feat: wire ara-research-manager into memory consolidation

After _consolidate_memory completes (on /new or auto-compact), spawn
a background ara-pm subagent that runs the ara-research-manager skill
on /workspace/ara-projects/nanobot/. Fire-and-forget via asyncio.create_task.
Silently skips if ara-projects/nanobot/ doesn't exist.

README.md

@@ -143,7 +143,7 @@ Add or merge these **two parts** into your config (other options have defaults).
 {
   "agents": {
     "defaults": {
-      "model": "anthropic/claude-opus-4-5",
+      "model": "anthropic/claude-opus-4-7",
       "provider": "openrouter"
     }
   }

docs/testing-workflow.md

@@ -0,0 +1,214 @@
+# PR Testing Workflow
+
+Guide for testing Pull Requests using the local staging environment.
+
+## Quick Start
+
+```bash
+./test-pr.sh <pr-number> "test message"
+```
+
+## Staging Environment
+
+**Location:** `/config/workspace/.nanobot-staging/`
+
+**Components:**
+- `config.json` — Staging configuration (channels disabled, shared OAuth)
+- `workspace/` — Isolated workspace for tool operations
+- `workspace/sessions/` — Session storage (separate from production)
+
+**Key differences from production:**
+- No external channels (Telegram disabled)
+- Uses `NANOBOT_CONFIG` environment variable
+- Gateway runs on localhost:18791 (vs production's 18790)
+- `restrictToWorkspace: true` for safety
+
+## Testing a PR
+
+### Method 1: Helper Script (Recommended)
+
+```bash
+# Test PR with default message
+./test-pr.sh 31
+
+# Test with custom message
+./test-pr.sh 31 "test the hidden message feature"
+```
+
+**What it does:**
+1. Fetches PR from `wylab` remote (force updates if branch exists)
+2. Checks out PR branch locally
+3. Installs in editable mode with `uv pip install -e .`
+4. Runs test with staging config via `NANOBOT_CONFIG` env var
+5. Leaves branch checked out for further testing
+
+**After testing:**
+```bash
+git checkout main  # Return to main branch
+```
+
+### Method 2: Manual Testing
+
+```bash
+# 1. Fetch and checkout PR
+cd /config/workspace/nanobot-oauth-port/nanobot-fork
+git fetch wylab pull/<N>/head:pr-<N>
+git checkout pr-<N>
+
+# 2. Install in editable mode
+uv pip install -e .
+
+# 3. Test with staging config
+NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json \
+  .venv/bin/nanobot agent -m "test message"
+
+# 4. For multi-turn testing
+NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json \
+  .venv/bin/nanobot agent  # Interactive mode
+
+# 5. Return to main
+git checkout main
+```
+
+### Method 3: Gateway Validation
+
+Test that gateway starts without errors:
+
+```bash
+NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json \
+  .venv/bin/nanobot gateway
+
+# Kill with Ctrl+C when validated
+```
+
+## Verifying Cache Behavior
+
+To verify prompt caching works correctly (important for performance):
+
+```bash
+# Enable logs to see cache metrics
+NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json \
+  .venv/bin/nanobot agent --logs -m "Turn 1: list files"
+
+# Look for cache metrics in output:
+# - cache_write: New cache entries created
+# - cache_read: Tokens read from cache
+```
+
+**What to look for:**
+- Turn 1: High `cache_write`, moderate `cache_read`
+- Turn 2+: Low `cache_write`, high `cache_read` (reusing cache)
+- `cache_read` should increase across turns as context grows
+
+**Example healthy pattern:**
+```
+Turn 1: cache_write=354 cache_read=3563
+Turn 2: cache_write=255 cache_read=3917  ← Same as Turn 1 end
+Turn 3: cache_write=113 cache_read=4172  ← Growing with context
+```
+
+## Session Management
+
+### Clear session for fresh test
+
+```bash
+rm -f /config/workspace/.nanobot-staging/workspace/sessions/cli_direct.jsonl
+```
+
+### View session contents
+
+```bash
+cat /config/workspace/.nanobot-staging/workspace/sessions/cli_direct.jsonl | jq
+```
+
+### Check for specific features (e.g., hidden signatures)
+
+```bash
+cat /config/workspace/.nanobot-staging/workspace/sessions/cli_direct.jsonl | grep "_hidden_sig"
+```
+
+## Common Testing Scenarios
+
+### Test tool execution
+
+```bash
+./test-pr.sh 31 "List all Python files in the current directory"
+```
+
+### Test multi-turn conversation
+
+```bash
+NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json \
+  .venv/bin/nanobot agent
+
+# Then interact naturally:
+> list files in current directory
+> how many python files are there?
+> what's the total size?
+```
+
+### Test error handling
+
+```bash
+./test-pr.sh 31 "Try to read a file that doesn't exist: /nonexistent.txt"
+```
+
+### Test with thinking mode
+
+The staging config has `thinking_budget: 10000` enabled by default, so all tests use extended thinking.
+
+## Troubleshooting
+
+### "No API key configured" error
+
+- **Cause:** `NANOBOT_CONFIG` env var not set
+- **Fix:** Ensure you're using `NANOBOT_CONFIG=/config/workspace/.nanobot-staging/config.json`
+
+### "Module not found" after checkout
+
+- **Cause:** Need to reinstall after switching branches
+- **Fix:** Run `uv pip install -e .` after checkout
+
+### Changes not applying
+
+- **Cause:** Using cached `.pyc` files
+- **Fix:** Clear pycache: `find . -type d -name __pycache__ -exec rm -rf {} + 2>/dev/null || true`
+
+### Session has stale data
+
+- **Cause:** Previous test left session data
+- **Fix:** `rm /config/workspace/.nanobot-staging/workspace/sessions/cli_direct.jsonl`
+
+## Best Practices
+
+1. **Clear session between PR tests** to avoid cross-contamination
+2. **Test with tool use** to trigger agentic behavior (not just simple Q&A)
+3. **Check cache metrics** for performance-sensitive PRs
+4. **Run with `--logs`** to see detailed behavior during development
+5. **Return to main** after testing to avoid accidental commits on PR branches
+
+## Integration with CI/CD
+
+The staging environment is currently manual-only. Future enhancements:
+
+- [ ] Automated PR testing via Gitea Actions
+- [ ] Cache validation in CI pipeline
+- [ ] Multi-PR parallel testing using git worktrees
+- [ ] Regression test suite against production behavior
+
+## File Locations Reference
+
+| Path | Purpose |
+|------|---------|
+| `/config/workspace/nanobot-oauth-port/nanobot-fork/` | Local nanobot repository |
+| `/config/workspace/.nanobot-staging/` | Staging environment root |
+| `/config/workspace/.nanobot-staging/config.json` | Staging configuration |
+| `/config/workspace/.nanobot-staging/workspace/` | Staging workspace |
+| `/config/workspace/.nanobot-staging/workspace/sessions/` | Session storage |
+| `/config/workspace/nanobot-oauth-port/nanobot-fork/test-pr.sh` | Helper script |
+
+## Related Documentation
+
+- [nanobot README](../README.md) - Main project documentation
+- [CLAUDE.md](../CLAUDE.md) - Development guide for Claude Code
+- [config/schema.py](../nanobot/config/schema.py) - Configuration schema

