Update to OloLand DD

{
  "description": "Use when the user asks to run the OloLand ololand-dd-replay-run workflow. Re-run an existing OloLand agent run against the current harness + skills (or byte-for-byte against the original methodology bytes). Admin / compliance role required. The regulator-grade reproducibility surface — used to demonstrate that a graded output can be reconstructed from snapshots, and to A/B the current prompt against last week's data.",
  "included_files": [
    {
      "relative_path": "agents/openai.yaml",
      "size_in_bytes": 435
    }
  ],
  "name": "ololand-dd-replay-run",
  "skill_md_contents": "---\nname: ololand-dd-replay-run\ndescription: \"Use when the user asks to run the OloLand ololand-dd-replay-run workflow.\\\n  \\ Re-run an existing OloLand agent run against the current harness + skills (or\\\n  \\ byte-for-byte against the original methodology bytes). Admin / compliance role\\\n  \\ required. The regulator-grade reproducibility surface \\u2014 used to demonstrate\\\n  \\ that a graded output can be reconstructed from snapshots, and to A/B the current\\\n  \\ prompt against last week's data.\"\n---\n\n\n# Workflow for ololand-dd-replay-run\n\nFollow the OloLand command instructions below when the user asks for `ololand-dd-replay-run` or the equivalent workflow.\n\n# ololand-dd-replay-run\n\nYou are enqueueing a replay of an existing OloLand agent run. The original run's `root_prompt_snapshot` is replayed against either the *current* harness + skill catalog (`current_skills`) or the *original* methodology bytes (`byte_for_byte`).\n\n**This command requires admin or compliance role.** Non-privileged users will get `error_code: \"insufficient_role\"`.\n\n## Required inputs\n\n1. **run_id** — the agent_runs row to replay. Confirm from the user that this is the run they want replayed; replays cost credits and a Celery worker.\n\nOptional:\n- **mode** — `current_skills` (default) or `byte_for_byte`. See \"When to use which mode\" below.\n- **comment** — short reviewer note stored on the replay (e.g. \"Re-run after grader fix 2026-05-17\"). Strongly recommended for audit trail.\n\n## Action\n\n1. (Optional but recommended) Call `get_agent_run` first to confirm the run is replayable. The user should see at least: `agent_name`, `model_version`, `status`, `started_at`, `entry_point`. If `root_prompt_snapshot` is not present, the replay will fail with `error_code: \"no_snapshot\"` — surface that pre-emptively.\n\n2. Call `replay_agent_run` with `run_id`, `mode`, optional `comment`.\n\n3. On success, surface:\n   ```\n   Replay queued\n     Replay run id:  <replay_run_id>\n     Parent run id:  <parent_run_id>\n     Status:         queued\n     Compare:        <comparison_url>\n   ```\n   Tell the user to poll `get_agent_run <replay_run_id>` for completion. Typical end-to-end latency: 30–180 seconds depending on the original run's iteration count.\n\n4. On error, surface the `error_code` verbatim and the explanation:\n   - `not_found` — wrong run_id.\n   - `access_denied` — run belongs to a different company.\n   - `insufficient_role` — user is not admin/compliance.\n   - `no_snapshot` — original run pre-dates snapshot capture (pre-2026-05-14 runs).\n   - `bad_request` — invalid mode.\n   - `dispatch_failed` — Celery rejection; the replay row is marked error.\n\n## When to use which mode\n\n| Mode | Use when |\n|---|---|\n| `current_skills` | \"Would the *current* skill catalog have produced a different (better) answer for this run?\" Validates a methodology improvement against historical data. Default for ad-hoc reviews. |\n| `byte_for_byte` | \"Reproduce this run identically.\" Used for regulator-grade reproducibility — the auditor needs to see that the exact same skills + prompt produce the same output. Requires `skill_pack_snapshot` and `subagent_definition_snapshot` on the original run. |\n\n## When NOT to use\n\n- For exploratory \"what if\" prompt changes — author a new run instead, don't replay an old one.\n- To \"fix\" a wrong answer in a customer-facing output — replay creates a new audit-trail entry; the original stands. If the original is wrong, surface a claim correction via `unsupported OpenAI claim-correction workflow (see unsupported-workflows.json)`.\n- For live in-flight runs — replay is only defined for runs in terminal status (`completed`, `error`). Queue / running runs aren't replayable; finish them first.\n\n## Audit-trail discipline\n\nAlways pass `comment` when you can. The replay row is the regulator's tie between your reasoning (\"why did you re-run this?\") and the output. Sample comments:\n\n- `\"A/B test of v2 source-hierarchy skill against PR #1525 baseline\"`\n- `\"Compliance review per SR-2026-04-22; reproducing run for the regulator\"`\n- `\"Reviewer flagged inconsistent EBITDA; re-running with byte_for_byte\"`\n\nCompanion commands: `ololand-dd-inspect-run`, `ololand-dd-regulator-export`.\n\n\n## OpenAI execution requirements\n\nUse the connected MCP server's current tool schemas. Do not invent tool arguments, job-status endpoints, inputs, or outputs. If a referenced tool is unavailable, explain the limitation and request the needed connection or data. Invoke writes only when the user has requested the corresponding action. Do not initiate purchases or payments. An unrun, unsupported, or unavailable procedure is a gap, never a passing result.\n\n\nOpenAI cannot execute local lifecycle hooks. Use the packaged explicit setup/review workflow when requested. Server-side MCP auditing remains authoritative; preserve evidence blockers and human-only approval gates.\n\nClaim-correction submission is unsupported in this OpenAI upload because the current handler records surface=cowork. Do not invoke submit_agent_claim_correction from OpenAI. Explain the limitation and retain proposed correction text for human review without claiming submission. See unsupported-workflows.json for the excluded canonical workflow.\n"
}