SuperAPI AI Directory
Partner中文AI newsFavorites
Skillshermes-config-optmzr

hermes-config-optmzr

Office & productivity

View repository

Updated 10-16 09:18

About this skill

Hermes Configuration Optimization

Optimize the Hermes configuration to reduce token consumption and improve Chinese trigger efficiency.

Optimization Process

1. Back up the current configuration

cp -r ~/.hermes/routing ~/.hermes/routing-backup-$(date +%Y%m%d-%H%M%S)

2. Routing file consolidation

Check for duplicates:

  • unified-index.md and skill-index.md typically overlap by 30-40%
  • Delete unified-index.md and unified-index.json

Streamline skill-router.md:

  • Remove line number markers (such as 28|, 29|）
  • Merge similar trigger words
  • Use a more compact format
  • Goal: compress from 13KB+ to 6-8KB

Verify:

ls -la ~/.hermes/routing/*.md
du -sh ~/.hermes/routing/*.md

3. Chinese localization of trigger words

Identify pure English trigger words:

# 找出只有英文触发词的 skill
import re
pattern = r'\|\s*`([^`]+)`\s*\|[^|]*\|[^|]*\|\s*([^|]+)\s*\|'
# 检查是否有中文字符
has_chinese = bool(re.search(r'[\u4e00-\u9fff]', triggers))

Add Chinese equivalent words:

  • findmy → 查找设备, 定位, 找手机
  • codex → 代码生成, AI编程, 代码代理
  • excalidraw → 手绘图, 流程图, 白板
  • github-auth → GitHub认证, gh登录, git凭据

Three-file synchronized update process(when adding a new skill or adding Chinese trigger words to an existing skill):

  1. skill-index.json — primary data source. Use a Python script to batch update the triggers array, deduplicate and sort, then write back.
  2. skill-router.md — routing table. Update the trigger word list in the corresponding category section.
  3. skill-index.md — table index. Update the trigger word column in the corresponding category table.

Complete process for batch adding Chinese trigger words:

import json

# 1. 读取 skill-index.json
with open('~/.hermes/routing/skill-index.json', 'r') as f:
    skills = json.load(f)

# 2. 检查哪些 skills 没有中文触发词
no_chinese = []
for skill in skills:
    triggers = skill.get('triggers', [])
    has_cn = any(any('\u4e00' <= c <= '\u9fff' for c in t) for t in triggers)
    if not has_cn:
        no_chinese.append(skill['name'])

print(f"没有中文触发词的 skills ({len(no_chinese)}):")
for s in sorted(no_chinese):
    print(f"  - {s}")

# 3. 定义中文触发词映射
chinese_triggers = {
    'apple-notes': ['苹果备忘录', '备忘录', 'Apple备忘录', '笔记'],
    'apple-rmndrs': ['苹果提醒', '提醒事项', 'Apple提醒', '待办'],
    # ... 为每个 skill 添加
}

# 4. 批量更新
for skill in skills:
    if skill['name'] in chinese_triggers:
        existing_triggers = skill.get('triggers', [])
        new_triggers = chinese_triggers[skill['name']]
        all_triggers = list(set(existing_triggers + new_triggers))
        skill['triggers'] = sorted(all_triggers)

# 5. 写回
with open('~/.hermes/routing/skill-index.json', 'w') as f:
    json.dump(skills, f, indent=2, ensure_ascii=False)

Key pitfalls:

  • Skill names in skill-index.json must match exactly; first use a script to check which skills exist and which are missing
  • Missing skills need a complete entry added first (name/category/grade/priority/security_review/supersedes/triggers/description)
  • The skill-index.md table row format is inconsistent (some start with | and some start with ||); when patching, include enough context to avoid multiple matches
  • MCP-type skills such as open-computer-use are classified under mcp category, not apple or red-teaming
  • After batch updating, skill-router.md and skill-index.md need to be updated in sync
  • Finally, remember to commit to the git repo

4. Compression threshold and context window calibration

Core finding (2026-04-27): Hermes's auto-compress trigger point looks at 主模型 context_length × compression.threshold, not at auxiliary.compression.extra_body.context_window。

Determination formula：

  • Trigger threshold = max(int(model.context_length * compression.threshold), 64000)
  • Example:model.context_length: 2000000 and compression.threshold: 0.5 → approximately 1000000 tokens before compress is triggered

Correct way to change it：

  1. First look at compression.threshold(the common default is 0.5）
  2. Then look at the actual context source of the main model:
    • model.context_length
    • or custom_providers[].models[<exact model name>].context_length
  3. If the current model name has an alias/suffix (such as gpt-5.4(xhigh)），custom_providers.models must write theexact same namekey; writing only gpt-5.4 is not stable enough
  4. When the custom endpoint /models is inaccurate or missing, directly write the integer in model.context_length to avoid Hermes falling back to an incorrect default value

Do not confuse：

  • auxiliary.compression.extra_body.context_window: only affects the summary model request body parameters
  • auxiliary.compression.context_length: only used for auxiliary compression model feasibility check/override
  • What truly determines when auto-compress happens is still the main model's context_length

5. Memory compression (architecture chain format)

Core finding (2026-04-24): flat entries → architecture chain, one tree-shaped dependency replaces 10+ scattered entries.

