docs(agents): summarize the OpenCode ecosystem (closes #100)

chapters/ai-tools/opencode-ecosystem.qmd

@@ -0,0 +1,304 @@
+[OpenCode](https://opencode.ai) [@opencode_repo]
+is the open-source coding agent harness (MIT-licensed, about 206,000 GitHub stars, measured 2026-09-09)
+that the earlier OpenCode sections of this chapter build on:
+running it against local models (@sec-ai-opencode-ollama),
+against OpenRouter (@sec-ai-opencode-openrouter),
+and under the Oh My OpenCode multi-agent framework (@sec-ai-oh-my-opencode).
+Around the harness itself sits a community ecosystem of plugins, clients, and agent bundles,
+which the OpenCode maintainers index on a single documentation page [@opencode_ecosystem].
+This section maps that page (as of 2026-09-09) and calls out the entries most relevant to us.
+Where the site already covers a project, this section points there instead of repeating it.
+
+#### How the ecosystem is organized
+
+The ecosystem page groups its entries into three lists [@opencode_ecosystem]:
+
+- **Plugins** (38 entries):
+  JavaScript or TypeScript modules that OpenCode loads at startup and that hook into its lifecycle events
+  (tool execution, file edits, session compaction, permissions, notifications, and so on).
+  A plugin is named in the `plugin` key of `opencode.json` as an npm package,
+  or dropped as a file into `.opencode/plugins/` (project) or `~/.config/opencode/plugins/` (global);
+  npm packages are fetched automatically with Bun [@opencode_plugins].
+- **Projects** (11 entries):
+  clients and integrations built on the OpenCode server API or SDK ---
+  editor front ends, web and desktop apps, a Discord bot, an Obsidian plugin, and an extension manager.
+- **Agents** (2 entries):
+  bundles of agent definitions, prompts, and commands that reshape how OpenCode plans and executes work.
+
+The issue that prompted this section also asked about **providers**.
+The ecosystem page has no provider list;
+model providers are configured in `opencode.json` and documented separately,
+and the two first-party paid options are covered under "OpenCode Zen and OpenCode Go" below.
+
+A recurring caveat applies to almost every entry:
+these projects are independent, and most carry an explicit
+"not built by the OpenCode team and not affiliated with OpenCode" disclaimer
+(for example `ocx` [@ocx] and `opencode-worktree` [@opencode_worktree]).
+Listing on the ecosystem page is a courtesy index, not an endorsement or a security review.
+Every plugin runs with the same permissions as OpenCode itself,
+so treat adding one the way you would treat adding any unaudited dependency
+(see @sec-ai-benefits-hazards).
+
+#### Notable plugins
+
+Activity figures below (stars, last push, latest release) were read from the GitHub API on 2026-09-09.
+Several repositories have moved since the ecosystem page was written;
+the current location is given where it differs.
+
+**Context and token management**
+
+- [`opencode-dynamic-context-pruning`](https://github.com/Opencode-DCP/opencode-dynamic-context-pruning) [@opencode_dcp]
+  (now under the `Opencode-DCP` organization; AGPL-3.0; about 4,200 stars; release `v3.1.15` on 2026-08-16).
+  Reduces token spend by deduplicating repeated tool calls,
+  purging the inputs of failed tool calls after a configurable number of turns,
+  and letting the model compress stale conversation content into summaries.
+  The README reports prompt-cache hit rates of roughly 85% with the plugin versus 90% without,
+  and says development focus has shifted to a separate tool called Sleev.
+  *Useful to us?* Yes, for long ARDI sessions on paid metered providers;
+  the AGPL license is irrelevant for a locally-run tool but worth knowing.
+- [`opencode-morph-plugin`](https://github.com/morphllm/opencode-morph-plugin)
+  (Morph, the vendor; MIT; 85 stars; release `v2.0.17` on 2026-09-07)
+  and its community predecessor [`opencode-morph-fast-apply`](https://github.com/JRedeker/opencode-morph-fast-apply)
+  (MIT; 170 stars).
+  Both route edits through Morph's hosted "Fast Apply" model to cut edit latency,
+  which means a second paid API and a second party seeing your code.
+  *Useful to us?* Not by default; the privacy cost outweighs the speed gain for research code.
