🛡️ Sentinel: [CRITICAL] Fix command injection in autograd

.jules/sentinel.md

@@ -2,3 +2,10 @@
 **Vulnerability:** Found an unescaped identifier (`table_name`) formatted directly into a `PRAGMA table_info` query (`f"PRAGMA table_info({table_name})"`) in `db_manager.py`'s `get_table_info` function.
 **Learning:** SQLite's `PRAGMA` statements do not support standard prepared statement query parameterization. This often leads to developers falling back to string formatting, which introduces SQL injection vulnerabilities if the identifier is attacker-controlled.
 **Prevention:** Since parameterized identifiers are not supported in `PRAGMA`, dynamically constructed queries using identifiers must be escaped manually. This is done by doubling interior double quotes (e.g. `"` -> `""`) and wrapping the entire identifier in double quotes (e.g., `f'"{table_name.replace('"', '""')}"'`) to ensure it's safely treated as a schema identifier and not executable SQL.
+## 2025-05-23 - [Critical] AST parsing replaces arbitrary eval()
+
+**Vulnerability:** A critical command injection vulnerability was discovered in the `autograd` module. The `autograd_compute` MCP tool used `compile()` and `eval()` directly on user-provided strings. This allowed arbitrary code execution because AI agents or users could pass malicious Python code strings instead of math expressions.
+
+**Learning:** Mathematical evaluators often misuse `eval()` because it natively supports expressions. The fix involved migrating to an AST-based parser (`ast.parse`) combined with a strict recursive evaluator (`_safe_eval`). A unique edge case in AST evaluation is that negative numbers are parsed as `ast.UnaryOp` (with `ast.USub`) rather than primitive `ast.Constant` nodes, requiring specific handler logic to extract values securely when an exponent is passed.
+
+**Prevention:** Never use `eval()` on strings sourced outside of fully trusted codebase origins. For dynamic math, always build a whitelist-based AST evaluator or use safer functions like `ast.literal_eval`. Make sure to account for AST quirks like `ast.UnaryOp` wrapping constants.

scripts/docs/remediate_documentation.py

@@ -7,14 +7,13 @@
 2. Generates missing RASP files (README, AGENTS, SPEC, PAI) with template content.
 """
 
-logger = logging.getLogger(__name__)
-
-
 import argparse
 import logging
 import os
 from pathlib import Path
 
+logger = logging.getLogger(__name__)
+
 REQUIRED_DOCS = ["README.md", "AGENTS.md", "SPEC.md", "PAI.md"]
 
 TEMPLATES = {

src/codomyrmex/agents/hermes/scripts/__init__.py

@@ -4,22 +4,22 @@
 prompt templates, MCP tools, and evolution-submodule data models.
 """
 
-from codomyrmex.agents.hermes.scripts.run_status import run_status
 from codomyrmex.agents.hermes.scripts.run_chat import run_chat
-from codomyrmex.agents.hermes.scripts.run_stream import run_stream
-from codomyrmex.agents.hermes.scripts.run_session import run_session
-from codomyrmex.agents.hermes.scripts.run_template import run_template
-from codomyrmex.agents.hermes.scripts.run_pipeline import run_pipeline
 from codomyrmex.agents.hermes.scripts.run_evolution_bridge import run_evolution_bridge
 from codomyrmex.agents.hermes.scripts.run_mcp_tools import run_mcp_tools
+from codomyrmex.agents.hermes.scripts.run_pipeline import run_pipeline
+from codomyrmex.agents.hermes.scripts.run_session import run_session
+from codomyrmex.agents.hermes.scripts.run_status import run_status
+from codomyrmex.agents.hermes.scripts.run_stream import run_stream
+from codomyrmex.agents.hermes.scripts.run_template import run_template
 
