{
 "schema_version": 1,
 "recorded_at": "2026-09-30",
 "subject": "Whether the building-agentskills loader Skill is discovered when installed and loads before the first edit on authoring, audit and tool-description prompts (toolboxmd/building-agentskills #17)",
 "sources": {
  "repository": "toolboxmd/building-agentskills",
  "issue": "https://github.com/toolboxmd/building-agentskills/issues/17",
  "refs": {
   "main": "53f1864",
   "loader": "f8db18e",
   "desc": "eec057d"
  },
  "docs": [
   {
    "url": "https://code.claude.com/docs/en/plugins-reference",
    "read": "2026-09-30"
   },
   {
    "url": "https://learn.chatgpt.com/docs/build-skills",
    "read": "2026-09-30"
   }
  ]
 },
 "hosts": [
  {
   "host": "claude-code",
   "version": "2.1.284",
   "model": "claude-opus-5-5",
   "effort": "medium",
   "install": "--plugin-dir",
   "confinement": "Bash sandbox denyRead of real home and keychain link"
  },
  {
   "host": "codex",
   "version": "0.159.0",
   "model": "gpt-6-astra",
   "effort": "low",
   "install": "local marketplace, codex plugin add",
   "confinement": "permission profile: root read, project roots and tmpdir write, real home none"
  },
  {
   "host": "grok",
   "version": "1.0.44",
   "model": "grok-4.7",
   "effort": "medium",
   "install": "grok plugin install --trust",
   "confinement": "sandbox profile extending workspace, kernel deny of real home entries except ~/.grok"
  },
  {
   "host": "opencode",
   "version": "1.18.33",
   "model": "opencode-go/muse-spark-1.3-contributor",
   "effort": "",
   "install": "skills.paths",
   "confinement": "external_directory denied except the copy and the repository"
  }
 ],
 "runs": {
  "total": 84,
  "benchmark_batch": 68,
  "desc_remeasure": 16,
  "excluded_no_output": 0
 },
 "scoring": "fired = loader loaded and any routed page read successfully before the first edit; negative clean = neither opened; refused reads do not count",
 "files": [
  {
   "path": "case-studies/evidence/2026-09-30-loader-trigger-benchmark/records.jsonl",
   "lines": 84,
   "sha256": "77992e19f5a201bcff7978b58f6505b42e926ee691bbf2627a0939045fbb1766",
   "role": "one anonymized record per run",
   "runs": 84
  },
  {
   "path": "case-studies/evidence/2026-09-30-loader-trigger-benchmark/summary.md",
   "lines": 58,
   "sha256": "798e1cb4f422e6d6931d3936de12f4499055b7964ff23d5e4c377335a541762d",
   "role": "tables regenerated by trigger_test.py --summary records.jsonl"
  },
  {
   "path": "case-studies/evidence/2026-09-30-loader-trigger-benchmark/discovery.txt",
   "lines": 140,
   "sha256": "4a6d663900fd4dfb40115b5dc1dcde6d71bf29e8bd358717c39a8e2b11519f07",
   "role": "host listings for main and the fix, from discover.py"
  },
  {
   "path": "case-studies/evidence/2026-09-30-loader-trigger-benchmark/trigger_test.py",
   "lines": 876,
   "sha256": "ab89573ae82ed054435885c205a51a650af20a3cdb60435c082f5e3af50bfc59",
   "role": "harness, scorer and summary"
  },
  {
   "path": "case-studies/evidence/2026-09-30-loader-trigger-benchmark/discover.py",
   "lines": 89,
   "sha256": "eb2b92810cee3e71da8c26b21e78a9ae56d2de382541792df73b29278705ac70",
   "role": "discovery check"
  }
 ],
 "amendments": [
  "tool case added at the planner's request after #4",
  "Grok Build and OpenCode added by the planner at 3 runs per cell",
  "records scrubbed once after the batch with trigger_test.scrub; summary unchanged"
 ],
 "limits": [
  "3 to 5 runs per cell, one fixture, one model per host",
  "control arm is 5 pooled runs per host (2, 2, 1 per prompt), below 5 per prompt and arm; no per-prompt comparison",
  "control arm on Claude Code can browse the --add-dir copy",
  "desc arm lacks #4's tool-description row",
  "OpenCode confinement is a permission rule, not an OS sandbox"
 ]
}