@wendellresearch-ai/benchwarmers 0.2.0 → 0.2.1

package/suite/tasks/claim-list-body/fixtures/gateway.py

@@ -1,235 +1,235 @@
-"""The bench method as an OpenAI-compatible endpoint, at http://127.0.0.1:8123/v1.
-
-    python gateway.py
-
-Ask for the model `bench/panel` and the prompt goes to the panel `route.py`
-would seat — primary, second opinion, two explorers — in parallel. The reply
-is every answer, each labelled by seat and model with a call id, including the
-seats that failed or came back empty: the panel's make-up is evidence about
-the roster, and a summary naming only the good answer hides it.
-
-    bench/panel           Mercury Decide picks the task shape, else general
-    bench/panel-review    the shape given: review, code, design, enumeration, …
-
-A request body may add `"with": ["<alias>"]` to seat a model in an explorer
-chair, as `route.py --with` does.
-
-THE LOOP
---------
-Every seat is recorded in `metrics/offloads.jsonl` as an unjudged offload the
-moment it answers, so the explorer seats rotate even when nobody judges. The
-caller then checks the answers against the real code — agreement between seats
-is not a verdict, they fail together — and records one per call id:
-
-    python metrics/offload.py verdict <call id> --useful --note "..."
-    POST /bench/verdict {"call": "<call id>", "useful": true, "note": "..."}
-
-The verdict replaces the unjudged entry, and the next panel is seated from it.
-
-WHAT IT IS NOT
---------------
-A single-model proxy. Tools such as Roo Code call Open WebUI directly; this
-exists only for the multi-model work. Seats are called with `llm -m`, exactly
-as offloads always were, so endpoints, headers and per-model options all come
-from the same registry. Full prompts and answers go to `runs/gateway/`,
-gitignored. Bound to 127.0.0.1, like the dashboard.
-"""
-
-from __future__ import annotations
-
-import datetime as dt
-import json
-import os
-import pathlib
-import secrets
-import subprocess
-import sys
-import time
-from concurrent.futures import ThreadPoolExecutor
-from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
-
-ROOT = pathlib.Path(__file__).resolve().parent
-PORT = 8123
-LOG_DIR = ROOT / "runs" / "gateway"
-SEAT_TIMEOUT = 900
-sys.path.insert(0, str(ROOT / "metrics"))
-sys.path.insert(0, str(ROOT / "suite"))
-import decide  # noqa: E402
-import offload  # noqa: E402
-import route  # noqa: E402
-
-SEATS = [("primary", "primary"), ("second", "second opinion"),
-         ("explorer", "explorer"), ("explorer2", "explorer 2")]
-# What the shape picker is asked to choose between. Each option carries a
-# description because bare names mislead it: Mercury Decide called "list every
-# place serve.py writes a file" `code` at 0.93 when given names alone.
-SHAPES = {
-    "general": "A general question or task that fits none of the other kinds.",
-    "design": "Architecture, process or approach decisions: weighing options, an ADR, a plan.",
-    "prose": "Writing or editing documentation or other text.",
-    "review": "Judging whether given code is correct: finding bugs, or where a guarantee breaks.",
-    "code": "Writing new code or changing existing code.",
-    "enumeration": "Listing every place something happens in given code or files.",
-    "quick": "A short, bounded task: classify, extract or summarise a small input.",
-}
-CONFIDENT = 0.6
-LARGE = 8000          # characters; route.py keeps models that fail big prompts off big tasks
-
-
-def flatten(messages: list[dict]) -> tuple[str, str]:
-    """(system, prompt) for `llm`. A panel question is one prompt; a longer
-    conversation is kept as labelled turns rather than silently truncated."""
-    def text(m):
-        c = m.get("content") or ""
-        return c if isinstance(c, str) else "\n".join(
-            p.get("text", "") for p in c if isinstance(p, dict))
-    system = "\n\n".join(text(m) for m in messages if m.get("role") == "system")
-    turns = [m for m in messages if m.get("role") != "system"]
-    if len(turns) == 1:
-        return system, text(turns[0])
-    return system, "\n\n".join(f"{m.get('role', 'user')}: {text(m)}" for m in turns)
-
-
-def pick_shape(prompt: str) -> tuple[str, str]:
-    """(shape, how it was chosen). Falls back to general on any doubt."""
