fix: extend JSON repair to browser tool and CLI (#862)

- repair malformed browser command envelopes before treating stdout as non-JSON
- repair eval/image payloads in browser_tool.py
- repair CLI parsing for vision and /cron tool responses
- add focused regression tests for browser and CLI repair paths

agent/context_snapshots.py

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

cli.py

@@ -589,6 +589,7 @@ from tools.terminal_tool import set_sudo_password_callback, set_approval_callbac
 from tools.skills_tool import set_secret_capture_callback
 from hermes_cli.callbacks import prompt_for_secret
 from tools.browser_tool import _emergency_cleanup_all_sessions as _cleanup_all_browsers
+from utils import repair_and_load_json
 
 # Guard to prevent cleanup from running multiple times on exit
 _cleanup_done = False
@@ -3569,7 +3570,11 @@ class HermesCLI:
                 result_json = _asyncio.run(
                     vision_analyze_tool(image_url=str(img_path), user_prompt=analysis_prompt)
                 )
-                result = _json.loads(result_json)
+                result = repair_and_load_json(
+                    result_json,
+                    default={},
+                    context="cli_image_analysis",
+                ) if isinstance(result_json, str) else {}
                 if result.get("success"):
                     description = result.get("analysis", "")
                     enriched_parts.append(
@@ -4960,7 +4965,14 @@ class HermesCLI:
         from tools.cronjob_tools import cronjob as cronjob_tool
 
         def _cron_api(**kwargs):
-            return json.loads(cronjob_tool(**kwargs))
+            result = repair_and_load_json(
+                cronjob_tool(**kwargs),
+                default=None,
+                context="cli_cron_command",
+            )
+            if isinstance(result, dict):
+                return result
+            return {"success": False, "error": "Invalid JSON from cronjob tool"}
 
         def _normalize_skills(values):
             normalized = []

docs/hindsight-local-eval.md

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

plugins/memory/hindsight/__init__.py

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

run_agent.py

@@ -604,8 +604,6 @@ class AIAgent:
         checkpoint_max_snapshots: int = 50,
         pass_session_id: bool = False,
         persist_session: bool = True,
-        context_snapshots_enabled: bool | None = None,
-        context_snapshots_dir: str | None = None,
     ):
         """
         Initialize the AI Agent.
@@ -1131,43 +1129,6 @@ class AIAgent:
         except Exception:
             _agent_cfg = {}
 
-        def _is_enabled(value):
-            if isinstance(value, bool):
-                return value
-            return str(value).strip().lower() in {"1", "true", "yes", "on"}
-
-        _debug_cfg = _agent_cfg.get("debug", {}) if isinstance(_agent_cfg, dict) else {}
-        if not isinstance(_debug_cfg, dict):
-            _debug_cfg = {}
-        _snapshot_cfg = _debug_cfg.get("context_snapshots", {})
-        if not isinstance(_snapshot_cfg, dict):
-            _snapshot_cfg = {}
-        _snapshots_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS")
-        _snapshots_dir_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS_DIR")
-        if context_snapshots_enabled is None:
-            if _snapshots_env is not None:
-                self._context_snapshots_enabled = _is_enabled(_snapshots_env)
-            else:
-                self._context_snapshots_enabled = _is_enabled(_snapshot_cfg.get("enabled", False))
-        else:
-            self._context_snapshots_enabled = bool(context_snapshots_enabled)
-        self._context_snapshots_dir = (
-            context_snapshots_dir
-            or _snapshots_dir_env
-            or _snapshot_cfg.get("dir")
-            or None
-        )
-        try:
-            from agent.context_snapshots import ContextSnapshotRecorder
-            self._context_snapshot_recorder = ContextSnapshotRecorder(
-                session_id=self.session_id,
-                enabled=self._context_snapshots_enabled,
-                base_dir=self._context_snapshots_dir,
-            )
-        except Exception as _snapshot_err:
-            logger.debug("Context snapshot recorder init failed: %s", _snapshot_err)
-            self._context_snapshot_recorder = None
-
         # Persistent memory (MEMORY.md + USER.md) -- loaded from disk
         self._memory_store = None
         self._memory_enabled = False
@@ -8183,17 +8144,12 @@ class AIAgent:
         # Use original_user_message (clean input) — user_message may contain
         # injected skill content that bloats / breaks provider queries.
         _ext_prefetch_cache = ""
-        _memory_provider_prompt_cache = ""
         if self._memory_manager:
             try:
                 _query = original_user_message if isinstance(original_user_message, str) else ""
                 _ext_prefetch_cache = self._memory_manager.prefetch_all(_query) or ""
             except Exception:
                 pass
-            try:
-                _memory_provider_prompt_cache = self._memory_manager.build_system_prompt() or ""
-            except Exception:
-                pass
 
