nabla-cli 0.6.4__tar.gz → 0.7.0__tar.gz

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nabla-cli
-Version: 0.6.4
+Version: 0.7.0
 Summary: Call-site harvester: scan OpenAI call sites, record real traffic, emit site profiles
 License: MIT
 Requires-Python: >=3.11
@@ -162,6 +162,23 @@ returns a goal description, implicit constraints, and difficulty
 dimensions with cited evidence. That's the goal line printed above, and
 it's what lands in the profile's `goal` block.
 
+### `nabla jeremy`
+
+An alternative to `build` with the same inputs and site handling: instead
+of the full site profile, it emits `<site>.jeremy.json` — a seed spec for
+a problem generator:
+
+```json
+{
+  "goal": "You are generating new problems to test and eventually fine-tune a small student model on the following task: ...",
+  "system_prompt": "<the site's own system prompt>",
+  "examples": [ {"problem": "...", "solution": "..."}, ... ]
+}
+```
+
+The goal text is composed deterministically from the same inferred goal
+and constraints that `build` uses; the examples are the captured pairs.
+
 ## Providers
 
 `nabla init` supports three providers. Pick whichever you have a key

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/README.md

@@ -148,6 +148,23 @@ returns a goal description, implicit constraints, and difficulty
 dimensions with cited evidence. That's the goal line printed above, and
 it's what lands in the profile's `goal` block.
 
+### `nabla jeremy`
+
+An alternative to `build` with the same inputs and site handling: instead
+of the full site profile, it emits `<site>.jeremy.json` — a seed spec for
+a problem generator:
+
+```json
+{
+  "goal": "You are generating new problems to test and eventually fine-tune a small student model on the following task: ...",
+  "system_prompt": "<the site's own system prompt>",
+  "examples": [ {"problem": "...", "solution": "..."}, ... ]
+}
+```
+
+The goal text is composed deterministically from the same inferred goal
+and constraints that `build` uses; the examples are the captured pairs.
+
 ## Providers
 
 `nabla init` supports three providers. Pick whichever you have a key

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/nabla/__init__.py

@@ -1,3 +1,3 @@
 """nabla — call-site harvester CLI (pre-Phase-0)."""
 
-__version__ = "0.6.4"
+__version__ = "0.7.0"

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/nabla/cli.py

@@ -132,6 +132,29 @@ def _build_parser() -> argparse.ArgumentParser:
     p_profile.add_argument("--out", default=None,
                            help="output path (default: <site>.profile.json)")
 
+    p_jeremy = sub.add_parser(
+        "jeremy",
+        help="assemble a seed spec (goal + system_prompt + problem/solution examples)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=(
+            "an alternative to `build`: same inputs, but emits <site>.jeremy.json —\n"
+            "a generator-framed goal, the site's system prompt, and the captured\n"
+            "pairs as {problem, solution} reference examples.\n\n"
+            "examples:\n"
+            "  nabla jeremy                       auto: pick site, find <site>.captured.jsonl\n"
+            "  nabla jeremy path/to/repo --site extract_invoice --captured captured.jsonl\n"
+        ),
+    )
+    p_jeremy.add_argument("repo", nargs="?", default=".", help="path to target repo (default: current directory)")
+    p_jeremy.add_argument(
+        "--site",
+        help="site symbol (auto-picked when omitted and the repo has exactly one supported call site)",
+    )
+    p_jeremy.add_argument("--captured", "--examples", dest="examples",
+                          help="captured JSONL from `nabla capture` (default: <site>.captured.jsonl)")
+    p_jeremy.add_argument("--out", default=None,
+                          help="output path (default: <site>.jeremy.json)")
+
     return parser
 
 
@@ -193,6 +216,7 @@ def main(argv: list[str] | None = None) -> int:
         "record": _cmd_record,    # legacy alias
         "build": _cmd_profile,
         "profile": _cmd_profile,  # legacy alias
+        "jeremy": lambda a: _cmd_profile(a, mode="jeremy"),
     }
     try:
         return commands[args.command](args)
@@ -778,21 +802,23 @@ def _find_captured_file(site) -> Path | None:
     )
 
