pod
Other formats agents might prefer:
markdownjsonllms.txt

Agent? You probably want markdown or json, or Pod over MCP.

Reported issues for Flameox

Pod holds 21 of 100 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to Flameox.

Most discussed

Publish Flameox as a discoverable MCP server in the MCP Registry

Current limitation

Flameox exposes MCP functionality through its maintained package distributions, but the repository does not currently publish MCP Registry metadata or an automated registry release path. Users therefore cannot discover the server through the official MCP Registry and may need to configure it manually.

Desired outcome

Publish Flameox under the registry name io.github.morluto/flameox, choosing and documenting the canonical package path:

  • npm ownership verification v

Read the thread · 2026-08-01 · closed · 4 comments

Bound and isolate high-cardinality Memray extraction

Problem

Memray extraction runs synchronously in the Flameox control process, eagerly materializes provider record sets and all normalized frame aggregates, and declares no record/frame/stack/time/memory limits. A modest native profile can block without progress for minutes; a high-cardinality profile can consume unbounded control-process memory and generate unbounded evidence rows.

Real performance evidence

A real Slime GAE profile had:

  • native artifact size: 5,100,561 bytes;
  • Memray

Read the thread · 2026-08-29 · closed · 3 comments

Concurrent read-only analyses fail after DuckDB configuration is locked

Before submitting

  • I searched the open and closed issues for a related report.
  • I have removed credentials, tokens, private paths, and other sensitive data.

Observed behavior

Two read-only MCP analysis calls against the same registered run can interfere with each other. Running analyze_memory and analyze_hotspots concurrently caused one analysis to fail with a raw DuckDB configuration error.

Expected behavior is for independent read-only analyses to be concurrency-safe,

Read the thread · 2026-08-29 · closed · 3 comments

[Security][Architecture] Do not let agent-approved adapters execute inside the Flameox control process

Summary

Flameox explicitly documents Python entry-point adapters as part of the trusted computing base because loading an entry point executes arbitrary Python code. At the same time, the MCP-facing prepare_adapter flow allows an agent to approve an installed third-party adapter identity, after which AdapterRegistry.load_approved() calls entry_point.load() inside the Flameox control process.

That collapses the security boundary the rest of Flameox works hard to maintain. Once impo

Read the thread · 2026-08-10 · closed · 3 comments

Use a producer-qualified Memray reader for captured profiles

Problem

Flameox can capture a Memray profile with the workload's exact Memray environment and then fail to extract that same profile because extraction imports a different, older Memray reader from the Flameox control environment.

The capture advertises the producer as supported and registers a valid artifact, but the producer/reader compatibility is not actually bound end to end.

Real reproduction

Against a fresh current-main workspace for Slime:

  1. `capture run memray --workload sli

Read the thread · 2026-08-29 · closed · 2 comments

Route missing Perfetto extraction through managed capability setup

Before submitting

  • I searched the open and closed issues for a related report.
  • I have removed credentials, tokens, private paths, and other sensitive data.

Observed behavior

When Perfetto extraction needs the managed Trace Processor and it is absent, the MCP recovery path does not route the caller to Flameox's managed capability setup.

Reproduction

From a workspace without a prepared Perfetto runtime, preserve or import a Torch trace that requires extract_perfetto,

Read the thread · 2026-08-29 · closed · 2 comments

Make the documented AIPerf setup path callable through MCP

Before submitting

  • I searched the open and closed issues for a related report.
  • I have removed credentials, tokens, private paths, and other sensitive data.

Observed behavior

Flameox's inference planner and built-in AIPerf adapter direct callers to start_capability_setup(adapters=["aiperf"]), but the MCP tool schema does not accept aiperf. A configured inference scenario can therefore reach a permanent manual-recovery dead end even though AIPerf is declared as a managed

Read the thread · 2026-08-29 · closed · 2 comments

Expose GPU correctness and kernel adapters through capability discovery

Before submitting

  • I searched the open and closed issues and pull requests for a related proposal.
  • I have removed credentials, tokens, private paths, and other sensitive data.

Area

Adapters and providers

Current limitation

Flameox implements and documents GPU-kernel adapters that do not appear in the live list_capabilities result. An agent therefore cannot discover whether Nsight Compute, Compute Sanitizer, Triton compiler capture, or CuTe compiler capture is availa

Read the thread · 2026-08-29 · closed · 2 comments

Most recent

Expose bounded process output for failed-run recovery

Problem

When a captured workload fails, Flameox preserves stdout/stderr as immutable process_output artifacts but provides no supported CLI, MCP tool, or resource for reading even a bounded text preview. The run projection sets process.stderr/process.stdout to null, and the artifact resource is explicitly metadata-only.

This leaves the diagnostic required for recovery inside Flameox while making it inaccessible through Flameox's public agent surfaces.

Reproduction

Tested from F

Read the thread · 2026-08-29 · closed · 1 comment

fix(workloads): reject oracle declarations without argv before capture planning

Before submitting

  • I searched the open and closed issues for a related report.
  • I have removed credentials, tokens, private paths, and other sensitive data.
  • I ran the reproduction below.

Evidence basis

  • Runtime reproduction
  • Source or test analysis
  • Documentation or contract mismatch

Observed behavior

A workload can declare an oracle with a strength but no argv. Flameox accepts the configuration and produces a capture plan, then fails only when executing the

Read the thread · 2026-08-29 · closed · 1 comment

Add a CLI handoff for executing an already reviewed capture plan

Problem

The capture CLI can preview and issue a server-owned capture plan, but it provides no way to execute that exact reviewed plan. capture run always plans again and immediately consumes the replacement without accepting the preview's token or expected plan identity.