         while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call:
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
@@ -8261,8 +8217,6 @@ class AIAgent:
             # However, providers like Moonshot AI require a separate 'reasoning_content' field
             # on assistant messages with tool_calls. We handle both cases here.
             api_messages = []
-            _current_api_user_message = ""
-            _current_memory_context_block = ""
             for idx, msg in enumerate(messages):
                 api_msg = msg.copy()
 
@@ -8277,15 +8231,12 @@ class AIAgent:
                         _fenced = build_memory_context_block(_ext_prefetch_cache)
                         if _fenced:
                             _injections.append(_fenced)
-                            _current_memory_context_block = _fenced
                     if _plugin_user_context:
                         _injections.append(_plugin_user_context)
                     if _injections:
                         _base = api_msg.get("content", "")
                         if isinstance(_base, str):
                             api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections)
-                    if isinstance(api_msg.get("content"), str):
-                        _current_api_user_message = api_msg["content"]
 
                 # For ALL assistant messages, pass reasoning back to the API
                 # This ensures multi-turn reasoning context is preserved
@@ -8320,13 +8271,7 @@ class AIAgent:
                 from agent.privacy_filter import PrivacyFilter
                 pf = PrivacyFilter()
                 # Sanitize messages before they reach the provider
-                _pf_result = pf.sanitize_messages(api_messages)
-                if isinstance(_pf_result, tuple):
-                    api_messages, _pf_report = _pf_result
-                    if getattr(pf, "last_report", None) is None:
-                        pf.last_report = _pf_report
-                else:
-                    api_messages = _pf_result
+                api_messages = pf.sanitize_messages(api_messages)
                 if pf.last_report and pf.last_report.had_redactions:
                     logger.info(f"Privacy Filter: Redacted sensitive data from turn payload. Details: {pf.last_report.summary()}")
             except Exception as e:
@@ -8397,27 +8342,6 @@ class AIAgent:
                     new_tcs.append(tc)
                 am["tool_calls"] = new_tcs
 
-            if self._context_snapshot_recorder:
-                try:
-                    self._context_snapshot_recorder.record_call(
-                        api_call_count,
-                        system_prompt=effective_system,
-                        memory_provider_system_prompt=_memory_provider_prompt_cache,
-                        memory_prefetch_raw=_ext_prefetch_cache,
-                        memory_context_block=_current_memory_context_block,
-                        api_user_message=_current_api_user_message,
-                        api_messages=api_messages,
-                        metadata={
-                            "model": self.model,
-                            "provider": self.provider,
-                            "platform": self.platform or "",
-                            "api_mode": self.api_mode,
-                            "memory_providers": [p.name for p in getattr(self._memory_manager, "providers", [])],
-                        },
-                    )
-                except Exception as _snapshot_err:
-                    logger.debug("Context snapshot capture failed: %s", _snapshot_err)
-
             # Calculate approximate request size for logging
             total_chars = sum(len(str(msg)) for msg in api_messages)
             approx_tokens = estimate_messages_tokens_rough(api_messages)