+
+**Subscription and quota bridges**
+
+These plugins let OpenCode consume a chat subscription you already pay for instead of API credits.
+They are among the most-starred plugins on the page and also the most fragile,
+since they depend on undocumented OAuth flows that the subscription vendors can close at any time:
+
+- [`opencode-openai-codex-auth`](https://github.com/numman-ali/opencode-openai-codex-auth) [@opencode_codex_auth]
+  (Numman Ali; about 2,200 stars; release `v4.4.0` on 2026-01-09; no push since then).
+  Uses the ChatGPT Plus/Pro OAuth flow to expose the GPT and Codex model presets,
+  and its README restricts it to "personal development use" with your own subscription.
+- [`opencode-gemini-auth`](https://github.com/jenslys/opencode-gemini-auth)
+  (MIT; about 1,700 stars; release `v1.4.16` on 2026-05-21).
+  The same idea for a Gemini plan.
+- [`opencode-antigravity-auth`](https://github.com/NoeFabris/opencode-antigravity-auth) [@opencode_antigravity_auth]
+  (MIT; about 11,000 stars; **archived** on GitHub as of 2026-09-09).
+  Authenticated against Google Antigravity's model pool with a Google account,
+  rotating across accounts to stay under quota.
+  Its own README states that using it "violates Google's Terms of Service"
+  and that your account "may be suspended or permanently banned".
+  The similar [`opencode-google-antigravity-auth`](https://github.com/shekohex/opencode-google-antigravity-auth)
+  (375 stars) is archived as well.
+
+*Useful to us?* No.
+Our lab policy is to use delegation budgets through their supported routes
+(the `codex` CLI on the ChatGPT plan, OpenCode Go or Zen, OpenRouter credits),
+and a bridge whose README concedes a terms-of-service violation is a liability, not a saving.
+The Codex and Gemini bridges are less clearly prohibited but share the fragility.
+
+**Sandboxing, isolation, and worktrees**
+
+- [`opencode-daytona`](https://github.com/daytona/integrations/tree/main/packages/opencode-plugin) [@opencode_daytona]
+  (Daytona, the vendor; Apache-2.0; release `opencode-plugin-v0.192.1` on 2026-09-03).
+  Runs each session in a hosted Daytona sandbox,
+  syncs the sandbox to a local branch named `opencode/<n>`,
+  and produces live preview links when a server starts.
+  Requires a Daytona account and API key.
+  *Useful to us?* Maybe, for untrusted-code experiments;
+  otherwise the firewall and container setups in this chapter's Firewall and Network Configuration section cover the same ground.
+- [`opencode-devcontainers`](https://github.com/athal7/opencode-devcontainers)
+  (MIT; 222 stars; release `v0.5.1` on 2026-08-24; pushed 2026-09-08).
+  Runs several devcontainer instances at once, one per branch, with auto-assigned ports.
+  *Useful to us?* Yes, for repositories that already ship a `.devcontainer/`.
+- [`opencode-worktree`](https://github.com/kdcokenny/opencode-worktree) [@opencode_worktree]
+  (MIT; 726 stars; no tagged release; pushed 2026-08-17).
+  Adds `worktree_create` and `worktree_delete` tools that create an isolated git worktree
+  under `~/.local/share/opencode/worktree/`, sync configured files into it,
+  and open a terminal with OpenCode running there.
+  *Useful to us?* Yes; it is the OpenCode counterpart of the per-agent worktrees our Claude Code sessions use.
+
+**Orchestration and workflow bundles**
+
+- [`oh-my-opencode`](https://github.com/code-yeongyu/oh-my-openagent)
+  (now `oh-my-openagent`; about 68,800 stars; release `v5.0.0-beta.51` on 2026-09-09).
+  Covered in @sec-ai-oh-my-opencode.
+- [`opencode-workspace`](https://github.com/kdcokenny/opencode-workspace) [@opencode_workspace]
+  (MIT; 586 stars; no tagged release).
+  A one-install bundle of 16 components:
+  four plugins (workspace management, async delegation, notifications, git isolation),
+  two npm plugins, three MCP servers (documentation, web search, code search),
+  four specialist agents (researcher, coder, scribe, reviewer),
+  four skill modules, and one command interface.
