BatchJob Async Job Skill

Use this skill when the user wants to run or manage batch jobs through the BatchJob service.

Required Environment

• BATCHJOB_BASE_URL
• BATCHJOB_BEARER_TOKEN

All HTTP requests must include:

-H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}"
-H "Content-Type: application/json"

API Endpoints

• POST /v1/batch/files:upload
• POST /v1/batch/jobs:precheck
• POST /v1/batch/jobs
• GET /v1/batch/jobs/{job_id}
• GET /v1/batch/jobs?page=1&page_size=10&status=...
• POST /v1/batch/jobs/{job_id}:cancel

Automation Policy (Default)

• Always run in full-auto mode.
• Do not ask user for file_id first.
• Resolve file source from current message/context, then upload automatically when needed.
• Ask follow-up questions only when no readable file source can be obtained.
• Accepted input file formats for upload: jsonl, csv, xlsx, xls (BatchJob normalizes to internal JSONL).
• For jsonl, each line must be either:
  • canonical Vertex format (contents + optional generationConfig)
  • simple prompt format (prompt + optional aspect_ratio / image_urls, where image_urls must be publicly reachable URLs)
• If user gives only model mode after a file message, treat it as confirmation and continue automatically.

Guardrails (Must Follow)

• Do not auto-retry by creating a second job unless user explicitly asks.
• Do not auto-rewrite dataset format after a terminal failure unless user explicitly asks.
• Before upload, auto-normalization is allowed only once for known safe mappings (e.g. simple prompt JSONL -> Vertex JSONL).
• Upload failure is terminal for this run: do not continue to precheck/submit when upload fails.
• Never submit when row_count <= 0.
• Do not fetch or parse output_summary_url automatically; only do it when user asks for detailed failure reason.
• After reaching terminal status (completed, failed, partially_failed), stop execution and return summary immediately.

JSONL Compatibility Rule (Important)

BatchJob internal execution expects each JSONL line to be a VertexGeminiImageRequest shape.

• Canonical line format:
  • contents[0].parts[0].text contains prompt text
  • generationConfig.imageConfig.aspectRatio is optional
  • generationConfig.responseModalities should include IMAGE and TEXT
• Non-canonical simple JSONL like {"prompt":"...","aspect_ratio":"1:1"} is acceptable; server will normalize it to Vertex format.
• If JSONL has neither contents nor prompt, stop and ask user to provide valid data.
• Explicitly unsupported (must reject before submit):
  • OpenAI Batch style lines containing method + url + body (for example /v1/chat/completions payload).
  • This schema will fail with Vertex error: at least one contents field is required.

Dataset Output Rule (Important)

When user asks you to generate a template/demo file for BatchJob image tasks:

• Prefer CSV headers: prompt,aspect_ratio,image_urls
• Or JSONL line format with prompt / aspect_ratio / image_urls
• If image_urls is provided, it must be publicly accessible (http/https).
• Do NOT generate OpenAI batch envelope (custom_id + method + url + body) as final upload file

User Template Response (Important)

When user asks for format/template, return:

1. A short explanation (prompt required, aspect_ratio/image_urls optional, and image_urls must be public URLs).
2. One copyable CSV snippet (default).
3. Optionally one JSONL snippet (simple prompt schema).
4. Local template file paths (if available):
   • /home/node/.openclaw/workspace/templates/batchjob-input-template.csv
   • /home/node/.openclaw/workspace/templates/batchjob-input-template.jsonl
   • /home/node/.openclaw/workspace/templates/batchjob-format-guide.md

Do not output OpenAI batch envelope examples in template replies.

File Source Resolver (Strict Order)

1. Existing file_id:
   • if provided, skip upload.
2. Public file_url (http:// or https://):
   • download to temp local file, then upload.
3. Explicit local file_path:
   • if readable, upload.
4. Inbound attachment local path from channel/runtime context:
   • examples: /tmp/..., MEDIA:<path>, /tmp/openclaw-media/....
   • if readable, upload.
5. Channel private file token/object (no local path, no public URL):
   • if runtime has a channel adapter that can download attachment bytes, use it and upload.
   • if not available, enter fallback interaction.