tests/agent/test_context_snapshots.py

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

tests/cli/test_cli_json_repair.py

@@ -0,0 +1,62 @@
+import sys
+import types
+from unittest.mock import patch
+
+
+def _stub_auxiliary_client():
+    stub = types.ModuleType("agent.auxiliary_client")
+    stub.call_llm = lambda *args, **kwargs: None
+    stub.resolve_provider_client = lambda *args, **kwargs: (None, None)
+    stub.get_text_auxiliary_client = lambda *args, **kwargs: (None, None)
+    stub.async_call_llm = lambda *args, **kwargs: None
+    stub.extract_content_or_reasoning = lambda *args, **kwargs: ""
+    stub._OR_HEADERS = {}
+    stub._get_task_timeout = lambda *args, **kwargs: 30
+    sys.modules["agent.auxiliary_client"] = stub
+
+
+def _stub_vision_tools(vision_analyze_tool):
+    stub = types.ModuleType("tools.vision_tools")
+    stub.vision_analyze_tool = vision_analyze_tool
+    sys.modules["tools.vision_tools"] = stub
+
+
+def test_preprocess_images_with_vision_repairs_malformed_json(tmp_path):
+    _stub_auxiliary_client()
+    from cli import HermesCLI
+
+    cli_obj = HermesCLI.__new__(HermesCLI)
+    image_path = tmp_path / "test.png"
+    image_path.write_bytes(b"fake-image-bytes")
+
+    async def fake_vision(**kwargs):
+        return "{'success': true, 'analysis': 'Recovered image description',}"
+
+    _stub_vision_tools(fake_vision)
+    result = HermesCLI._preprocess_images_with_vision(
+        cli_obj,
+        "Describe this",
+        [image_path],
+        announce=False,
+    )
+
+    assert "Recovered image description" in result
+    assert "Describe this" in result
+    assert str(image_path) in result
+
+
+def test_handle_cron_command_repairs_malformed_json(capsys):
+    _stub_auxiliary_client()
+    from cli import HermesCLI
+
+    cli_obj = HermesCLI.__new__(HermesCLI)
+    malformed_result = """{'success': true, 'jobs': [{'job_id': 'job-1234567890ab', 'name': 'Nightly Check', 'state': 'scheduled', 'schedule': 'every 1h', 'repeat': 'forever', 'prompt_preview': 'Check server status', 'skills': ['blogwatcher',], 'next_run_at': '2026-04-22T01:00:00Z',},],}"""
+
+    with patch("tools.cronjob_tools.cronjob", return_value=malformed_result):
+        HermesCLI._handle_cron_command(cli_obj, "/cron list")
+
+    out = capsys.readouterr().out
+    assert "Scheduled Jobs:" in out
+    assert "job-1234567890ab" in out
+    assert "Nightly Check" in out
+    assert "blogwatcher" in out

