feat(gds-games): parameterizable reactive_decision_agent() — include_outcome and include_feedback flags

[rororowyourboat]

Summary

reactive_decision_agent() in ogs/dsl/library.py hard-codes a full 5-game loop (Context Builder → History → Policy → Reactive Decision → Outcome) with 3 contravariant feedback flows. This makes it unusable for multi-agent patterns where:

• The Outcome game lives in a shared Decision Router (not per-agent)
• Feedback flows span across agents and must be wired at the pattern level, not inside each agent

As a result, every multi-agent use case must re-implement the 4-game open-loop chain from scratch as a client-side factory — duplicating ~85 lines of composition logic that belongs upstream.

Proposed API

from typing import Literal, overload

@overload
def reactive_decision_agent(
    name: str = "Reactive Decision Agent",
    include_outcome: Literal[True] = True,
    include_feedback: Literal[True] = True,
) -> FeedbackLoop: ...

@overload
def reactive_decision_agent(
    name: str = "Reactive Decision Agent",
    include_outcome: bool = ...,
    include_feedback: bool = ...,
) -> FeedbackLoop | SequentialComposition: ...

def reactive_decision_agent(
    name: str = "Reactive Decision Agent",
    include_outcome: bool = True,
    include_feedback: bool = True,
) -> FeedbackLoop | SequentialComposition:
    ...

Behaviour matrix

include_outcome	include_feedback	Returns	Games
True (default)	True (default)	FeedbackLoop	CB → Hist → Pol → RD → Out + 3 feedback flows (current behaviour, unchanged)
False	True	FeedbackLoop	CB → Hist → Pol → RD + empty or partial feedback
True	False	SequentialComposition	CB → Hist → Pol → RD → Out (no feedback wrap)
False	False	SequentialComposition	CB → Hist → Pol → RD (open loop, no Outcome, no feedback)

The False / False case is the open-loop agent chain needed by multi-agent patterns.

Files affected

• packages/gds-games/ogs/dsl/library.py — modify reactive_decision_agent()
• packages/gds-games/ogs/dsl/__init__.py — no export changes needed (already exported)

Motivation

Multi-agent patterns (bilateral negotiation, N-party agreement) all follow the structure:

(open_loop_agent("Agent 1") | open_loop_agent("Agent 2")) >> router

where router contains the Outcome game and feedback is wired at the pattern level for all agents simultaneously. With this change, open_loop_agent(prefix) becomes reactive_decision_agent(name=prefix, include_outcome=False, include_feedback=False) — no client-side reimplementation needed.

Effort estimate

Low — additive change to an existing function, backward-compatible (default behaviour unchanged).

[rororowyourboat]

Implemented in v0.3.0 — reactive_decision_agent() supports include_outcome and include_feedback flags. Tested in tests/test_new_features.py::TestReactiveDecisionAgentFlags.