The MCP surface exposes execute_capture_plan, so the application capability already exists; the CLI omits the handoff.

Reproduction

Current main:

flameox capture plan memray --workload slime-gae-sanit

[Read the thread](https://github.com/morluto/flameox/issues/357) · 2026-08-29 · closed · 1 comment

### Warn and bound high-volume whole-entrypoint Torch profiler captures

## Problem

Flameox allows high-volume Torch profiler options over an entire entrypoint without a strong admission warning or bounded first-pass recommendation. Combining CUDA/CPU activities, stack capture, shape recording, and memory profiling can create hundreds of megabytes of trace data and gigabytes of profiler memory for a tiny workload.

The existing artifact and RSS limits eventually bound the process, but they do not help users design a minimally perturbing capture before execution.

##

[Read the thread](https://github.com/morluto/flameox/issues/346) · 2026-08-29 · closed · 1 comment

### Support bounded trace windows over normalized Nsight Systems events

## Problem

Flameox normalizes Nsight Systems CUDA, kernel, memory, and NVTX events into timestamped `trace.event` evidence, but its bounded `trace window` workflow is hard-wired to the Perfetto Trace Processor. Users can aggregate an Nsight Systems trace with `analyze accelerator-launches` but cannot inspect the exact events in a suspicious time interval through Flameox.

## Evidence from a real Slime trace

After working around the independently reported nullable-NVTX parser defect (#339), Fla

[Read the thread](https://github.com/morluto/flameox/issues/340) · 2026-08-29 · closed · 1 comment

### Define a durable run-scoped semantic envelope alongside native artifacts

### Before submitting

- [x] I searched open and closed issues and pull requests for a related discussion.
- [x] I removed credentials, private paths, and other sensitive data.
- [x] I separated confirmed constraints from the recommendation below.

### Area

Storage and evidence; MCP; adapters and providers.

### Evidence basis

Runtime reproduction; source analysis; documented Flameox contracts.

### Decision needed

Should every captured run persist a bounded, typed semantic envelope containin

[Read the thread](https://github.com/morluto/flameox/issues/338) · 2026-08-29 · closed · 1 comment

### Bind Compute Sanitizer version into capture and XML validation

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

Flameox discovers the installed Compute Sanitizer version, but capture planning and XML validation do not bind that version into the produced evidence. A clean sanitizer result can therefore remain `inconclusive` because the validator sees an unknown producer version.

### Reproduction

The capability i

[Read the thread](https://github.com/morluto/flameox/issues/330) · 2026-08-29 · closed · 1 comment

### Make Compute Sanitizer first-pass captures explicitly bounded

### Before submitting

- [x] I searched the open and closed issues and pull requests for a related proposal.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Area

Adapters and providers

### Current limitation

The Compute Sanitizer adapter already exposes `launch_count`, but a first-pass capture defaults to `launch_count=0` (unlimited). Planning does not make that risk prominent or choose a bounded probe, and timeout recovery does not suggest using the ex

[Read the thread](https://github.com/morluto/flameox/issues/329) · 2026-08-29 · closed · 1 comment

### Honor the global --workspace override during init

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

`flameox --workspace <path> init` reports success but initializes the command-local/default workspace instead of the global override. This is particularly confusing for automation because other commands honor the global option and the successful exit gives no indication that initialization happened else

[Read the thread](https://github.com/morluto/flameox/issues/326) · 2026-08-29 · closed · 1 comment

### Prevent shared artifact digests from expanding exact run analysis scope

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

Exact-run analysis can expand beyond the selected run when identical artifact bytes are registered to multiple runs. Artifact registration lookup brings evidence from another run into scope even though the caller supplied one run ID.

### Reproduction

1. Register one content digest under two distinct r

[Read the thread](https://github.com/morluto/flameox/issues/325) · 2026-08-29 · closed · 1 comment

### Distinguish empty process snapshots from successfully observed zero activity

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

An empty process snapshot is represented as a successful observation of zero matching activity even when the response does not establish that process visibility was complete.

### Reproduction

Run the process-snapshot tool while exercising a Slime workload and receive an empty process collection. The r

[Read the thread](https://github.com/morluto/flameox/issues/323) · 2026-08-29 · closed · 1 comment

### Preserve bounded process output when detached capture is cancelled

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

Cancelling a detached capture records cancellation but does not publish stdout and stderr already collected within the configured bounds as retrievable process-output evidence.

### Reproduction

1. Start a detached capture for a workload that prints a marker and remains active.
2. Wait for the marker, 

[Read the thread](https://github.com/morluto/flameox/issues/322) · 2026-08-29 · closed · 1 comment

### Bind comparison identity to exact inputs instead of the global corpus head

### Before submitting

- [x] I searched the open and closed issues for a related report.
- [x] I have removed credentials, tokens, private paths, and other sensitive data.

### Observed behavior

The same comparison request over the same frozen run sets and unchanged measurement rows receives a new `comparison_id` whenever unrelated evidence advances the corpus commit.

This makes comparison references unstable and defeats idempotent recording even though the numerical result and all declared in

[Read the thread](https://github.com/morluto/flameox/issues/321) · 2026-08-29 · closed · 1 comment

The remaining reports are on [the project's issue tracker](https://github.com/morluto/flameox/issues).