Stallhunt

Stallhunt

Stallhunt is a Linux-first command-line performance triage tool.

Repository: https://github.com/guillem/stallhunt

Traditional tools such as top, htop, iotop, vmstat, and iostat expose measurements. They are excellent tools, but the human operator still has to answer the harder question:

What is actually constraining useful work right now, who is suffering, and who is probably responsible?

Stallhunt aims to automate that reasoning.

Install

Requirements:

• Linux 4.20 or newer with procfs mounted; readable PSI files under /proc/pressure are required for pressure verdicts,
• Rust 1.85 or newer for source builds.

See docs/install.md for cargo install, release tarballs, and the support matrix.

From a clone:

cargo install --path .stallhunt

Bare stallhunt runs a default 10-second hunt. On a terminal it renders a compact, color-coded report; piped or redirected output (stallhunt | cat, >file, CI) is unchanged plain text. Use --json for the full structured evidence, --verbose to expand the compact report's collapsed caveats back to full text, or --no-color (also NO_COLOR=1) to disable color without changing the layout:

stallhunt --jsonstallhunt hunt --duration 30sstallhunt hunt --verbose

Capture and replay a normalized observation:

stallhunt record --duration 10s --output incident.jsonstallhunt replay incident.jsonstallhunt redact incident.json --output incident.redacted.json

Follow finding lifecycle for a bounded number of rolling windows. On a terminal this opens a full-screen TUI (q quit, arrows/jk select, Enter/Space show or hide detail, PageUp/PageDown/Home/End scroll, h/? help). At 120×30 or larger it also shows the selected host/cgroup's six process-role lists beside the lifecycle panels; piped output or --json remain append-only text/JSON and carry the same scoped roles:

stallhunt watch --interval 2s --count 3

Serve Model Context Protocol tools over stdio so coding agents can query diagnoses directly — a resident sampler keeps a rolling view of recent pressure for instant answers (for Claude Code: claude mcp add stallhunt -- stallhunt mcp; see docs/mcp-server.md):

stallhunt mcp [--interval 2s] [--no-sampler]

Source packaging for an OpenAI local plugin, a Linux MCPB extension for compatible clients, and official MCP Registry metadata lives in the repository. These packages deliberately keep diagnosis local instead of hosting Stallhunt on a machine other than the one being investigated. See docs/directory-distribution.md.

Generate shell completions:

stallhunt completions bash > ~/.local/share/bash-completion/completions/stallhuntstallhunt completions zsh > ~/.local/share/zsh/site-functions/_stallhunt

Recording output paths are not overwritten unless --force is supplied. Ten-second hunts are the normal diagnostic path; sub-second observations are telemetry smoke tests and do not receive healthy or pressure verdicts. See docs/development.md for validation and opt-in acceptance commands.

Core idea

The primary abstraction is lost time, not utilization.

High utilization is not automatically a problem. A machine using 95% of its RAM may be perfectly healthy. A CPU at 70% utilization may still have latency-sensitive work suffering from scheduler contention. The project therefore focuses on evidence of stalled progress:

• CPU scheduler pressure,
• I/O stalls,
• memory pressure/reclaim,
• lock contention,
• network-related waits,
• eventually deeper blocking chains.

Example output shape:

$ stallhunt
SYSTEM HEALTH: DEGRADED
1. CPU scheduling contention                         SEVERE   Impact:    23.4% pressure during observation   Victims:   postgres [4812], nginx [5120]   Suspects:  rustc [9231], ffmpeg [9401]   Confidence: high
   Evidence:     CPU PSI some avg10:          23.4%     run queue latency estimate:  elevated     rustc CPU consumption:       735%     postgres runnable delay:     3.8s / 10s
2. Block I/O contention                              MODERATE   Device:    nvme0n1   Victim:    postgres [4812]   Suspect:   restic [7712]   Confidence: medium
Memory: no significant pressure detected.High memory occupancy alone is not treated as a bottleneck.

This output is aspirational; the project will reach it incrementally.

Product principles

1. Diagnose, do not merely display.
2. Measure stalled progress whenever possible.
3. Separate observation from inference.
4. Show evidence for every diagnosis.
5. Express uncertainty explicitly.
6. Remain useful without eBPF.
7. Stay cheap enough to run on a stressed system.
8. Treat Git as the complete project memory.

Initial scope

The first useful release targets Linux and focuses on:

• CPU scheduling contention,
• memory pressure,
• block I/O pressure,
• per-process attribution where Linux exposes enough evidence,
• cgroup/systemd-aware grouping when practical,
• human-readable terminal output,
• versioned machine-readable JSON (the pre-1.0 shape may evolve),
• bounded observation windows,
• deterministic offline fixture/replay analysis.

Later releases may add:

• eBPF-based off-CPU analysis,
• futex/lock contention,
• syscall/blocking attribution,
• network queue/socket diagnosis,
• dependency/wait graphs,
• richer cgroup/container analysis.

Repository map

.├── AGENTS.md├── CHANGELOG.md├── Cargo.toml├── LICENSE-APACHE├── LICENSE-MIT├── README.md├── src/│   ├── analysis.rs│   ├── cgroup.rs│   ├── cli.rs│   ├── cpu.rs│   ├── duration_us.rs│   ├── io.rs│   ├── main.rs│   ├── mcp/│   ├── memory.rs│   ├── observe.rs│   ├── psi.rs│   ├── record.rs│   ├── render.rs│   ├── report.rs│   ├── style.rs│   ├── tui/│   └── watch.rs├── tests/│   ├── cgroup_acceptance.rs│   ├── cli.rs│   ├── cpu_acceptance.rs│   ├── io_acceptance.rs│   ├── mcp.rs│   ├── memory_acceptance.rs│   └── fixtures/│       ├── cpu/│       └── proc-*└── docs/    ├── README.md    ├── install.md    ├── product.md    ├── architecture.md    ├── data-model.md    ├── telemetry.md    ├── inference-engine.md    ├── cli-ux.md    ├── security-privileges.md    ├── testing.md    ├── development.md    ├── codex-workflow.md    ├── experiments.md    ├── references.md    ├── roadmap.md    ├── status.md    ├── glossary.md    └── decisions/

Start with AGENTS.md, then docs/README.md.

Privacy Policy

Stallhunt reads bounded local Linux telemetry and does not independently send it over the network. MCP clients may transmit tool results according to their own data policies. See the full Stallhunt privacy policy for the data, storage, sharing, retention, and contact disclosures.

License

Dual-licensed under MIT or Apache-2.0, at your option.

Current state

For the current milestone, implemented capabilities, validation, known limits, and the next recommended task, see docs/status.md. Planned sequencing remains in docs/roadmap.md.