Borrowing it
Nothing to install: this file belongs to ucb-bar/merlin. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/ucb-bar/merlin/main/.claude/skills/capsule-run-shape/SKILL.mdgit clone --depth 1 https://github.com/ucb-bar/merlinWrote 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/ucb-bar/merlin/capsule-run-shape)<a href="https://agentmods.dev/skills/ucb-bar/merlin/capsule-run-shape"><img src="https://agentmods.dev/badge/skills/ucb-bar/merlin/capsule-run-shape/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/ucb-bar/merlin/capsule-run-shape"><img src="https://agentmods.dev/badge/skills/ucb-bar/merlin/capsule-run-shape.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 | $0.00053 | $0.02174 |
| Opus 5.5 | $0.00021 | $0.00870 |
| Sonnet 5.5 | $0.00011 | $0.00435 |
| Haiku 4.5 | $0.00005 | $0.00217 |
Grade A, and why
capsule-run-shape 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 15d 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 — 132 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Capsule-bench run shape
The shape
--schedule continuous with a long --round-timeout (e.g. 43200). Rounds are not a terminator:
the run stops on evidence (converged / plateaued) or a declared budget, and the post-freeze
public+hidden L3 grade still runs, so a formal success is reachable.
Not --continuous. That is the legacy single-session path: it re-grades under one session but skips
the post-freeze public+hidden grade, returns 1, and hardcodes formal_complete=False. Measured
2026-09-01: both gemmini sessions closed after ~1.5 h at 18/33 with grades=2 when the agent stopped,
well inside a 12 h round timeout.
agent works continuously
├─ background grader every --grade-interval (900s): grades a SNAPSHOT COPY → qa/verdict.json
├─ capsule ladder is per-capsule, cheapest-first: L2 pass → L3 in the SAME grade
└─ tier_promote: a loop verdict enqueues a cert job at once
→ capsule 2's L3 runs while the agent is still on capsule 1's L2
→ record_cert writes the outcome back against the digest that earned it
Grading a snapshot copy is what makes it safe to grade while the agent works: the agent cannot mutate
what is being measured, and the grade cannot be contaminated mid-flight. _qa_work/cand_NN/ is that
copy; _qa_work/runs_NN/ holds the per-capsule verdicts for the same grade.
Why it is gated
Promotion is wrapped in a try/except so it can never gate a run. That is correct — but it means a
broken promotion is indistinguishable from an idle one. Five defects hid in this single path on
merlincirct_arm4_func_20260901_v4, and none looked like a failure:
| defect | how it presented |
|---|---|
enqueued sim="contract", rejected on any bespoke-sim target |
"nothing needed promoting" |
verilator slot dir was a fixed /tmp path owned by another user → broker crashed |
"the L3 infrastructure crashed" |
forwarded --tiers to agent_selfcheck.py, which had no such flag |
"no verdict produced" |
broker ran children with stdout/stderr=DEVNULL |
no diagnostic existed at all |
| a completed promotion's result was never written back | the capsule is "still pending" |
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.
- 15d ago First seen · 132 lines · 53 tokens per session scan A a76a6b7863ad
capsule-run-shape is a skill published in the GitHub repository ucb-bar/merlin (21 stars, last pushed yesterday), licensed Apache-2.0. It adds 53 tokens to every session and 2,174 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 2026-10-14.
Other skills, from other repositories
research-engineer
An uncompromising Academic Research Engineer. Operates with absolute scientific rigor, objective criticism, and zero flair. Focuses on theoretical correctness, formal verification, and optimal implementation across any required technology.
tika-eval-compare
Compare extracts from two Tika builds over a corpus to detect regressions in content, encoding, exceptions, and embedded-document handling. Use for "compare before/after extracts", "eval this change against the corpus".
cua-driver
Use Cua Driver for desktop or browser tasks that are awkward or unavailable through Bash/APIs, or when the user explicitly wants GUI interaction: app testing, visual bug reproduction, form filling, calendar entry, screenshots, and demo recording. Also covers Cua setup; not OpenAI Codex Computer Use or web research.
jetson-validate-image
Use after jetson-flash-image to run static BSP checks, on-target smoke/regression tests on a flashed DUT, or both. Not for build or flash steps. Triggers: validate bsp, on-target validation.
atmos-validation
Validate Atmos projects, components, arbitrary JSON Schema inputs, EditorConfig, and GitHub Actions; use affected-file selection and native CI annotations.
skill-benchmark
Benchmark AI skill effectiveness by measuring implementation quality against legacy constraints.