Syncing Agent Output Preferences

Syncing Agent Output Preferences

Overview

When the user rejects an output contract, fix the durable instruction surfaces, not just the current reply. Check Hermes memory plus host-agent docs for Codex, Claude Code, and OpenClaw.

When to Use

Use when the user says:

• "不要这种格式", "别再用 []", "修系统提示词和记忆"
• Codex/Claude/OpenClaw/Hermes output differs
• a style preference should persist across agent hosts

Do not use for one-off formatting requests unless the user asks to remember or unify it.

Workflow

1. Load autonomous-ai-agents/hermes-agent before Hermes config/prompt work.
2. If routing/skills are involved, load software-development/skill-routing-gov.
3. Find active prompt surfaces first:
   • Hermes repo/current project: nearest AGENTS.md / CLAUDE.md
   • Codex: ~/.codex/AGENTS.md
   • Claude Code: ~/.claude/CLAUDE.md
   • OpenClaw: ~/.openclaw/workspace/AGENTS.md
4. Search only targeted config/workspace roots first. Avoid broad home-directory scans unless needed.
5. Remove or rewrite the offending rule in every active surface where it exists.
6. Save a compact user memory only if the preference is durable.
7. Verify by re-reading the edited section and searching the active host roots for the rejected phrase.
8. Report which hosts were changed or already clean.

Known Host Files

• Hermes current project: AGENTS.md in the active repo or project.
• Codex global: ~/.codex/AGENTS.md.
• Claude Code global: ~/.claude/CLAUDE.md.
• OpenClaw workspace: ~/.openclaw/workspace/AGENTS.md.

References

• references/bracketed-audit-output-cleanup.md records the concrete cleanup pattern for rejected bracketed audit fields across Hermes, Codex, Claude Code, and OpenClaw.

Verification Commands

rg -n "Mandatory Output Protocol|TARGET PARAMETER/ASSET|CURRENT CONTEXT|ACTIONABLE TELEMETRY" \
  ~/.codex ~/.claude ~/.openclaw/workspace ~/.hermes/hermes-agent -g '*.md'

Expected result after removing the bracketed audit format: no matches in active host prompt files.

Pitfalls

• Do not treat Hermes memory as the only source; host agents read their own prompt docs.
• Do not claim Codex/Claude/OpenClaw are fixed without checking their actual files.
• Do not scan all of ~/Library for prompt text; macOS protected paths create noise and timeouts.
• Do not add a new rigid output schema while removing another one.