Threads, commands and labels on the new loop (M2.3)
CI / checks (push) Skipped
CI / checks (pull_request) Successful in 3m45s

- Commands that name their item ("go #14", "abandon #14", "stop #14")
  apply to it wherever they're said; go, approve and abandon alone at top
  level are asked about (L5). stop is never taken as an answer.
- A plain-language reply to a restatement is read by GoAheadReader, one
  model call that agrees only on a clear yes; questions never reach it. The
  agent says "taking that as a go-ahead on rev-N" first, bound to the
  revision on screen at arrival (L3, L5, D1).
- Labels steer (D1): `bug` goes ahead after tuning.loop.bug_grace_s
  (default 600 s) unless the operator replies; `discuss` isn't offered
  restate until the operator has spoken; `bug`, `discuss` and `design` add
  to the driver's instructions. A restatement asks no questions.

Tests: label and command unit tests, the property walk with labelled
issues, plain yeses, named gos and time passing (grace_only_unanswered,
discuss_talks_first), and e2e tests for a bug proceeding after its grace
period and a plain-language go-ahead.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPQ14L3LikyefCMFBqEbKE
This commit is contained in:
dongarandClaude Opus 5.5 committed 2026-10-04 16:13:22 +00:00
1 parent a504526da2
commit 4d9c1d7529
13 files changed
+793 -35

No files matched your search

+81
View File
@@ -0,0 +1,81 @@
"""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_dev.canonical import StoredEvent, ToolCall, UserMessage
from nisse_dev.llm import LLMClient, LLMError
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) -> bool:
if is_question(reply):
return False
prompt = f"The proposal:\n\n{proposal}\n\nThe operator's reply:\n\n{reply}"
try:
produced = self._client.complete(
[
StoredEvent(
id=0,
ts="",
event=UserMessage(text=prompt, source="go-ahead reader"),
actor="orchestrator",
)
],
tools=[VERDICT],
model=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"
)