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 skillberry-ai/cap-evolve --skill hill-climbgit clone --depth 1 https://github.com/skillberry-ai/cap-evolveWrote 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/skillberry-ai/cap-evolve/hill-climb)<a href="https://agentmods.dev/skills/skillberry-ai/cap-evolve/hill-climb"><img src="https://agentmods.dev/badge/skills/skillberry-ai/cap-evolve/hill-climb.svg" alt="Measured on agentmods" height="20"></a>- NVIDIA SkillSpector pass
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 | $0.00120 | $0.02318 |
| Opus 5 | $0.00060 | $0.01159 |
| Sonnet 5 | $0.00024 | $0.00464 |
| Haiku 4.5 | $0.00012 | $0.00232 |
Grade A, and why
hill-climb 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 8d 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 — 152 lines — stays where its author put it; the contents beside it link to each section on GitHub.
hill-climb — one loop, three focus schedules
Greedy search over candidates: the parent is always the run's current best, and a
child replaces it only by clearing the val significance gate. The test split is
never touched here — that is finalize.
Requires baseline first. Without --resume the loop reads the seed's val
result from <run-dir>/baseline.json (scripts/run.py:121-122); with no run state
it raises FileNotFoundError: no run state at .../state.json. --resume instead
reads the current best's val from its stored rollouts, falling back to
baseline.json when the run has no best yet (run.py:118-122).
One iteration, end to end
This is the mechanism the other algorithm skills vary; they describe only their
differences and point back here. One iteration is harness.run_step
(core/cap_evolve/harness.py:1409):
- Pick the parent — always the current best (
harness.py:2224,run_dir.candidate_dir(run_dir.best_id)). Copy it towork/<cand_id>/; the optimizer edits that copy in place, so the parent is never mutated (harness.py:1454-1458). - Build the prompt. The parent's val per-task rows are split into
always-failing / flaky / infra-errored / solid (
harness.py:1775-1795), rendered as the failure index plus an explicit protect these passing ids block (harness.py:1803-1879), and substituted into the project's optimizer-instructions template.--focusnarrows which failures are emphasized; nothing else changes. - Inject context and memory. Full trajectories, capability guidance, and the four
cross-iteration files land in the workdir (
harness.py:1062-1094) — seereferences/run-step.md. - Optimize. The optimizer command mutates the workdir. A crash is caught, logged,
and left as an unchanged copy of the parent, so the gate simply rejects it — a
wasted iteration, not a dead run (
harness.py:1486-1500). - Evaluate on val only (
harness.py:1516), at--n-trialstrials per task. - Gate. With per-task data on both sides the paired test is chosen automatically:
accept iff mean per-task Δ >
k·SE of those paired deltas (harness.py:1524-1532).--no-regressionadds a second, harder condition on top. - Commit. Every candidate is snapshotted — accepted and rejected — so any
iteration can be diffed (
harness.py:1557); the version store commits it (harness.py:1609-1612). Only an accepted candidate callsset_bestand becomes the next parent (harness.py:1558-1559); a rejected one is filed in the rejected memory that feeds the next prompt (harness.py:1607-1608).
What ships with it
7 files 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.
- 8d ago First seen · 152 lines · 120 tokens per session scan A e4738eab24c8
hill-climb is a skill published in the GitHub repository skillberry-ai/cap-evolve (54 stars, last pushed yesterday), licensed Apache-2.0. It adds 120 tokens to every session and 2,318 once invoked, about $0.0006 per session on Opus 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 2026-08-30.
Other skills, from other repositories
foundry-config-setup
Resolve missing setup caused by a hardcoded Foundry project endpoint or model in a sample. Use when a sample fails because it uses a placeholder/hardcoded projectendpoint (for example "https://your-project.services.ai.azure.com") or a hardcoded model instead of reading them from the environment.
jetson-diagnostic
Read-only Jetson health snapshot for identity, memory, GPU, thermal, power, storage, services, and top processes.
nemo-automodel-launcher-config
Configure NeMo AutoModel job launches for interactive runs, Slurm clusters, and SkyPilot cloud execution.
dogfood
Systematically explore and test a mobile app on iOS/Android with agent-device to find bugs, UX issues, and other problems. Use when asked to dogfood, QA, exploratory test, find issues, bug hunt, or test this app on mobile.
mochi-remind
Handle due reminders — notify the user with natural language and mark them done.
neuron-evaluation-engineer
Create and run AI evaluations with datasets, assertions, and output drivers in Neuron AI. Use this skill whenever the user mentions evaluation, testing AI systems, creating evaluators, dataset-driven testing, assertion-based validation, or wants to measure AI system performance. Also trigger for tasks involving…