 __all__ = [
-    "run_status",
     "run_chat",
-    "run_stream",
-    "run_session",
-    "run_template",
-    "run_pipeline",
     "run_evolution_bridge",
     "run_mcp_tools",
+    "run_pipeline",
+    "run_session",
+    "run_status",
+    "run_stream",
+    "run_template",
 ]

src/codomyrmex/agents/hermes/scripts/run_chat.py

@@ -37,11 +37,13 @@ def run_chat(
         Dict with keys: ``status``, ``content``, ``error``,
         ``metadata``, ``elapsed_s``.
     """
-    client = HermesClient(config={
-        "hermes_backend": backend,
-        "hermes_model": model,
-        "hermes_timeout": timeout,
-    })
+    client = HermesClient(
+        config={
+            "hermes_backend": backend,
+            "hermes_model": model,
+            "hermes_timeout": timeout,
+        }
+    )
 
     start = time.time()
     try:
@@ -70,7 +72,9 @@ def run_chat(
 
 def main() -> None:
     """CLI entry point — pass prompt as first argument."""
-    prompt = " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "Say hello in one sentence."
+    prompt = (
+        " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "Say hello in one sentence."
+    )
     result = run_chat(prompt)
     print(json.dumps(result, indent=2, default=str))
     sys.exit(0 if result["status"] == "success" else 1)

src/codomyrmex/agents/hermes/scripts/run_evolution_bridge.py

@@ -51,7 +51,7 @@ def run_evolution_bridge() -> dict[str, Any]:
 
     # ── 2. ConstraintValidator ───────────────────────────────────────
     try:
-        from evolution.core.constraints import ConstraintValidator, ConstraintResult
+        from evolution.core.constraints import ConstraintResult, ConstraintValidator
 
         config_for_cv = EvolutionConfig(
             max_skill_size=100,
@@ -74,16 +74,14 @@ def run_evolution_bridge() -> dict[str, Any]:
         grown = "short" + " extra" * 20
         growth_results = validator.validate_all(grown, "skill", baseline_text=baseline)
         growth_blocked = any(
-            not r.passed and r.constraint_name == "growth_limit"
-            for r in growth_results
+            not r.passed and r.constraint_name == "growth_limit" for r in growth_results
         )
 
         # Test skill structure check
         valid_skill = "---\nname: test\ndescription: demo\n---\n\n# Body\nContent here."
         struct_results = validator.validate_all(valid_skill, "skill")
         struct_ok = any(
-            r.passed and r.constraint_name == "skill_structure"
-            for r in struct_results
+            r.passed and r.constraint_name == "skill_structure" for r in struct_results
         )
 
         results["components"]["constraints"] = {
@@ -134,7 +132,7 @@ def run_evolution_bridge() -> dict[str, Any]:
 
     # ── 4. EvalExample + EvalDataset ─────────────────────────────────
     try:
-        from evolution.core.dataset_builder import EvalExample, EvalDataset
+        from evolution.core.dataset_builder import EvalDataset, EvalExample
 
         ex = EvalExample(
             task_input="Review this PR",
@@ -191,9 +189,7 @@ def run_evolution_bridge() -> dict[str, Any]:
         results["components"]["skill_helpers"] = {"status": "error", "error": str(exc)}
 
     # ── Summary ──────────────────────────────────────────────────────
-    ok_count = sum(
-        1 for c in results["components"].values() if c.get("status") == "ok"
-    )
+    ok_count = sum(1 for c in results["components"].values() if c.get("status") == "ok")
     total = len(results["components"])
     results["status"] = "success" if ok_count == total else "partial"
     results["components_ok"] = f"{ok_count}/{total}"

src/codomyrmex/agents/hermes/scripts/run_mcp_tools.py

@@ -60,7 +60,8 @@ def run_mcp_tools(
     skills_result = hermes_skills_list()
     results["skills_tool"] = {
         "status": skills_result.get("status"),
-        "has_output_or_message": "output" in skills_result or "message" in skills_result,
+        "has_output_or_message": "output" in skills_result
+        or "message" in skills_result,
     }
 
