Skip to content

Free, open-source AI video generator for Mac. Run LTX-2.5 and MiniMax H3 (Hailuo) locally on Apple Silicon: text-to-video and image-to-video with sound, lip-sync, character LoRA training, songs from lyrics (YuE2) and image editing (Qwen-Image-Edit). One-click install via Pinokio. Offline: no cloud, no API key, no NVIDIA GPU.

fix(agent/runtime): cap submit_shot/submit_shots volume per turn (H4)

mrbizarro39c032b4190d
diff --git a/agent/runtime.py b/agent/runtime.py
index 1d3e1ca..4390233 100644
--- a/agent/runtime.py
+++ b/agent/runtime.py
@@ -37,6 +37,14 @@ from agent import engine, tools, prompts
MAX_STEPS_PER_TURN = 30 # safety cap on tool-loop iterations per user turn
+# Per-turn cap on total shots queued via submit_shot / submit_shots. Counts
+# every individual shot — one submit_shot call = +1, one submit_shots call =
+# +len(shots). Hitting the cap returns a structured error to the agent and
+# does NOT dispatch the tool, so a prompt-injected or runaway agent can't
+# burn hours of GPU time queuing 50 LTX renders in a single turn before the
+# user has a chance to intervene. Counter resets at the start of each
+# run_turn() call. (Security review H4, 2026-05-08.)
+MAX_SUBMITS_PER_TURN = 8
SESSION_VERSION = 1
@@ -349,6 +357,10 @@ def run_turn(session: Session, user_message: str | None,
# Clear finished flag — a new user message resets the loop.
session.finished = False
+ # Per-turn submit budget — see MAX_SUBMITS_PER_TURN. Local to this
+ # run_turn() call: a fresh user message always gets a fresh budget.
+ submits_used = 0
+
# 2. Drive the loop.
for step_i in range(max_steps):
# Call engine
@@ -391,7 +403,32 @@ def run_turn(session: Session, user_message: str | None,
"tool": tool_name, "args": tool_args, "step": step_i,
}))
- result_obj = tools.dispatch(tool_name, tool_args, panel_ops, session.tool_state)
+ # Enforce the per-turn submit budget BEFORE dispatch. We refuse the
+ # call (and synthesize an error result for the agent) the moment
+ # this call would push the running total past the cap. Partial
+ # batches are not allowed: a submit_shots(shots=[a,b,c,d]) when 6
+ # have already been queued is rejected wholesale rather than
+ # silently truncated, since truncation would surprise the agent.
+ submit_count = 0
+ if tool_name == "submit_shot":
+ submit_count = 1
+ elif tool_name == "submit_shots":
+ shots = tool_args.get("shots") if isinstance(tool_args, dict) else None
+ submit_count = len(shots) if isinstance(shots, list) else 1
+
+ if submit_count and submits_used + submit_count > MAX_SUBMITS_PER_TURN:
+ result_obj = {
+ "ok": False,
+ "error": (
+ f"submit cap reached: agent already queued "
+ f"{submits_used} shots this turn (limit "
+ f"{MAX_SUBMITS_PER_TURN}). Use multiple turns to "
+ f"queue more, or call finish."
+ ),
+ }
+ else:
+ submits_used += submit_count
+ result_obj = tools.dispatch(tool_name, tool_args, panel_ops, session.tool_state)
yield emit(TurnEvent("tool_result", {
"tool": tool_name, "result": result_obj, "step": step_i,
}))