Skip to content

OpenCode Community Workflows, Skills, and Videos

OpenCode Community Workflows, Skills, and Videos

Section titled “OpenCode Community Workflows, Skills, and Videos”

Researched 2026-10-20. OpenCode’s useful community advantage is the ability to customize model routing, agent roles, tools, and context management. The material reviewed here does not establish a general coding-quality win over current Claude Code or Codex. It supplies concrete workflows worth testing.

The research checked creator repositories, published-package revisions, primary documentation, and selected English caption passages from four YouTube videos. Plugins were inspected as source; they were not installed or live-qualified on the audited Mac. Dates and version boundaries matter: the earlier videos show older OpenCode and competing tools.

For current native V2, use opencode --auto for automatic approval of asks and opencode --continue to reopen the last session. Combine them as opencode --auto --continue. For a specific session, use opencode session list, then opencode --session SESSION_ID; a scripted fork uses opencode run --session SESSION_ID --fork "Try another approach".

Explicit deny rules remain active under --auto. The old --attach and --dangerously-skip-permissions examples in videos are not 2.1.0 flags. See the checked native flag table for headless output, model selection and private-server options, and the four-harness comparison for Claude, Codex and Vibe equivalents. This flag check is dated 2026-10-21; the research observations below retain their original date.

Use it when a project benefits from selecting provider/model pairs for different jobs, editing the orchestration layer, or adding a context-management plugin. Keep requirements, source discovery, implementation, and review as distinct responsibilities when that improves the task. A simple change can stay with one agent.

These practices are not exclusive to OpenCode. Current Codex subagents support role-specific models and reasoning effort, although a custom agent’s model_provider does not override the parent. Claude subagents also support model selection and tool restrictions. Portable skills can improve all three harnesses. Compare actual workflows rather than an old feature checklist.

ProjectUseful materialFit for an initial experiment
Oh My OpenCode SlimOrchestrator and specialists, model presets, worktree guidance, repository maps, verification planningMost directly relevant orchestration candidate for native OpenCode; compare with a minimal baseline first
Dynamic Context PruningSelective compression, duplicate-call cleanup, protected contextUseful for tool-heavy sessions; measure cache misses and recollection failures as well as tokens removed
SuperpowersBrainstorming, plans, debugging, TDD, verification, worktreesChoose a focused workflow; current source has an explicit V2 adapter
Matt Pocock’s skillsRequirements interviews, specs, tickets, implementation, reviewBorrow small composable skills and their dependencies across harnesses
Addy Osmani’s agent-skillsSpec-driven development, incremental implementation, debugging, code qualityUseful portable engineering guidance with explicit OpenCode setup notes
Omer Hamerman’s dotfilesConcrete lead, requirements, implementation, testing, and editor setupStudy the role boundaries; adapt older tool/config fields to current V2
OpenSpecVersioned proposals, specifications, and tasks with per-tool adaptersUseful when durable requirements and cross-harness continuity matter

OpenCode Slim: a credible candidate, not a free upgrade

Section titled “OpenCode Slim: a credible candidate, not a free upgrade”

The registry’s latest version was 3.0.5 at inspection. Its published source includes a V2 server export and a compatibility guide. That guide records live verification against OpenCode 2.0.7 and an older plugin baseline; this does not establish current 2.1.0 compatibility on our machine.

Its bundled skills include codemap, verification-planning, step-back, deepwork, and worktrees. Those are useful ideas even without adopting the entire plugin: map the relevant source, plan the evidence, and reassess repeated failed patches. Slim also changes the agent catalog and default routing, so inspect the resulting roles rather than assuming native defaults remain identical.

The inspected published version was 3.2.0. Its entry point includes a V2 setup adapter even though legacy plugin peers remain in package metadata. Source inspection is stronger evidence than guessing compatibility from the dependency name, but still does not prove runtime behavior.

The maintainer documents the cache trade-off: pruning changes message prefixes, which can invalidate cached input. A smaller prompt can cost more if discounted cache reads become fresh input. Compression may also lose useful detail. Begin with measured manual compression and preserve requirements and verification evidence outside the conversation.

Portable skills need their supporting files

Section titled “Portable skills need their supporting files”

Matt’s current grill-me delegates to grilling, and grill-with-docs loads grilling plus domain-modeling. Copying only the small wrapper does not provide the underlying workflow. Its TDD skill also references codebase-design. Adapt the dependency set and tool vocabulary together.

Addy’s OpenCode setup guide warns that individual skill copies may omit shared references. For example, incremental implementation refers to a shared definition of done. That guide also says the optional .opencode/commands directory is not included on the reviewed main branch. Check paths before copying commands. Do not replace a project’s AGENTS.md with the skill repository’s own development instructions.

Superpowers’ current OpenCode instructions distinguish V1 and V2, identify a V2 minimum, and document pinning and cache behavior. Use that maintained path rather than an old symlink tutorial. Its process can also add unnecessary ceremony to small tasks; select proportionate steps.

