kdense-arbor

name

kdense-arbor

description

Autonomously improve a real artifact (code, training recipe, agent harness, data pipeline, prompt) against an objective and an evaluator, using Hypothesis Tree Refinement (HTR) from the Arbor paper. Use this whenever someone wants to iteratively optimize something over many experiments without overfitting — e.g. "get my model's eval score up", "improve this agent/harness", "tune this pipeline", "beat the baseline on this benchmark", "run a search over approaches and keep the best", "do an MLE-bench / Kaggle-style optimization", or any long-horizon "make this artifact better and don't just memorize the dev set" task. Trigger it even when the user doesn't say "Arbor" or "hypothesis tree" but describes repeated experiment-and-evaluate loops, branching exploration of competing ideas, or worries about a dev/test gap. Runs Claude itself as the coordinator with subagent executors in isolated git worktrees; for the standalone `arbor` CLI tool see references/arbor-upstream.md.

Kdense Arbor

Use this Codex adapter for the pinned upstream scientific skill.

Catalog status

• Status: inactive
• Source: kdense
• Audit reasons: executable-content
• Upstream license: MIT license
• Inventory: ../../inventory.json
• Upstream instructions: ../../../vendor/scientific-agent-skills/skills/arbor/SKILL.md

Procedure

1. Resolve the plugin root as three directories above this skill directory.
2. Read the matching inventory record and verify its status, reasons, source path, and pinned revision.
3. Do not read the upstream instructions until the user has seen every audit reason and given explicit acknowledgement. Acknowledgement permits instruction inspection only.
4. Treat upstream frontmatter as source metadata, not as Codex configuration. Follow the instruction body only where it is compatible with available Codex tools and the current user-approved plan.
5. Treat tools named for another agent runtime as capability requests. Map them to an available Codex tool only when the behavior is equivalent; otherwise report the unavailable capability.
6. Ask before credentials, package installation, new network hosts, remote compute, writes, paid services, destructive actions, or imported executable code.
7. Use $science-provenance for produced scientific outputs and $science-review before presenting scientific claims.

Boundaries

• This wrapper does not grant credentials, network access, package installation, or script execution.
• Do not change an inactive record to active from inside a research task.
• Preserve failures, uncertainty, source identifiers, and licensing constraints.