chess-evolve: path to (Task, Workflow) → outer loop with zero adapter code

[colehurwitz]

Goal

chess-evolve (the reference factory-as-library demo) should define only two things:

task = PositionTask()
workflow = build_position_eval_workflow()
result = SwarmEngine.from_task(task, workflow, budget=100).run()

No adapter, no monkey-patches, no workspace hacks. Today we need a ChessSwarmEvaluator adapter (~30 lines) because of gaps in remote-factory. This issue tracks the path to eliminating it.

Current adapter and why each piece exists

class ChessSwarmEvaluator(SwarmEvaluator):
    def evaluate(self, workflow, ...):
        _prime_workspace(workspace)              # ← Gap A
        _runner.invoke_agent = _make_move_invoke(inner)  # ← Gap B + C
        loop = InnerLoop(project_dir=workspace, workflow=workflow)
        record = loop.step()
        _runner.invoke_agent = _orig              # ← Gap B cleanup
        return EvalResult(score=record.score_end)

Gap A: Workspace priming (_prime_workspace)

Creates .factory/chess/ dir and placeholder board_state.md, memory.md before the executor runs — otherwise the executor blocks on reads that don't exist yet.

Fix (chess-evolve side): Move priming into PositionTask.setup() — the Task contract already owns workspace preparation. No factory change needed. Can fix now.

Gap B: Agent backend swap (invoke_agent monkey-patch)

Swaps factory.agents.runner.invoke_agent (spawns claude CLI) for chess-evolve's _sdk_invoke_agent (Haiku API). Restored in finally.

Fix (remote-factory): WorkflowExecutor should accept an optional agent_fn parameter. Tracked in #1488 (Gap #1).

Gap C: Output persistence (_make_move_invoke wrapper)

Writes agent output to .factory/chess/move.md on disk after every call. Needed because WorkflowExecutor keeps output in memory only, but PositionTask.verify() reads from disk.

Fix (remote-factory): Executor should write agent output to paths declared in node.writes automatically. Akash's Factory Artifact Format design (Port-level write validation) is the upstream fix. Tracked in #1488 (Gap #2).

Sequencing

1. Now (chess-evolve): Fix Gap A — Task.setup() handles priming
2. Upstream: Fix Gap B — WorkflowExecutor(agent_fn=...) eliminates monkey-patch
3. Upstream: Fix Gap C — executor writes to node.writes automatically (artifact format)
4. Upstream: Add SwarmEngine.from_task(task, workflow, budget) convenience API
5. Final: Delete ChessSwarmEvaluator. evolution.py becomes 5 lines.

End state

from chess_evolve.tasks import PositionTask
from chess_evolve.pipeline import build_position_eval_workflow
from factory.outer_loop import SwarmEngine

def main():
    task = PositionTask()
    workflow = build_position_eval_workflow()
    result = SwarmEngine.from_task(task, workflow, budget=100).run()
    print(f"Best score: {result.best_score}")

Related

• (Task, Workflow) → outer loop: 3 execution-layer fixes to close the gap #1488 — Factory-as-library gaps (all 6 gaps)
• feat: DataNode graph primitive — first-class data loading in workflows #1483 — DataNode PR (_step_with_data_node)
• Factory Artifact Format (Akash's design) — upstream fix for Gap C

🤖 Generated with Claude Code

[colehurwitz]

Moved to colehurwitz/chess-evolve#7 — this is a chess-evolve integration roadmap, not a remote-factory issue. The upstream gaps it depends on are tracked in #1488.