jianzhichun/emerge/emerge-forward-synthesis

Emerge Forward Synthesis

Purpose

Turn successful icc_exec WAL samples into one deterministic pipeline candidate. The Python daemon is only a coordinator; you are the lead agent doing distillation.

Inputs

Use the job payload from forward_synthesis_pending:

• normalized_intent: target connector.mode.pipeline_name
• samples: successful exec samples with code, args, optional result
• connector NOTES.md and synthesis_hints.yaml when included

Rules

1. Compare all samples. Literals that vary with sample args should become __args["name"].
2. Be conservative. If unsure whether a literal is a parameter, keep it constant.
3. Remove dead code: no print, debug logging, timing noise, or exploratory branches.
4. Do not introduce new dependencies unless connector notes already require them.
5. For read mode, assign __result. For write mode, assign __action.
6. Generate a stronger verify strategy. Prefer required_fields, type checks, and row-shape checks over bool(rows).
7. Validate the candidate through icc_exec with no_replay=true, then write the pending artifact yourself using the normal file-editing tools.

Output JSON

Use this shape in your working notes and pending artifact rationale:

{
  "connector": "mock",
  "mode": "read",
  "pipeline_name": "sheet",
  "code": "__result = [{'file': __args['filename']}]",
  "confidence": 0.91,
  "rationale": "filename varied across samples, so it was parameterized; sheet name was constant.",
  "verify_strategy": {
    "required_fields": ["file"]
  }
}

Few-Shot Patterns

Varying file path

Samples use /tmp/a.xlsx, /tmp/b.xlsx, /tmp/c.xlsx and each sample args has filename.

Output code should use __args["filename"].

Constant file path

All samples use /tmp/q3.xlsx and there is no input arg explaining it.

Keep /tmp/q3.xlsx literal. 保守优先: do not invent __args["filename"].

Required fields

If every successful __result row has name and area, include:

{"verify_strategy": {"required_fields": ["name", "area"]}}