Benchmark from OpenCode: headless server + multi-surface arch, plugin bus, agents-as-md, on-chain share

[mega123-art]

OpenCode postmortem → what AgentNet should benchmark.

OpenCode (sst/opencode — 175k★, MIT, TypeScript, v1.17.7) already solved the multi-surface problem on our roadmap (VSCode / CLI / mobile parity + cloud + mobile UI) with a client/server split: opencode serve is a headless HTTP/OpenAPI server, and every surface (TUI, desktop, web, IDE, mobile) is a thin client driving it. It also ships a typed SDK (@opencode-ai/sdk), a plugin event-bus, agents/subagents-as-markdown, a granular permission model, md custom-commands, and MCP (local + remote + OAuth).

Below: a ranked list of what to benchmark, one comment per item, each mapped to our code (packages/core/src/runtime/spawn.ts, the skill-market MCP surface, the approval channel) and our existing plans (plans/hermes-skill-minting.md, plans/ai-power-features.md).
React 👍 on the comments we should adopt.

Sources (all fetched live, June 2026): https://opencode.ai/docs/server/ · /sdk · /plugins · /agents · /permissions · /commands · /mcp-servers · /cli · /share

[mega123-art]

1 — Headless server + OpenAPI + SDK (highest leverage — adopt first)

opencode serve = headless HTTP server, OpenAPI at /doc; TUI launches its own bg server; many clients drive one server. Ships @opencode-ai/sdk (TS): create/fork sessions, send prompts, subscribe to SSE events, control the TUI (append/submit prompt, toasts, dialogs). Endpoints incl. session create/delete/fork/share, messages, files, tools/agents, auth/OAuth, TUI-control.

benchmark: wrap our runtime (spawn.ts Engine) in a serve HTTP/OpenAPI layer → VSCode/CLI/mobile become thin clients instead of each reimplementing the engine. TUI-control endpoints = drive a desktop session from the phone. This is the structural fix for our parity + mobile + cloud items.

[mega123-art]

2 — Plugin event-bus

Plugins = TS modules on an event bus: tool.execute.before/after, permission.asked/replied, session.created/compacted/idle/diff, file.edited, message.*, tui.*. Can add custom tools, block actions (e.g. deny .env read), inject shell env, customize compaction.

** benchmark:** replace our scattered approval / memory-inject / skill-cast hooks inside spawn.ts with a clean hook bus; let bought skills ship plugin hooks, not just SKILL.md text.

[mega123-art]

3 — Agents & subagents as markdown

Primary agents (build = full, plan = read-only) + subagents (general / explore / scout), defined as md in .opencode/agents/ with per-agent model, prompt, permission, temperature, mode. @mention to invoke a subagent.

** benchmark:** our skills are already md — a skill could be a subagent (model+prompt+tools+perm). Feeds the cross-model fusion idea (ai-power-features.md §10): subagents = the parallel competing agents.

[mega123-art]

4 — Granular permission model

Three-tier allow | ask | deny, keyed per tool (read/edit/bash/task/skill/webfetch/…) AND per-path AND bash-pattern wildcards ("git *", "npm *"); ~/$HOME expansion; agent-level override; .env denied by default; approval choices once | always | reject.

benchmark: upgrade our imperative canUseTool (spawn.ts:117) to this declarative, path/pattern-scoped model. Cleaner cards, fewer prompts, per-agent policy.

[mega123-art]

5 — Custom commands as markdown

Commands = md in .opencode/commands/ (test.md → /test); YAML frontmatter + prompt body; $ARGUMENTS/$1.., live shell injection !`cmd`, file refs @file, per-command agent / model / subtask: true.

Benchmark: align our skills/slash-commands with this format; subtask+agent ties commands to subagents (comment 3) and fusion.

[mega123-art]

6 — MCP: local + remote + OAuth, per-agent

MCP servers: local (stdio command) and remote (http) with OAuth via RFC 7591 dynamic client registration; tools auto-surfaced alongside built-ins; enable/disable globally or per-agent under agent.<name>.tools.

Benchmark: this is the pattern to finish wiring our marketplace MCP (skill-market/index.ts) into the runtime (see hermes-skill-minting.md §2 — the unwired mcpServers option), incl. per-agent tool scoping to avoid context bloat.

[mega123-art]

7 — Headless run + JSON + export/import

opencode run "prompt" --format json (non-interactive, stream JSON events), --attach http://localhost:4096 (skip cold boot), --session/--continue, export/import sessions as JSON, stats, db.

Benchmark: gives us CI/automation + scheduled cloud agents for free once the server (comment 1) exists. Foundation for the cloud-integration roadmap item.

[mega123-art]

8 — Session share — but on-chain, one better

/share → public URL opncd.ai/s/<id>, syncs history to their servers; auto/manual/disabled + self-host option.

Benchmark the UX, beat the backend: AgentNet shares = wallet-owned, permanent, on-chain (tie to our sessions + reviews/notes), not a vendor server. Shareable agent-session links = a growth/virality loop with our ownership moat.

[mega123-art]

9 — Session fork / branch

Sessions support fork (+ undo/redo, export/import).

Benchmark: forking a run = branching a conversation → directly enables agent-level fusion (run both engines, branch + judge) and DGM-style lineage for evolving skills (ai-power-features.md §5/§10).