-    picker = next((a for a, d in decide.load().items() if "choice" in d["question_types"]), None)
-    if not picker:
-        return "general", "no decision model onboarded"
-    question = {"shape": {"type": "choice",
-                          "instructions": "What kind of task does this prompt ask for?",
-                          "criteria": SHAPES}}
-    try:
-        answer = decide.ask(picker, prompt[:20000], question)["answers"]["shape"]
-    except Exception as exc:                           # noqa: BLE001
-        return "general", f"{picker} failed ({decide.error_text(exc)[:80]})"
-    conf = answer.get("confidence", 0)
-    if answer.get("choice") not in SHAPES or conf < CONFIDENT:
-        return "general", f"{picker} unsure ({answer.get('choice')} at {conf:.2f})"
-    return answer["choice"], f"{picker}, {conf:.2f}"
-
-
-def ask_seat(alias: str, system: str, prompt: str) -> dict:
-    cmd = ["llm", "-m", alias] + (["-s", system] if system else [])
-    started = time.monotonic()
-    try:
-        done = subprocess.run(cmd, input=prompt, capture_output=True, text=True,
-                              encoding="utf-8", errors="replace", timeout=SEAT_TIMEOUT,
-                              env={**os.environ, "PYTHONIOENCODING": "utf-8"})
-        answer, error = done.stdout.strip(), (done.stderr.strip() if done.returncode else "")
-    except subprocess.TimeoutExpired:
-        answer, error = "", f"no answer in {SEAT_TIMEOUT}s"
-    status = "error" if error else ("empty" if not answer else "ok")
-    return {"status": status, "answer": answer, "error": error[-500:],
-            "seconds": round(time.monotonic() - started, 1)}
-
-
-def run_panel(model: str, messages: list[dict], nominated: list[str] | None = None) -> str:
-    system, prompt = flatten(messages)
-    if model == "bench/panel":
-        shape, how = pick_shape(prompt)
-    else:
-        shape, how = model.removeprefix("bench/panel-"), "asked for by name"
-    # `with` in the request body is route.py's --with: an explorer seat, never
-    # the primary or second.
-    p = route.panel(shape, large=len(prompt) > LARGE, nominated=nominated or None)
-    seats = [(key, label, p[key]["model"]) for key, label in SEATS if p.get(key)]
-    if not seats:
-        raise LookupError(f"route.py seated nobody for shape {shape}")
-    panel_id = f"p{dt.datetime.now():%Y%m%d-%H%M%S}-{secrets.token_hex(2)}"
-    with ThreadPoolExecutor(max_workers=len(seats)) as pool:
-        results = list(pool.map(lambda s: ask_seat(s[2], system, prompt), seats))
-
-    task = " ".join(prompt.split())[:90] + f" (bench/panel, {shape})"
-    LOG_DIR.mkdir(parents=True, exist_ok=True)
-    record = {"panel": panel_id, "at": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
-              "model": model, "shape": shape, "shape_by": how, "system": system,
-              "prompt": prompt, "seats": []}
-    out = [f"Bench panel {panel_id} · shape {shape} ({how}) · {len(seats)} seats\n"]
-    for (key, label, alias), r in zip(seats, results):
-        call = f"{panel_id}.{key}"
-        offload.add(alias, task, None, f"{label}; {r['status']} in {r['seconds']}s"
-                    + (f": {r['error'][:120]}" if r["error"] else ""), call)
-        record["seats"].append({"call": call, "seat": label, "alias": alias, **r})
-        body = r["answer"] or (f"_error: {r['error'][:300]}_" if r["error"] else
-                               "_empty reply — usually a token budget, not incapacity_")
-        out.append(f"## {label} — {alias}\n`{call}` · {r['status']} · {r['seconds']}s\n\n{body}\n")
-    with (LOG_DIR / "calls.jsonl").open("a", encoding="utf-8") as fh:
-        fh.write(json.dumps(record, ensure_ascii=False) + "\n")
-    out.append("---\nRecord a verdict per seat once checked against the code:\n"
-               "`python metrics/offload.py verdict <call id> --useful|--not-useful --note \"...\"`")
-    return "\n".join(out)
-
-
-def models() -> list[str]:
-    return ["bench/panel"] + [f"bench/panel-{s}" for s in SHAPES]
-
-
-class Handler(BaseHTTPRequestHandler):
-    def _json(self, code: int, data: dict) -> None:
-        body = json.dumps(data).encode()
-        self.send_response(code)
-        self.send_header("Content-Type", "application/json")
-        self.send_header("Content-Length", str(len(body)))
-        self.end_headers()
-        self.wfile.write(body)
-
-    def _error(self, code: int, message: str) -> None:
-        self._json(code, {"error": {"message": message, "type": "bench_gateway", "code": code}})
-
-    def do_GET(self):                                  # noqa: N802
-        if self.path.rstrip("/") in ("/v1/models", "/models"):
-            self._json(200, {"object": "list", "data": [
-                {"id": m, "object": "model", "owned_by": "model-bench"} for m in models()]})
-        else:
-            self._error(404, "GET /v1/models is the only GET here")
-
-    def do_POST(self):                                 # noqa: N802
-        try:
-            body = json.loads(self.rfile.read(int(self.headers.get("Content-Length") or 0)) or b"{}")