nanobot/agent/context.py

@@ -11,6 +11,7 @@ from loguru import logger
 from nanobot.agent.memory import MemoryStore
 from nanobot.agent.memory_mem0 import Mem0MemoryStore, HAS_MEM0
 from nanobot.agent.skills import SkillsLoader
+from nanobot.agent.visibility import compute_signature
 
 
 class ContextBuilder:
@@ -226,12 +227,14 @@ visibility markers will be rejected."""
         Returns:
             Updated message list.
         """
-        messages.append({
+        msg: dict[str, Any] = {
             "role": "tool",
             "tool_call_id": tool_call_id,
             "name": tool_name,
-            "content": result
-        })
+            "content": result,
+            "_hidden_sig": compute_signature(result if isinstance(result, str) else ""),
+        }
+        messages.append(msg)
         return messages
     
     def add_assistant_message(
@@ -254,13 +257,14 @@ visibility markers will be rejected."""
             Updated message list.
         """
         msg: dict[str, Any] = {"role": "assistant", "content": content or ""}
-        
+
         if tool_calls:
             msg["tool_calls"] = tool_calls
-        
+            msg["_hidden_sig"] = compute_signature(content or "")
+
         # Thinking models reject history without this
         if reasoning_content:
             msg["reasoning_content"] = reasoning_content
-        
+
         messages.append(msg)
         return messages

nanobot/agent/loop.py

@@ -11,7 +11,7 @@ from loguru import logger
 
 from nanobot.bus.events import InboundMessage, OutboundMessage
 from nanobot.bus.queue import MessageBus
-from nanobot.providers.base import LLMProvider
+from nanobot.providers.base import LLMProvider, LongContextError
 from nanobot.agent.context import ContextBuilder
 from nanobot.agent.tools.registry import ToolRegistry
 from nanobot.agent.tools.filesystem import ReadFileTool, WriteFileTool, EditFileTool, ListDirTool
@@ -221,7 +221,7 @@ class AgentLoop:
             return self._quota_cache["model"]
 
         # Default models
-        OPUS = "claude-opus-4-6"
+        OPUS = "claude-opus-4-7"
         SONNET = "claude-sonnet-4-6"
         TOLERANCE = 1.17  # 17% overage triggers downgrade
 
@@ -417,12 +417,35 @@ class AgentLoop:
 
             # Call LLM
             logger.debug(f"Calling LLM with model={selected_model}, provider.thinking_budget={self.provider.thinking_budget}")
-            response = await self.provider.chat(
-                messages=messages,
-                tools=self.tools.get_tools(),  # Pass tool objects for beta flag extraction
-                model=selected_model,
-                context_management=self.CONTEXT_MANAGEMENT,
-            )
+            try:
+                response = await self.provider.chat(
+                    messages=messages,
+                    tools=self.tools.get_tools(),  # Pass tool objects for beta flag extraction
+                    model=selected_model,
+                    context_management=self.CONTEXT_MANAGEMENT,
+                )
+            except LongContextError:
+                logger.warning("Long context 429 — auto-consolidating session")
+                await self._consolidate_memory(session, archive_all=False)
+                # Apply trim immediately (normally deferred to end of turn)
+                checkpoint = getattr(session, '_trim_checkpoint', None)
+                if checkpoint is not None:
+                    old_size = len(session.messages)
+                    session.messages = session.messages[checkpoint:]
+                    session._trim_checkpoint = None
+                    self.sessions.save(session)
+                    logger.info(f"Emergency trim: {old_size} -> {len(session.messages)} messages")
+                    # Rebuild messages from trimmed session
+                    messages = self.context.build_messages(
+                        history=session.get_history(),
+                        current_message=current_message,
+                        media=msg.media if msg.media else None,
+                        channel=msg.channel,
+                        chat_id=msg.chat_id,
+                    )
+                    turn_start = len(messages)
+                    continue  # Retry LLM call with shorter context
+                raise  # No trim happened — can't recover
 
             # Handle tool calls
             if response.has_tool_calls:
@@ -672,12 +695,32 @@ class AgentLoop:
         while iteration < self.max_iterations:
             iteration += 1
 
-            response = await self.provider.chat(
-                messages=messages,
-                tools=self.tools.get_tools(),  # Pass tool objects for beta flag extraction
-                model=selected_model,
-                context_management=self.CONTEXT_MANAGEMENT,
-            )
+            try:
+                response = await self.provider.chat(
+                    messages=messages,
+                    tools=self.tools.get_tools(),  # Pass tool objects for beta flag extraction
+                    model=selected_model,
+                    context_management=self.CONTEXT_MANAGEMENT,
+                )
+            except LongContextError:
+                logger.warning("Long context 429 in system handler — auto-consolidating")
+                await self._consolidate_memory(session, archive_all=False)
+                checkpoint = getattr(session, '_trim_checkpoint', None)
+                if checkpoint is not None:
+                    old_size = len(session.messages)
+                    session.messages = session.messages[checkpoint:]
+                    session._trim_checkpoint = None
+                    self.sessions.save(session)
+                    logger.info(f"Emergency trim: {old_size} -> {len(session.messages)} messages")
+                    messages = self.context.build_messages(
+                        history=session.get_history(),
+                        current_message=msg.content,
+                        channel=origin_channel,
+                        chat_id=origin_chat_id,
+                    )
+                    turn_start = len(messages)
+                    continue
+                raise
 
             if response.has_tool_calls:
                 tool_call_dicts = [
@@ -1008,9 +1051,33 @@ Respond with ONLY valid JSON, no markdown fences."""
             session.messages = self._trim_to_clean_boundary(session.messages, keep_count) if keep_count else []
             self.sessions.save(session)
             logger.info(f"Memory consolidation done, session trimmed to {len(session.messages)} messages")
+            self._spawn_ara_research_manager()
         except Exception as e:
             logger.error(f"Memory consolidation failed: {e}")
 
+    def _spawn_ara_research_manager(self) -> None:
+        """Fire-and-forget: update the nanobot ARA trace after memory consolidation.
+
+        Uses the SpawnTool which already has channel/chat_id context from the
+        last message. Silently skips if ARA path doesn't exist or spawn fails.
+        """
+        import os
+        ara_path = os.path.join(self.workspace, "ara-projects", "nanobot")
+        if not os.path.isdir(ara_path):
+            return
+        spawn_tool = self.tools.get("spawn")
+        if not isinstance(spawn_tool, SpawnTool):
+            return
+        task = (
+            f"Read /root/.nanobot/workspace/skills/ara-research-manager/SKILL.md "
+            f"and run the per-turn procedure on the nanobot ARA at {ara_path}/. "
+            f"Capture any research events from this session."
+        )
+        asyncio.create_task(
+            spawn_tool.execute(task=task, label="ara-pm", model="claude-haiku-4-5")
+        )
+        logger.debug("ARA research-manager task spawned")
+
     async def process_direct(
         self,
         content: str,

nanobot/agent/memory_mem0.py

@@ -175,8 +175,9 @@ class Mem0MemoryStore:
             response = await provider.chat(
                 messages=extraction_messages,
                 model=model,
-                max_tokens=2000,
+                max_tokens=16384,
                 temperature=0.3,
+                thinking_budget=0,
             )
             text = (response.content or "").strip()
             if text.startswith("```"):
@@ -211,16 +212,23 @@ class Mem0MemoryStore:
 
         stored = 0
         for fact in facts:
+            # Normalize: LLM may return dicts like {"fact": "...", "date": "..."} or plain strings
+            if isinstance(fact, dict):
+                fact_text = fact.get("fact", fact.get("text", str(fact)))
+            else:
+                fact_text = str(fact)
+            if not fact_text.strip():
+                continue
             try:
                 self.memory.add(
-                    fact,
+                    fact_text,
                     user_id=user_id,
                     infer=False,
                     metadata=metadata if metadata else None,
                 )
                 stored += 1
             except Exception as e:
-                logger.error(f"Failed to store fact '{fact[:50]}...': {e}")
+                logger.error(f"Failed to store fact '{str(fact_text)[:50]}...': {e}")
 
         logger.info(f"Stored {stored}/{len(facts)} facts for user {user_id}")
 

nanobot/agent/visibility.py

@@ -4,11 +4,19 @@
 import hmac
 import hashlib
 import re
-from typing import Tuple
 
