harness: a sandboxed child that dies without sending escapes as a bare EOFError, so the agent is told 'TOOL_ERROR:' and nothing else

[Shashankss1205]

What happens

SandboxedExecutor.run guards the wait with poll() and then reads:

if not parent_conn.poll(spec.timeout_seconds):
 self._kill(proc)
 raise SandboxViolation(f"tool {spec.name!r} exceeded its {spec.timeout_seconds}s timeout")
kind, payload = parent_conn.recv()

poll() returns true when the pipe is readable, and a closed pipe is readable. So a child that died without sending — os._exit, a segfault, an OOM-kill, anything that kills the process rather than raising inside spec.fn — falls straight through into recv(), which raises a bare EOFError:

import os, tempfile
from grapharc.harness import ToolSpec
from grapharc.harness.executor import SandboxedExecutor

def suicidal(**kwargs):
 os._exit(1)

ex = SandboxedExecutor(workspace=tempfile.mkdtemp())
ex.run(ToolSpec(name="suicidal", description="", fn=suicidal), {})

RAISED builtins.EOFError: ''
does the message name the tool? False

Why it matters

run has two documented failure shapes and this is neither: SandboxViolation for a confinement breach, and RuntimeError(f"tool {name!r} failed: ...") for the tool's own exception. An EOFError with an empty string for a message is not attributable to anything.

Downstream is where it bites. AgentNode's tool-call loop catches it under the blanket clause:

except Exception as exc: # noqa: BLE001 — a broken tool must not end the run
 status, detail = ToolCallStatus.ERROR, f"TOOL_ERROR: {exc}"

str(EOFError('')) is '', so the model is handed literally:

TOOL_ERROR:

Nothing after the colon. The model is told its call failed and given no way to tell why, which tool, or whether retrying could help — and the same text goes into the trace, so the audit trail records a failure it cannot explain either.

This is the same shape as a defect this repo already closed: "a phase the budget curtailed reported nothing at all … so a curtailed phase wrote [budget_exhausted] with nothing after it." Same empty-message failure, one layer down.

The _CALL_SHAPE_ERROR hint machinery right below that except also can't fire, since it pattern-matches on message text and there is no message.

Where in the code

• grapharc/harness/executor.py — SandboxedExecutor.run, the poll() / recv() pair
• grapharc/harness/executor.py — _child_main's finally: conn.close(), and the ("ok"|"violation"|"error") protocol the parent expects
• grapharc/harness/agent.py — the blanket except Exception that renders TOOL_ERROR: {exc}
• grapharc/harness/core.py — Harness.call, the single caller of executor.run

What to change

Catch EOFError around recv() and raise the failure shape callers already handle: a RuntimeError naming the tool and saying the child exited without sending, with the exit code — a negative one being the signal that killed it, which is what distinguishes an OOM-kill from a deliberate _exit.

It should stay a RuntimeError rather than becoming a SandboxViolation: a child dying is not evidence it tried to escape confinement, and a violation is a specific accusation that ends up in the trace as one.

Acceptance criteria

• A sandboxed tool whose child dies without sending raises an error naming the tool, not a bare EOFError
• A signal death is distinguishable from a clean non-zero exit in the message
• It is not reported as a SandboxViolation
• The normal return path, the tool-raises path and the timeout path are unchanged
• uv run pytest green, uv run ruff check . clean