mirror of
https://github.com/magnus919/agent-skills.git
synced 2026-09-16 14:06:27 +03:00
Magnus Hedemark
3a4aad40bf
Spawns a panel of expert agents to debate any question with convergence-aware iteration and typed synthesis output. - PydanticAI + PydanticGraph-based Python package - 5-phase debate protocol: compose → premortem → position → cross-examine (iterative, eval-driven) → synthesis - Convergence detection: confidence dispersion, argument novelty, concession rate — stops when diminishing returns - Typed output schemas (Pydantic models) for every phase - CLI tool with markdown and JSON output modes - Custom persona file support - Bootstrap detection: sys.executable -m pip install fallback - agentskills.io compatible SKILL.md with triggers Signed-off-by: Magnus Hedemark <magnus919@pm.me>
113 lines
3.7 KiB
Python
113 lines
3.7 KiB
Python
"""Convergence detection — evaluates debate state to decide when to stop."""
|
|
|
|
import math
|
|
from agent_council.state import CouncilState, RoundMetrics
|
|
|
|
|
|
def compute_round_metrics(state: CouncilState) -> RoundMetrics:
|
|
"""Compute convergence metrics from the current round's data."""
|
|
if not state.cross_examination_rounds:
|
|
return RoundMetrics(
|
|
round=state.round_number,
|
|
mean_confidence=0.0,
|
|
dispersion=0.0,
|
|
new_arguments=0,
|
|
concessions_made=0,
|
|
)
|
|
|
|
current_round = state.cross_examination_rounds[-1]
|
|
confidences = []
|
|
concessions = 0
|
|
total_arguments_before = set()
|
|
|
|
# Count arguments from all prior rounds for novelty detection
|
|
for r in state.cross_examination_rounds[:-1]:
|
|
for ce in r.values():
|
|
if ce.remaining_disagreements:
|
|
total_arguments_before.update(ce.remaining_disagreements)
|
|
if ce.new_evidence_needed:
|
|
total_arguments_before.update(ce.new_evidence_needed)
|
|
|
|
new_arguments = 0
|
|
for ce in current_round.values():
|
|
if ce.updated_confidence is not None:
|
|
confidences.append(ce.updated_confidence)
|
|
if ce.concessions:
|
|
concessions += len(ce.concessions)
|
|
if ce.remaining_disagreements:
|
|
for arg in ce.remaining_disagreements:
|
|
if arg not in total_arguments_before:
|
|
new_arguments += 1
|
|
|
|
mean_conf = sum(confidences) / len(confidences) if confidences else 0.0
|
|
dispersion = (
|
|
math.sqrt(sum((c - mean_conf) ** 2 for c in confidences) / len(confidences))
|
|
if confidences
|
|
else 0.0
|
|
)
|
|
|
|
return RoundMetrics(
|
|
round=state.round_number,
|
|
mean_confidence=round(mean_conf, 3),
|
|
dispersion=round(dispersion, 3),
|
|
new_arguments=new_arguments,
|
|
concessions_made=concessions,
|
|
)
|
|
|
|
|
|
def should_stop(state: CouncilState, metrics: RoundMetrics) -> str:
|
|
"""Evaluate whether the debate should stop.
|
|
|
|
Returns one of:
|
|
- "converged" — dispersion below threshold, confidence stable
|
|
- "diminishing_returns" — nothing new is surfacing
|
|
- "genuine_disagreement" — dispersion widened, positions hardened
|
|
- "continue" — run another round
|
|
"""
|
|
# Hard cap
|
|
if state.round_number >= state.max_rounds:
|
|
return "max_rounds"
|
|
|
|
# Need at least 2 rounds to compare
|
|
if len(state.cross_examination_rounds) < 2:
|
|
return "continue"
|
|
|
|
prior = state.cross_examination_rounds[-2]
|
|
prior_confs = [
|
|
ce.updated_confidence
|
|
for ce in prior.values()
|
|
if ce.updated_confidence is not None
|
|
]
|
|
current_confs = [
|
|
ce.updated_confidence
|
|
for ce in state.cross_examination_rounds[-1].values()
|
|
if ce.updated_confidence is not None
|
|
]
|
|
|
|
if not prior_confs or not current_confs:
|
|
return "continue"
|
|
|
|
prior_mean = sum(prior_confs) / len(prior_confs)
|
|
current_mean = sum(current_confs) / len(current_confs)
|
|
|
|
# Converged: dispersion below threshold
|
|
if metrics.dispersion < state.convergence_threshold:
|
|
# Still check if anything changed — settled means done
|
|
if abs(current_mean - prior_mean) < 0.03:
|
|
return "converged"
|
|
return "continue"
|
|
|
|
# Diminishing returns: no new arguments, no concessions
|
|
if metrics.new_arguments == 0 and metrics.concessions_made == 0:
|
|
return "diminishing_returns"
|
|
|
|
# Genuine disagreement: dispersion widened and no one moved
|
|
if (
|
|
metrics.dispersion > state.convergence_threshold * 1.5
|
|
and metrics.concessions_made == 0
|
|
and metrics.new_arguments == 0
|
|
):
|
|
return "genuine_disagreement"
|
|
|
|
return "continue"
|