菜单
Derive compact eval rows from command-attributed trace trajectories.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | skill-command-trajectory-eval-v1 |
| description | Derive compact eval rows from command-attributed trace trajectories. |
| mode | agent-skill |
| requires | ["ot dataset","ot trace get","ot trace map"] |
Build a compact, publish-safe evaluation row projection from the raw
skill-command-trajectories evidence source.
The source evidence is intentionally verbose: it keeps full steps, Trace Map nodes, and raw trace/bucket references so attribution can be audited. This workflow is the projection layer: it emits the smaller row shape an eval harness actually needs:
{
"source_trace_id": "raw:raw-row-000001",
"source_unit_id": "raw-unit:raw-row-000001",
"summary": "scout-research: research Voyage AI...",
"eval_task_type": "command_trajectory_attribution",
"source_raw_row_id": "raw-row-000001",
"project_slug": "2026-05-01-hackathon-...",
"session_id": "...",
"command_name": "/scout-research",
"command_source": "claude_slash_command",
"command_intent": "research Voyage AI...",
"command_intent_source": "command_args",
"expected_skill_name": "scout-research",
"expected_trajectory_start_step": 1,
"expected_trajectory_end_step": 11,
"expected_excluded_steps_json": "[]",
"expected_has_write_operations": true,
"expected_write_operation_count": 1,
"expected_files_modified_json": "[\"kb/br/01-voyage-ai-code-mode-data-interface.md\"]",
"expected_tools_json": "[\"Bash\", \"Read\", \"Write\"]",
"label_confidence": "high",
"source_has_injected_body_args": false,
"source_skill_body_mismatch": false,
"limitations_json": "[]",
"evidence_refs_json": "{\"trace\":\"...\",\"unit\":\"...\"}"
}
python ~/.opentraces/workflows/skill-command-trajectory-eval-v1/scripts/build_rows.py \
--output "$OT_DATASET_OUTPUT"
The script defaults to:
~/.opentraces/datasets/skill-command-trajectories/data/train.jsonl
Pass --source <path> to project a different raw command trajectory dataset.
This row shape supports deterministic checks:
Raw evidence stays in the private evidence source/bucket. Eval rows carry only
compact labels and the source raw row ordinal for audit. The private evidence
source retains the exact trace ids, unit ids, and ot:// refs.
Share agent traces to open datasets on HuggingFace Hub. Use this skill when the user mentions OpenTraces, trace capture, Trace Trails, workflow-built datasets, dataset review, or publishing reviewed dataset rows.
Share agent traces to open datasets on HuggingFace Hub. Use this skill when the user mentions OpenTraces, trace capture, Trace Trails, workflow-built datasets, dataset review, or publishing reviewed dataset rows.
Project already-captured traces into scored-rollout rows for the SkillOpt optimizer.
Per-commit trace lineage rows for a branch context, deterministic.
Land every site-related fix on main, deploy opentraces.ai to Vercel production, and verify it works across all pages (desktop + mobile) and the email-capture flow, fixing and re-deploying until green. Runs as a goal: land -> deploy -> verify -> triage -> repeat until safe. Use when the user says "deploy site", "deploy to vercel", "push site", "ship the site", "deploy-site", or "land and verify the site". For a full coordinated release that also bumps the version and publishes the Python packages, use /release-pack instead.
Take one or more GitHub issues and drive each to a review-ready pull request, holding one belief throughout: an issue is a hypothesis, not a spec. The pipeline VERIFIES BEFORE IT TRUSTS and runs TWO ADVERSARIES before it commits: multi-agent investigation of the real cause (flagging any "landmine" — a cited function or file:line that does not exist) and the blast radius, a plan whose test must go RED BEFORE GREEN, two adversarial gates on that plan (a self/ethos checklist + a pure codex adversarial pass), implementation expressed as a forged `/goal` (via goal-forge) whose verify loop runs the otbox acceptance journey, a green suite + docs sync, a PR that `Closes #<n>`, and a FINAL codex adversarial review of the diff. Honesty, not theater: it reports unmet acceptance criteria and invalid issues rather than faking a fix. The human merges. Everything happens in a per-issue git worktree; multiple issues each get their own worktree + branch + PR. Use when the user says "issue-to-pr", "solve issue", "fix issue #N"