docs(rfc): RFC-0058 Dynamic Workflow Capability

[Million-mo]

Summary

Proposes RFC-0058: a new DynamicWorkflowCapability that lets the LLM author Python scripts to orchestrate subagents at scale, inspired by Claude Code dynamic workflows and pi-dynamic-workflows.

Motivation

AgentPool currently has two orchestration models:

• Static graph workflows (YAML graph: section) — topology fixed at config-load time, cannot adapt at runtime
• Dynamic team mode (RFC-0055) — LLM-driven via tool calls, but every intermediate result consumes the parent context

Neither handles tasks requiring 50+ parallel agents or iterative loops with unknown iteration counts without context degradation. This RFC addresses that gap with a script-based orchestration model where the plan lives in code, not in the LLM context window.

Key Design Decisions

• Script language: Python (not JavaScript like Claude Code) — AgentPool is Python-native; ast module provides robust sandboxing without a JS runtime dependency
• Sandbox: ast-level allowlist (no import, open, eval, dunder access) — sufficient for cooperative-LLM threat model
• Primitives: agent(), parallel(), pipeline(), phase(), log(), budget — matching Claude Code / pi-dynamic-workflows proven API
• Integration: Standard AbstractCapability subclass, registered via wolfharness.capabilities entry-point group
• Resumability: In-session journal-based resume (cross-session deferred to v2)

Options Analyzed

Option	Verdict
Python ast sandbox	Recommended — expressive, no deps, Python-native
RestrictedPython library	Marginal safety gain, adds external dep + 3.13 risk
Declarative YAML/JSON IR	Safest but cannot express loops/conditionals

Related

• RFC-0055 (Dynamic Team Mode — complementary, different scale)
• RFC-0042 (Unified Lifecycle Architecture — Journal, SnapshotStore)
• openspec/specs/static-graph-workflows/ (current static workflow baseline)

Checklist

• All 12 RFC sections complete per template
• 3 options analyzed with trade-offs
• Evaluation criteria defined with weights
• Technical design with file layout, data models, event types
• Security analysis with threat model
• Implementation plan with 5 phases / 4 weeks
• 6 open questions documented