By name: omo-orchestration-bridge description: >- Bridge Oh My OpenAgent (OMO) discipline-agent workflows into Kimi Code CLI. Maps OMO commands (ultrawork, /start-work, resume boulder) to Kimi-native Agent tool orchestration using existing .github/agents/ definitions and .sisyphus/ boulder state. Triggers: ultrawork, /start-work, resume boulder, omo, atlas, prometheus, sisyphus, oracle, team mode, parallel agents, boulder workflow, plan execution.
omo-orchestration-bridge
Core rule: OMO is an OpenCode plugin; Kimi Code CLI is a different runtime. This skill translates OMO patterns into Kimi-native actions using the Agent tool, existing .github/agents/ definitions, and .sisyphus/ state. No OpenCode installation required.
When to use this skill
- User says
ultrawork,/start-work,resume boulder, or any OMO trigger phrase - Need to run Prometheus-style planning before implementation
- Need Atlas-style task delegation to specialist subagents
- Need parallel discipline agents (Team Mode emulation)
- Need Oracle-style compliance verification against a plan
- Need to read or update
.sisyphus/boulder.jsonstate
When NOT to use this skill
- Simple single-file edits → use Direct delivery, no OMO overhead
- OpenCode-specific features (hash-anchored edits, LSP tools) → unavailable in Kimi
OMO-to-Kimi mapping
| OMO Pattern | Kimi Equivalent |
|---|---|
ultrawork / ulw |
Skill trigger → IntentGate classification → auto-orchestrate |
| Prometheus (planner) | Agent with subagent_type="plan" reading .github/agents/prometheus.agent.md |
| Atlas (work manager) | Primary agent reads plan + boulder.json, delegates via Agent tool |
| Sisyphus (task worker) | Agent with subagent_type="coder" and full task spec |
| Oracle (auditor) | Agent with subagent_type="explore" doing read-only verification |
| Team Mode (parallel) | Multiple Agent calls with run_in_background=true |
| IntentGate | Pre-flight: classify intent as research / implement / investigate / fix |
| Todo Enforcer | Primary agent uses SetTodoList + idle-check before every delegation |
| Ralph Loop | Self-referential: after subagent returns, ask "is this 100% done?" |
Trigger quick-reference
| Trigger | Kimi Action |
|---|---|
ultrawork / ulw |
Classify intent → plan → delegate → execute → audit |
/start-work |
Prometheus generates plan → Momus reviews → Atlas delegates |
resume boulder |
Read boulder.json → find next pending → delegate |
atlas |
Primary agent plays work manager role |
prometheus |
Spawn plan subagent to generate .sisyphus/plans/ |
sisyphus |
Spawn coder subagent with full task spec |
oracle |
Spawn explore subagent for read-only compliance audit |
team mode |
Launch ≤4 parallel Agent calls with run_in_background=true |
Quick examples
- Start a new boulder: Read
.github/agents/prometheus.agent.md→ generate plan → save to.sisyphus/plans/→ initboulder.json→ delegate Task 1 to Sisyphus subagent. - Resume existing boulder: Read
.sisyphus/boulder.json→ find next unchecked task → delegate to Sisyphus with full spec from plan. - Run Oracle audit: Read plan → compare completed tasks against acceptance criteria → search for Must NOT Have violations → report APPROVE/REJECT per task.
Troubleshooting
"OMO plugin not found" OMO cannot be installed as a plugin in Kimi Code CLI. Use this skill instead — it provides the same workflow patterns via Kimi-native tools.
Subagent loses context between tasks
Use resume with the same agent_id instead of creating fresh subagents. Append session IDs to boulder.json.session_ids.
Plan file is stale after edits
Only mark checkboxes (- [x]) in plan files. Never rewrite task specs during execution — update boulder.json state instead.
Parallel agents overload context
Team Mode emulation uses background agents. Limit to 3-4 parallel subagents. Use TaskOutput to poll results.
Intent misclassification When ambiguous, default to Prometheus planning (kickoff lane) rather than Direct delivery.
Review checklist
Before completing an OMO-bridged task:
-
boulder.jsonstate matches actual progress - Plan checkboxes reflect completed work
- Evidence files saved to
.sisyphus/evidence/ - Subagent sessions resumed (not recreated) for follow-ups
- Oracle audit run for Tier 2+ work before sign-off
- Ralph Loop check performed: "Is this 100% done?"
- All parallel subagents terminated or completed
Related files
- references/agent-registry.md — Agent-to-subagent mapping and prompt templates
- references/boulder-protocol.md —
boulder.jsonschema, plan format, state rules - references/discipline-workflows.md — Full Prometheus→Atlas→Sisyphus→Oracle workflows
.github/agents/— Canonical agent definitions (source of truth).sisyphus/— Boulder state, plans, evidence directory