Better OpenCode Efficient Agents

Dev-friendly pack: eight namespaced OpenCode V2 agents—one primary workspace mutator (efficient/ship) and seven focused subagents. No model pins, framework dependencies, or vendor lock-in. Tested against OpenCode v2.0.25.

Quick links: Prompt install · Validate · Agents · Usage analysis

---

At a glance

What	How
Install and activate	Paste the three-line installation prompt into OpenCode
Try them	opencode run --agent efficient/ship "Inspect this repo and propose the smallest verified change."
Check before use	./scripts/check.sh and ./scripts/test-install.sh

---

Why this exists

A read-only, anonymized analysis refreshed on October 13, 2026 covered 164 sessions: 92 roots and 72 children across 14 projects. Testing or verification appeared in 57 roots, review or audit in 45, implementation in 44, and operations or deployment in 37. Of 53 roots that requested verification, 13 lacked recognizable command evidence; 13 of all 92 roots lacked a final parent conclusion. Missing targets or resources caused 47 of 126 failed V2 tool calls.

The smallest response is to discover before acting, keep one implementation owner, use narrow specialists, verify explicitly, and finish with evidence. These aggregate results describe role fit, not the effectiveness of this pack. See docs/usage-analysis.md for the complete counts, method, and heuristic limits.

---

Agents

Agent	Mode	Scope
efficient/ship	primary	Implements, integrates, verifies, and reports; sole editor in this pack
efficient/research	subagent	Code paths, current documentation, patterns, and bounded evidence
efficient/quality	subagent	Correctness, security, regressions, tests, and release readiness
efficient/ui-mobile	subagent	Web/mobile UX, accessibility, responsive behavior, and native conventions
efficient/data	subagent	Schemas, migrations, APIs, transactions, privacy, and rollback safety
efficient/seo	subagent	Crawlability, metadata, schema, performance, content, and visibility
efficient/opencode	subagent	OpenCode agents, skills, tools, permissions, configuration, and integrations
efficient/release-ops	subagent	Approval-gated artifacts, deployments, remote changes, and rollback checks

Specialists are skill-first: each loads the narrowest applicable installed skill, then applies its focused instructions. The pack does not bundle skills or pretend an unavailable integration exists.

These namespaced roles supplement OpenCode's built-in build, plan, general, and explore agents; they do not remove or redefine them. Keep using built-ins for general-purpose work. Choose an efficient/* role when its tighter scope and permission boundary fit the task. Within this pack, ship owns workspace changes and parent synthesis; specialists return evidence rather than edits.

Permission boundaries

All agents start with deny-all, preserve approval prompts for *.env and *.env.*, allow *.env.example, and inherit the active model.

Agent class	Edit	Shell / execute	Skills	External directory
efficient/ship	allow	ask	allow	ask
Six read-only specialists	deny	deny	allow	deny
efficient/release-ops	deny	ask	allow	ask

The six read-only specialists are research, quality, ui-mobile, data, seo, and opencode. Some may fetch public sources where their scope requires it, but none can edit or run arbitrary commands.

Do not use efficient/release-ops under auto-approval. Its shell, provider, deployment, publishing, and other remote mutations must remain individually approval-gated.

---

Prompt install

Copy and paste these three lines into OpenCode:

Install and configure the OpenCode V2 agent pack from https://github.com/anasfik/better-opencode-efficient-agents globally for my user.
Clone or update the repository, run its checks and `./scripts/install.sh --global`, and preserve unrelated agents plus any backup the installer creates.
Run `opencode reload` and `opencode debug agents`, verify all eight `efficient/*` agents are available, then report what changed and any failed check.

---

Validate

./scripts/check.sh        # exact roster, modes, permissions, model inheritance, privacy, syntax
./scripts/test-install.sh # install, upgrade, backup, namespace preservation, locking, rollback
opencode reload
opencode debug agents
opencode run --agent efficient/ship \
  "Inspect this repo and propose the smallest verified change."

---

Safety model

• Only efficient/ship can edit workspace files.
• Only ship and release-ops expose shell or execute; both default to ask. Choose Allow once when prompted, never save broad approval patterns, and never use auto-approval for release-ops.
• Every specialist can load applicable skills without gaining edit access.
• Release, deploy, publish, promote, destructive, account, and production actions require confirmation immediately before the mutation.
• The parent owns synthesis, checks, and the final conclusion. Child success is not parent completion.
• This repository must contain no credentials, session IDs, personal paths, raw prompts, or model pins.

---

Compatibility

OpenCode V2 has no agent package registry. This repository uses namespaced Markdown agents (.opencode/agents/<namespace>/*.md), producing IDs such as efficient/ship. It works with any compatible OpenCode V2 CLI; v2.0.25 is the tested version.

---

What's not in the core pack

• Framework, platform, and vendor expertise belongs in installed skills and integrations, loaded only when relevant.
• Vendor-specific APIs belong in their MCP or integration, not a default agent.
• Session-history mining stays private; public packs should contain only anonymous aggregate evidence.

---

License

MIT. Security reports: follow SECURITY.md. Contributions: CONTRIBUTING.md.