Add opencode support

[bhavesh80]

Add opencode support

Bee now runs on opencode the same way it runs on Claude Code: 35 subagents, 12 commands, 20 skills, full SDD lifecycle (brainstorm → spec → architecture → slice loop → review). No runtime bridge, no JS plugin — just generated markdown twins under bee/.opencode/ and a per-file symlink installer.

Demo

30s walkthrough: fresh opencode session → /bee-brainstorm → /bee-sdd → SDD complete with 29 passing tests, all driven through the opencode web UI on Kimi K2.6.

https://github.com/incubyte/ai-plugins/releases/download/untagged-02385493534a294c71e2/bee-opencode-demo.mp4

The demo builds a small CLI URL shortener (tinyurl) end-to-end: brainstorm decides on SHA-256 hex slugs + functional split, SDD triages it as SMALL, slice-coder/slice-tester/sdd-verifier deliver two slices, and the final test suite is green.

What ships

Path	What it is
bee/scripts/generate-opencode-{agents,commands}.py	Generators that transform Claude-side bee/{agents,commands}/*.md into opencode twins under bee/.opencode/. Single source of truth stays on the Claude side.
bee/.opencode/agents/*.md (×35)	Generated agent definitions. Frontmatter rewritten to opencode's schema (description, mode: subagent, category). Body has skill-load preamble, bee:foo → bee-foo cross-references, and Claude-specific tool names mapped to opencode equivalents.
bee/.opencode/commands/*.md (×12)	Generated slash commands. Same transformations.
bee/.opencode/bin/install.sh	Idempotent installer. Per-file symlinks for agents and commands (so they install as bee-<name> cleanly), directory symlinks for skills and scripts. Refuses to overwrite anything not already pointing into the plugin.
docs/opencode-install.md	User-facing install guide, including the one real gotcha (opencode caches config at server startup — restart after install).
docs/opencode-port-findings.md	Full findings doc — opencode's discovery surfaces, the foot-guns hit (frontmatter name: clobbering filename), and the runtime gotchas. Useful background for the next person extending this.

How it works (the short version)

Opencode discovers agents/commands/skills from a small set of fixed config directories (~/.config/opencode/, walked-up .opencode/, $OPENCODE_CONFIG_DIR). Plugin paths are not scanned. So instead of fighting the discovery model, we symlink each markdown file individually from the plugin into ~/.config/opencode/{agent,command,skill}/. Agents and commands get a bee- prefix from the filename (cross-references like Task(subagent_type=bee:foo) are rewritten to bee-foo by the generator); skills keep their bare names.

The Claude-side files in bee/{agents,commands}/ remain the single source of truth. Re-run the generator after editing.

Full discovery walk-through, foot-guns hit, and rationale: see docs/opencode-port-findings.md.

Notes on the model side

The demo was built using Kimi K2.6 through opencode's kimi-for-coding provider. It executed the full SDD pipeline — context-gather, slice-coder, slice-tester, sdd-verifier — with parallel subagent dispatch working correctly and returning clean structured output. The interesting bit isn't really the model: it's that Bee's prompts (skill-loading discipline, agent role separation, AC-driven verification) translate cleanly to whatever model the developer is using. With detailed instructions, an open-source model holds its own on a multi-step structured task.

Bundled refactor: review-orchestration

bee/commands/review.md was 175 lines of orchestration logic. Extracted the body (scope detection, parallel agent spawn, merging, hotspot-prioritized roadmap, dedup rules, output format) into a new bee/skills/review-orchestration/SKILL.md. The command is now ~36 lines of wiring: load skill, spawn 7 specialists, follow skill instructions.

This is the same pattern Bee already uses for brainstorming, clean-code, etc. — keeping commands thin and putting the methodology in skills.

Bundled because it's small, doesn't conflict with the opencode work, and shipping two PRs for one branch felt like extra ceremony. Happy to split if reviewers prefer.

Tested

End-to-end on opencode web (default model: Kimi K2.6):

• ✅ bash bee/.opencode/bin/install.sh symlinks 35 agents, 12 commands, 20 skills, scripts dir
• ✅ Restart opencode → /bee-help runs and lists all commands/skills
• ✅ /bee-brainstorm runs full diverge/converge loop with question tool
• ✅ /bee-sdd "build a CLI URL shortener..." triages, codes, tests, verifies — 29/29 tests passing on the demo project
• ✅ Subagent dispatch via Task(subagent_type=bee-context-gatherer) resolves and runs (the original blocker)
• ✅ /bee-review (the refactored one) loads review-orchestration skill and spawns 7 reviewers in parallel

Claude Code side: unchanged. Source-of-truth files in bee/{agents,commands}/ were edited only where the change benefits both platforms (e.g., the platform-aware skill-loading instructions in help.md).

Out of scope (intentional v1)

• No JS plugin (bee/.opencode/plugins/*.js). The opencode Hooks interface has no agent/command registration hook, so a plugin can't help discovery. v2 could add one for skill-aware autoload or intercepting subagent dispatch — open question, not blocking v1.
• Skill names are bare (e.g. clean-code, not bee-clean-code). Renaming would mean ~50 cross-reference updates across both Claude and opencode files. Cosmetic inconsistency accepted; help text was updated to be platform-aware.
• No tests. The generator and installer are simple enough to verify by inspection + the end-to-end smoke test above. If reviewers want pytest coverage of the generator transformations, easy to add.