-        except ValueError:
-            return self._error(400, "body is not JSON")
-        path = self.path.rstrip("/")
-        if path == "/bench/verdict":
-            try:
-                e = offload.verdict(str(body.get("call", "")), body.get("useful"), body.get("note", ""))
-            except KeyError as exc:
-                return self._error(404, exc.args[0])
-            return self._json(200, {"ok": True, "recorded": e})
-        if path not in ("/v1/chat/completions", "/chat/completions"):
-            return self._error(404, "POST /v1/chat/completions or /bench/verdict")
-        model = body.get("model", "")
-        if model not in models():
-            return self._error(404, f"{model!r} is not served here; this endpoint only runs"
-                                    f" panels: {', '.join(models())}")
-        try:
-            content = run_panel(model, body.get("messages") or [], body.get("with"))
-        except LookupError as exc:
-            return self._error(503, str(exc))
-        reply_id, created = f"chatcmpl-{secrets.token_hex(8)}", int(time.time())
-        if not body.get("stream"):
-            return self._json(200, {
-                "id": reply_id, "object": "chat.completion", "created": created, "model": model,
-                "choices": [{"index": 0, "finish_reason": "stop",
-                             "message": {"role": "assistant", "content": content}}]})
-        # A panel finishes all at once, so a streaming client gets it as one chunk.
-        self.send_response(200)
-        self.send_header("Content-Type", "text/event-stream")
-        self.end_headers()
-        for delta, finish in (({"role": "assistant", "content": content}, None), ({}, "stop")):
-            chunk = {"id": reply_id, "object": "chat.completion.chunk", "created": created,
-                     "model": model, "choices": [{"index": 0, "delta": delta, "finish_reason": finish}]}
-            self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
-        self.wfile.write(b"data: [DONE]\n\n")
-
-    def log_message(self, fmt, *args):                 # quieter than the default
-        sys.stderr.write(f"{self.log_date_time_string()} {fmt % args}\n")
-
-
-def main() -> None:
-    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
-    print(f"bench gateway on http://127.0.0.1:{PORT}/v1 — models: {', '.join(models())}")
-    ThreadingHTTPServer(("127.0.0.1", PORT), Handler).serve_forever()
-
-
-if __name__ == "__main__":
-    main()
+"""The bench method as an OpenAI-compatible endpoint, at http://127.0.0.1:8123/v1.
+
+    python gateway.py
+
+Ask for the model `bench/panel` and the prompt goes to the panel `route.py`
+would seat — primary, second opinion, two explorers — in parallel. The reply
+is every answer, each labelled by seat and model with a call id, including the
+seats that failed or came back empty: the panel's make-up is evidence about
+the roster, and a summary naming only the good answer hides it.
+
+    bench/panel           Mercury Decide picks the task shape, else general
+    bench/panel-review    the shape given: review, code, design, enumeration, …
+
+A request body may add `"with": ["<alias>"]` to seat a model in an explorer
+chair, as `route.py --with` does.
+
+THE LOOP
+--------
+Every seat is recorded in `metrics/offloads.jsonl` as an unjudged offload the
+moment it answers, so the explorer seats rotate even when nobody judges. The
+caller then checks the answers against the real code — agreement between seats
+is not a verdict, they fail together — and records one per call id:
+
+    python metrics/offload.py verdict <call id> --useful --note "..."
+    POST /bench/verdict {"call": "<call id>", "useful": true, "note": "..."}
+
+The verdict replaces the unjudged entry, and the next panel is seated from it.
+
+WHAT IT IS NOT
+--------------
+A single-model proxy. Tools such as Roo Code call Open WebUI directly; this
+exists only for the multi-model work. Seats are called with `llm -m`, exactly
+as offloads always were, so endpoints, headers and per-model options all come
+from the same registry. Full prompts and answers go to `runs/gateway/`,
+gitignored. Bound to 127.0.0.1, like the dashboard.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import json
+import os
+import pathlib
+import secrets
+import subprocess
+import sys
+import time
+from concurrent.futures import ThreadPoolExecutor
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+ROOT = pathlib.Path(__file__).resolve().parent
+PORT = 8123
+LOG_DIR = ROOT / "runs" / "gateway"
+SEAT_TIMEOUT = 900
+sys.path.insert(0, str(ROOT / "metrics"))
+sys.path.insert(0, str(ROOT / "suite"))
+import decide  # noqa: E402
+import offload  # noqa: E402
+import route  # noqa: E402
+
+SEATS = [("primary", "primary"), ("second", "second opinion"),
+         ("explorer", "explorer"), ("explorer2", "explorer 2")]
+# What the shape picker is asked to choose between. Each option carries a
+# description because bare names mislead it: Mercury Decide called "list every
+# place serve.py writes a file" `code` at 0.93 when given names alone.
