Summary

Three OMO-inspired improvements bundled for atomic review:

1. delegate_task model_hint parameter (OMO 11.3, P0) — let parent route child to different model using short names like haiku / opus or full IDs.
2. hermes_orchestrator boulder + ultragoal (OMO 11.2, P1) — persistent evidence chain for every delegated task.
3. Task-type keyword routing (OMO 11.1, P1) — auto-inject system-prompt templates based on goal keywords like /research, /critic, etc.

Why

Borrowed from Sisyphus Labs' OMO tool (lazyclaudecode / oh-my-openagent) — addresses three real gaps in delegate_task:

• No model layering — all subagents inherit parent's model; can't cheaply route "explore this dir" to haiku
• No persistence — long-task intermediate state is lost; no way to resume or audit
• No orchestration templates — every task needs a hand-written system prompt; no reusable patterns

Changes

• tools/delegate_tool.py — add model_hint parameter + task-type keyword detection (~+200 lines)
• hermes_orchestrator.py (new) — OrchestratorContext + boulder + ultragoal + audit (590 lines)
• hermes_task_templates.py (new) — keyword regex + 6 task-type templates (320 lines)
• tests/test_hermes_orchestrator.py (new) — 22 unit tests
• tests/test_hermes_task_templates.py (new) — 27 unit tests
• tests/tools/test_delegate_model_hint.py (new) — 20 unit tests

Test Results

$ pytest tests/test_hermes_task_templates.py \
        tests/test_hermes_orchestrator.py \
        tests/tools/test_delegate_model_hint.py \
        tests/tools/test_delegate.py
204 passed, 1 warning in 18.68s

Plus full delegate regression: 177 passed (was 155, +27 new templates +22 orchestrator = 204).

Backward Compatibility

✅ All new features are opt-in:

• model_hint=None → identical behavior to before
• hermes_orchestrator.py is a separate module (not auto-imported in delegate_task)
• hermes_task_templates.py is wrapped in try/except ImportError — if missing, delegate_task works as before

Usage Examples

1. model_hint (OMO 11.3)

# Parent on opus, child on haiku (1/20 cost)
delegate_task(
    goal="扫一下 ~/.hermes/skills/ 找 contract",
    model_hint="haiku",
    toolsets=["file", "terminal"],
)

# Batch 混搭
delegate_task(tasks=[
    {"goal": "读 README 总结", "model_hint": "haiku"},
    {"goal": "对比方案给主推荐", "model_hint": "opus"},
])

2. boulder + ultragoal (OMO 11.2)

from hermes_orchestrator import OrchestratorContext

with OrchestratorContext(
    "扫目录找 contract",
    model="claude-haiku-4-5",
    success_criteria=["找到所有 contract 相关 skill 路径"],
    anti_criteria=["不要修改文件"],
) as ctx:
    ctx.append_step(tool="terminal", args_summary="ls ~/.hermes/skills")
    ctx.complete_step(result_summary="找到 2 个文件")
    ctx.record_evidence(
        source="工具输出",
        content="ls 返回 contract-review/, contract-review-intent/",
        citation="terminal: ls ~/.hermes/skills/",
    )
# audit.md 自动生成 at ~/.hermes/orchestrator/ultragoal/{task_id}/audit.md

The source field (一手/二手/推断/工具输出/用户输入) addresses the Mavis paper's "二手 source passed as firsthand" failure mode — different source types get distinct emoji markers in audit.md so reviewers can spot goal drift at a glance.

3. Task-type keyword routing (OMO 11.1)

# 自动检测触发词 → 注入模板 + 推荐 model
delegate_task(goal="/research 当前 AI agent 生态")        # → sonnet + 调研模板
delegate_task(goal="/critic 审查这个方案")                  # → opus + Momus 7维挑刺模板
delegate_task(goal="扫一下 skills/ 找 contract")           # → haiku + 只读侦察模板
delegate_task(goal="/implement 做一个 CLI 工具")           # → opus + 5步实施模板

Caller-supplied model_hint always wins:

delegate_task(goal="/critic 方案", model_hint="haiku")  # haiku 优先

Supported Keywords

Thread Safety

• hermes_orchestrator: threading.Lock() around all file ops, atomic writes (.tmp + rename)
• hermes_task_templates: pure functions, no I/O
• delegate_task: model_hint additions are pure value-passing, no shared state

Related

• multi-agent-harness/SKILL.md 第十一节 11.1/11.2/11.3 路线图
• OMO 工具调研报告 (/tmp/omo_research.html)
• Skill 路线图 (本地): ~/.hermes/skills/multi-agent-harness/SKILL.md

Generated with Claude Code