morluto/flameox

View on GitHub ↗

Runtime evidence that helps agents trace, profile, and burn down hotspots in application and native code, GPU kernels, and inference stacks.

34 ★1 forksPythonUpdated 18d ago

What you need to know

Local runtime evidence recorder for coding agents investigating performance, memory, execution, concurrency, and reliability. It connects profilers/benchmarks/trace processors (pyperf, py-spy, Memray, Perfetto, torch.profiler, Nsight, etc.) to a bounded, immutable evidence record and runs discriminating experiments that support or refute hypotheses.

Install

npx flameox@latest setup
uv sync --extra dev && uv run flameox init . && uv run flameox status

Usage

  • Ask your agent to 'Initialize Flameox in this project and list the available profiling capabilities', then restart the client and reopen the project.
  • Define named workloads and experiments in flameox.toml (declared scalars, no shell expansion); the MCP configure_workload tool validates and writes the canonical definition without executing it.
  • Plan a capture (resolves every executable once) then run it; execution revalidates the binding instead of searching PATH again.
  • Use read-only analysis (analyze hotspots/scaling/compare/memory) and only `analyze record` or `findings record` to make a durable claim.

Key features

  • Captures native artifacts with provenance from pyperf, py-spy, pytest-reportlog, coverage.py, Memray, Perfetto, torch.profiler, Nsight Systems/Compute, ROCprofiler, Compute Sanitizer, NVBench and more.
  • Randomized complete-blocks experiments with declared estimands, polarity, practical thresholds and confidence levels, retaining trial order, failures, cancellations and exclusions.
  • Content-addressed evidence in .diagnostics/ with a SQLite control plane, immutable Parquet generations and a rebuildable catalog.duckdb cache.
  • Bounded task-shaped operations over the CLI and MCP server (no raw shell strings, SQL or artifact bytes); reports missing evidence rather than substituting weaker sources.
  • Dry-run GC by default, reversible trash, and a validate/validate --full chain that never repairs evidence.

Best for

Teams that want a trustworthy, reproducible local evidence trail for coding-agent performance and reliability experiments without sending data off-machine.

Caveats

  • Requires Python 3.12+ and a committed uv.lock.
  • Evidence-source availability depends on host, permissions, installed extras and adapter; a profile does not by itself establish a performance/correctness conclusion.
  • Trusted-local capture does not enforce containment for child processes (it records the limitation); workloads may access the network unless active containment denies it.
  • MIT licensed.
Platforms: LocalClients: Any supported MCP client · CLI

Reviewed 2026-08-18

Topics

benchmarkingcoding-agentscordisdebuggingdeveloper-toolsdshdsh-plugingpu-profilinglocal-firstmcpmcp-servermodel-context-protocolperformanceperformance-analysisperformance-engineeringperformance-regressionprofilerprofilingpythonruntime-analysis
Stars
34★
Forks
1
Language
Python
License
MIT
Created
2026-07-25
Last push
2026-08-17