{
  "schema_version": 1,
  "recorded_at": "2026-09-29",
  "subject": "Whether agents read the procedure an operations routing row links before the row's action (toolboxmd/agentsmd #164, #174, #182)",
  "sources": {
    "repository": "toolboxmd/agentsmd",
    "commit": "06b7af6",
    "commit_url": "https://github.com/toolboxmd/agentsmd/tree/06b7af6/docs/work",
    "version": "14.5.0",
    "issues": [
      "https://github.com/toolboxmd/agentsmd/issues/164",
      "https://github.com/toolboxmd/agentsmd/issues/174",
      "https://github.com/toolboxmd/agentsmd/issues/182"
    ],
    "pull_requests": [
      {
        "number": 168,
        "merge_commit": "038d40b",
        "version": "14.2.0",
        "role": "harness, confinement, reworded first-edit rows, #164 benchmark"
      },
      {
        "number": 179,
        "merge_commit": "3fa82c0",
        "version": "14.4.0",
        "role": "session-start pointer, prose before-clause, successful-read scoring, no-edit verdicts, plugin and config-link reads"
      },
      {
        "number": 181,
        "merge_commit": "b24f0a4",
        "version": "14.4.1",
        "role": "Bash allowed on Claude, verification cell measured, shell-read rule"
      },
      {
        "number": 183,
        "merge_commit": "588d713",
        "version": "14.4.2",
        "role": "Claude Code Bash sandbox, live v14.4.0 canary record"
      },
      {
        "number": 184,
        "merge_commit": "06b7af6",
        "version": "14.5.0",
        "role": "routing-row sweep, verification row, widened pointer, instrument fixes"
      }
    ],
    "records": [
      {
        "path": "docs/work/164-trigger-audit/records.jsonl",
        "lines": 600,
        "sha256": "7a85cff2dab045f60aca5ef08251305e4b71c38572a021aa13c1b5a495fea3b3",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/164-trigger-audit/records.jsonl",
        "runs": 600,
        "role": "#164 four-host benchmark plus an earlier pilot; the pilot is out of scope"
      },
      {
        "path": "docs/work/174-routing-injection/records.jsonl",
        "lines": 180,
        "sha256": "59c8f11dfd31583cbfddd4a386f42fefa3cd7bac855b31b5654734a84dd64b12",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/records.jsonl",
        "runs": 180,
        "role": "#174 first-edit cases, fixed harness"
      },
      {
        "path": "docs/work/174-routing-injection/midtask.jsonl",
        "lines": 40,
        "sha256": "5c5abad859aa0e1422894b4e09f4139667db9b652086fa7f5573c340b63f953c",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/midtask.jsonl",
        "runs": 40,
        "role": "#174 mid-task cases on Claude Code and OpenCode, fixed harness"
      },
      {
        "path": "docs/work/174-routing-injection/canary.jsonl",
        "lines": 4,
        "sha256": "2a770e8026e46de4a00e1433ea68d64f6adb9efc483a4d421cc39c4754218e9b",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/canary.jsonl",
        "runs": 4,
        "role": "post-install canary on live v14.4.0"
      },
      {
        "path": "docs/work/174-routing-injection/records-prefix.jsonl",
        "lines": 520,
        "sha256": "9ac66ea73e9f126fbcca63412dd9f5866a4845bd542694383e68b40a43820d8d",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/records-prefix.jsonl",
        "runs": 520,
        "role": "superseded pre-fix batch"
      },
      {
        "path": "docs/work/174-routing-injection/midtask-prefix.jsonl",
        "lines": 60,
        "sha256": "71ab2931c56b9b9a834a1184520190a1787ed77bafa1cff3465a582ecd728bb2",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/midtask-prefix.jsonl",
        "runs": 60,
        "role": "superseded pre-fix mid-task batch; Codex and Grok cells still cited"
      },
      {
        "path": "docs/work/182-routing-sweep/records.jsonl",
        "lines": 309,
        "sha256": "323797dc5a40d046586e0147ec9e497b9c748d932f50fbc78ae92af20fedbc72",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/182-routing-sweep/records.jsonl",
        "runs": 309,
        "role": "#182 routing-row sweep and other-host check"
      }
    ],
    "results": [
      {
        "path": "docs/work/164-trigger-audit/results.md",
        "lines": 255,
        "sha256": "e89c0ce1d860c0f89456e562e30090de63cdbfd0f5e5f685d04a7911f095804a",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/164-trigger-audit/results.md"
      },
      {
        "path": "docs/work/174-routing-injection/results.md",
        "lines": 172,
        "sha256": "3d73d62a0ccd996e09ac979b7e1ac35ee02a3c04620ade906146fef52adbb291",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/174-routing-injection/results.md"
      },
      {
        "path": "docs/work/182-routing-sweep/results.md",
        "lines": 140,
        "sha256": "0bba29204a8678a9475eef6bdc7e71c9990025d0619c1856a498d1ed0e78672d",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/182-routing-sweep/results.md"
      }
    ],
    "harness": [
      {
        "path": "docs/work/164-trigger-audit/trigger_test.py",
        "lines": 1145,
        "sha256": "bfce6ade53ae06c253ea70b5b80026e8f6b69a426372a4f9561047d45db75d0a",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/docs/work/164-trigger-audit/trigger_test.py"
      },
      {
        "path": "tests/test_trigger_harness.py",
        "lines": 363,
        "sha256": "196de1d57de569f8d38eff8077dd8b0a17a57ff225bd84557c42a9887ad72d3d",
        "url": "https://github.com/toolboxmd/agentsmd/blob/06b7af6/tests/test_trigger_harness.py"
      }
    ],
    "comments": [
      "https://github.com/toolboxmd/agentsmd/issues/182#issuecomment-5897168360"
    ]
  },
  "run_counts": {
    "total_records": 1713,
    "in_scope_note": "The pilot rows in 164 records.jsonl and the pilot-model rows in records-prefix.jsonl are out of scope and not reported."
