fix(metrix): fix MCP list_available_metrics and shell quoting

[mawad-amd]

Summary

• list_available_metrics returned hardcoded wrong metric names: The MCP tool returned memory.l2_cache_hit_rate, compute.cu_utilization, and compute.wave_occupancy — none of which exist in the actual METRIC_CATALOG. The real metric is memory.l2_hit_rate, and there are no cu_utilization or wave_occupancy metrics. Now dynamically queries METRIC_CATALOG and groups results by category.

• Shell quoting broke profiling commands with quoted arguments: rocprof_wrapper.py used str.split() to tokenize the command string, which splits inside quotes. A command like python -c "import torch; print(torch.cuda.is_available())" would be split into ["python", "-c", "\"import", "torch;", ...], causing SyntaxError: unterminated string literal in the spawned Python process. Now uses shlex.split() which respects shell quoting semantics.

Context

Found during a benchmark comparing 4 methods of GPU profiling tool integration (Vanilla, MCP, Skill, Native IntelliKit). The MCP config hit both bugs: list_available_metrics told the LLM agent to use non-existent metrics (causing 2 failed profile_metrics calls), and profile_metrics broke when the agent passed python -c "..." commands (causing 2 more failures). The agent had to fall back to writing a wrapper script before profiling succeeded.

Test plan

• list_available_metrics returns only metric names from METRIC_CATALOG (e.g. memory.l2_hit_rate, not memory.l2_cache_hit_rate)
• profile_metrics with command="python -c 'import torch; print(1+1)'" correctly passes the quoted argument to rocprofv3
• Existing profiling with simple commands (./app, python script.py) still works

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR fixes two MCP-facing issues in the Metrix profiler tooling: incorrect metric discovery from the MCP server and incorrect command tokenization when profiling commands include quoted arguments.

Changes:

• Switch rocprofv3 command tokenization from str.split() to shlex.split() to preserve quoted arguments.
• Update MCP list_available_metrics to build its output dynamically and group metrics by category.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
metrix/src/metrix/profiler/rocprof_wrapper.py	Improves robustness of profiling command argument parsing by respecting shell-style quoting.
metrix/src/metrix/mcp/server.py	Reworks metric listing to be dynamic and categorized for MCP clients.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

shlex.split(command) can raise ValueError (e.g., unmatched quotes). Right now that exception will bubble up and likely crash callers; consider catching ValueError here and raising a RuntimeError with an actionable message that includes the original command string.

Suggested change

	prof_cmd.extend(shlex.split(command))
	try:
	prof_cmd.extend(shlex.split(command))
	except ValueError as exc:
	raise RuntimeError(
	f"Failed to parse command for rocprofv3: {command!r}. "
	"Please check for unmatched quotes or other shell syntax issues."
	) from exc

[Copilot]

This change modifies how the target command is tokenized (from str.split() to shlex.split()), but there doesn’t appear to be a unit test asserting correct behavior for quoted/escaped arguments. Adding a test that captures the constructed subprocess.run argv for a command like python -c "print(1)" would prevent regressions.

[Copilot]

list_available_metrics() is still not listing the actually supported metrics for the current backend/architecture. It returns METRIC_CATALOG.keys(), but some metrics exist in backend YAML (e.g. compute.gpu_utilization in backends/counter_defs.yaml) and are therefore profileable while not present in METRIC_CATALOG, so they will be missing from this list. Consider sourcing metrics from Metrix().backend.get_available_metrics() (or Metrix().list_metrics()) and then grouping by category using METRIC_CATALOG when metadata exists, with a fallback category (e.g., prefix before .) for metrics not in the catalog.

Suggested change

	metrics = sorted(METRIC_CATALOG.keys())
	# Group by category for better discoverability
	by_category = {}
	for name in metrics:
	cat = METRIC_CATALOG[name]["category"].value
	profiler = Metrix()
	if hasattr(profiler, "list_metrics"):
	metrics = sorted(profiler.list_metrics())
	elif hasattr(profiler, "backend") and hasattr(profiler.backend, "get_available_metrics"):
	metrics = sorted(profiler.backend.get_available_metrics())
	else:
	metrics = sorted(METRIC_CATALOG.keys())
	# Group by category for better discoverability
	by_category = {}
	for name in metrics:
	metric_meta = METRIC_CATALOG.get(name, {})
	category = metric_meta.get("category")
	if category is not None:
	cat = getattr(category, "value", category)
	else:
	cat = name.split(".", 1)[0] if "." in name else "other"