 
-def _cmd_profile(args) -> int:
+def _cmd_profile(args, mode: str = "profile") -> int:
     from .ui import rail, rail_end, rail_start
 
-    rail_start(_out, "nabla build")
+    command = "jeremy" if mode == "jeremy" else "build"
+    noun = "seed spec(s)" if mode == "jeremy" else "profile(s)"
+    rail_start(_out, f"nabla {command}")
     sites = scan(args.repo)
 
-    # Multi-site: no --site/--captured, several supported sites — build a
-    # profile for every site that has captured pairs on disk.
+    # Multi-site: no --site/--captured, several supported sites — build the
+    # artifact for every site that has captured pairs on disk.
     if args.site is None and args.examples is None:
         supported = [s for s in sites
                      if is_supported(resolve_schema(s, args.repo)[1], s.flags)]
         if len(supported) > 1:
             with_captures = [s for s in supported if _find_captured_file(s) is not None]
             if not with_captures:
-                print(f"nabla build: no captured pairs found for any of the "
+                print(f"nabla {command}: no captured pairs found for any of the "
                       f"{len(supported)} supported sites. Run first: nabla capture {args.repo}",
                       file=sys.stderr)
                 return 2
@@ -801,11 +827,11 @@ def _cmd_profile(args) -> int:
             for i, site in enumerate(with_captures):
                 if i:
                     rail(_out)
-                rc = _build_one(args, site) or rc
+                rc = _build_one(args, site, mode=mode) or rc
             if skipped:
                 rail(_out, f"skipped (no captured pairs): {', '.join(skipped)}", style="yellow")
             rail(_out)
-            rail_end(_out, f"done — built {len(with_captures)} profile(s)", style="green")
+            rail_end(_out, f"done — built {len(with_captures)} {noun}", style="green")
             return rc
 
     site, err = _resolve_site(args, sites)
@@ -813,16 +839,32 @@ def _cmd_profile(args) -> int:
         return err
     if args.site:
         rail(_out, f"site {site.symbol} · {site.file}")
-    rc = _build_one(args, site, final_rail=True)
+    rc = _build_one(args, site, final_rail=True, mode=mode)
     return rc
 
 
-def _build_one(args, site, final_rail: bool = False) -> int:
+def _compose_seed_goal(goal) -> str:
+    """Generator-framed goal text for the seed spec, composed in pure Python
+    from the induced GoalBlock — no extra LLM call, no invented facts."""
+    parts = [
+        "You are generating new problems to test and eventually fine-tune a "
+        "small student model on the following task: " + goal.description.rstrip(".") + "."
+    ]
+    if goal.constraints:
+        parts.append("Every generated problem's solution must satisfy: "
+                     + " ".join(c.rstrip(".") + "." for c in goal.constraints))
+    parts.append("Match the style, difficulty range, and output format of the "
+                 "reference examples below.")
+    return " ".join(parts)
+
+
+def _build_one(args, site, final_rail: bool = False, mode: str = "profile") -> int:
     from .induce import induce
     from .prompt_recon import reconstruct
     from .schema_resolver import resolve_schema as _resolve
     from .ui import rail, rail_end, rail_ok
 
+    command = "jeremy" if mode == "jeremy" else "build"
     examples_path = args.examples
     if examples_path is None:
         found = _find_captured_file(site) or (
@@ -831,7 +873,7 @@ def _build_one(args, site, final_rail: bool = False) -> int:
             capture_cmd = f"nabla capture {args.repo}"
             if args.site:
                 capture_cmd += f" --site {site.symbol}"
-            print(f"nabla build: no captured pairs found ({site.symbol}.captured.jsonl). "
+            print(f"nabla {command}: no captured pairs found ({site.symbol}.captured.jsonl). "
                   f"Run first: {capture_cmd}", file=sys.stderr)
             return 2
         examples_path = str(found)
@@ -857,21 +899,37 @@ def _build_one(args, site, final_rail: bool = False) -> int:
     else:
         goal = induce(template, schema, examples)
 