tests/plugins/memory/test_hindsight_provider.py

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

tests/run_agent/test_run_agent_context_snapshots.py

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

tests/test_hindsight_local_eval_doc.py

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

tests/tools/test_browser_json_repair.py

@@ -0,0 +1,108 @@
+import io
+import json
+import sys
+import types
+from unittest.mock import MagicMock, patch
+
+
+def _stub_auxiliary_client():
+    stub = types.ModuleType("agent.auxiliary_client")
+    stub.call_llm = lambda *args, **kwargs: None
+    stub.resolve_provider_client = lambda *args, **kwargs: (None, None)
+    stub.get_text_auxiliary_client = lambda *args, **kwargs: (None, None)
+    stub.async_call_llm = lambda *args, **kwargs: None
+    stub.extract_content_or_reasoning = lambda *args, **kwargs: ""
+    stub._OR_HEADERS = {}
+    stub._get_task_timeout = lambda *args, **kwargs: 30
+    sys.modules["agent.auxiliary_client"] = stub
+
+
+def test_run_browser_command_repairs_malformed_stdout_envelope(tmp_path):
+    _stub_auxiliary_client()
+    from tools.browser_tool import _run_browser_command
+
+    mock_proc = MagicMock()
+    mock_proc.returncode = 0
+    mock_proc.wait.return_value = 0
+    fake_session = {
+        "session_name": "test-session",
+        "session_id": "test-id",
+        "cdp_url": None,
+    }
+    malformed_stdout = "{'success': true, 'data': {'url': 'https://example.com',},}"
+
+    def fake_open(path, mode="r", *args, **kwargs):
+        path = str(path)
+        if path.endswith("_stdout_navigate"):
+            return io.StringIO(malformed_stdout)
+        if path.endswith("_stderr_navigate"):
+            return io.StringIO("")
+        raise FileNotFoundError(path)
+
+    with (
+        patch("tools.browser_tool._find_agent_browser", return_value="/usr/bin/agent-browser"),
+        patch("tools.browser_tool._get_session_info", return_value=fake_session),
+        patch("tools.browser_tool._socket_safe_tmpdir", return_value=str(tmp_path)),
+        patch("tools.browser_tool._merge_browser_path", side_effect=lambda p: p),
+        patch("tools.interrupt.is_interrupted", return_value=False),
+        patch("subprocess.Popen", return_value=mock_proc),
+        patch("os.open", return_value=99),
+        patch("os.close"),
+        patch("os.unlink"),
+        patch("builtins.open", side_effect=fake_open),
+    ):
+        result = _run_browser_command("task-1", "navigate", ["https://example.com"])
+
+    assert result["success"] is True
+    assert result["data"]["url"] == "https://example.com"
+
+
+def test_agent_browser_eval_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import _browser_eval
+
+    with patch(
+        "tools.browser_tool._run_browser_command",
+        return_value={"success": True, "data": {"result": "{'items': ['a', 'b',],}"}},
+    ):
+        result = json.loads(_browser_eval("document.body.innerText", task_id="test"))
+
+    assert result["success"] is True
+    assert result["result"] == {"items": ["a", "b"]}
+    assert result["result_type"] == "dict"
+
+
+def test_camofox_eval_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import _camofox_eval
+
+    with (
+        patch("tools.browser_camofox._ensure_tab", return_value={"tab_id": "tab-1", "user_id": "user-1"}),
+        patch("tools.browser_camofox._post", return_value={"result": "{'count': 3,}"}),
+    ):
+        result = json.loads(_camofox_eval("2+1", task_id="test"))
+
+    assert result["success"] is True
+    assert result["result"] == {"count": 3}
+    assert result["result_type"] == "dict"
+
+
+def test_browser_get_images_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import browser_get_images
+
+    with patch(
+        "tools.browser_tool._run_browser_command",
+        return_value={
+            "success": True,
+            "data": {
+                "result": "[{\"src\": \"https://example.com/cat.png\", \"alt\": \"cat\",}]"
+            },
+        },
+    ):
+        result = json.loads(browser_get_images(task_id="test"))
+
+    assert result["success"] is True
+    assert result["count"] == 1
+    assert result["images"] == [{"src": "https://example.com/cat.png", "alt": "cat"}]
+    assert "warning" not in result

tools/browser_tool.py

@@ -67,6 +67,7 @@ from typing import Dict, Any, Optional, List
 from pathlib import Path
 from agent.auxiliary_client import call_llm
 from hermes_constants import get_hermes_home
+from utils import repair_and_load_json
 
 try:
     from tools.website_policy import check_website_access
@@ -1171,8 +1172,12 @@ def _run_browser_command(
             return {"success": False, "error": f"Browser command '{command}' returned no output"}
 
         if stdout_text:
-            try:
-                parsed = json.loads(stdout_text)
+            parsed = repair_and_load_json(
+                stdout_text,
+                default=None,
+                context=f"browser_{command}_stdout",
+            )
+            if isinstance(parsed, dict):
                 # Warn if snapshot came back empty (common sign of daemon/CDP issues)
                 if command == "snapshot" and parsed.get("success"):
                     snap_data = parsed.get("data", {})
@@ -1181,35 +1186,35 @@ def _run_browser_command(
                                        "Possible stale daemon or CDP connection issue. "
                                        "returncode=%s", returncode)
                 return parsed
-            except json.JSONDecodeError:
-                raw = stdout_text[:2000]
-                logger.warning("browser '%s' returned non-JSON output (rc=%s): %s",
-                               command, returncode, raw[:500])
 
