Progressive-disclosure reflector for OpenCode. Hermes automatically analyzes your conversations to discover reusable patterns, workflows, and domain knowledge, then generates AI skills (SKILL.md) and behavioral rules — turning hard-won experience into permanent team knowledge.
Hermes never installs anything directly. All proposals go to a staging directory for human review before activation.
| Mode | Trigger | Scope |
|---|---|---|
| Manual | /reflect or "reflect" | Current session — ad-hoc analysis |
| Batch | Weekly automatic | All idle sessions — cross-session pattern discovery |
- Skills (
SKILL.md) — Step-by-step workflows and domain expertise. Agents load these automatically when conditions match. - Rules — Concise behavioral principles and environment facts. Staged for human review.
- Staging workflow — All proposals land in
~/.config/opencode/hermes-staging/for you to approve or reject via the/skill-refreshcommand.
Hermes categorizes findings into four types:
| Type | Classification | Output |
|---|---|---|
| Workflow | "How to do X" — step-by-step procedures | SKILL.md |
| Domain Knowledge | "How X works" — in-depth expertise | SKILL.md |
| Context Knowledge | "In this env, X is true" — environment facts | Rule |
| Universal Principle | "Always / Never do X" — behavioral rules | Rule |
{
"plugin": ["opencode-hermes"]
}Add command entries for /reflect and /skill-refresh:
{
"plugin": ["opencode-hermes"],
"command": {
"reflect": {
"template": "Call hermes_nudge_now to start background reflection.",
"description": "Analyze conversation for patterns and auto-create skills"
},
"skill-refresh": {
"template": "逐 group 处理 hermes-staging/ 中各待审项。",
"description": "评估 hermes-staging 待审 proposals"
}
}
}Restart OpenCode after making changes.
Create ~/.config/opencode/hermes.json to override the reflection model:
{
"reflection_model": {
"providerID": "anthropic",
"modelID": "claude-sonnet-4-5",
"variant": "max"
}
}If not configured, Hermes uses the session's current model.
/reflect
Hermes performs a two-step analysis:
- Step 1 — Lightweight scan: Quick pass to identify candidate rounds
- Step 2 — Deep analysis: Detailed extraction of skills and rules
Results appear as a toast notification and proposals land in hermes-staging/.
/skill-refresh
This walks you through staged proposals grouped by batch, lets you rate each one, and decide whether to keep, delete, or merge.
Hermes automatically runs a weekly batch scan of all completed sessions that have been idle for 48+ hours. Results appear as a Windows dialog (or terminal notification) when ready.
| File / Directory | Purpose |
|---|---|
~/.config/opencode/hermes-state.json | Persistent state (analyzed rounds, session skills, logs) |
~/.config/opencode/hermes.json | Optional model config |
~/.config/opencode/hermes-staging/ | Staging area for human review |
- Tracking: Hermes listens to
chat.messageevents to track the active session - Analysis pipeline: When triggered, it fetches session messages, parses them into structured rounds (user turns → tool calls → assistant responses), and runs a two-pass AI analysis
- Background sessions: Analysis runs in dedicated background sessions (
<title>#reflect) so it doesn't pollute your main conversation - Deduplication: Only new rounds since last analysis are processed each time
- Staging: Results are written to
hermes-staging/organized by date and batch, never installed automatically
- Requires OpenCode SDK with
client.session.*API access - Batch mode is Windows-only for the completion popup (can be disabled by removing the
spawnblock) - Model dependent: Quality of analysis depends on the reflection model's reasoning capability
MIT