| Capability | Owner agent | Skill(s) | Skill load default | Artifact / output | Gate |
|---|---|---|---|---|---|
| Feature / plan | architect |
architect-plan, architect-review |
Parent: sub-skills by mode; Task children: load: full, minimal, or auto per agents/architect.md |
.plan/<type>.<slug>.md via scribe |
User → orchestrate |
| Stage execution | orchestrate |
orchestrate-execution, orchestrate-recovery |
Parent: sub-skills by situation; Task children: load: full, minimal, or auto per agents/orchestrate.md |
Graded stages, completion prompt | verifier |
| Worktree env copies (opt-in) | worktree-env |
worktree-env |
Parent: load: full once per bootstrap; trust canonical evidence |
Copy report with wt_root, main_root, per-file is_regular_file |
preflight (verify only) |
| Env bootstrap / repair (opt-in) | preflight |
preflight |
Parent: load: full on preflight Task |
Ready or one recommended_env_fix; repair pass for deps/runtime/indexing |
orchestrate sets env_gate_passed |
| Backend / generic code | developer |
developer |
auto (tiered triggers in agent; parent overrides) |
Completion report | verifier |
| Frontend | frontend-dev |
frontend-dev |
auto |
Completion report | verifier |
| HTML prototype | ux-dev |
ux-dev |
auto |
.prototype/<slug>/ |
verifier |
| Review (planning) | review |
review + optional security-reviewer, performance-reviewer, doc-reviewer |
auto (effective load: full for review; specialists default load: full) |
Review markdown to parent | Architect + scribe |
| CodeRabbit gate (completion) | orchestrate → review |
code-review (+ review on load: full) |
Parent: load: full on CodeRabbit Task; exactly once per artifact/feature |
CODERABBIT_GATE + full finding inventory + per-item local resolutions; feature completion summary |
After last verifier / queue exhausted; coderabbit review --agent --base develop by default; before final push/PR, difficulty gates, and architect — not per issue and not after remediation |
| Evidence check | verifier |
verifier |
auto |
Verdict + evidence | Orchestrate |
| Docker compose build/test + optional review URL (Sysbox sibling) | developer, frontend-dev, verifier (probe: preflight); orchestrate instructs Tasks only — does not load the skill; menu (2) = sandbox feature build/refresh (no issue queue) |
docker-sandbox |
Load when Compose/Docker or web review needed (sandbox: preferred|required / execution_mode: sandbox_feature_build from parent, compose in test_commands, or preflight sandbox: ready + compose file). Soft-skip when unavailable unless required. Not Cloudflare Workers Sandbox. |
sandbox exec logs; optional https://{slug}.{apex} |
OPENCODE_SANDBOX_ENABLED (expose: localhost publish + host tunnel public hostname + OPENCODE_SANDBOX_REVIEW_DNS); rebuild OpenCode image after CONFIG_REF change |
| Docs generation | document |
document |
auto |
Content to scribe |
Architect |
| Ship / hotfix / TDD | (user-chosen agent with skill allowed) | ship, hotfix, debug-fix, tdd |
User-chosen | Git / PR | User confirms each step |
Other read-only specialists (debugger, refactor, designer, strategist, helper, senior-dev, mentor, vision, scribe) define Execution readiness and Auto-load triggers in their agent files; parents still pass load: full|minimal|auto on every Task. worktree-env and preflight are narrow startup subagents (bash + allow-listed skills only; no app-code edits).
- Per-agent model routing in
opencode.json. - MCP:
mcpjungleandclaude-context(command path unchanged). MCPJungle manages Context7, Cloudflare API, Cloudflare Docs, anddocs-mcp-serverupstream authentication. Narrow Cloudflare skills:cloudflare,wrangler,workers-best-practices(from cloudflare/skills). strategist,preflight, architect plan/review split.