-    profile = build_profile(site, args.repo, examples, goal, recorded_from=recorded_from)
-    out = (args.out if final_rail else None) or f"{site.symbol}.profile.json"
-    Path(out).write_text(json.dumps(profile, indent=2), encoding="utf-8")
-    supported = profile["classification"]["supported"]
-    rail(_out)
-    rail_ok(_out, f"wrote {out}")
-    rail(_out, f"site         {site.symbol} · {profile['site_id'][:16]}…")
-    rail(_out, f"supported    {supported}")
-    rail(_out, f"examples     {len(examples)} ({recorded_from})")
-    rail(_out, f"goal         {goal.description[:90]}")
-    dims = ", ".join(d["name"] for d in goal.difficulty_dimensions[:4])
-    rail(_out, f"difficulty   {dims}")
+    if mode == "jeremy":
+        seed = {
+            "goal": _compose_seed_goal(goal),
+            "system_prompt": template.system_prompt or "",
+            "examples": [
+                {"problem": e.prompt, "solution": e.completion} for e in examples
+            ],
+        }
+        out = (args.out if final_rail else None) or f"{site.symbol}.jeremy.json"
+        Path(out).write_text(json.dumps(seed, indent=2), encoding="utf-8")
+        rail(_out)
+        rail_ok(_out, f"wrote {out}")
+        rail(_out, f"site         {site.symbol}")
+        rail(_out, f"examples     {len(examples)} ({recorded_from})")
+        rail(_out, f"goal         {seed['goal'][:90]}")
+    else:
+        profile = build_profile(site, args.repo, examples, goal, recorded_from=recorded_from)
+        out = (args.out if final_rail else None) or f"{site.symbol}.profile.json"
+        Path(out).write_text(json.dumps(profile, indent=2), encoding="utf-8")
+        supported = profile["classification"]["supported"]
+        rail(_out)
+        rail_ok(_out, f"wrote {out}")
+        rail(_out, f"site         {site.symbol} · {profile['site_id'][:16]}…")
+        rail(_out, f"supported    {supported}")
+        rail(_out, f"examples     {len(examples)} ({recorded_from})")
+        rail(_out, f"goal         {goal.description[:90]}")
+        dims = ", ".join(d["name"] for d in goal.difficulty_dimensions[:4])
+        rail(_out, f"difficulty   {dims}")
     if len(examples) < 20:
         rail(_out, f"only {len(examples)} examples (< 20) — enough for a look, but a "
-                   "handoff-quality profile wants 20+ real-traffic pairs", style="yellow")
+                   "handoff-quality artifact wants 20+ real-traffic pairs", style="yellow")
     if final_rail:
         rail(_out)
         rail_end(_out, f"done — {out} is the handoff artifact", style="green")

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/nabla_cli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: nabla-cli
-Version: 0.6.4
+Version: 0.7.0
 Summary: Call-site harvester: scan OpenAI call sites, record real traffic, emit site profiles
 License: MIT
 Requires-Python: >=3.11
@@ -162,6 +162,23 @@ returns a goal description, implicit constraints, and difficulty
 dimensions with cited evidence. That's the goal line printed above, and
 it's what lands in the profile's `goal` block.
 