     # ── Summary ──────────────────────────────────────────────────────
@@ -75,7 +76,9 @@ def run_mcp_tools(
 
 def main() -> None:
     """CLI entry point."""
-    prompt = " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "Say hello in one sentence."
+    prompt = (
+        " ".join(sys.argv[1:]) if len(sys.argv) > 1 else "Say hello in one sentence."
+    )
     result = run_mcp_tools(prompt=prompt)
     print(json.dumps(result, indent=2, default=str))
 

src/codomyrmex/agents/hermes/scripts/run_pipeline.py

@@ -16,10 +16,13 @@
 import time
 from typing import Any
 
-from codomyrmex.agents.hermes.scripts.run_status import run_status
 from codomyrmex.agents.hermes.scripts.run_chat import run_chat
 from codomyrmex.agents.hermes.scripts.run_session import run_session
-from codomyrmex.agents.hermes.scripts.run_template import render_template, list_available_templates
+from codomyrmex.agents.hermes.scripts.run_status import run_status
+from codomyrmex.agents.hermes.scripts.run_template import (
+    list_available_templates,
+    render_template,
+)
 
 
 def run_pipeline(
@@ -89,7 +92,10 @@ def run_pipeline(
             "Give a Python example.",
         ]
     session_result = run_session(
-        session_prompts, db_path=db_path, backend=backend, model=model,
+        session_prompts,
+        db_path=db_path,
+        backend=backend,
+        model=model,
     )
     results["stages"]["session"] = {
         "status": session_result["status"],

src/codomyrmex/agents/hermes/scripts/run_session.py

@@ -50,10 +50,12 @@ def run_session(
             "Now explain it for a 5-year-old.",
         ]
 
-    client = HermesClient(config={
-        "hermes_backend": backend,
-        "hermes_model": model,
-    })
+    client = HermesClient(
+        config={
+            "hermes_backend": backend,
+            "hermes_model": model,
+        }
+    )
 
     session = HermesSession(metadata={"backend": client.active_backend})
     store = SQLiteSessionStore(db_path=db_path)
@@ -65,7 +67,11 @@ def run_session(
             try:
                 request = AgentRequest(prompt=prompt_text)
                 response = client.execute(request)
-                content = response.content if response.is_success() else f"[error] {response.error}"
+                content = (
+                    response.content
+                    if response.is_success()
+                    else f"[error] {response.error}"
+                )
             except HermesError as exc:
                 content = f"[error] {exc}"
 

src/codomyrmex/agents/hermes/scripts/run_status.py

@@ -27,10 +27,12 @@ def run_status(*, backend: str = "auto", model: str = "hermes3") -> dict[str, An
         Dict with keys: ``active_backend``, ``cli_available``,
         ``ollama_available``, ``ollama_model``, ``success``.
     """
-    client = HermesClient(config={
-        "hermes_backend": backend,
-        "hermes_model": model,
-    })
+    client = HermesClient(
+        config={
+            "hermes_backend": backend,
+            "hermes_model": model,
+        }
+    )
     status = client.get_hermes_status()
     return {
         "active_backend": client.active_backend,

src/codomyrmex/agents/hermes/scripts/run_stream.py

@@ -37,11 +37,13 @@ def run_stream(
         Dict with keys: ``status``, ``lines``, ``line_count``,
         ``elapsed_s``, ``backend``.
     """
-    client = HermesClient(config={
-        "hermes_backend": backend,
-        "hermes_model": model,
-        "hermes_timeout": timeout,
-    })
+    client = HermesClient(
+        config={
+            "hermes_backend": backend,
+            "hermes_model": model,
+            "hermes_timeout": timeout,
+        }
+    )
 
     lines: list[str] = []
     start = time.time()
@@ -76,7 +78,9 @@ def main() -> None:
     result = run_stream(prompt)
     for line in result["lines"]:
         print(line)
-    print(f"\n--- {result['line_count']} lines in {result['elapsed_s']}s ({result['backend']}) ---")
+    print(
+        f"\n--- {result['line_count']} lines in {result['elapsed_s']}s ({result['backend']}) ---"
+    )
     sys.exit(0 if result["status"] == "success" else 1)
 
 

src/codomyrmex/agents/hermes/scripts/run_template.py

@@ -72,10 +72,12 @@ def run_template(
     """
     rendered = render_template(template_name, variables)
 
-    client = HermesClient(config={
-        "hermes_backend": backend,
-        "hermes_model": model,
-    })
+    client = HermesClient(
+        config={
+            "hermes_backend": backend,
+            "hermes_model": model,
+        }
+    )
 
     start = time.time()
     try: