Mac: short Hermes side tasks on Claude Haiku 5.5

The smart-approval guardian asks for a one-word verdict with reasoning off
and max_tokens=16 (title generation: 64). GLM-5.3 Flash's reasoning is
mandatory on OpenRouter, so Hermes floors it to effort=low, the model
spends the budget thinking and returns nothing (finish_reason=length), and
every flagged command escalated to the user. Move approval, titles, memory
query rewrite and MCP sampling to Claude Haiku 5.5 (reasoning off by
default); compression, curator and the mail monitor stay on GLM.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
PhilandClaude Opus 5.5 committed 2026-11-11 08:47:28 -07:00
1 parent 420df90827
commit 9a9f0f0bd5
2 files changed
+19 -13

No files matched your search

+17 -11
View File
@@ -59,19 +59,25 @@ hermes_settings:
# /model deepseek or /model haiku in Telegram; add --once for one reply.
model_aliases.deepseek: {model: deepseek/deepseek-v4.1-flash, provider: openrouter}
model_aliases.haiku: {model: anthropic/claude-haiku-5.5, provider: openrouter}
# Side tasks would otherwise run on the chat model ("auto"). GLM-5.3 Flash
# is the cheapest model with credible Hermes scores; its slow first token
# doesn't matter here. Vision, voice and /review stay on the chat model.
# Side tasks would otherwise run on the chat model ("auto"). Vision, voice
# and /review stay on the chat model.
# Short-answer tasks turn reasoning off and allow only a few tokens (the
# approval guardian: 16, titles: 64). GLM-5.3 Flash can't turn reasoning off
# (OpenRouter: reasoning mandatory), spends the budget thinking and returns
# nothing, so the guardian escalated every command to you. Haiku 5.5 has
# reasoning off by default.
auxiliary.approval.provider: openrouter
auxiliary.approval.model: anthropic/claude-haiku-5.5
auxiliary.title_generation.provider: openrouter
auxiliary.title_generation.model: anthropic/claude-haiku-5.5
auxiliary.memory_query_rewrite.provider: openrouter
auxiliary.memory_query_rewrite.model: anthropic/claude-haiku-5.5
auxiliary.mcp.provider: openrouter
auxiliary.mcp.model: anthropic/claude-haiku-5.5
# Longer tasks: GLM-5.3 Flash is the cheapest model with credible Hermes
# scores; its slow first token doesn't matter here.
auxiliary.compression.provider: openrouter
auxiliary.compression.model: z-ai/glm-5.3-flash
auxiliary.approval.provider: openrouter
auxiliary.approval.model: z-ai/glm-5.3-flash
auxiliary.title_generation.provider: openrouter
auxiliary.title_generation.model: z-ai/glm-5.3-flash
auxiliary.mcp.provider: openrouter
auxiliary.mcp.model: z-ai/glm-5.3-flash
auxiliary.memory_query_rewrite.provider: openrouter
auxiliary.memory_query_rewrite.model: z-ai/glm-5.3-flash
auxiliary.curator.provider: openrouter
auxiliary.curator.model: z-ai/glm-5.3-flash
auxiliary.monitor.provider: openrouter
+2 -2
View File
@@ -28,8 +28,8 @@ Manages cogitator-02, an Apple Silicon Mac on the tailnet (`mac.yml`,
and the dashboard.
Covers update backups, timezone, not borrowing Claude Code's login, sandbox
memory, disabled toolsets, skill-write approval, deny rules, streaming,
voice-note transcription, and models: chat on GPT-6.1 Sol, side tasks and
cron on GLM-5.3 Flash, subagents on DeepSeek V4.1 Flash, `/model` aliases.
voice-note transcription, and models: chat on GPT-6.1 Sol, short side tasks
(approval, titles) on Claude Haiku 5.5, longer ones and cron on GLM-5.3 Flash, subagents on DeepSeek V4.1 Flash, `/model` aliases.
Keys in `hermes_settings_force` are set with `--force` (Hermes reads
`agent.reasoning_effort` but `config set` refuses it, hermes-agent #140221).
- **MCP servers and secrets** (also tag `hermes-settings`): MCP servers are