+SHAPES = {
+    "general": "A general question or task that fits none of the other kinds.",
+    "design": "Architecture, process or approach decisions: weighing options, an ADR, a plan.",
+    "prose": "Writing or editing documentation or other text.",
+    "review": "Judging whether given code is correct: finding bugs, or where a guarantee breaks.",
+    "code": "Writing new code or changing existing code.",
+    "enumeration": "Listing every place something happens in given code or files.",
+    "quick": "A short, bounded task: classify, extract or summarise a small input.",
+}
+CONFIDENT = 0.6
+LARGE = 8000          # characters; route.py keeps models that fail big prompts off big tasks
+
+
+def flatten(messages: list[dict]) -> tuple[str, str]:
+    """(system, prompt) for `llm`. A panel question is one prompt; a longer
+    conversation is kept as labelled turns rather than silently truncated."""
+    def text(m):
+        c = m.get("content") or ""
+        return c if isinstance(c, str) else "\n".join(
+            p.get("text", "") for p in c if isinstance(p, dict))
+    system = "\n\n".join(text(m) for m in messages if m.get("role") == "system")
+    turns = [m for m in messages if m.get("role") != "system"]
+    if len(turns) == 1:
+        return system, text(turns[0])
+    return system, "\n\n".join(f"{m.get('role', 'user')}: {text(m)}" for m in turns)
+
+
+def pick_shape(prompt: str) -> tuple[str, str]:
+    """(shape, how it was chosen). Falls back to general on any doubt."""
+    picker = next((a for a, d in decide.load().items() if "choice" in d["question_types"]), None)
+    if not picker:
+        return "general", "no decision model onboarded"
+    question = {"shape": {"type": "choice",
+                          "instructions": "What kind of task does this prompt ask for?",
+                          "criteria": SHAPES}}
+    try:
+        answer = decide.ask(picker, prompt[:20000], question)["answers"]["shape"]
+    except Exception as exc:                           # noqa: BLE001
+        return "general", f"{picker} failed ({decide.error_text(exc)[:80]})"
+    conf = answer.get("confidence", 0)
+    if answer.get("choice") not in SHAPES or conf < CONFIDENT:
+        return "general", f"{picker} unsure ({answer.get('choice')} at {conf:.2f})"
+    return answer["choice"], f"{picker}, {conf:.2f}"
+
+
+def ask_seat(alias: str, system: str, prompt: str) -> dict:
+    cmd = ["llm", "-m", alias] + (["-s", system] if system else [])
+    started = time.monotonic()
+    try:
+        done = subprocess.run(cmd, input=prompt, capture_output=True, text=True,
+                              encoding="utf-8", errors="replace", timeout=SEAT_TIMEOUT,
+                              env={**os.environ, "PYTHONIOENCODING": "utf-8"})
+        answer, error = done.stdout.strip(), (done.stderr.strip() if done.returncode else "")
+    except subprocess.TimeoutExpired:
+        answer, error = "", f"no answer in {SEAT_TIMEOUT}s"
+    status = "error" if error else ("empty" if not answer else "ok")
+    return {"status": status, "answer": answer, "error": error[-500:],
+            "seconds": round(time.monotonic() - started, 1)}
+
+
+def run_panel(model: str, messages: list[dict], nominated: list[str] | None = None) -> str:
+    system, prompt = flatten(messages)
+    if model == "bench/panel":
+        shape, how = pick_shape(prompt)
+    else:
+        shape, how = model.removeprefix("bench/panel-"), "asked for by name"
+    # `with` in the request body is route.py's --with: an explorer seat, never
+    # the primary or second.