Architecture chain template：

元：[核心行为规则]→[次级规则]→[三级规则]
├─红线：[不可做]
├─重试：[失败策略]
├─[领域A]链：[约束]→[执行]→[验证]
└─[领域B]链：[约束]→[执行]→[验证]

Example：

元：长忆从严→直做勿问→效驱>轮询→全量不挑。
├─红线：勿删cron/自动化，清前确认存废。
├─重试：始后败不重（断线除外），败即弃。
├─cron链：模型驱→禁脚本→deliver=feishu→冲突必解→sniper参数分离不与长驻争。
├─Hansa链：只收不挪(无FluxA权)→cron走MCP→状态机→答题agent推理。
└─Quest链：禁Twitter/X→禁视频→Medium/Dev.to/gist可。

USER.md optimization：

  • 10+ flat entries → 1 architecture chain
  • Classical Chinese compression, each entry ≤20 characters
  • Store only hard rules; delete project details

MEMORY.md optimization：

  • Keep only cross-project stable technical facts
  • Delete all project details → belongs in repo docs / session_search
  • Goal: < 500 bytes

5. Skills cleanup

# 1. 查大小排序
du -sh ~/.hermes/skills/* | sort -rh | head -20

# 2. 统计每个目录 skill 数
for d in ~/.hermes/skills/*/; do
  count=$(find "$d" -name "SKILL.md" | wc -l)
  size=$(du -sh "$d" | cut -f1)
  echo "$count skills, $size: $(basename $d)"
done | sort -t, -k1 -rn

Deletion criteria：

  • Auto-installed large packages (>1M and >10 skills, never manually loaded)
  • Cloned repos (containing .git directory)
  • Empty directories (0 skills)
  • Domains the user will not use

Boundary confirmation before deletion (2026-04-25 retrospective)：

  • First list an inventory by "functional domain," then delete; do not misjudge "has duplicates" as "all can be deleted."
  • Tools like OpenChronicle can only be benchmarked against the same functional domain: timeline memory / activity replay / session retrieval / personal behavior archiving.
  • Do not casually delete cross-domain productivity skills (such as coding agents like codex / opencode / claude-code) unless the user explicitly includes that domain in the cleanup scope.
  • When the user says "kill the feature library/conflicting items," interpret it by default as: only delete conflicting items in the same function as the target tool; it is not a global deduplication and massive cleanup.
  • Before destructive deletion, first provide a "list to be deleted + reasons for keeping"; when the scope is unconfirmed, prefer only adding routing and do not delete.

6. Routing reference cleanup (must be done after deleting skills)

⚠️ skill-index.json top level is an array [{skill}, ...], not a dict {"skills": [...]}

Python cleanup (sed is unreliable for multi-line sections)：

import json, re

# skill-index.json
data = json.load(open('skill-index.json'))
data = [s for s in data if s.get('category') not in deleted_cats
        and s.get('name') not in deleted_skills]
json.dump(data, open('skill-index.json', 'w'), indent=2)

# skill-index.md — 删除 section + table rows
content = open('skill-index.md').read()
for cat in deleted_cats:
    pattern = rf'## {re.escape(cat)}\n.*?(?=\n## |\Z)'
    content = re.sub(pattern, '', content, flags=re.DOTALL)
for skill in deleted_skills:
    lines = content.split('\n')
    content = '\n'.join(l for l in lines if skill not in l)
open('skill-index.md', 'w').write(content)

# skill-router.md — 删除对应行
lines = open('skill-router.md').readlines()
lines = [l for l in lines if not any(s in l for s in deleted_skills)]
open('skill-router.md', 'w').writelines(lines)

Verification：

grep -c "DELETED_SKILL" skill-router.md skill-index.md
# 期望：0

Expected benefits

Optimization item Before optimization After optimization Savings
routing file 90KB+ 50-60KB 30-40%
skill-router.md 13KB 6-8KB 50%+
English trigger words 30+ <15 50%+
USER.md 1900 bytes 1200 bytes 37%

Token savings: about 2000-3000 tokens per conversation

Verification steps

  1. Check file size
  2. Verify trigger word matching
  3. Test routing functionality
  4. Monitor token usage

Risks and Pitfalls

  • Deleting unified-index.md may cause some routes to fail
  • Added Chinese trigger words may not be used by the user
  • Backups are needed for rollback
  • After deleting skills, routing references must be cleaned up in sync（skill-router.md, skill-index.md, skill-index.json）
  • skill-index.json is an array, not a dict: top level [{...}], not {"skills": [...]}
  • sed is unreliable for multi-line sections: use Python regex to clean skill-index.md
  • Use write_file for MEMORY.md, use the memory() tool for the user profile: two independent systems
  • First take the current state of repo/AGENTS as authoritative: the actual landing point of routing must be verified on site; do not treat ~/.hermes/hermes-agent/routing/、~/.hermes/routing/, project-internal .hermes/routing/ as one of them as always true. If the current task already has a project-internal .hermes/routing/, prioritize updating its official file; do not first write the official routing to /tmp/* and then stop there.
  • memory replace requires an exact match of old_text: Multiple entries cannot be replaced at once; they must be removed one by one and then new entries added
  • The architecture chain format is superior to a flat list: A tree-like dependency relationship saves more tokens than 10+ loose entries and is logically clearer
  • When modifying config.yaml , first read the exact context before patching: containing api_key、base_url, inline map (such as providers: {}) YAML is very prone to accidentally affecting adjacent keys due to fuzzy matching; do not directly replace based on redacted fragments or memory