  },
  "hosts": [
    {
      "host": "Claude Code",
      "version": "2.1.284",
      "models": [
        {
          "model": "claude-opus-5-5",
          "effort": "medium"
        }
      ]
    },
    {
      "host": "Codex",
      "version": "0.159.0",
      "models": [
        {
          "model": "gpt-6-astra",
          "effort": "low"
        },
        {
          "model": "gpt-6-luna",
          "effort": "high",
          "benchmarks": [
            "#164"
          ]
        }
      ]
    },
    {
      "host": "Grok Build",
      "version": "1.0.44",
      "models": [
        {
          "model": "grok-4.7",
          "effort": "medium"
        }
      ]
    },
    {
      "host": "OpenCode",
      "version": "1.18.33",
      "models": [
        {
          "model": "opencode-go/muse-spark-1.3-contributor",
          "effort": "default variant"
        }
      ]
    }
  ],
  "regenerate": "cd <extracted>/docs/work/<folder> && python3 ../164-trigger-audit/trigger_test.py --summary <records>.jsonl",
  "counting_rules": {
    "successful_read": "A read counts only when the tool call is not in permission_denials, not answered by an error (Read tools), and not an OpenCode error part. Shell calls fail only on a permission denial (from #181).",
    "fired": "--summary counts a required file as fired when it was read before the moment and the moment happened.",
    "read_noaction": "#182 results.md also counts read-noaction runs (read, then no action) as read before the action; the case study shows both counts.",
    "both_counts": [
      {
        "cell": "#182 main finalization, repository-setup, reconciliation",
        "summary": "0/3 each",
        "results_md": "3/3 each"
      },
      {
        "cell": "#182 fix2 finalization, repository-setup, reconciliation",
        "summary": "0/3 each",
        "results_md": "3/3 each"
      },
      {
        "cell": "#182 main project-direction",
        "summary": "2/3",
        "results_md": "3/3"
      },
      {
        "cell": "#182 fix2 project-direction",
        "summary": "2/3",
        "results_md": "3/3"
      },
      {
        "cell": "#182 main delivery",
        "summary": "1/3",
        "results_md": "3/3"
      },
      {
        "cell": "#182 Codex gpt-6-astra verification on 10ece7c",
        "summary": "3/5",
        "results_md": "5/5"
      }
    ]
  },
  "exclusions": [
    "The #164 pilot batch that escaped into the canonical checkout, discarded by AgentsMD; the pilot model's scores are not reported.",
    "Discarded or aborted batches listed in the #164 results, not present in the records.",
    "The pre-fix #174 batch (records-prefix.jsonl) and the Claude Code and OpenCode cells of midtask-prefix.jsonl: invalid because refused reads counted and no-edit runs scored as fired.",
    "#182 batches dropped for wrong moments or collapsed record keys, rerun before recording."
  ],
  "canaries": [
    {
      "version": "14.4.0",
      "result": "4/4",
      "source": "docs/work/174-routing-injection/canary.jsonl",
      "note": "--summary raises KeyError on this file (no source or login_files_changed field); counted from verdict fields"
    },
    {
      "version": "14.5.0",
      "result": "4/4",
      "source": "https://github.com/toolboxmd/agentsmd/issues/182#issuecomment-5897168360"
    }
  ],
  "interpretation_limits": [
    "3 to 10 runs per cell, one model per host, one fixture: the results rank wordings on these prompts and are not variance estimates.",
    "#164 records carry no failed_calls; its OpenCode cells were never rescored under successful-read rules.",
    "Opus base in #174 may be 5/20 rather than 3/20 under the non-zero-exit rule (per #174 results.md; the kept call lists show one such read before the first edit).",
    "Codex and Grok mid-task cells in #174 come from the pre-fix batch.",
    "Moments are tool-call patterns, so a regex decides what counts as the action.",
    "The benchmarks are evidence, not release gates; behavioral Live Verification stays with the user."
  ]
}