dongarandClaude Opus 5.5
dd060566b2
nisse-core records each model call of every agent into a file beside its state store (<state>.calls.db), for the status page (L7, X6): - the request body exactly as sent (never headers), the raw response with its reasoning, the model asked and the one that answered, time, tokens and llama-server's own prompt reuse (timings.cache_n); - its work item, role, run and purpose: Driver.run and nisse-dev's delegation loop open runs, nested; compaction, expert consults, the classifier and the go-ahead reader label their calls; - provenance, worked out in BudgetedClient by comparing the events at each stage: each event sent with its log id and content as sent, which a compaction replaced, which fitting shortened (by how much) and dropped. Content is stored once by hash, compressed. It is diagnostic only: the agent never reads it, a failed write is logged once and recording stops while the call goes on, and the file is deleted when the agent starts, so its layout never migrates; CallReader refuses a layout it doesn't know. While recording, the system prompt ends with one sentence saying so. `call_record: false` in either agent's YAML turns it off. The benchmark records each run beside that run's own state store, and `--keep-runs DIR` keeps both for the page. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J5Utyfzw7gF7UhjRDEtjTk
94 lines
3.4 KiB
Python
94 lines
3.4 KiB
Python
"""Reading a plain-language reply as a go-ahead (L5, D1).
|
|
|
|
"Yep, that's it" in reply to a restatement may be taken as `go`, because a
|
|
wrong reading costs only time. One short model call decides, and it decides
|
|
conservatively: only a clear yes counts. A question is never a go-ahead and
|
|
never reaches the model (L5: questions never change state).
|
|
|
|
The driver's model reads for now; the classifier role takes over in M2.6.
|
|
"""
|
|
|
|
from typing import Any
|
|
|
|
from nisse_core.log.canonical import StoredEvent, ToolCall, UserMessage
|
|
from nisse_core.models.calls import purpose
|
|
from nisse_core.models.llm import LLMClient, LLMError
|
|
|
|
|
|
def command_form(text: str) -> str:
|
|
"""How a message reads as a command: case, surrounding whitespace, runs
|
|
of inner whitespace and trailing '.' or '!' do not matter, so a phone's
|
|
"Go." is still a go-ahead (#64). Only for matching; what the operator
|
|
wrote is what gets recorded and passed on."""
|
|
return " ".join(text.strip().rstrip(".!").split()).lower()
|
|
|
|
|
|
VERDICT = {
|
|
"type": "function",
|
|
"function": {
|
|
"name": "verdict",
|
|
"description": "Your reading of the operator's reply.",
|
|
"parameters": {
|
|
"type": "object",
|
|
"properties": {
|
|
"answer": {
|
|
"type": "string",
|
|
"enum": ["yes", "no", "unclear"],
|
|
"description": "yes: a clear go-ahead on the proposal as it "
|
|
"stands. no: it asks for changes, disagrees or says to wait. "
|
|
"unclear: anything else.",
|
|
}
|
|
},
|
|
"required": ["answer"],
|
|
},
|
|
},
|
|
}
|
|
|
|
SYSTEM = """You read one reply from an operator to a software agent. The agent
|
|
put a proposal up for agreement; decide whether the reply clearly agrees to the
|
|
proposal as it stands, so the agent may start. A reply that asks for any
|
|
change, asks a question, adds a condition, or is about something else is not a
|
|
go-ahead. When in doubt, it is not. Answer with the verdict tool."""
|
|
|
|
|
|
def is_question(text: str) -> bool:
|
|
return text.rstrip().endswith("?")
|
|
|
|
|
|
class GoAheadReader:
|
|
def __init__(self, client: LLMClient, model: str) -> None:
|
|
self._client = client
|
|
self._model = model
|
|
|
|
def agrees(self, proposal: str, reply: str, model: str = "") -> bool:
|
|
"""``model``: the work item's pinned classifier (L9)."""
|
|
if is_question(reply):
|
|
return False
|
|
prompt = f"The proposal:\n\n{proposal}\n\nThe operator's reply:\n\n{reply}"
|
|
try:
|
|
with purpose("go-ahead reading", role="classifier"):
|
|
produced = self._client.complete(
|
|
[
|
|
StoredEvent(
|
|
id=0,
|
|
ts="",
|
|
event=UserMessage(text=prompt, source="go-ahead reader"),
|
|
actor="orchestrator",
|
|
)
|
|
],
|
|
tools=[VERDICT],
|
|
model=model or self._model,
|
|
system=SYSTEM,
|
|
)
|
|
except LLMError:
|
|
return False # unread: the driver reads it as an answer instead
|
|
return any(_yes(event) for event in produced)
|
|
|
|
|
|
def _yes(event: Any) -> bool:
|
|
return (
|
|
isinstance(event, ToolCall)
|
|
and event.tool == "verdict"
|
|
and event.arguments.get("answer") == "yes"
|
|
)
|