 SECRET_KEY = "nanobot_visibility_secret_key_v1"
 
 
+def compute_signature(content: str) -> str:
+    """Compute HMAC signature for content (hex string, no prefix)."""
+    return hmac.new(
+        SECRET_KEY.encode(),
+        content.encode(),
+        hashlib.sha256
+    ).hexdigest()[:8]
+
+
 def sign_content(content: str) -> str:
     """
     Sign content with HMAC and prepend marker.
@@ -19,15 +27,11 @@ def sign_content(content: str) -> str:
     Returns:
         Content with signed visibility marker: "[HIDDEN:{sig}] {content}"
     """
-    sig = hmac.new(
-        SECRET_KEY.encode(),
-        content.encode(),
-        hashlib.sha256
-    ).hexdigest()[:8]
+    sig = compute_signature(content)
     return f"[HIDDEN:{sig}] {content}"
 
 
-def verify_signature(marked_content: str) -> Tuple[bool, str]:
+def verify_signature(marked_content: str) -> tuple[bool, str]:
     """
     Verify HMAC signature and extract clean content.
 
@@ -44,12 +48,7 @@ def verify_signature(marked_content: str) -> Tuple[bool, str]:
         return False, marked_content
 
     claimed_sig, content = match.groups()
-    expected_sig = hmac.new(
-        SECRET_KEY.encode(),
-        content.encode(),
-        hashlib.sha256
-    ).hexdigest()[:8]
-
+    expected_sig = compute_signature(content)
     is_valid = hmac.compare_digest(claimed_sig, expected_sig)
     return is_valid, content
 

nanobot/config/schema.py

@@ -220,7 +220,7 @@ class AgentDefaults(Base):
     """Default agent configuration."""
 
     workspace: str = "~/.nanobot/workspace"
-    model: str = "anthropic/claude-opus-4-5"
+    model: str = "anthropic/claude-opus-4-7"
     provider: str = "auto"  # Provider name (e.g. "anthropic", "openrouter") or "auto" for auto-detection
     max_tokens: int = 8192
     temperature: float = 0.1

nanobot/providers/__init__.py

@@ -1,6 +1,6 @@
 """Provider module exports."""
 
-from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.base import LLMProvider, LLMResponse, LongContextError, ToolCallRequest
 from nanobot.providers.litellm_provider import LiteLLMProvider
 from nanobot.providers.openai_codex_provider import OpenAICodexProvider
 from nanobot.providers.anthropic_oauth import AnthropicOAuthProvider

nanobot/providers/anthropic_oauth.py

@@ -10,8 +10,8 @@ from typing import Any
 import httpx
 from loguru import logger
 
-from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
-from nanobot.providers.oauth_utils import get_auth_headers
+from nanobot.providers.base import LLMProvider, LLMResponse, LongContextError, ToolCallRequest
+from nanobot.providers.oauth_utils import get_auth_headers, get_claude_code_system_prefix
 
 
 class AnthropicOAuthProvider(LLMProvider):
@@ -27,7 +27,7 @@ class AnthropicOAuthProvider(LLMProvider):
     def __init__(
         self,
         oauth_token: str,
-        default_model: str = "claude-opus-4-5",
+        default_model: str = "claude-opus-4-7",
         api_base: str | None = None,
         thinking_budget: int = 0,
     ):
@@ -51,8 +51,8 @@ class AnthropicOAuthProvider(LLMProvider):
     def _normalize_model(model: str) -> str:
         """Normalize model name for the Anthropic API.
 
-        Anthropic model IDs use hyphens (claude-sonnet-4-5), but users often
-        write dots (claude-sonnet-4.5). Normalize so both work.
+        Anthropic model IDs use hyphens (claude-sonnet-4-6), but users often
+        write dots (claude-sonnet-4.6). Normalize so both work.
         """
         return model.replace(".", "-")
 
@@ -377,7 +377,14 @@ class AnthropicOAuthProvider(LLMProvider):
             payload["temperature"] = temperature
 
         if system:
-            payload["system"] = [{"type": "text", "text": system, "cache_control": {"type": "ephemeral", "ttl": "1h"}}]
+            payload["system"] = [
+                {"type": "text", "text": get_claude_code_system_prefix()},
+                {"type": "text", "text": system, "cache_control": {"type": "ephemeral", "ttl": "1h"}},
+            ]
+        else:
+            payload["system"] = [
+                {"type": "text", "text": get_claude_code_system_prefix()},
+            ]
 
         if tools:
             cached_tools = list(tools)
@@ -424,70 +431,114 @@ class AnthropicOAuthProvider(LLMProvider):
 
         import asyncio
         import time as _time
-        _t0 = _time.monotonic()
-        try:
-            response = await client.post(
-                self._get_api_url(),
-                headers=headers,
-                json=payload,
-            )
-        except httpx.ConnectTimeout:
-            elapsed = _time.monotonic() - _t0
-            logger.error(f"ConnectTimeout after {elapsed:.1f}s — running diagnostics")
-            await self._diagnose_connectivity()
-            await self._reset_client()
-            raise
-        except httpx.PoolTimeout:
-            elapsed = _time.monotonic() - _t0
-            logger.error(f"PoolTimeout after {elapsed:.1f}s — resetting client")
-            await self._reset_client()
-            raise
-        except (httpx.ConnectError, httpx.TimeoutException) as e:
-            elapsed = _time.monotonic() - _t0
-            logger.error(f"{type(e).__name__} after {elapsed:.1f}s")
-            raise
-        elapsed = _time.monotonic() - _t0
-        if elapsed > 30:
-            logger.warning(f"Anthropic API slow response: {elapsed:.1f}s")
 