Resolver output must be normalized to one of:

• file_id
• file_path (readable local file)

Execution Flow

1. Confirm model and mode; if missing, use safe defaults (model=google/gemini-2.5-flash-image, mode=fast) and tell user.
2. Resolve file source using resolver above.
3. If resolver gives file_path, upload via POST /v1/batch/files:upload to get file_id.
   • For .jsonl, inspect a few non-empty lines first:
     • if contents exists, upload as-is.
     • if only prompt/aspect_ratio/image_urls exists, upload as-is (server will normalize).
     • if method + url + body exists, stop and ask user to switch to BatchJob schema.
     • if structure is unknown, stop and ask user for valid schema.
   • Backward compatibility fallback:
   • if upload fails with unsupported file type for csv/xlsx/xls, convert once to JSONL and retry upload once.
   • do not retry more than once.
   • If upload returns validation error (unsupported schema, no valid data rows), stop immediately and return fix guidance.
4. Run precheck with record_count (prefer uploaded file row_count).
5. Submit job with file_id.
6. Poll job status until terminal (completed, failed, partially_failed) with bounded timeout:
   • interval: 5 seconds
   • max polls: 12 (about 60 seconds)
   • if still non-terminal after max polls: return current status and job_id, then stop.
7. Return concise summary with job_id, status, progress, and output_summary_url.

When user only asks for estimate, stop at precheck and do not submit.

Fallback Interaction (Only When Needed)

Use this when resolver cannot read file bytes from current channel/context:

我拿到了“文件引用”，但当前运行环境无法直接读取该附件内容。请任选其一： 1) 直接发一个可公网下载的 URL 2) 提供本机可读路径（如 /tmp/xxx.csv） 3) 先把文件上传到 BatchJob，给我 file_id

If the current channel supports resending as direct attachment path in context, also ask user to resend once.

File Source Playbook

A) Public URL -> Local Temp File

FILE_URL="https://example.com/input.jsonl"
EXT="${FILE_URL##*.}"
FILE_PATH="$(mktemp "/tmp/batchjob-input.XXXXXX.${EXT:-jsonl}")"
curl -fL --retry 3 --connect-timeout 10 "$FILE_URL" -o "$FILE_PATH"

A2) JSONL Schema Sanity Check

Use this before upload to avoid wrong schema submission.

SRC_JSONL="/tmp/input.jsonl"
FIRST_LINE="$(grep -m1 -v '^[[:space:]]*$' "$SRC_JSONL")"
echo "$FIRST_LINE" | jq -e '
  (has("contents")) or
  (has("prompt")) and ( (has("method")|not) and (has("url")|not) and (has("body")|not) )
' >/dev/null || {
  echo "JSONL schema invalid for BatchJob: requires contents or prompt-only schema"
  exit 1
}

B) Feishu/Channel Attachment Path

If message context already includes local attachment path, treat it as FILE_PATH directly.

FILE_PATH="/tmp/openclaw-media/your-uploaded-file.jsonl"

If only a channel token/link is provided but no downloadable URL and no local path, try channel adapter download first. If adapter is unavailable, use fallback interaction.

Curl Templates

FILE_PATH="/path/to/input.jsonl"
FILE_NAME="$(basename "${FILE_PATH}")"
test -f "${FILE_PATH}" || { echo "文件 ${FILE_PATH} 不存在"; exit 1; }
FILE_CONTENT_B64="$(base64 < "${FILE_PATH}" | tr -d '\n')"

curl -sS "${BATCHJOB_BASE_URL}/v1/batch/files:upload" \
  -H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}" \
  -H "Content-Type: application/json" \
  -d "{\"filename\":\"${FILE_NAME}\",\"mode\":\"fast\",\"content\":\"${FILE_CONTENT_B64}\"}"

curl -sS "${BATCHJOB_BASE_URL}/v1/batch/jobs:precheck" \
  -H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{"record_count": 100, "model": "google/gemini-2.5-flash-image", "mode": "fast"}'

curl -sS "${BATCHJOB_BASE_URL}/v1/batch/jobs" \
  -H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{"file_id": "your-file-id", "model": "google/gemini-2.5-flash-image", "mode": "fast"}'

curl -sS "${BATCHJOB_BASE_URL}/v1/batch/jobs/${JOB_ID}" \
  -H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}"

curl -sS "${BATCHJOB_BASE_URL}/v1/batch/jobs/${JOB_ID}:cancel" \
  -H "Authorization: Bearer ${BATCHJOB_BEARER_TOKEN}" \
  -H "Content-Type: application/json" \
  -d '{"reason":"user requested cancellation"}'