-                if command == "screenshot":
-                    stderr_text = (stderr or "").strip()
-                    combined_text = "\n".join(
-                        part for part in [stdout_text, stderr_text] if part
+            raw = stdout_text[:2000]
+            logger.warning("browser '%s' returned non-JSON output (rc=%s): %s",
+                           command, returncode, raw[:500])
+
+            if command == "screenshot":
+                stderr_text = (stderr or "").strip()
+                combined_text = "\n".join(
+                    part for part in [stdout_text, stderr_text] if part
+                )
+                recovered_path = _extract_screenshot_path_from_text(combined_text)
+
+                if recovered_path and Path(recovered_path).exists():
+                    logger.info(
+                        "browser 'screenshot' recovered file from non-JSON output: %s",
+                        recovered_path,
                     )
-                    recovered_path = _extract_screenshot_path_from_text(combined_text)
+                    return {
+                        "success": True,
+                        "data": {
+                            "path": recovered_path,
+                            "raw": raw,
+                        },
+                    }
 
-                    if recovered_path and Path(recovered_path).exists():
-                        logger.info(
-                            "browser 'screenshot' recovered file from non-JSON output: %s",
-                            recovered_path,
-                        )
-                        return {
-                            "success": True,
-                            "data": {
-                                "path": recovered_path,
-                                "raw": raw,
-                            },
-                        }
-
-                return {
-                    "success": False,
-                    "error": f"Non-JSON output from agent-browser for '{command}': {raw}"
-                }
+            return {
+                "success": False,
+                "error": f"Non-JSON output from agent-browser for '{command}': {raw}"
+            }
         
         # Check for errors
         if returncode != 0:
@@ -1777,10 +1782,11 @@ def _browser_eval(expression: str, task_id: Optional[str] = None) -> str:
     # is valid JSON, parse it so the model gets structured data.
     parsed = raw_result
     if isinstance(raw_result, str):
-        try:
-            parsed = json.loads(raw_result)
-        except (json.JSONDecodeError, ValueError):
-            pass  # keep as string
+        parsed = repair_and_load_json(
+            raw_result,
+            default=raw_result,
+            context="browser_eval_result",
+        )
 
     return json.dumps({
         "success": True,
@@ -1801,10 +1807,11 @@ def _camofox_eval(expression: str, task_id: Optional[str] = None) -> str:
         raw_result = resp.get("result") if isinstance(resp, dict) else resp
         parsed = raw_result
         if isinstance(raw_result, str):
-            try:
-                parsed = json.loads(raw_result)
-            except (json.JSONDecodeError, ValueError):
-                pass
+            parsed = repair_and_load_json(
+                raw_result,
+                default=raw_result,
+                context="camofox_eval_result",
+            )
 
         return json.dumps({
             "success": True,
@@ -1904,26 +1911,29 @@ def browser_get_images(task_id: Optional[str] = None) -> str:
     if result.get("success"):
         data = result.get("data", {})
         raw_result = data.get("result", "[]")
-        
-        try:
-            # Parse the JSON string returned by JavaScript
-            if isinstance(raw_result, str):
-                images = json.loads(raw_result)
-            else:
-                images = raw_result
-            
-            return json.dumps({
-                "success": True,
-                "images": images,
-                "count": len(images)
-            }, ensure_ascii=False)
-        except json.JSONDecodeError:
-            return json.dumps({
-                "success": True,
-                "images": [],
-                "count": 0,
-                "warning": "Could not parse image data"
-            }, ensure_ascii=False)
+
+        warning = None
+        if isinstance(raw_result, str):
+            images = repair_and_load_json(
+                raw_result,
+                default=None,
+                context="browser_get_images_result",
+            )
+        else:
+            images = raw_result
+
+        if not isinstance(images, list):
+            images = []
+            warning = "Could not parse image data"
+
+        payload = {
+            "success": True,
+            "images": images,
+            "count": len(images),
+        }
+        if warning:
+            payload["warning"] = warning
+        return json.dumps(payload, ensure_ascii=False)
     else:
         return json.dumps({
             "success": False,