-        # Dump rate limit headers for analysis
-        try:
-            import datetime
-            import os
-            header_dump = {
-                "timestamp": datetime.datetime.now(datetime.UTC).isoformat(),
-                "status_code": response.status_code,
-                "model": payload.get("model"),
-                "headers": dict(response.headers),
-            }
-            dump_path = "/root/.nanobot/workspace/api_headers.jsonl"
-            with open(dump_path, "a") as f:
-                f.write(json.dumps(header_dump) + "\n")
-            # Capture rate limit state for quota-based model switching
-            hdrs = response.headers
-            rate_limit_state = {
-                "updated_at": datetime.datetime.utcnow().isoformat(),
-                "model": payload.get("model"),
-                "weekly_all_models": float(hdrs["anthropic-ratelimit-unified-7d-utilization"]) if hdrs.get("anthropic-ratelimit-unified-7d-utilization") else None,
-                "weekly_sonnet": float(hdrs["anthropic-ratelimit-unified-7d_sonnet-utilization"]) if hdrs.get("anthropic-ratelimit-unified-7d_sonnet-utilization") else None,
-                "session_5h": float(hdrs["anthropic-ratelimit-unified-5h-utilization"]) if hdrs.get("anthropic-ratelimit-unified-5h-utilization") else None,
-                "weekly_reset": int(hdrs["anthropic-ratelimit-unified-7d-reset"]) if hdrs.get("anthropic-ratelimit-unified-7d-reset") else None,
-                "session_reset": int(hdrs["anthropic-ratelimit-unified-5h-reset"]) if hdrs.get("anthropic-ratelimit-unified-5h-reset") else None,
-                "binding_limit": hdrs.get("anthropic-ratelimit-unified-representative-claim"),
-                "sonnet_fallback": hdrs.get("anthropic-ratelimit-unified-fallback"),
-            }
-            state_path = "/root/.nanobot/workspace/memory/rate_limits.json"
-            os.makedirs(os.path.dirname(state_path), exist_ok=True)
-            with open(state_path, "w") as f:
-                json.dump(rate_limit_state, f, indent=2)
-        except Exception as e:
-            logger.warning("Rate limit header capture failed: {}", e)
+        max_retries = 3
+        base_delay = 2.0  # seconds
 
-        if response.status_code != 200:
-            error_text = response.text
-            raise Exception(f"Anthropic API error {response.status_code}: {error_text}")
+        for attempt in range(max_retries + 1):
+            _t0 = _time.monotonic()
+            try:
+                response = await client.post(
+                    self._get_api_url(),
+                    headers=headers,
+                    json=payload,
+                )
+            except httpx.ConnectTimeout:
+                elapsed = _time.monotonic() - _t0
+                logger.error(f"ConnectTimeout after {elapsed:.1f}s (attempt {attempt+1}/{max_retries+1})")
+                if attempt == 0:
+                    await self._diagnose_connectivity()
+                await self._reset_client()
+                if attempt < max_retries:
+                    delay = base_delay * (2 ** attempt)
+                    logger.info(f"Retrying in {delay:.1f}s...")
+                    await asyncio.sleep(delay)
+                    continue
+                raise
+            except httpx.PoolTimeout:
+                elapsed = _time.monotonic() - _t0
+                logger.error(f"PoolTimeout after {elapsed:.1f}s (attempt {attempt+1}/{max_retries+1})")
+                await self._reset_client()
+                if attempt < max_retries:
+                    delay = base_delay * (2 ** attempt)
+                    logger.info(f"Retrying in {delay:.1f}s...")
+                    await asyncio.sleep(delay)
+                    continue
+                raise
+            except (httpx.ConnectError, httpx.TimeoutException) as e:
+                elapsed = _time.monotonic() - _t0
+                logger.error(f"{type(e).__name__} after {elapsed:.1f}s (attempt {attempt+1}/{max_retries+1})")
+                if attempt < max_retries:
+                    delay = base_delay * (2 ** attempt)
+                    logger.info(f"Retrying in {delay:.1f}s...")
+                    await asyncio.sleep(delay)
+                    continue
+                raise
+            elapsed = _time.monotonic() - _t0
+            if elapsed > 30:
+                logger.warning(f"Anthropic API slow response: {elapsed:.1f}s")
 
-        return response.json()
+            # Dump rate limit headers for analysis
+            try:
+                import datetime
+                import os
+                header_dump = {
+                    "timestamp": datetime.datetime.now(datetime.UTC).isoformat(),
+                    "status_code": response.status_code,
+                    "model": payload.get("model"),
+                    "headers": dict(response.headers),
+                }
+                dump_path = "/root/.nanobot/workspace/api_headers.jsonl"
+                with open(dump_path, "a") as f:
+                    f.write(json.dumps(header_dump) + "\n")
+                # Capture rate limit state for quota-based model switching
+                hdrs = response.headers
+                rate_limit_state = {
+                    "updated_at": datetime.datetime.utcnow().isoformat(),
+                    "model": payload.get("model"),
+                    "weekly_all_models": float(hdrs["anthropic-ratelimit-unified-7d-utilization"]) if hdrs.get("anthropic-ratelimit-unified-7d-utilization") else None,
+                    "weekly_sonnet": float(hdrs["anthropic-ratelimit-unified-7d_sonnet-utilization"]) if hdrs.get("anthropic-ratelimit-unified-7d_sonnet-utilization") else None,
+                    "session_5h": float(hdrs["anthropic-ratelimit-unified-5h-utilization"]) if hdrs.get("anthropic-ratelimit-unified-5h-utilization") else None,
+                    "weekly_reset": int(hdrs["anthropic-ratelimit-unified-7d-reset"]) if hdrs.get("anthropic-ratelimit-unified-7d-reset") else None,
+                    "session_reset": int(hdrs["anthropic-ratelimit-unified-5h-reset"]) if hdrs.get("anthropic-ratelimit-unified-5h-reset") else None,
+                    "binding_limit": hdrs.get("anthropic-ratelimit-unified-representative-claim"),
+                    "sonnet_fallback": hdrs.get("anthropic-ratelimit-unified-fallback"),
+                }
+                state_path = "/root/.nanobot/workspace/memory/rate_limits.json"
+                os.makedirs(os.path.dirname(state_path), exist_ok=True)
+                with open(state_path, "w") as f:
+                    json.dump(rate_limit_state, f, indent=2)
+            except Exception as e:
+                logger.warning("Rate limit header capture failed: {}", e)
+
+            # Retry on 5xx server errors and 429 rate limits
+            if response.status_code >= 500 or response.status_code == 429:
+                error_text = response.text
+                logger.warning(f"Anthropic API {response.status_code} (attempt {attempt+1}/{max_retries+1}): {error_text[:200]}")
+
+                # Long context 429 — retrying won't help, need to trim context
+                if response.status_code == 429 and "long context" in error_text.lower():
+                    raise LongContextError(f"Context too long for current plan: {error_text[:200]}")
+
+                if attempt < max_retries:
+                    if response.status_code == 429:
+                        retry_after = response.headers.get("retry-after")
+                        delay = float(retry_after) if retry_after else base_delay * (2 ** attempt)
+                    else:
+                        delay = base_delay * (2 ** attempt)
+                    logger.info(f"Retrying in {delay:.1f}s...")
+                    await asyncio.sleep(delay)
+                    continue
+                raise Exception(f"Anthropic API error {response.status_code}: {error_text}")
+
+            if response.status_code != 200:
+                error_text = response.text
+                raise Exception(f"Anthropic API error {response.status_code}: {error_text}")
+
+            return response.json()
+
+        # Should not reach here, but just in case
+        raise Exception("Exhausted all retry attempts")
 
     async def chat(
         self,
@@ -506,7 +557,7 @@ class AnthropicOAuthProvider(LLMProvider):
         if "/" in model:
             model = model.split("/")[-1]
 