+  Its sibling [`opencode-background-agents`](https://github.com/kdcokenny/opencode-background-agents) [@opencode_background_agents]
+  (MIT; 382 stars) provides `delegate()`, `delegation_read()`, and `delegation_list()` tools
+  whose results persist to disk under `~/.local/share/opencode/delegations/`,
+  so they survive context compaction;
+  only read-only sub-agents may run in the background,
+  because background sessions sit outside OpenCode's undo and branching system.
+  *Useful to us?* Worth a trial as a lighter alternative to Oh My OpenCode;
+  the disk-persisted delegation results address the same loss-on-compaction problem
+  our lab notebook convention exists for.
+- [`micode`](https://github.com/vtemian/micode) (MIT; 483 stars) and
+  [`octto`](https://github.com/vtemian/octto) (MIT; 493 stars),
+  from the same author:
+  a brainstorm-plan-implement workflow with session continuity,
+  and a browser UI that turns an agent's clarifying questions into multi-question forms.
+  *Useful to us?* Possibly `octto`, which does for OpenCode what `AskUserQuestion` does in Claude Code.
+- [`opencode-conductor`](https://github.com/derekbar90/opencode-conductor)
+  (Apache-2.0; 129 stars; last push 2026-03-02)
+  ports the Context, Spec, Plan, Implement lifecycle described in @sec-ai-conductor-extension to OpenCode.
+  *Useful to us?* Only if that lifecycle is adopted; the port looks dormant.
+- [`plannotator`](https://github.com/backnotprop/plannotator)
+  (Apache-2.0; about 8,600 stars; release `v0.27.12` on 2026-09-03)
+  is a visual plan-and-diff annotation tool with an OpenCode plugin;
+  it is agent-agnostic and also targets Claude Code.
+  *Useful to us?* Yes, for reviewing agent plans before implementation;
+  it addresses the plan-review step directly.
+- [`opencode-goal-plugin`](https://github.com/willytop8/OpenCode-goal-plugin)
+  (MIT; 252 stars; release `v0.10.0` on 2026-09-07)
+  adds a session-scoped `/goal` that keeps an objective in context
+  and auto-continues until "evidence-gated completion".
+  *Useful to us?* Maybe, for unattended ARDI-style loops.
+
+**Safety and observability**
+
+- [`opencode-vibeguard`](https://github.com/inkdust2021/opencode-vibeguard) [@opencode_vibeguard]
+  (MIT; 198 stars; no tagged release; last push 2026-03-01).
+  Replaces configured secrets and personal data with placeholders of the form `__VG_<CATEGORY>_<hash>__`
+  before each request leaves for the model provider,
+  restores them locally when output completes,
+  and restores them again before tool execution so shell commands still work.
+  *Useful to us?* Yes in principle, for work touching participant data,
+  but the redaction list is hand-configured and the project has been quiet for six months;
+  test it before relying on it.
+- [`opencode-shell-strategy`](https://github.com/JRedeker/opencode-shell-strategy)
+  (MIT; 136 stars) and [`opencode-pty`](https://github.com/shekohex/opencode-pty) (MIT; 571 stars)
+  tackle the same hang: an agent running a TTY-dependent command with no terminal attached.
+  The first injects instructions to use non-interactive flags;
+  the second gives the agent a real pseudo-terminal it can write to and read from.
+  *Useful to us?* Yes; the hang they fix is one we hit regularly with `R` and `quarto preview`.
+- [`opencode-sentry-monitor`](https://github.com/stolinski/opencode-sentry-monitor) (53 stars),
+  [`opencode-wakatime`](https://github.com/angristan/opencode-wakatime) (198 stars),
+  and [`opencode-helicone-session`](https://github.com/H2Shami/opencode-helicone-session) (16 stars)
+  send traces or usage to Sentry, Wakatime, and Helicone respectively.
+  *Useful to us?* No; we do not run those services.
+
+**Notifications and quality of life**
+
+Four plugins do desktop notifications
+(`opencode-notifier`, 810 stars, is the most-used; `opencode-notify`, `opencode-notificator`),
+and the rest of the list is small conveniences:
+`opencode-md-table-formatter` cleans up model-generated Markdown tables,
+`opencode-zellij-namer` names Zellij sessions,
+`opencode-supermemory` (about 1,600 stars) adds cross-session memory through the hosted Supermemory service,
+and `opencode-scheduler` runs recurring jobs through `launchd` or `systemd`.
