Daniel01410's picture
ctx_compress artifact with harness+escalation
770a262 verified
Raw History Blame Contribute Delete
3.12 kB
from __future__ import annotations
from pathlib import Path
from typing import Any
import httpx
from harness.adapters import Adapter
PROMPTS = Path(__file__).resolve().parent / "prompts"
TRACE_LIMIT = 1200
RELAY_TIMEOUT_S = 100.0
def _trim(text: str, limit: int) -> str:
return text if len(text) <= limit else text[-limit:]
def _tail(session: str, limit: int) -> str:
"""Keep the end of the session, where the current work is. Cheaper than a full redo."""
if limit <= 0 or len(session) <= limit:
return session
cut = session[-limit:]
start = cut.find("\n[")
return "<session>\n[… earlier steps omitted …]\n" + (cut[start + 1 :] if start >= 0 else cut)
def repair(
task: dict[str, Any],
draft: Any,
trace: str,
errors: str,
*,
adapter: Adapter,
relay_url: str,
ticket: str,
model: str | list[str],
tail_chars: int = 0,
max_tokens: int = 3072,
) -> Any:
"""Ask a frontier model through the relay to repair the draft.
Only the end of the session is sent when tail_chars is set, which keeps the
repair cheaper than a full frontier redo (the escalation score). Whatever
comes back goes through the same deterministic finish as the harness, so the
headings, facts and budget always hold. Any relay failure returns the draft.
"""
draft_text = adapter.as_text(draft)
finish = getattr(adapter, "finish", None)
session = str(task.get("user") or "")
view = dict(task)
view["user"] = _tail(session, tail_chars)
budget = max_tokens
if finish is not None:
from harness.adapters.context_compression import budget_of
budget = min(max_tokens, budget_of(task))
system = adapter.render(PROMPTS, "system", view)
user = adapter.render(
PROMPTS,
"repair",
view,
draft=draft_text,
errors=_trim(errors or "no verifier output", 400),
trace=_trim(trace or "", TRACE_LIMIT),
budget=str(budget),
)
models = [model] if isinstance(model, str) else list(model)
text = ""
for name in [m for m in models if m]:
body = {
"model": name,
"max_tokens": budget,
"temperature": 0,
"messages": [{"role": "system", "content": system}, {"role": "user", "content": user}],
}
try:
with httpx.Client(timeout=RELAY_TIMEOUT_S) as http:
response = http.post(
relay_url.rstrip("/") + "/v1/chat/completions",
json=body,
headers={"x-mt-ticket": ticket},
)
except Exception:
continue
if response.status_code != 200:
continue
try:
choices = response.json().get("choices") or [{}]
text = str((choices[0].get("message") or {}).get("content") or "")
except Exception:
text = ""
if text.strip():
break
if finish is not None:
answer, _note = finish(text or draft_text, task)
return answer
return adapter.clean(text) or draft