-        # Normalize dots to hyphens (claude-sonnet-4.5 -> claude-sonnet-4-5)
+        # Normalize dots to hyphens (claude-sonnet-4.6 -> claude-sonnet-4-6)
         model = self._normalize_model(model)
 
         system, prepared_messages = self._prepare_messages(messages)
@@ -539,6 +590,8 @@ class AnthropicOAuthProvider(LLMProvider):
                 beta_flags=beta_flags,
             )
             return self._parse_response(response)
+        except LongContextError:
+            raise  # Let caller handle context trimming
         except Exception as e:
             logger.exception("Exception in chat():")
             error_msg = f"{type(e).__name__}: {str(e)}" if str(e) else f"{type(e).__name__} (no message)"

nanobot/providers/base.py

@@ -28,6 +28,11 @@ class LLMResponse:
         return len(self.tool_calls) > 0
 
 
+class LongContextError(Exception):
+    """Raised when the API rejects a request due to long context limits."""
+    pass
+
+
 class LLMProvider(ABC):
     """
     Abstract base class for LLM providers.

nanobot/providers/litellm_provider.py

@@ -37,7 +37,7 @@ class LiteLLMProvider(LLMProvider):
         self, 
         api_key: str | None = None, 
         api_base: str | None = None,
-        default_model: str = "anthropic/claude-opus-4-5",
+        default_model: str = "anthropic/claude-opus-4-7",
         extra_headers: dict[str, str] | None = None,
         provider_name: str | None = None,
     ):
@@ -187,7 +187,7 @@ class LiteLLMProvider(LLMProvider):
         Args:
             messages: List of message dicts with 'role' and 'content'.
             tools: Optional list of tool definitions in OpenAI format.
-            model: Model identifier (e.g., 'anthropic/claude-sonnet-4-5').
+            model: Model identifier (e.g., 'anthropic/claude-sonnet-4-6').
             max_tokens: Maximum tokens in response.
             temperature: Sampling temperature.
         

nanobot/providers/oauth_utils.py

@@ -36,3 +36,11 @@ def get_auth_headers(token: str, is_oauth: bool = False) -> dict[str, str]:
         headers["x-api-key"] = token
 
     return headers
+
+
+def get_claude_code_system_prefix() -> str:
+    """Get the required system prompt prefix for OAuth tokens.
+
+    Anthropic requires this identity declaration for OAuth auth.
+    """
+    return "You are a Claude agent, built on Anthropic's Claude Agent SDK."

nanobot/session/manager.py

@@ -59,13 +59,22 @@ class Session:
         trimming old tool chains safely at token thresholds, so we send the full
         history and let the server decide what to drop.
 
+        Messages with ``_hidden_sig`` get a ``[HIDDEN:{sig}]`` prefix applied to
+        their content so the model knows the user never saw them.  The prefix is
+        applied at read time (not stored in content) to preserve prompt-cache
+        stability: the same prefixed string is produced every turn.
+
         Returns:
             List of messages in LLM format (API-relevant fields only).
         """
-        return [
-            {k: v for k, v in m.items() if k in self._API_FIELDS and v is not None}
-            for m in self.messages
-        ]
+        out: list[dict[str, Any]] = []
+        for m in self.messages:
+            msg = {k: v for k, v in m.items() if k in self._API_FIELDS and v is not None}
+            sig = m.get("_hidden_sig")
+            if sig and isinstance(msg.get("content"), str):
+                msg["content"] = f"[HIDDEN:{sig}] {msg['content']}"
+            out.append(msg)
+        return out
     
     def clear(self) -> None:
         """Clear all messages and reset session to initial state."""
@@ -178,6 +187,25 @@ class SessionManager:
                 f.write(json.dumps(msg, ensure_ascii=False) + "\n")
 
         self._cache[session.key] = session
+        self._append_audit(session)
+
+    def _append_audit(self, session: Session) -> None:
+        """Append session state to an audit log (append-only, rotated monthly)."""
+        now = datetime.now()
+        safe_key = safe_filename(session.key.replace(":", "_"))
+        audit_path = self.sessions_dir / f"{safe_key}.audit.{now:%Y-%m}.jsonl"
+        try:
+            with open(audit_path, "a", encoding="utf-8") as f:
+                marker = {
+                    "_type": "save_marker",
+                    "timestamp": now.isoformat(),
+                    "message_count": len(session.messages),
+                }
+                f.write(json.dumps(marker, ensure_ascii=False) + "\n")
+                for msg in session.messages:
+                    f.write(json.dumps(msg, ensure_ascii=False) + "\n")
+        except Exception as e:
+            logger.warning("Audit log write failed for {}: {}", session.key, e)
     
     def invalidate(self, key: str) -> None:
         """Remove a session from the in-memory cache."""

tests/test_anthropic_oauth.py

@@ -10,14 +10,14 @@ def provider():
     """Create provider with test OAuth token."""
     return AnthropicOAuthProvider(
         oauth_token="sk-ant-oat01-test-token",
-        default_model="claude-opus-4-5"
+        default_model="claude-opus-4-7"
     )
 
 
 def test_provider_init(provider):
     """Provider should initialize with OAuth token."""
     assert provider.oauth_token == "sk-ant-oat01-test-token"
-    assert provider.default_model == "claude-opus-4-5"
+    assert provider.default_model == "claude-opus-4-7"
 
 
 def test_provider_uses_bearer_auth(provider):

tests/test_config_oauth_integration.py

@@ -13,7 +13,7 @@ def test_oauth_token_injected_into_config(tmp_path, monkeypatch):
     # Create a minimal config file (no api key set)
     config_path = tmp_path / "config.json"
     config_path.write_text(json.dumps({
-        "agents": {"defaults": {"model": "anthropic/claude-opus-4-5"}},
+        "agents": {"defaults": {"model": "anthropic/claude-opus-4-7"}},
         "providers": {"anthropic": {"apiKey": ""}}
     }))
 

tests/test_long_context_recovery.py

@@ -0,0 +1,78 @@
+"""Test auto-consolidation on long context 429 errors."""
+
+import pytest
+from unittest.mock import AsyncMock, MagicMock
+
+from nanobot.providers.base import LongContextError, LLMResponse
+
+
+def test_long_context_error_is_exception():
+    """LongContextError should be a distinct exception class."""
+    err = LongContextError("too long")
+    assert isinstance(err, Exception)
+    assert str(err) == "too long"
+
+
+@pytest.mark.asyncio
+async def test_provider_raises_long_context_error_on_long_context_429():
+    """Provider should raise LongContextError immediately for long-context 429s."""
+    from nanobot.providers.anthropic_oauth import AnthropicOAuthProvider
+
+    provider = AnthropicOAuthProvider(
+        oauth_token="sk-ant-oat01-test-token",
+        default_model="claude-sonnet-4-6",
+    )
+
+    mock_response = MagicMock()
+    mock_response.status_code = 429
+    mock_response.text = '{"type":"error","error":{"type":"rate_limit_error","message":"Extra usage is required for long context requests."}}'
+    mock_response.headers = {}
+
+    mock_client = AsyncMock()
+    mock_client.post.return_value = mock_response
+    provider._client = mock_client
+
+    with pytest.raises(LongContextError, match="Context too long"):
+        await provider._make_request(
+            messages=[{"role": "user", "content": "hello"}],
+        )
+
+    # Should NOT retry — only one call
+    assert mock_client.post.call_count == 1
+
+
+@pytest.mark.asyncio
+async def test_provider_retries_normal_429():
+    """Provider should still retry normal 429s (not long-context)."""
+    from nanobot.providers.anthropic_oauth import AnthropicOAuthProvider
+
+    provider = AnthropicOAuthProvider(
+        oauth_token="sk-ant-oat01-test-token",
+        default_model="claude-sonnet-4-6",
+    )
+
+    rate_limit_response = MagicMock()
+    rate_limit_response.status_code = 429
+    rate_limit_response.text = '{"type":"error","error":{"type":"rate_limit_error","message":"Rate limit exceeded"}}'
+    rate_limit_response.headers = {}
+
+    success_response = MagicMock()
+    success_response.status_code = 200
+    success_response.headers = {}
+    success_response.json.return_value = {
+        "content": [{"type": "text", "text": "ok"}],
+        "stop_reason": "end_turn",
+        "usage": {"input_tokens": 1, "output_tokens": 1},
+    }
+
+    mock_client = AsyncMock()
+    mock_client.post.side_effect = [rate_limit_response, success_response]
+    provider._client = mock_client
+
+    result = await provider._make_request(
+        messages=[{"role": "user", "content": "hello"}],
+    )
+
+    # Should have retried and succeeded
+    assert mock_client.post.call_count == 2
+    assert result["stop_reason"] == "end_turn"