The original project is now Oh My OpenAgent. Current main describes OmO running on senpi, its Pi fork. OpenCode adapter code still exists, and recent release notes describe OpenCode 2 loading, but installing the current standalone product is a different choice from adding an OpenCode plugin.

Old videos, forks, development branches, package names, and standalone editions can describe different products. The inspected registry package oh-my-opencode was 5.1.40. Check the exact release and adapter instructions before changing an existing setup. This research does not qualify its migration tooling or carry provider credentials into it.

DevOps Toolbox: a practical personal workflow

Section titled “DevOps Toolbox: a practical personal workflow”

My Opencode Workflow As A Senior Engineer, 2026-03-13, 19:16. Start at 5:47 for agent roles, 13:41 for the team, and 15:30 for limitations.

Omer describes a lead coordinating requirements, backend implementation, tests, and review. He links his dotfiles, allowing the prompts to be inspected. He also says teams cost more tokens, oversight becomes awkward, and simple work should remain simple. His reported experience is personal and provisional, not a controlled comparison. The role boundaries are more transferable than the older configuration syntax.

Darren Builds AI: watch the correction loop

Section titled “Darren Builds AI: watch the correction loop”

Oh My OpenCode Is Actually Insane, 2026-01-01, 23:56. Watch the application build, then 20:30 onward for his caution about simple tasks and remaining corrections.

The captioned demo includes a deployed app whose create-account flow fails, followed by missing frontend work despite an implemented API. It illustrates the need for browser acceptance and iteration. The creator advises against this plugin for simple jobs. It is a useful historical orchestration demonstration; it does not establish that the current OpenCode/plugin pair outperforms other harnesses. Its old authentication setup should not be copied into a current installation.

mattpocock/skills: A complete AI Coding workflow, end-to-end, 2026-07-16, 17:17. Useful sections start at 8:31, 10:08, and 13:51.

This demonstration uses Claude Code, and the creator explicitly describes the skills as portable. The transferable sequence is clarification, a durable spec, manageable tickets when needed, implementation, and review against the spec. His numeric context “smart zone” is a personal heuristic, not a universal model limit. The current skill repo has evolved since the recording; resolve today’s dependencies.

OpenCode Review: Because I Hit Claude Code Usage Limits, 2026-01-11, 24:22. See 11:15 for the workflow and 22:55 for the conclusion.

The creator prefers OpenCode’s interface and customization over the then-current Codex CLI, while still praising Claude’s planning. The conclusion does not recommend switching away from Claude on quality grounds. LSP and competing-agent capability claims are version-bound: native OpenCode V2 currently has no active LSP tooling. This is a historical user-experience comparison, not current benchmark evidence.

A useful counterexample is the public same-model harness benchmark. It includes a runner, fixtures, results, routing corrections, and limitations. Its tested OpenCode configurations were slower or failed some cases while other harnesses completed them. The study is small, environment-specific, and not pristine defaults; some early Claude runs used the wrong upstream model and are separated from corrected runs. Child-session token accounting also needs care. We inspected the repository and did not reproduce its results.

That evidence is enough to reject a blanket superiority claim, not to rank every current harness. A high cache-hit percentage can coexist with high total context throughput. Another creator’s edit-format experiment shows why the tool interface matters, with gains and regressions varying by model. It is not an OpenCode-versus-Codex benchmark.

Keep the working Claude-through-Go lane as a reference. Compare native OpenCode on the same compatible Go model, then add one portable workflow or one plugin at a time. If comparing default Codex/Claude models too, label that as a whole-system comparison rather than a harness-only result.

Use a seeded bug fix, an already-fixed no-op task, a review with known defects, and a small UI change with browser acceptance. Hold the baseline, task, account, reasoning/output settings, relevant instructions, and execution boundary as close as practical. Record differences instead of claiming identical conditions when tools or prompts differ.

Measure accepted results, scope mistakes, time, total parent-and-child input/output, cache reads/writes, quota/spend, retries, and human correction. Verify the actual served model so fallback routing cannot contaminate the comparison. Repeat promising cases before choosing a default.

The initial adoption order is a recommendation from this research: focused portable skills first, Slim as a separate native experiment, and DCP only after collecting context/cache measurements. Keep one integration owner and start with one or two workers.

Published revisions inspected: Slim 3.0.5 (fc76b4fc), DCP 3.2.0 (d6379815), and the OmO registry metadata for 5.1.40. Repository snapshots included Matt’s skills (5beb35b0), Addy’s skills (aeac41d4), Superpowers (b739a8af), Omer’s dotfiles (97b2213a), OpenSpec (c04f30e5), and the harness benchmark (5f4a6e3c). These are research provenance, not installed-version recommendations.

Recent changes

  1. docs(wiki): add coding agent approval and session flag references @sac
  2. docs(wiki): add source-checked OpenCode community workflows @sac