+### `nabla jeremy`
+
+An alternative to `build` with the same inputs and site handling: instead
+of the full site profile, it emits `<site>.jeremy.json` — a seed spec for
+a problem generator:
+
+```json
+{
+  "goal": "You are generating new problems to test and eventually fine-tune a small student model on the following task: ...",
+  "system_prompt": "<the site's own system prompt>",
+  "examples": [ {"problem": "...", "solution": "..."}, ... ]
+}
+```
+
+The goal text is composed deterministically from the same inferred goal
+and constraints that `build` uses; the examples are the captured pairs.
+
 ## Providers
 
 `nabla init` supports three providers. Pick whichever you have a key

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "nabla-cli"
-version = "0.6.4"
+version = "0.7.0"
 description = "Call-site harvester: scan OpenAI call sites, record real traffic, emit site profiles"
 readme = "README.md"
 license = { text = "MIT" }

{nabla_cli-0.6.4 → nabla_cli-0.7.0}/tests/test_cli_ux.py

@@ -517,3 +517,72 @@ def test_scan_skips_build_and_dist_dirs(tmp_path, monkeypatch, capsys):
     out = capsys.readouterr().out
     assert "1 call site found" in out
     assert "copy_of_site" not in out
+
+
+# ---------------------------------------------------------------------------
+# nabla jeremy — seed-spec sibling of build
+# ---------------------------------------------------------------------------
+
+def _fake_goal(*a, **k):
+    from nabla.contract import GoalBlock
+
+    return GoalBlock(
+        description="Solve a grade-school word problem step by step.",
+        constraints=["final answer is a bare number", "no invented constants"],
+        difficulty_dimensions=[{"name": "d", "levels": ["a", "b"], "evidence": "e"}],
+    )
+
+
+def test_jeremy_single_site_writes_seed_spec(tmp_path, monkeypatch, capsys):
+    repo = _make_single_site_repo(tmp_path)
+    monkeypatch.chdir(tmp_path)
+    pair = {"input_slots": {"ocr_text": "r"}, "prompt": "A farmer has 6 crates...",
+            "completion": "The answer is 57.",
+            "tokens": {"in": 5, "out": 5}, "latency_ms": 10, "_source_kind": "samples"}
+    (tmp_path / "extract_receipt.captured.jsonl").write_text(
+        json.dumps(pair) + "\n", encoding="utf-8")
+    monkeypatch.setattr("nabla.induce.induce", _fake_goal)
+
+    rc = main(["jeremy", str(repo)])
+
+    assert rc == 0
+    seed = json.loads((tmp_path / "extract_receipt.jeremy.json").read_text(encoding="utf-8"))
+    assert set(seed) == {"goal", "system_prompt", "examples"}
+    assert seed["goal"].startswith("You are generating new problems")
+    assert "grade-school word problem" in seed["goal"]
+    assert "final answer is a bare number" in seed["goal"]
+    assert "Extract the receipt fields" in seed["system_prompt"]
+    assert seed["examples"] == [
+        {"problem": "A farmer has 6 crates...", "solution": "The answer is 57."}
+    ]
+    assert "extract_receipt.jeremy.json" in capsys.readouterr().out
+
+
+def test_jeremy_multi_site(tmp_path, monkeypatch, capsys):
+    repo = _make_two_site_repo(tmp_path)
+    monkeypatch.chdir(tmp_path)
+    pair = {"input_slots": {}, "prompt": "p", "completion": "s",
+            "tokens": {"in": 1, "out": 1}, "latency_ms": 1, "_source_kind": "samples"}
+    for symbol in ("extract_receipt", "extract_menu"):
+        (tmp_path / f"{symbol}.captured.jsonl").write_text(
+            json.dumps(pair) + "\n", encoding="utf-8")
+    monkeypatch.setattr("nabla.induce.induce", _fake_goal)
+
+    rc = main(["jeremy", str(repo)])
+
+    assert rc == 0
+    assert (tmp_path / "extract_receipt.jeremy.json").exists()
+    assert (tmp_path / "extract_menu.jeremy.json").exists()
+    assert "built 2 seed spec(s)" in capsys.readouterr().out
+
+
+def test_jeremy_without_captures_points_at_capture(tmp_path, monkeypatch, capsys):
+    repo = _make_single_site_repo(tmp_path)
+    monkeypatch.chdir(tmp_path)
+
+    rc = main(["jeremy", str(repo)])
+
+    assert rc == 2
+    err = capsys.readouterr().err
+    assert "nabla jeremy" in err
+    assert "nabla capture" in err