tests/test_mem0_extract_facts.py

@@ -0,0 +1,129 @@
+"""Test mem0 fact extraction calls provider with thinking disabled."""
+
+import pytest
+from unittest.mock import AsyncMock, MagicMock
+from pathlib import Path
+
+from nanobot.providers.base import LLMResponse
+
+
+@pytest.fixture
+def mock_provider():
+    provider = AsyncMock()
+    provider.chat = AsyncMock(return_value=LLMResponse(
+        content='{"facts": ["user likes Python", "user works on nanobot"]}',
+        finish_reason="end_turn",
+    ))
+    return provider
+
+
+@pytest.fixture
+def mem0_store(tmp_path):
+    """Create a Mem0MemoryStore with mocked mem0 dependency."""
+    # We can't import Mem0MemoryStore at module level because it requires
+    # the mem0 package. Instead, we test extract_facts as a standalone method
+    # by constructing a minimal instance.
+    try:
+        from nanobot.agent.memory_mem0 import Mem0MemoryStore
+        store = Mem0MemoryStore(workspace=tmp_path)
+        return store
+    except ImportError:
+        pytest.skip("mem0 not installed")
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_passes_thinking_budget_zero(mock_provider):
+    """extract_facts must pass thinking_budget=0 to provider.chat().
+
+    Without this, the provider inherits its instance default (e.g. 10000),
+    causing the model to spend tokens on thinking instead of outputting JSON.
+    """
+    try:
+        from nanobot.agent.memory_mem0 import Mem0MemoryStore
+    except ImportError:
+        pytest.skip("mem0 not installed")
+
+    # Create a minimal instance without full mem0 init
+    store = object.__new__(Mem0MemoryStore)
+    store.custom_prompt = "Extract facts as JSON: "
+
+    messages = [
+        {"role": "user", "content": "I like Python programming"},
+        {"role": "assistant", "content": "That's great! Python is versatile."},
+    ]
+
+    facts = await store.extract_facts(messages, mock_provider, "claude-sonnet-4-6")
+
+    # Verify provider.chat was called with thinking_budget=0
+    mock_provider.chat.assert_called_once()
+    call_kwargs = mock_provider.chat.call_args.kwargs
+    assert call_kwargs["thinking_budget"] == 0
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_returns_parsed_facts(mock_provider):
+    """extract_facts should parse JSON response into a list of fact strings."""
+    try:
+        from nanobot.agent.memory_mem0 import Mem0MemoryStore
+    except ImportError:
+        pytest.skip("mem0 not installed")
+
+    store = object.__new__(Mem0MemoryStore)
+    store.custom_prompt = "Extract facts as JSON: "
+
+    messages = [
+        {"role": "user", "content": "I like Python programming"},
+    ]
+
+    facts = await store.extract_facts(messages, mock_provider, "claude-sonnet-4-6")
+
+    assert facts == ["user likes Python", "user works on nanobot"]
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_handles_empty_response():
+    """extract_facts should return empty list when provider returns no content."""
+    try:
+        from nanobot.agent.memory_mem0 import Mem0MemoryStore
+    except ImportError:
+        pytest.skip("mem0 not installed")
+
+    provider = AsyncMock()
+    provider.chat = AsyncMock(return_value=LLMResponse(
+        content="",
+        finish_reason="end_turn",
+    ))
+
+    store = object.__new__(Mem0MemoryStore)
+    store.custom_prompt = "Extract facts as JSON: "
+
+    messages = [{"role": "user", "content": "Hello there"}]
+
+    facts = await store.extract_facts(messages, provider, "claude-sonnet-4-6")
+
+    assert facts == []
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_skips_empty_messages():
+    """extract_facts should return empty list when all messages have empty content."""
+    try:
+        from nanobot.agent.memory_mem0 import Mem0MemoryStore
+    except ImportError:
+        pytest.skip("mem0 not installed")
+
+    provider = AsyncMock()
+
+    store = object.__new__(Mem0MemoryStore)
+    store.custom_prompt = "Extract facts as JSON: "
+
+    messages = [
+        {"role": "user", "content": ""},
+        {"role": "assistant", "content": ""},
+    ]
+
+    facts = await store.extract_facts(messages, provider, "claude-sonnet-4-6")
+
+    assert facts == []
+    # Provider should not be called when there's no content
+    provider.chat.assert_not_called()

