Synthlabs Curation
Apps & Automationsynthlabs-curationSynthlabs Curation is an Agent Skill for apps & automation published in the mkurman/zorai repository on GitHub: a SKILL.md instruction file an AI agent can load to do this task.
Use when SynthLabs dataset curation should run as repeatable background HTTP jobs instead of manual UI work.
How to use this skill
Bring this guide into your coding agent with a prompt tailored to the tool you use.
- Open your project in Codex.
- Copy and paste the prompt below.
- Review files and risks before approving any action.
Inspect the Agent Skill at https://github.com/mkurman/zorai/blob/HEAD/skills/synthlabs/synthlabs-curation/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/synthlabs-curation, but wait for my approval before writing files or running scripts.
Copying this prompt does not install or run the skill. Review third-party files before use.
SynthLabs Curation
Overview
Use this skill for SynthLabs dataset curation that runs through existing HTTP job routes instead of manual UI-first work. This is workflow guidance for an external app; do not assume zorai exposes dedicated SynthLabs tools.
When to Use
Use this skill when:
- the task is to autoscore, rewrite, remove items, or migrate reasoning content through background jobs,
- the operator wants a repeatable polling loop instead of row-by-row browser actions,
- the run should leave a verification trail in the active zorai session, task, or goal notes,
- or maintenance checks such as logs stats, tag cleanup, or orphan scans help validate the dataset state.
Do not use this skill when:
- the task is row-level manual review, verifier approval, data preview, or other visual inspection,
- provider credentials must be entered interactively in the SynthLabs UI,
- or the backend is not healthy yet and
synthlabs-setupstill needs to start or repair it.
Agent Rules
- Read current job or session state before starting a new curation job.
- Start one specific job, record its
jobId, then poll until it reachescompletedorfailed. - Treat autoscore and rewrite as AI-backed jobs that need a SynthLabs-compatible encrypted
apiKeyvalue. - Treat remove-items, migrate-reasoning, orphan checks, and most tag cleanup as non-AI maintenance that does not need encrypted model credentials.
- Route row-level manual review, verifier approval, and visual spot checks to
synthlabs-ui-operator.
Verification-First Loop
- Inspect existing work with
GET /api/jobsor session state so you do not start duplicate jobs. - Start exactly one curation route with the narrowest safe payload.
- Poll
GET /api/jobs/:iduntil the job iscompletedorfailed. - Review
progress,result, anderrorfields before declaring success. - Update the active zorai session, task, or goal notes with the session ID, job ID, filters used, final counts, and whether UI verification is still pending.
- If any rows need human judgment or visual approval, hand that follow-up to
synthlabs-ui-operatorinstead of continuing in blind HTTP mode.
Polling And Discovery
List recent jobs, optionally filtered by type or status:
curl -fsS "http://localhost:8787/api/jobs?type=rewrite&status=running&limit=20"
Fetch the full state of one job after start:
curl -fsS "http://localhost:8787/api/jobs/job_123"
Prefer GET /api/jobs/:id once you have a jobId; use GET /api/jobs for discovery, queue visibility, or to confirm whether a similar run is already active.
AI-Backed Jobs
POST /api/jobs/autoscore
Use autoscore when a session's items need model-based 1-5 scoring.
curl -fsS -X POST "http://localhost:8787/api/jobs/autoscore" \
-H "Content-Type: application/json" \
-d '{
"sessionId": "session_abc",
"provider": "openai",
"model": "gpt-4.1-mini",
"baseUrl": "https://api.openai.com/v1",
"apiKey": "<encrypted iv:ciphertext value>",
"limit": 100,
"offset": 0,
"concurrency": 1,
"maxRetries": 2,
"retryDelay": 2000,
"sleepMs": 500,
"force": false
}'
POST /api/jobs/rewrite
Use rewrite when the backend should improve query, reasoning, or answer fields in place.
curl -fsS -X POST "http://localhost:8787/api/jobs/rewrite" \
-H "Content-Type: application/json" \
-d '{
"sessionId": "session_abc",
"provider": "openai",
"model": "gpt-4.1-mini",
"baseUrl": "https://api.openai.com/v1",
"apiKey": "<encrypted iv:ciphertext value>",
"fields": ["reasoning", "answer"],
"limit": 50,
"concurrency": 1,
"sleepMs": 500
}'
Autoscore and rewrite both call the SynthLabs AI client and decrypt the submitted apiKey. Use an encrypted value compatible with the backend's VITE_API_KEY_SALT or API_KEY_SALT behavior. Do not send a plaintext provider key and do not invent your own encryption flow.
Non-AI Cleanup And Migration Jobs
POST /api/jobs/remove-items
Use remove-items for deterministic cleanup. This route does not need model credentials.
Dry-run a threshold cleanup first:
curl -fsS -X POST "http://localhost:8787/api/jobs/remove-items" \
-H "Content-Type: application/json" \
-d '{
"sessionId": "session_abc",
"scoreThreshold": 3,
"scoreField": "score",
"dryRun": true
}'
The route accepts exactly one removal method: indices or scoreThreshold, never both.
POST /api/jobs/migrate-reasoning
Use migrate-reasoning to move assistant <think> content into reasoning_content fields for a session. This route does not need model credentials.
curl -fsS -X POST "http://localhost:8787/api/jobs/migrate-reasoning" \
-H "Content-Type: application/json" \
-d '{
"sessionId": "session_abc",
"limit": 200,
"offset": 0,
"concurrency": 5,
"sleepMs": 100,
"force": false
}'
Optional Maintenance Surfaces
- Use
GET /api/logs/statsfor log-level summaries before or after a curation pass.
curl -fsS "http://localhost:8787/api/logs/stats?sessionUid=session_abc"
- Use tag management when session organization is part of the curation flow:
GET /api/tags,POST /api/tags,DELETE /api/tags/:uid,GET /api/sessions/:sessionUid/tags,POST /api/sessions/:sessionUid/tags, andDELETE /api/sessions/:sessionUid/tags. - Use
POST /api/orphans/checkwhen you need a background scan for orphaned logs.
curl -fsS -X POST "http://localhost:8787/api/orphans/check"
Common Mistakes
- Starting a second job before checking whether an equivalent one is already running.
- Treating autoscore or rewrite like plaintext-key routes when they require encrypted credentials.
- Using remove-items destructively before a dry run or before capturing the removal criteria in notes.
- Claiming success from the initial
{ "jobId": ... }response without pollingGET /api/jobs/:id. - Keeping row-level reviewer work in HTTP-only mode instead of handing it to
synthlabs-ui-operator.