A workflow for running a set of building-data questions against several answering agents and scoring their results. It first checks that the required data-server tools are actually connected.

Use it to run, score, compare, or reproduce evaluations of building-profile answers. It is for testing agent setups rather than answering building questions itself.

It prevents a test run from appearing successful when an agent is present but its data server is unavailable. It helps compare answering approaches using the same question set.

How DaveGold/mcp-metadata-demo configures Claude Code and Codex

Reuse

Borrowing it

Nothing to install: this file belongs to DaveGold/mcp-metadata-demo. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.

Copy the file

curl -O https://raw.githubusercontent.com/DaveGold/mcp-metadata-demo/main/.agents/skills/run-eval/SKILL.md

Clone the repo

git clone --depth 1 https://github.com/DaveGold/mcp-metadata-demo

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

README.md

[![agentmods](https://agentmods.dev/badge/skills/davegold/mcp-metadata-demo/run-eval/github.svg)](https://agentmods.dev/skills/davegold/mcp-metadata-demo/run-eval)

Your own site

<a href="https://agentmods.dev/skills/davegold/mcp-metadata-demo/run-eval"><img src="https://agentmods.dev/badge/skills/davegold/mcp-metadata-demo/run-eval/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

Your own site · 80×15

<a href="https://agentmods.dev/skills/davegold/mcp-metadata-demo/run-eval"><img src="https://agentmods.dev/badge/skills/davegold/mcp-metadata-demo/run-eval.svg" alt="Reviewed on agentmods" width="80" height="20"></a>

Per session 66 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.

When invoked 5,558 The whole file, excluding the scripts and references it only reads on demand.

Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.

Origin original No closer match found in the catalogue.

Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

Model	Per session	Once invoked
Fable 5.1	$0.00066	$0.05558
Opus 5.5	$0.00026	$0.02223
Sonnet 5.5	$0.00013	$0.01112
Haiku 4.5	$0.00007	$0.00556

Measured 1mo ago against content hash cff4b4deaba4, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

run-eval scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.agents/skills/run-eval/SKILL.md · 367 lines

How it starts

The opening of the file, as written. The whole thing — 367 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Run the eval

Before running

Check the arms' MCP TOOLS are connected. Confirm the mcp__eval-<arm>__get_building_profile tools are actually available to this session — not that the arm appears in .mcp.json, and not that the eval-<arm> agent appears in the agent list.

The agent list is NOT a proxy for this, and treating it as one will waste a whole run. Observed on 2026-09-21: after eval-inline was added mid-session, the harness picked up the new agent from .Codex/agents/ and announced it as available, while the MCP server stayed unconnected. Spawning that agent would have produced a subagent with zero tools, declining every question — a run that looks like a result and is an artefact of the harness.

Skills reload mid-session. The agent registry can reload mid-session. An MCP connection is established when the client connects — at process start, and again on any reconnect.

Observed on 2026-09-22, in one session, in both directions. The eval-inline-oneline arm was built, deployed and verified on the wire by raw HTTP. Minutes later the harness announced the new agent as available — and a ToolSearch for mcp__eval-inline-oneline__get_building_profile returned no matching tool in that same session. Agent present, tools absent, simultaneously. Spawning it there would have produced 60 runs of a subagent with no tools, declining every question, and the output would have looked like a result.

Then, later in the SAME session, the MCP client dropped and reconnected — and picked the new arm up. A live call to mcp__eval-inline-oneline__get_building_profile returned the correct payload.

CORRECTION, 2026-09-22. This step used to say MCP connections "do not" reload, and that a session predating an arm "will never reach it, and no amount of retrying fixes it." That is too strong and it is now known to be false. A reconnect picks up .mcp.json entries added since the session began, and a reconnect can happen on its own. A full restart is a fix, not the only fix.

What survives, and is the part that matters: the agent registry and the MCP connection move independently, so the agent list is never evidence. The window where an agent exists and its tools do not is real and was observed twice today. Do not infer either state from the other, in either direction.

Read the full file on GitHub · 367 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

1mo ago Changed cff4b4deaba4
1mo ago Changed · +27 lines 0fb7f703203e
1mo ago Changed · +21 lines 3daeed1db3a0
1mo ago Changed · +106 lines d1462687ad82
1mo ago First seen · 213 lines · 66 tokens per session scan A 8c114ab1f037

Subscribe to this mod's changes

run-eval is a skill published in the GitHub repository DaveGold/mcp-metadata-demo (5 stars, last pushed 11d ago), licensed MIT. It adds 66 tokens to every session and 5,558 once invoked, about $0.0003 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-21.

Testing ai-tooling bag claude energy-label kadaster mcp

run-eval next to 6 skills from other repositories
Mod	Type	Stars	Last push	Licence	Scan	Tokens	Works with
run-evalDaveGold/mcp-metadata-demo	Skill	5	11d ago	MIT	A	66	Claude Code, Codex
reviewing-pull-requestsd-oit/rust-2026-template	Skill	15	yesterday	MIT	A	132	Claude Code
architecture-diagramd-oit/rust-2026-template	Skill	15	yesterday	MIT	A	67	Claude Code, Codex
dora-reportd-oit/rust-2026-template	Skill	15	yesterday	MIT	A	52	Claude Code, Codex
atomic-commitd-oit/rust-2026-template	Skill	15	yesterday	MIT	A	76	Claude Code, Codex
privacy-firstd-oit/rust-2026-template	Skill	15	yesterday	MIT	A	72	Claude Code, Codex
task-decompositiond-oit/rust-2026-template	Skill	15	yesterday	MIT	A	56	Claude Code, Codex

mcp-metadata-demo: Skill for Claude Code