tests/test_message_visibility.py

@@ -0,0 +1,153 @@
+"""Tests for message visibility signing (hidden intermediate messages)."""
+
+import json
+from pathlib import Path
+
+from nanobot.agent.context import ContextBuilder
+from nanobot.agent.visibility import compute_signature, sign_content
+from nanobot.session.manager import Session
+
+
+class TestComputeSignature:
+    """Tests for compute_signature()."""
+
+    def test_returns_8_char_hex(self):
+        sig = compute_signature("hello")
+        assert len(sig) == 8
+        assert all(c in "0123456789abcdef" for c in sig)
+
+    def test_deterministic(self):
+        assert compute_signature("hello") == compute_signature("hello")
+
+    def test_different_content_different_sig(self):
+        assert compute_signature("hello") != compute_signature("world")
+
+    def test_sign_content_uses_compute_signature(self):
+        """sign_content should produce [HIDDEN:{compute_signature(content)}] prefix."""
+        content = "test message"
+        sig = compute_signature(content)
+        assert sign_content(content) == f"[HIDDEN:{sig}] {content}"
+
+
+class TestAddAssistantMessage:
+    """Tests for _hidden_sig in add_assistant_message()."""
+
+    def setup_method(self):
+        self.ctx = ContextBuilder(Path("/tmp"))
+
+    def test_intermediate_message_gets_hidden_sig(self):
+        msgs: list = []
+        tool_calls = [{"id": "tc1", "type": "function", "function": {"name": "test", "arguments": "{}"}}]
+        self.ctx.add_assistant_message(msgs, "thinking...", tool_calls)
+
+        assert msgs[0].get("_hidden_sig") is not None
+        assert msgs[0]["_hidden_sig"] == compute_signature("thinking...")
+
+    def test_final_message_no_hidden_sig(self):
+        msgs: list = []
+        self.ctx.add_assistant_message(msgs, "Here is the answer", None)
+
+        assert "_hidden_sig" not in msgs[0]
+
+    def test_empty_content_signed(self):
+        msgs: list = []
+        tool_calls = [{"id": "tc1", "type": "function", "function": {"name": "test", "arguments": "{}"}}]
+        self.ctx.add_assistant_message(msgs, None, tool_calls)
+
+        assert msgs[0]["_hidden_sig"] == compute_signature("")
+
+
+class TestAddToolResult:
+    """Tests for _hidden_sig in add_tool_result()."""
+
+    def setup_method(self):
+        self.ctx = ContextBuilder(Path("/tmp"))
+
+    def test_tool_result_gets_hidden_sig(self):
+        msgs: list = []
+        self.ctx.add_tool_result(msgs, "tc1", "read_file", "file contents here")
+
+        assert msgs[0]["_hidden_sig"] == compute_signature("file contents here")
+
+    def test_tool_result_non_string_content(self):
+        msgs: list = []
+        # Multipart content (e.g. image) is a list, not a string
+        self.ctx.add_tool_result(msgs, "tc1", "screenshot", [{"type": "text", "text": "ok"}])
+
+        assert msgs[0]["_hidden_sig"] == compute_signature("")
+
+
+class TestGetHistoryPrefix:
+    """Tests for get_history() applying [HIDDEN:sig] prefix."""
+
+    def test_hidden_sig_applied_at_read_time(self):
+        session = Session(key="test")
+        sig = compute_signature("thinking...")
+        session.messages = [
+            {"role": "assistant", "content": "thinking...", "tool_calls": [{}], "_hidden_sig": sig},
+        ]
+
+        history = session.get_history()
+        assert history[0]["content"] == f"[HIDDEN:{sig}] thinking..."
+        assert "_hidden_sig" not in history[0]
+
+    def test_no_prefix_without_hidden_sig(self):
+        session = Session(key="test")
+        session.messages = [
+            {"role": "assistant", "content": "Here is the answer"},
+        ]
+
+        history = session.get_history()
+        assert history[0]["content"] == "Here is the answer"
+
+    def test_tool_result_gets_prefix(self):
+        session = Session(key="test")
+        sig = compute_signature("file contents")
+        session.messages = [
+            {"role": "tool", "tool_call_id": "tc1", "name": "read", "content": "file contents", "_hidden_sig": sig},
+        ]
+
+        history = session.get_history()
+        assert history[0]["content"] == f"[HIDDEN:{sig}] file contents"
+
+    def test_roundtrip_jsonl(self, tmp_path):
+        """Write to session JSONL, reload, verify get_history() produces correct prefix."""
+        from nanobot.session.manager import SessionManager
+
+        workspace = tmp_path / "workspace"
+        workspace.mkdir()
+        mgr = SessionManager(workspace)
+
+        session = mgr.get_or_create("test:roundtrip")
+        sig = compute_signature("intermediate")
+        session.add_raw_message({
+            "role": "assistant",
+            "content": "intermediate",
+            "tool_calls": [{"id": "tc1", "type": "function", "function": {"name": "x", "arguments": "{}"}}],
+            "_hidden_sig": sig,
+        })
+        session.add_raw_message({
+            "role": "assistant",
+            "content": "final answer",
+        })
+        mgr.save(session)
+
+        # Reload from disk
+        mgr.invalidate("test:roundtrip")
+        reloaded = mgr.get_or_create("test:roundtrip")
+        history = reloaded.get_history()
+
+        assert history[0]["content"] == f"[HIDDEN:{sig}] intermediate"
+        assert history[1]["content"] == "final answer"
+
+    def test_idempotent_across_calls(self):
+        """Same prefix produced every call (cache stability)."""
+        session = Session(key="test")
+        sig = compute_signature("msg")
+        session.messages = [
+            {"role": "assistant", "content": "msg", "_hidden_sig": sig},
+        ]
+
+        h1 = session.get_history()
+        h2 = session.get_history()
+        assert h1[0]["content"] == h2[0]["content"]

tests/test_oauth_identity_block.py

@@ -0,0 +1,102 @@
+"""Test that the Anthropic OAuth identity block is always included in API requests."""
+
+import pytest
+from unittest.mock import AsyncMock, patch, MagicMock
+
+import httpx
+
+from nanobot.providers.anthropic_oauth import AnthropicOAuthProvider
+from nanobot.providers.oauth_utils import get_claude_code_system_prefix
+
+
+IDENTITY_TEXT = get_claude_code_system_prefix()
+
+
+@pytest.fixture
+def provider():
+    return AnthropicOAuthProvider(
+        oauth_token="sk-ant-oat01-test-token",
+        default_model="claude-opus-4-7",
+    )
+
+
+def _mock_response(status_code=200, json_data=None):
+    """Create a mock httpx.Response."""
+    resp = MagicMock(spec=httpx.Response)
+    resp.status_code = status_code
+    resp.headers = {}
+    resp.text = ""
+    resp.json.return_value = json_data or {
+        "content": [{"type": "text", "text": "ok"}],
+        "stop_reason": "end_turn",
+        "usage": {"input_tokens": 1, "output_tokens": 1},
+    }
+    return resp
+
+
+@pytest.mark.asyncio
+async def test_identity_block_present_with_system_prompt(provider):
+    """When a system prompt is provided, identity block is the first system block."""
+    mock_client = AsyncMock()
+    mock_client.post.return_value = _mock_response()
+    provider._client = mock_client
+
+    await provider._make_request(
+        messages=[{"role": "user", "content": "hello"}],
+        system="You are a helpful assistant.",
+    )
+
+    call_kwargs = mock_client.post.call_args
+    payload = call_kwargs.kwargs["json"] if "json" in call_kwargs.kwargs else call_kwargs[1]["json"]
+    system_blocks = payload["system"]
+
+    assert len(system_blocks) == 2
+    assert system_blocks[0]["type"] == "text"
+    assert system_blocks[0]["text"] == IDENTITY_TEXT
+    assert system_blocks[1]["text"] == "You are a helpful assistant."
+
+
+@pytest.mark.asyncio
+async def test_identity_block_present_without_system_prompt(provider):
+    """When no system prompt is provided, identity block is still included.
+
+    This is the critical fix: extract_facts and similar calls pass system=None,
+    but Anthropic requires the identity block for OAuth tokens.
+    """
+    mock_client = AsyncMock()
+    mock_client.post.return_value = _mock_response()
+    provider._client = mock_client
+
+    await provider._make_request(
+        messages=[{"role": "user", "content": "extract facts"}],
+        system=None,
+    )
+
+    call_kwargs = mock_client.post.call_args
+    payload = call_kwargs.kwargs["json"] if "json" in call_kwargs.kwargs else call_kwargs[1]["json"]
+    system_blocks = payload["system"]
+
+    assert len(system_blocks) == 1
+    assert system_blocks[0]["type"] == "text"
+    assert system_blocks[0]["text"] == IDENTITY_TEXT
+
+
+@pytest.mark.asyncio
+async def test_identity_block_present_with_empty_string_system(provider):
+    """Empty string system prompt should still include the identity block."""
+    mock_client = AsyncMock()
+    mock_client.post.return_value = _mock_response()
+    provider._client = mock_client
+
+    await provider._make_request(
+        messages=[{"role": "user", "content": "hello"}],
+        system="",
+    )
+
+    call_kwargs = mock_client.post.call_args
+    payload = call_kwargs.kwargs["json"] if "json" in call_kwargs.kwargs else call_kwargs[1]["json"]
+    system_blocks = payload["system"]
+
+    # Empty string is falsy, so should go through the else branch
+    assert len(system_blocks) == 1
+    assert system_blocks[0]["text"] == IDENTITY_TEXT

tests/test_provider_factory.py

@@ -9,7 +9,7 @@ def test_create_provider_oauth_token():
     """OAuth tokens should create AnthropicOAuthProvider."""
     provider = create_provider(
         api_key="sk-ant-oat01-test-token",
-        model="anthropic/claude-opus-4-5"
+        model="anthropic/claude-opus-4-7"
     )
     assert isinstance(provider, AnthropicOAuthProvider)
 
