Zig Coding Agents: Two Agents, One Language, Opposite Philosophies

Wintermolt has no loop detection beyond a MAX_ITERATIONS = 25 cap on the agentic loop. If the agent gets stuck repeating the same tool call with the same arguments, it will burn through all 25 iterations before stopping.

Zaica has the most sophisticated loop detection in this entire directory set: a Wyhash-based ring buffer of tool call signatures with 3-tier escalation (warning → stronger warning → force break). The same tool with different arguments or different output gets a different hash — so legitimate iteration isn't flagged as a loop.

This is a meaningful difference for real-world coding tasks, where agents commonly get stuck in file-edit-verify cycles on stubborn patterns.

Wintermolt manages state imperatively — the AgentLoop struct owns its history, config, storage, and tools, and mutates them directly through method calls.

Zaica uses a custom reactive state graph via zefx (Effector-inspired). Events trigger Store reducers, which trigger watchers in a two-phase flush. The status bar is a watcher on derived stores that ultimately trace back to token-received events. This is genuinely novel architecture for a coding agent — nothing else in this set uses a reactive engine.

Wintermolt has no equivalent to chain mode. Its skills system defines specialized agent roles, and its cron scheduler automates recurring commands, but there's no concept of a multi-step pipeline with per-step tool filtering.

Zaica's chain mode (.chain.md files) is a structured workflow system. Each step has a name, prompt with variable substitution ({task}, {previous}), optional tool filter, and max iterations. The chain computes the maximum risk level across all steps for a single permission prompt.

This is comparable to LangGraph-based workflows in DeerFlow but implemented in 528 lines of Zig instead of a full graph orchestration framework.

Wintermolt uses a ~/.wintermolt/.env file with environment variable conventions. Simple, familiar, but not layered — the last write wins.

Zaica implements a 6-layer JSON priority chain with deep object merging: comptime defaults → provider presets → user config → project config → env vars → CLI flags. API key resolution follows a separate 4-tier chain. The system prompt is also layered: base → provider-specific → ~/.config/zaica/ZAICA.md → ./ZAICA.md.

Zaica's config system is comparable in rigor to Qwen Code's five-layer model resolution — arguably more elaborate, given the deep JSON merge.

Wintermolt doesn't appear to implement output truncation beyond what the model's context window naturally limits. Large file reads or bash outputs go into the message history as-is.

Zaica implements a head/tail truncation strategy: for large outputs, it keeps the first half and last half, with a [WARNING: output truncated] marker in the middle. This is based on the observation that the most useful information in tool output is typically at the beginning (what was done) and the end (the result / error).