+    p = route.panel(shape, large=len(prompt) > LARGE, nominated=nominated or None)
+    seats = [(key, label, p[key]["model"]) for key, label in SEATS if p.get(key)]
+    if not seats:
+        raise LookupError(f"route.py seated nobody for shape {shape}")
+    panel_id = f"p{dt.datetime.now():%Y%m%d-%H%M%S}-{secrets.token_hex(2)}"
+    with ThreadPoolExecutor(max_workers=len(seats)) as pool:
+        results = list(pool.map(lambda s: ask_seat(s[2], system, prompt), seats))
+
+    task = " ".join(prompt.split())[:90] + f" (bench/panel, {shape})"
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    record = {"panel": panel_id, "at": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
+              "model": model, "shape": shape, "shape_by": how, "system": system,
+              "prompt": prompt, "seats": []}
+    out = [f"Bench panel {panel_id} · shape {shape} ({how}) · {len(seats)} seats\n"]
+    for (key, label, alias), r in zip(seats, results):
+        call = f"{panel_id}.{key}"
+        offload.add(alias, task, None, f"{label}; {r['status']} in {r['seconds']}s"
+                    + (f": {r['error'][:120]}" if r["error"] else ""), call)
+        record["seats"].append({"call": call, "seat": label, "alias": alias, **r})
+        body = r["answer"] or (f"_error: {r['error'][:300]}_" if r["error"] else
+                               "_empty reply — usually a token budget, not incapacity_")
+        out.append(f"## {label} — {alias}\n`{call}` · {r['status']} · {r['seconds']}s\n\n{body}\n")
+    with (LOG_DIR / "calls.jsonl").open("a", encoding="utf-8") as fh:
+        fh.write(json.dumps(record, ensure_ascii=False) + "\n")
+    out.append("---\nRecord a verdict per seat once checked against the code:\n"
+               "`python metrics/offload.py verdict <call id> --useful|--not-useful --note \"...\"`")
+    return "\n".join(out)
+
+
+def models() -> list[str]:
+    return ["bench/panel"] + [f"bench/panel-{s}" for s in SHAPES]
+
+
+class Handler(BaseHTTPRequestHandler):
+    def _json(self, code: int, data: dict) -> None:
+        body = json.dumps(data).encode()
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def _error(self, code: int, message: str) -> None:
+        self._json(code, {"error": {"message": message, "type": "bench_gateway", "code": code}})
+
+    def do_GET(self):                                  # noqa: N802
+        if self.path.rstrip("/") in ("/v1/models", "/models"):
+            self._json(200, {"object": "list", "data": [
+                {"id": m, "object": "model", "owned_by": "model-bench"} for m in models()]})
+        else:
+            self._error(404, "GET /v1/models is the only GET here")
+
+    def do_POST(self):                                 # noqa: N802
+        try:
+            body = json.loads(self.rfile.read(int(self.headers.get("Content-Length") or 0)) or b"{}")
+        except ValueError:
+            return self._error(400, "body is not JSON")
+        path = self.path.rstrip("/")
+        if path == "/bench/verdict":
+            try:
+                e = offload.verdict(str(body.get("call", "")), body.get("useful"), body.get("note", ""))
+            except KeyError as exc:
+                return self._error(404, exc.args[0])
+            return self._json(200, {"ok": True, "recorded": e})
+        if path not in ("/v1/chat/completions", "/chat/completions"):
+            return self._error(404, "POST /v1/chat/completions or /bench/verdict")
+        model = body.get("model", "")
+        if model not in models():
+            return self._error(404, f"{model!r} is not served here; this endpoint only runs"
+                                    f" panels: {', '.join(models())}")
+        try:
+            content = run_panel(model, body.get("messages") or [], body.get("with"))
+        except LookupError as exc:
+            return self._error(503, str(exc))
+        reply_id, created = f"chatcmpl-{secrets.token_hex(8)}", int(time.time())
+        if not body.get("stream"):
+            return self._json(200, {
+                "id": reply_id, "object": "chat.completion", "created": created, "model": model,
+                "choices": [{"index": 0, "finish_reason": "stop",
+                             "message": {"role": "assistant", "content": content}}]})
+        # A panel finishes all at once, so a streaming client gets it as one chunk.
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        for delta, finish in (({"role": "assistant", "content": content}, None), ({}, "stop")):
+            chunk = {"id": reply_id, "object": "chat.completion.chunk", "created": created,
+                     "model": model, "choices": [{"index": 0, "delta": delta, "finish_reason": finish}]}
+            self.wfile.write(f"data: {json.dumps(chunk)}\n\n".encode())
+        self.wfile.write(b"data: [DONE]\n\n")
+
+    def log_message(self, fmt, *args):                 # quieter than the default
+        sys.stderr.write(f"{self.log_date_time_string()} {fmt % args}\n")
+
+
+def main() -> None:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    print(f"bench gateway on http://127.0.0.1:{PORT}/v1 — models: {', '.join(models())}")
+    ThreadingHTTPServer(("127.0.0.1", PORT), Handler).serve_forever()
+
+
+if __name__ == "__main__":
+    main()