Reported issues for Runner MCP

Pod holds 18 of 18 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to Runner MCP.

Most discussed

Harden source installer against resolver stalls and ambiguous Python crashes

Summary

A clean Linux lab installation exposed an installer/runtime hardening gap in install.sh.

During the normal source-install path, the isolated venv was created successfully, but:

1. python -m pip install "$ROOT_DIR" first terminated with a real Python SIGSEGV while resolving/installing dependencies.
2. A later disposable reproduction of the combined dependency install did not crash, but entered a single-threaded ~100% CPU resolver spin with no log progress for multiple…

Read the thread · 2026-09-30 · closed · 31 comments

Qualify aifordable-lab host/guest after reproducible Python SIGSEGV

Incident

aifordable-lab still produces intermittent native Python SIGSEGV / rc=139 during fresh-process imports of the MCP server stack.

This issue tracks the remaining private host/guest qualification after installer hardening #194 and host-integrity feature #195 were completed.

Reproducible crash boundary

Known lab runtime:

• Ubuntu 24.04
• Python 3.12.3
• mcp 2.2.0
• mcp-types 2.2.0
• pydantic 2.13.5
• pydantic-core 2.46.5
• starlette 0.52.1
• httpx2 2.13.1
• cryptography…

Read the thread · 2026-09-30 · closed · 9 comments

Replace GitHub mailbox as primary runtime control transport

Architectural requirement: Runner MCP must not depend on GitHub for normal runtime control.

GitHub remains acceptable for source control, CI, release distribution, audit export and an optional degraded-mode mailbox fallback, but the primary command/result path must continue to operate when GitHub is unavailable or rate-limited.

Target architecture:

• Runner establishes an outbound-only authenticated connection to an AIfordable-owned control relay; no inbound listener on the private host. -…

Read the thread · 2026-09-28 · open · 9 comments

Extend host-integrity gate with hardware and storage fault evidence

Context

Lab qualification for Runner MCP is currently blocked by intermittent native Python SIGSEGVs. The existing Linux host-integrity adapter correctly detects recent fatal Python process evidence, but it does not classify likely underlying host faults.

Current evidence makes hardware/runtime instability a serious hypothesis:

• Python/pip and MCP imports have crashed at different points;
• pure source compile stress can remain stable;
• source and installed files were byte-identical in…

Read the thread · 2026-09-30 · closed · 7 comments

Harden replacement-host qualification for service-user runtimes and stale installed CLIs

Context from live qualification on aifordable-lab after replacing the unstable i9-14900KS with a temporary i5-12400F.

Observed operator/runtime mismatches:

• The canonical Runner MCP runtime is private under /home/aifordable-runner/.local/share/runner-mcp/venv.
• Running the #198 userspace probe as operator gerard reported Python executable not found or not executable even though the runtime existed; the service-user home permissions prevented traversal. The safe invocation had to run…

Read the thread · 2026-10-02 · closed · 6 comments

Activate and prove Runner Fabric coarse work-unit bridge end to end

Transport foundation is already merged through Runner MCP PR #111 as main commit 3ebd39981c358c307fa66afa32ad2f5fc32731cf; issue #109 is complete. Do not rebuild or widen that bridge.

Remaining goal

Activate and prove the existing coarse Runner Fabric boundary end to end only after the upstream Fabric authority/durability prerequisites are ready.

Target: agent -> Runner MCP fabric_run_work_unit -> private loopback Runner Fabric agent MCP -> trusted Fabric work-unit service

Runner…

Read the thread · 2026-09-27 · open · 6 comments

Prove recovery-capable self-update on the private host

Track the remaining live proof from Phase 3.8.11 without widening Runner MCP authority.

Preconditions

• Use the merged recovery-capable baseline from current main; reconcile live GitHub first.
• Do not expose hostnames, addresses, usernames, paths, tokens, service names or other private infrastructure in GitHub.
• Use only the existing bounded Runner MCP/operator interfaces. Do not add a general remote shell or package-manager control as a workaround.
• Keep the operator emergency-stop…

Read the thread · 2026-09-27 · open · 5 comments

Mailbox bridge: liveness heartbeat and stale-request recovery contract

Add a generic, infrastructure-neutral resilience contract for private GitHub mailbox watchers.

Goals:

• expose a sanitized liveness heartbeat without host/path/service details
• detect requests that remain unprocessed beyond a configurable age
• resume backlog processing after watcher restart without replaying completed request IDs
• preserve the existing replay ledger and fail-closed protocol
• define bounded retry/backoff for transport errors only; never retry arbitrary Runner MCP actions…

Read the thread · 2026-09-20 · closed · 4 comments

Most recent

Expose bounded Agent Bus durable-result convergence status

Context: AF-22.6 / Fabric #479 live qualification needs to verify that a replayed terminal result was acknowledged and removed from the private outbox without inspecting private filesystem paths or payloads.

Goal

Add a read-only local status surface for the Runner MCP-owned Agent Bus state directory.

