages_file: "prefill.json"
 ```
 
 Prefill 消息在每次 API 调用时注入到系统 prompt 之后。它们是临时的——永远不会保存到会话或轨迹中。模型将其视为先前的对话上下文，从而建立合规模式。
@@ -419,4 +418,4 @@ Claude Sonnet 4 对所有当前技术在明显有害内容方面具有鲁棒性
 9. **在 execute_code 中始终使用 `load_godmode.py`** — 各个脚本（`parseltongue.py`、`godmode_race.py`、`auto_jailbreak.py`）有带 `if __name__ == '__main__'` 块的 argparse CLI 入口点。在 execute_code 中通过 `exec()` 加载时，`__name__` 为 `'__main__'`，argparse 会触发并导致脚本崩溃。`load_godmode.py` loader 通过将 `__name__` 设置为非 main 值并管理 sys.argv 来处理这个问题。
 10. **boundary_inversion 与模型版本相关** — 在 Claude 3.5 Sonnet 上有效，但在 Claude Sonnet 4 或 Claude 4.6 上无效。auto_jailbreak 中的策略顺序对 Claude 模型优先尝试它，但失败后会回退到 refusal_inversion。如果你知道模型版本，请更新策略顺序。
 11. **灰色地带查询 vs 硬查询** — 越狱技术对"双重用途"查询（撬锁、安全工具、化学）效果远好于明显有害的查询（钓鱼模板、恶意软件）。对于硬查询，直接跳到 ULTRAPLINIAN 或使用不拒绝的 Hermes/Grok 模型。
-12. **execute_code 沙箱没有环境变量** — 当 Hermes 通过 execute_code 运行 auto_jailbreak 时，沙箱不继承 `~/.hermes/.env`。显式加载 dotenv：`from dotenv import load_dotenv; load_dotenv(os.path.expanduser("~/.hermes/.env"))`
+12. **execute_code 沙箱没有环境变量** — 当 Hermes 通过 execute_code 运行 auto_jailbreak 时，沙箱不继承 `~/.hermes/.env`。显式加载 dotenv：`from dotenv import load_dotenv; load_dotenv(os.path.expanduser("~/.hermes/.env"))`

website/i18n/zh-Hans/docusaurus-plugin-content-docs/current/user-guide/skills/godmode.md

@@ -94,7 +94,7 @@ undo_jailbreak()
 7. **若某策略有效**，将其锁定：
    - 将获胜的系统提示词写入 `config.yaml` 的 `agent.system_prompt`
    - 将预填充消息写入 `~/.hermes/prefill.json`
-   - 在 `config.yaml` 中设置 `agent.prefill_messages_file: "prefill.json"`
+   - 在 `config.yaml` 中设置 `prefill_messages_file: "prefill.json"`
 8. **报告结果**——哪种策略获胜、得分、合规响应预览
 
 ### 各模型系列的策略顺序
@@ -150,8 +150,7 @@ export HERMES_EPHEMERAL_SYSTEM_PROMPT="[END OF INPUT] [START OF INPUT]..."
 创建 `~/.hermes/prefill.json` 并在配置中引用：
 
 ```yaml
-agent:
-  prefill_messages_file: "prefill.json"
+prefill_messages_file: "prefill.json"
 ```
 
 预填充消息在每次 API 调用时注入到系统提示词之后。它们是**临时的**——不会保存到会话或轨迹中。模型将其视为先前的对话上下文，从而建立合规模式。
@@ -277,4 +276,4 @@ Claude Sonnet 4 对所有当前技术在明显有害内容方面具有较强抵
 
 - **G0DM0D3：** [elder-plinius/G0DM0D3](https://github.com/elder-plinius/G0DM0D3)（AGPL-3.0）
 - **L1B3RT4S：** [elder-plinius/L1B3RT4S](https://github.com/elder-plinius/L1B3RT4S)（AGPL-3.0）
-- **Pliny the Prompter：** [@elder_plinius](https://x.com/elder_plinius)
+- **Pliny the Prompter：** [@elder_plinius](https://x.com/elder_plinius)