+Vendor plugins from JFrog, Firecrawl, and Tavily wrap their own CLIs.
+*Useful to us?* A notifier, yes; the rest only with a specific need.
+
+#### Notable clients and integrations
+
+- [`CodeNomad`](https://github.com/NeuralNomadsAI/CodeNomad) [@codenomad]
+  (Neural Nomads AI; MIT; about 2,600 stars; release `v0.19.0` on 2026-08-24; pushed 2026-09-09).
+  Describes itself as "The AI Coding Cockpit for OpenCode":
+  a SolidJS front end with a Node.js server that wraps an OpenCode CLI already on your `PATH`.
+  It ships as Electron and Tauri desktop builds for macOS, Windows (x64 and ARM64), and Linux,
+  and as a password-protected local web server
+  (`npx @neuralnomads/codenomad --password <password> --launch`) for remote or browser access.
+  Features include multi-instance workspaces, session management, git worktrees,
+  voice input, a file browser, a command palette, and "SideCars" that embed local web tools as tabs.
+  Despite the ecosystem page's "Desktop, Web, Mobile and Remote" description,
+  the README documents no native mobile app; mobile access is through the web server.
+  *Useful to us?* Yes, as the most complete graphical front end for OpenCode
+  and a plausible answer to "I want several OpenCode sessions side by side without `tmux`".
+- [`OpenChamber`](https://github.com/openchamber/openchamber) [@openchamber]
+  (now under its own organization; MIT; about 9,700 stars; release `v1.22.2` on 2026-09-05).
+  A workspace for running and reviewing agent work on desktop, web, VS Code, iOS, and Android,
+  with session goals, a "multi-run" mode that runs the same task across up to five models,
+  a guided changes walkthrough for large diffs, scheduled tasks,
+  and an end-to-end encrypted "Private Relay" for remote connections.
+  Like CodeNomad it uses OpenCode as the engine and is not affiliated with the OpenCode team.
+  *Useful to us?* Yes; the multi-model run is a direct fit for our habit of getting a second model's review.
+- [`OpenWork`](https://github.com/different-ai/openwork) [@openwork]
+  (different-ai; about 23,400 stars; release `v0.18.44` on 2026-09-09).
+  Positions itself as "an open-source alternative to Claude Cowork and Codex" for macOS, Windows, and Linux,
+  sharing skills, MCP servers, and connected services across tools and machines.
+  Code outside `ee/` is MIT; the organizational control plane under `ee/` uses a separate license
+  that is free for up to five users.
+  *Useful to us?* Compare against the collaborative workspaces in @sec-ai-collaborative-workspaces;
+  the five-user free tier fits a lab, and the open core makes it auditable.
+- [`opencode.nvim`](https://github.com/nickjvandyke/opencode.nvim) [@opencode_nvim]
+  (Nick van Dyke; MIT; about 3,800 stars; release `v1.0.0` on 2026-08-20)
+  connects Neovim to a running OpenCode server (`opencode --port`),
+  injects editor context (cursor, selection, diagnostics) into prompts,
+  surfaces OpenCode's server-sent events as Neovim autocommands,
+  and reloads buffers when the agent edits files.
+  A second, unrelated [`sudo-tee/opencode.nvim`](https://github.com/sudo-tee/opencode.nvim)
+  (Apache-2.0; 937 stars) is a full Neovim front end rather than a bridge.
+  *Useful to us?* For the Neovim users in the lab, yes.
+- [`kimaki`](https://github.com/remorses/kimaki) [@kimaki]
+  (MIT; about 1,400 stars; release `kimaki@0.27.0` on 2026-09-01).
+  A Discord bot in which each project is a channel and each session a thread;
+  it queues messages, forks sessions, transcribes voice messages, shows diffs,
+  and maps Discord roles to permission controls.
+  *Useful to us?* No for now; our coordination runs through GitHub, not Discord.
+- [`portal`](https://github.com/hosenur/portal) [@opencode_portal]
+  (MIT; 798 stars; last push 2026-05-12) is a mobile-first web UI meant to be reached over Tailscale.
+  [`OpenCode-Obsidian`](https://github.com/mtymek/opencode-obsidian) (MIT; about 1,100 stars)
+  embeds OpenCode in Obsidian's sidebar.