Contract

• runner-mcp agent-bus convergence;
• report only bounded category + pending durable-result count;
• categories: not_initialized, clear, pending, invalid;
• never print…

Read the thread · 2026-10-02 · closed · 0 comments

Add bounded async Runner Fabric bootstrap job

Required to activate the GitHub-independent Agent Bus on an existing private Runner MCP host without adding generic remote shell/package authority.

Scope:

• bootstrap only canonical Blacksp1d3r/Runner-Fabric at one full lowercase commit;
• asynchronous durable job with start/status actions so mailbox/MCP request timeout cannot leave ambiguous state;
• source fetched into a private temporary root;
• build wheel with the installed Runner MCP Python using --no-deps --no-build-isolation; -…

Read the thread · 2026-09-29 · closed · 0 comments

Make offline wheel staging test respect test-runtime packaging capability

Live self-update diagnosis: target lint passes, but target unit suite fails only at test_real_project_wheel_can_stage_without_index.

The live custom unit profile runs in the project test venv, while real self-update wheel staging runs in the installed Runner MCP service venv. The project test venv is not guaranteed to contain the package build backend.

Fix:

• keep the real offline wheel staging test active when the test interpreter has the required packaging backend;
• skip that…

Read the thread · 2026-09-29 · closed · 1 comment

Fix self-update runtime status distribution identity

Live pre-bootstrap runtime_status still reports version: development although the installed distribution is aifordable-runner-mcp.

The self-update runtime status path still queries the legacy distribution name runner-mcp.

Fix that read-only status path to use the canonical installed distribution identity and add a regression test. No update/install authority changes.

Read the thread · 2026-09-29 · closed · 0 comments

Add local-only self-update baseline bootstrap

The first self-update on an already-installed Runner MCP host currently fails safely with bootstrap_required because no compatibility/install baseline has been recorded yet.

Add one local-only bootstrap command for existing trusted installations.

Required behavior:

• command is CLI-only; no MCP, mailbox, Agent Bus or Fabric mapping;
• requires canonical runner-mcp project configuration and non-production staging environment;
• requires a clean checkout at an explicit full lowercase commit…

Read the thread · 2026-09-29 · closed · 0 comments

Add local private Agent Bus configuration commands

Follow-up to #130.

Add local-only Agent Bus configure/status/remove commands.

Requirements:

• relay origin + runner subject validated before persistence;
• relay credential entered through hidden prompt or stdin only, never CLI argv;
• private env update remains atomic/locked and preserves unrelated secrets;
• status returns configured/not-configured only;
• remove deletes only Agent Bus relay values;
• no MCP or GitHub-mailbox mapping for configure/remove;
• no arbitrary local endpoint,…

Read the thread · 2026-09-29 · closed · 0 comments

Add managed Agent Bus worker lifecycle and make GitHub watcher fallback-only

Parent: #125.

Operationalize the new AIfordable-owned control channel on the private runner.

Required:

• add a managed agent-bus-worker lifecycle beside server/github-watcher/completion-watcher;
• worker starts from private config only;
• worker runs outbound-only; no inbound port;
• expose bounded status: installed/enabled/active/primary-connected/fallback-only/degraded;
• do not print relay credential or private endpoint details;
• keep GitHub watcher independently enable/disable-able; -…

Read the thread · 2026-09-29 · closed · 2 comments

Respect GitHub REST reset windows and reduce watcher polling pressure

A live private-host mailbox watcher exhausted the authenticated GitHub REST primary rate limit (5000/hour).

Observed behavior:

• GitHub returned HTTP 403 with X-RateLimit-Remaining: 0 and an explicit reset timestamp.
• The mailbox watcher continued retrying through the generic 2s/5s transport retry path and then polled again every 5 seconds.
• With a backlog, the watcher performs both request-head and compare calls per cycle, so a 5-second default can consume ~1440 REST calls/hour before…

Read the thread · 2026-09-28 · closed · 0 comments

Fix CLI package-version lookup after aifordable-runner-mcp rename

Live private-host upgrade from the legacy 0.1.0 install to exact current main exposed a packaging identity mismatch: the new aifordable-runner-mcp==0.1.2 distribution installed successfully and new runtime checks were present, but runner-mcp --version and status still reported 0.1.0 because package_version() queries importlib.metadata.version("runner-mcp"). The old distribution metadata remains in the reused venv.

Acceptance:

• query the canonical distribution name…

Read the thread · 2026-09-28 · closed · 0 comments

Bridge Runner Fabric coarse work-units into Runner MCP

Goal

Reduce agent dependence on many direct GitHub/Git/test tool calls by exposing Runner Fabric's coarse work-unit boundary through Runner MCP.

Product boundary remains unchanged:

• Runner Fabric owns orchestration, GitHub Gateway, work-unit state/policy and durable execution;
• Runner MCP is transport/safety only and must not absorb Fabric orchestration logic.

Tool surface

When a private Fabric endpoint is configured, expose only:

• fabric_run_work_unit
• fabric_get_work_unit -…

Read the thread · 2026-09-27 · closed · 0 comments

The remaining reports are on the project's issue tracker.