Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add D-Robotics/moss --skill reference-diffgit clone --depth 1 https://github.com/D-Robotics/mossWrote 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.
[](https://agentmods.dev/skills/d-robotics/moss/reference-diff)<a href="https://agentmods.dev/skills/d-robotics/moss/reference-diff"><img src="https://agentmods.dev/badge/skills/d-robotics/moss/reference-diff/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.
<a href="https://agentmods.dev/skills/d-robotics/moss/reference-diff"><img src="https://agentmods.dev/badge/skills/d-robotics/moss/reference-diff.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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 | $0.00047 | $0.01040 |
| Opus 5.5 | $0.00019 | $0.00416 |
| Sonnet 5.5 | $0.00009 | $0.00208 |
| Haiku 4.5 | $0.00005 | $0.00104 |
Grade A, and why
reference-diff 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 3d 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.
How it starts
The opening of the file, as written. The whole thing — 91 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Reference diffing (self-serve diagnosis)
scripts/refdiff.py compares two files; it knows nothing task-specific.
Model diff (MuJoCo XML)
python3 scripts/refdiff.py xml my_robot.xml reference/robot.xml
Prints, for every named structural element (body/site/joint/geom/actuator):
- elements missing from yours that the reference has (mount sites, joints, payload bodies are the high-signal ones),
- elements extra in yours,
- attribute deltas sorted by magnitude with per-component vector comparison
(
pos[1],quat[0], …) so one wrong axis is never hidden by a matching norm.
Use it to answer "what is structurally different?" before touching numbers.
Report diff (verifier checkpoints)
python3 scripts/refdiff.py report my_report.json reference/report.json
Prints every checkpoint where your score is below the reference, sorted by the gap, plus both stage tables and the top-line reward. Use it to answer "which checkpoint is actually costing points?" — fix those first; leave passing checkpoints alone.
Loop
- Run the benchmark, keep the verifier report.
refdiff report→ ranked checkpoint gaps.- Pick the largest gap;
refdiff xml→ ranked structural deltas that could explain it. - Change only what the diff and the scoring table justify.
- Re-run the public harness locally, then re-submit.
Why this matters (verified on RLE-Bench task08)
An operator doing this diff by hand found the two fixes that moved a score from 0.402 to 0.6386 (arm-mount height => static margin; diagonal mecanum pairing => validity). The tool surfaces the same class of findings mechanically — including ones a hand diff missed (a missing free joint, a missing payload body, wheel motor gains 12 vs 8) — so the agent can do this diagnosis itself.
Verified end-to-end with the agent driving: on task08, an iteration that was
given ONLY the tool and the reference artifacts (no operator diagnosis) called
refdiff.py xml six times, reasoned "the real divergence is the chassis ...
refdiff shows the mount offset at x=+0.04", adopted the reference frame/mount/
actuator settings, and scored 0.7603 — above the reference solution's 0.7564
(all validity gates 1.0, static margin 1.0, all five dynamic scenarios 1.0, S7
0.75). Two iterations earlier the same agent, still being hand-fed findings,
sat at 0.6386. The capability is what closed the last 0.12.
What ships with it
1 file beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 3d ago First seen · 91 lines · 47 tokens per session scan A c3221d9095c3
reference-diff is a skill published in the GitHub repository D-Robotics/moss (144 stars, last pushed yesterday), licensed MIT. It adds 47 tokens to every session and 1,040 once invoked, about $0.0002 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 2026-11-09.
reference-diff compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| reference-diffD-Robotics/moss | Skill | 144 | yesterday | MIT | A | 47 | Claude Code, Codex |
| systematic-debuggingopen-metadata/OpenMetadata | Skill | 16k | today | Apache-2.0 | A | 37 | Claude Code |
| log-error-digestzebbern/claude-code-guide | Skill | 4.7k | today | MIT | A | 71 | Claude Code, Codex |
| cwv-optimizeradobe/skills | Skill | 326 | today | Apache-2.0 | A | 112 | Claude Code |
| ci-fixersipyourdrink-ltd/bernstein | Skill | 1.2k | today | Apache-2.0 | A | 15 | Claude Code, Codex |
| diagnosing-missing-recordingsPostHog/posthog | Skill | 41k | today | none | A | 76 | Claude Code, Codex |
| triaging-web-analytics-supportPostHog/posthog | Skill | 41k | today | none | A | 151 | Claude Code, Codex |