+  *Useful to us?* Niche; CodeNomad and OpenChamber cover the remote-access case with more activity.
+- [`ocx`](https://github.com/kdcokenny/ocx) [@ocx]
+  (MIT; 942 stars; release `v2.0.15` on 2026-08-17).
+  An extension manager with portable, isolated profiles:
+  `ocx profile add` installs a profile from a registry,
+  `ocx oc -p <name>` launches OpenCode with it,
+  and components are copied into `.opencode/` (the shadcn model) rather than hidden in dependencies,
+  with SHA verification of what is installed.
+  *Useful to us?* Yes, if we standardize an OpenCode profile for the lab;
+  it is the closest thing to the plugin marketplace we use for Claude Code (@sec-ai-plugins-deep-dive).
+- [`ai-sdk-provider-opencode-sdk`](https://github.com/ben-vargas/ai-sdk-provider-opencode-sdk)
+  (MIT; 116 stars) exposes OpenCode's configured providers to the Vercel AI SDK,
+  and [`opencode-plugin-template`](https://github.com/zenobi-us/opencode-plugin-template) (archived)
+  was the scaffold for writing plugins.
+  *Useful to us?* Only when building on the SDK.
+
+#### Agent bundles
+
+- [`OpenAgentsControl`](https://github.com/darrenhinde/OpenAgentsControl)
+  (listed as `opencode-agents`; MIT; about 4,800 stars; release `v0.7.1` on 2026-01-30)
+  is a plan-first framework with approval-gated execution and built-in test, review, and validation steps.
+- [`agentic`](https://github.com/Cluster444/agentic)
+  (MIT; 638 stars; last push 2025-09-02) is a context-engineering toolkit that appears dormant.
+
+*Useful to us?* `OpenAgentsControl` overlaps heavily with Oh My OpenCode (@sec-ai-oh-my-opencode) and `opencode-workspace`;
+pick one such harness rather than layering them.
+
+#### OpenCode Zen and OpenCode Go
+
+The ecosystem page does not describe them, but two first-party paid services sit alongside the community projects:
+
+- **OpenCode Zen** [@opencode_zen] is the maintainers' own model gateway:
+  a curated set of models tested against coding-agent workloads,
+  billed pay-as-you-go per million tokens from a prepaid balance
+  (with optional auto-reload, by default $20 whenever the balance drops below $5).
+  As of 2026-09-09 it lists six free models under limited-time trials,
+  including `Big Pickle`, `MiMo-V2.5 Free`, and `Nemotron 3 Ultra Free`,
+  beside paid Claude, GPT, Gemini, Grok, Qwen, DeepSeek, Kimi, and GLM models.
+- **OpenCode Go** [@opencode_go] is a $10-per-month subscription to 30-plus open-weight coding models
+  (Qwen, DeepSeek, Kimi, GLM, MiMo, Grok, and others)
+  with usage caps expressed in dollar value: $12 per five hours, $30 per week, $60 per month.
+  When a cap is hit, an optional "Use balance" setting falls back to Zen credits.
+  Only one member per workspace can hold the subscription.
+
+*Useful to us?* Yes; OpenCode Go is already one of the lab's delegation budgets,
+and the Zen free tier is a zero-cost way to try a new open-weight model before routing work to it
+(compare the OpenRouter `:free` models in @sec-ai-opencode-openrouter).
+
+#### Summary
+
+| Need | Start with | Also see |
+|---|---|---|
+| Graphical or remote front end | `CodeNomad`, `OpenChamber` | `portal`, `OpenWork` |
+| Multi-agent orchestration | Oh My OpenCode (@sec-ai-oh-my-opencode) | `opencode-workspace`, `OpenAgentsControl` |
+| Cheaper tokens | OpenCode Go, Zen free models | `opencode-dynamic-context-pruning` |
+| Isolation per task | `opencode-worktree`, `opencode-devcontainers` | `opencode-daytona` |
+| Redaction of sensitive data | `opencode-vibeguard` | (none) |
+| Plan review before implementation | `plannotator`, `octto` | `opencode-conductor` |
+| Shareable lab configuration | `ocx` | (none) |
+
+Skip the subscription-bridge plugins.