# Hill Climb

> Runs a global hill-climb optimization loop where the parent is always the current best candidate and the val significance gate decides acceptance. Use as the algorithm for most runs. Pick how each iteration's reflection is focused with --focus all (whole train set), cyclic (one task at a time), or hardest-first (lowest-scoring tasks first). Replaces the former all-at-once, cyclic, and hardest-fir…

- **Type:** Skill
- **Install:** `agentstack add skill-skillberry-ai-cap-evolve-hill-climb`
- **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior
- **Seller:** [skillberry-ai](https://agentstack.voostack.com/s/skillberry-ai)
- **Installs:** 0
- **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills)
- **Latest version:** 0.1.0
- **License:** Apache-2.0
- **Upstream author:** [skillberry-ai](https://github.com/skillberry-ai)
- **Source:** https://github.com/skillberry-ai/cap-evolve/tree/main/skills/algorithms/hill-climb

## Install

```sh
agentstack add skill-skillberry-ai-cap-evolve-hill-climb
```

Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent.

## About

# hill-climb — one loop, three focus schedules

Every iteration: take the current best candidate, ask the optimizer to propose an
edit (its prompt emphasizes a *focus set* of train tasks), score the result on
val, and accept only if it clears the significance gate. The accepted candidate
becomes the new best. The test split is never touched here — that is `finalize`.

The three former hill-climb skills were byte-identical except for one constant;
they are now one skill with `--focus`:

| `--focus` | what each iteration emphasizes | when to use |
|---|---|---|
| `all` (default) | the whole train set — find the single edit that lifts the most tasks | broad capability gaps; the usual choice |
| `cyclic` | one train task at a time, cycling through them | many distinct, unrelated failure modes |
| `hardest-first` | train tasks ranked by baseline score ascending (lowest first), then cycling | a few very hard tasks dominate the gap |

Why only the focus changes: the parent-selection rule (always the current best),
the gate, and the honesty guarantees are identical across schedules — only the
*attention* differs. Keeping one loop means a fix to the gate or memory wiring
can never drift between variants.

## Inputs / outputs (manifest tokens)
- **needs:** `scores` + `traces` (the per-task val results to reflect on) and
  `candidate` (the parent to extend).
- **provides:** `candidate` (the accepted best).

## Standalone use

```bash
python scripts/run.py --run-dir .capevolve/run_X --project .capevolve/project \
  --optimizer 'python .../run-optimizer/scripts/run.py --name mock --workdir {workdir} --prompt {prompt}' \
  --focus hardest-first --max-iterations 10 --n-trials 4
```

`--resume` continues from the run's current best (reading its val from rollouts)
instead of the baseline. `--no-regression` adds a SWE-bench-style dual gate:
reject a candidate that breaks any val task the parent already passed, even if the
mean improves.

Back-compat: `--focus all-at-once` is accepted and treated as `all`.

## References
- `references/focus-schedules.md` — how each schedule builds its focus set.

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [skillberry-ai](https://github.com/skillberry-ai)
- **Source:** [skillberry-ai/cap-evolve](https://github.com/skillberry-ai/cap-evolve)
- **License:** Apache-2.0

Install and usage instructions live in the source repository linked above.

## Pricing

- **Free** — Free

## Security capabilities

Automated source analysis of v0.1.0 — what this tool can access:

- **Network access:** no
- **Filesystem access:** no
- **Shell / process execution:** no
- **Environment & secrets:** no
- **Dynamic code execution:** no

*"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.*


## Versions

- **0.1.0** — security scan: passed — Imported from the upstream source.

## Links

- Listing page: https://agentstack.voostack.com/l/skill-skillberry-ai-cap-evolve-hill-climb
- Seller: https://agentstack.voostack.com/s/skillberry-ai
- Browse the marketplace: https://agentstack.voostack.com/browse

---
Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.