@@ -18,7 +18,7 @@ def test_create_provider_regular_key():
     """Regular API keys should create LiteLLMProvider."""
     provider = create_provider(
         api_key="sk-ant-api03-regular-key",
-        model="anthropic/claude-opus-4-5"
+        model="anthropic/claude-opus-4-7"
     )
     assert isinstance(provider, LiteLLMProvider)
 
@@ -27,6 +27,6 @@ def test_create_provider_openrouter():
     """OpenRouter keys should create LiteLLMProvider."""
     provider = create_provider(
         api_key="sk-or-v1-xxx",
-        model="anthropic/claude-opus-4-5"
+        model="anthropic/claude-opus-4-7"
     )
     assert isinstance(provider, LiteLLMProvider)

tests/test_registry_oauth.py

@@ -5,14 +5,14 @@ from nanobot.providers.registry import should_use_oauth_provider
 
 def test_should_use_oauth_for_oat_token():
     """OAuth provider should be used for sk-ant-oat tokens."""
-    assert should_use_oauth_provider("sk-ant-oat01-xxx", "anthropic/claude-opus-4-5") is True
+    assert should_use_oauth_provider("sk-ant-oat01-xxx", "anthropic/claude-opus-4-7") is True
     assert should_use_oauth_provider("sk-ant-oat01-xxx", "claude-sonnet-4") is True
 
 
 def test_should_not_use_oauth_for_regular_key():
     """Regular API keys should not use OAuth provider."""
-    assert should_use_oauth_provider("sk-ant-api03-xxx", "claude-opus-4-5") is False
-    assert should_use_oauth_provider("sk-or-v1-xxx", "anthropic/claude-opus-4-5") is False
+    assert should_use_oauth_provider("sk-ant-api03-xxx", "claude-opus-4-7") is False
+    assert should_use_oauth_provider("sk-or-v1-xxx", "anthropic/claude-opus-4-7") is False
 
 
 def test_should_not_use_oauth_for_non_anthropic():

tests/test_session_audit_log.py

@@ -0,0 +1,137 @@
+"""Test SessionManager audit log functionality."""
+
+import json
+
+import pytest
+
+from nanobot.session.manager import Session, SessionManager
+
+
+@pytest.fixture
+def session_manager(tmp_path):
+    return SessionManager(workspace=tmp_path)
+
+
+@pytest.fixture
+def session():
+    s = Session(key="telegram:12345")
+    s.add_message("user", "Hello")
+    s.add_message("assistant", "Hi there!")
+    return s
+
+
+def test_save_creates_audit_file(session_manager, session):
+    """SessionManager.save() should create a monthly audit log file."""
+    session_manager.save(session)
+
+    audit_files = list(session_manager.sessions_dir.glob("*.audit.*.jsonl"))
+    assert len(audit_files) == 1
+    assert "telegram_12345.audit." in audit_files[0].name
+
+
+def test_audit_file_contains_save_marker(session_manager, session):
+    """Audit log should start with a save_marker line containing metadata."""
+    session_manager.save(session)
+
+    audit_files = list(session_manager.sessions_dir.glob("*.audit.*.jsonl"))
+    lines = audit_files[0].read_text().strip().split("\n")
+
+    marker = json.loads(lines[0])
+    assert marker["_type"] == "save_marker"
+    assert marker["message_count"] == 2
+    assert "timestamp" in marker
+
+
+def test_audit_file_contains_all_messages(session_manager, session):
+    """Audit log should contain all session messages after the save marker."""
+    session_manager.save(session)
+
+    audit_files = list(session_manager.sessions_dir.glob("*.audit.*.jsonl"))
+    lines = audit_files[0].read_text().strip().split("\n")
+
+    # Line 0 = save_marker, lines 1-2 = messages
+    assert len(lines) == 3
+    msg1 = json.loads(lines[1])
+    msg2 = json.loads(lines[2])
+    assert msg1["role"] == "user"
+    assert msg1["content"] == "Hello"
+    assert msg2["role"] == "assistant"
+    assert msg2["content"] == "Hi there!"
+
+
+def test_audit_file_is_append_only(session_manager, session):
+    """Multiple saves should append to the same audit file, not overwrite."""
+    session_manager.save(session)
+
+    # Add another message and save again
+    session.add_message("user", "How are you?")
+    session_manager.save(session)
+
+    audit_files = list(session_manager.sessions_dir.glob("*.audit.*.jsonl"))
+    assert len(audit_files) == 1  # Same file
+
+    lines = audit_files[0].read_text().strip().split("\n")
+
+    # First save: 1 marker + 2 messages = 3 lines
+    # Second save: 1 marker + 3 messages = 4 lines
+    # Total: 7 lines
+    assert len(lines) == 7
+
+    # Both save markers present
+    markers = [json.loads(l) for l in lines if json.loads(l).get("_type") == "save_marker"]
+    assert len(markers) == 2
+    assert markers[0]["message_count"] == 2
+    assert markers[1]["message_count"] == 3
+
+
+def test_audit_preserves_message_fields(session_manager):
+    """Audit log should preserve all message fields including reasoning_content."""
+    session = Session(key="test:preserve")
+    session.add_raw_message({
+        "role": "assistant",
+        "content": "thinking response",
+        "reasoning_content": [{"type": "thinking", "thinking": "deep thoughts"}],
+        "timestamp": "2026-03-22T12:00:00",
+    })
+
+    session_manager.save(session)
+
+    audit_files = list(session_manager.sessions_dir.glob("*.audit.*.jsonl"))
+    lines = audit_files[0].read_text().strip().split("\n")
+
+    msg = json.loads(lines[1])
+    assert msg["reasoning_content"] == [{"type": "thinking", "thinking": "deep thoughts"}]
+
+
+def test_audit_failure_does_not_break_save(session_manager, session, tmp_path):
+    """If audit logging fails, the main session save should still succeed.
+
+    _append_audit has its own try/except, so internal failures are caught.
+    We simulate a realistic failure by making the sessions dir read-only
+    for audit file creation.
+    """
+    # First save works (creates both session file and audit file)
+    session_manager.save(session)
+
+    path = session_manager._get_session_path(session.key)
+    assert path.exists()
+
+    # Remove audit files and make a blocking file at the audit path
+    # so the next audit open("a") fails
+    for af in session_manager.sessions_dir.glob("*.audit.*.jsonl"):
+        af.unlink()
+
+    # Create a directory where the audit file should be — open() will fail
+    from datetime import datetime
+    now = datetime.now()
+    bad_path = session_manager.sessions_dir / f"telegram_12345.audit.{now:%Y-%m}.jsonl"
+    bad_path.mkdir()
+
+    # Second save should succeed despite audit failure
+    session.add_message("user", "another message")
+    session_manager.save(session)
+
+    # Session file should still be written correctly
+    with open(path) as f:
+        first_line = json.loads(f.readline())
+        assert first_line["_type"] == "metadata"