Repository navigation
v1.15.0
Agent orchestration layer
Every skill (33) now has an agent counterpart (41 agents total, including 9 phase-isolated splits — Reviewer/Critic, Generator/Evaluator, and other adversarial or parallel splits) generated the same way skills are, installable into any harness that supports it.
New:
agents/*/AGENT.md— canonical agent roster, 41 total (32 thin stubs + 9 hand-split for genuine isolation value: adversarial independence or real parallelism)workflows/*.md— 3 reusable multi-agent recipe docs- Per-harness generators for Claude Code, OpenCode, Hermes, and ZCode — each respecting that harness's own model defaults and nested-delegation capabilities instead of assuming Claude Code's
- New CLI install targets:
claude-agents,opencode-agents,zcode-agents - ZCode support (skills + agents), via the agentskills.io-compatible frontmatter confirmed against a real ZCode-created agent file
Fixed (found via live testing, not just static inspection):
- OpenCode granted
task: true(nested delegation) to every subagent by default, including ones deliberately built leaf-only — confirmed live viaopencode debug agent <name>, fixed by settingpermission.taskexplicitly per agent - ZCode failed to load an agent when a Claude-Code-specific
model:alias was copied over verbatim — model field is now Claude-Code-only across every generator opencode-agentssilently overwrote hand-customized entries in a consumer'sopencode.jsonwith no distinct log line — now logs Added/Updated/Unchanged per agentzcode-agentshad no env-var override, so its own CI idempotency test was silently writing into the real runner$HOME— addedZCODE_AGENTS_DIR
Researched and documented, not just assumed:
- Claude Code nested delegation — confirmed via official docs, default max depth 3
- Hermes nested delegation — off by default (
MAX_DEPTH=1), confirmed by readingdelegate_tool.pysource directly; generated templates now warn when an agent needsdelegation.max_spawn_depthraised - ZCode nested delegation — a real headless CLI path exists (
zcode.cjs --prompt) but was left unverified by explicit user choice, since testing it would spend real Z.AI account credits
See the README's "Agent orchestration" section for the full leaf/orchestrator design and per-harness details.
Full changelog: v1.1.0...v1.15.0