flameox

Bounded local runtime evidence for coding agents.

Flameox coordinates profilers, benchmark tools, trace processors, and direct local targets. It gives an agent a short path from an explicit native artifact or live command to bounded evidence, while keeping preservation optional.

Version 0.2 is a clean break. There is no workspace to initialize, no flameox.toml, no SQLite control plane, no durable job to poll, and no parent directory discovery. Existing artifacts remain usable by passing their exact paths and formats to analyze; old .diagnostics state is not migrated.

Quick start

uv sync --extra dev --extra memory --extra trace --extra cpu
uv run flameox mcp inspect
uv run flameox analyze artifact.preview /absolute/path/to/artifact.json
uv run flameox capture --provider direct --cwd "$PWD" -- python benchmark.py

The MCP server has no workspace or project binding:

uv run flameox mcp serve

Run the short global setup wizard through npm:

npx flameox@latest setup

Setup detects Claude Code, Cursor, OpenCode, Codex, Gemini CLI, and Google Antigravity, then asks which clients should use Flameox. It preserves unrelated client configuration and writes a Python 3.12 uvx launcher pinned to the exact Flameox release that ran setup. Restart or reconnect changed clients afterward. For automation, pass --client codex --yes, repeat --client, or use --all --yes; --dry-run reports the same global paths without writing them. Detection is never automation consent.

Explicit --provider selections prepare the exact version-pinned uvx environment in the saved launcher by resolving it once into uvx's cache; they do not create a persistent global uv tool installation. Each invocation declares the complete managed provider set for that launcher rather than adding to remembered state. Use --timeout-seconds for a slow cold resolution. System and vendor tools are diagnosed with external install guidance. Setup never initializes or mutates a project.

Authority model

explicit artifact paths / typed direct target
                    │
                    ▼
         bounded process-lifespan runtime
             │                │
             ▼                ▼
       inline evidence   session scratch/cache
                              │
                       explicit preservation
                              │
                              ▼
             user Flameox data directory

Analysis and unpreserved capture make no durable Flameox writes. Capture artifacts stay in bounded session scratch until preservation, least-recently-used eviction, or server shutdown. An evicted analysis_id returns EXPIRED_SESSION_ANALYSIS; preserve conclusions before relying on them. The first preserve_evidence call creates the user-level Flameox data directory and stores native bytes and a canonical evidence bundle by SHA-256. FLAMEOX_DATA_DIR overrides the platform default for isolation or another storage location. Flameox never edits project Git files.

The console-retention default is bounded diagnostics in memory, with explicit omission counts. Keep native artifacts when needed; retain full console output on disk only when it is the evidence, an oracle needs it, or the caller requests it. Preservation alone does not request full logs. See console retention.

Workload time and RSS budgets are optional: use target.budget in MCP or --workload-budget in CLI capture. They do not inherit analysis-worker limits; cancellation and storage protection remain active when no workload budget is set.

The agent owns hypotheses and narrative findings in its own notes. Flameox owns only observed inputs, effective requests, execution provenance, typed evidence, coverage, truncation, limitations, and optional immutable preservation.

MCP interface

The server exposes actual evidence operations for client-side tool search instead of hiding its capabilities behind discover, inspect, or generic analyze(capability_id, arguments) calls. There are 26 read-only analysis tools, 20 executing capture tools, and three lifecycle tools. For example:

analyze_cpu_hotspots       capture_cpu_hotspots
analyze_cpu_callers        capture_cpu_callers
analyze_gpu_launches       capture_gpu_launches
analyze_benchmark_compare  capture_benchmark_summary
analyze_kernel_validation  capture_sanitizer_failures
prepare_providers          preserve_evidence          query_evidence

Each tool advertises its capability-specific options and compatible providers in its input schema. Analysis and capture have separate names and annotations because reading an artifact and executing a target are materially different effects. Tool search happens in the MCP client; Flameox does not require an additional catalog-search call.

It exposes one resource template, flameox://evidence/{evidence_id}, for the digest-bound, redacted projection of the canonical immutable manifest. Full argv, environment values, working directories, and host paths remain available only through explicit local manifest inspection. Native artifact bytes are deliberately not available as MCP resources.

Direct capture accepts an argv array, an explicit absolute cwd, bounded environment overrides, a typed compatible-provider variant, capability-specific options, an explicit single/experiment choice, and limits as top-level tool arguments. There is no generic request or arguments envelope. Shell strings are never accepted. Work remains owned by the live MCP request, so SDK progress and cancellation apply directly; there are no detached or restart-surviving tasks.

Managed external collectors such as py-spy execute from Flameox's uvx environment. In-process collectors such as coverage.py and Memray are verified in, and run with, the workload's declared Python interpreter. Flameox does not substitute one Python runtime for the other. When a capture reports a missing managed provider, prepare_providers prepares its version-pinned uvx environment and returns that same launcher for reconnection. The agent supplies the complete provider list it wants in that launcher; Flameox does not merge it with prior calls. Preparation does not modify the running MCP process. When the client must reconnect, the result returns a typed next_action with kind: "reconnect_mcp", an agent-facing message, and the launcher to use. The managed provider IDs are aiperf, memray, otlp, perfetto, py-spy, and torch. Host tools, drivers, and permissions are never installed or changed; the same result reports their setup guidance.

Comparison is intentionally a two-stage workflow. Flameox captures representative baseline and candidate summaries separately, optionally preserves them, and then passes both artifacts to an analyze_*_compare tool. There are no capture_*_compare tools: experiment capture measures cases and reports an effect, but it is not a substitute for comparing explicit native artifacts.

Evidence quality

An investigation still follows:

symptom → capture or explicit artifact → bounded evidence → hypothesis
        → discriminating experiment → supported, refuted, or inconclusive finding

A profile supports exploration, not causality. Confirmatory claims require a representative target, declared metric and estimand, compatible identities, preserved samples, a practical threshold, and an appropriate semantic oracle.

See architecture, storage and evidence, interfaces, runtime safety, and investigations for the contracts.

Development

Flameox requires Python 3.12 or newer and uses the committed uv.lock.

uv run ruff check src tests tools
uv run ruff format --check src tests tools
uv run mypy src tests tools
uv run lint-imports
uv run pytest -q

The project is licensed under the MIT License.