OpenCode: reimplemented Claude Code hooks/plugin as OpenCode plugin + skill + commands

[cl0ckt0wer]

Summary

Reimplemented agentmemory's 12 Claude Code hooks + 4 skills as an OpenCode plugin, skill, and commands. This gives non-Claude-Code users the auto-capture pipeline: tool calls → observations, session lifecycle, compaction memory injection, and manual recall/remember commands.

What was built

1. OpenCode Plugin (~/.config/opencode/plugins/agentmemory-capture.ts)

Claude hook	OpenCode hook	agentmemory API
SessionStart	session.created	POST /session/start
Stop	session.idle	POST /summarize
SessionEnd	session.deleted	POST /session/end
PostToolUse	tool.execute.after	POST /observe
PostToolUseFailure	session.error	POST /observe
UserPromptSubmit	message.updated (user role)	POST /observe
PreCompact	experimental.session.compacting	GET /context → output.context

2. Session Recall Skill

Updated skills/agentmemory-setup/SKILL.md to instruct the agent to call memory_smart_search and memory_lessons_recall at session start, with example queries.

3. Commands

Created .opencode/commands/recall.md and remember.md — wrap memory_smart_search and memory_save MCP tools with usage guidance.

Gaps (not yet implemented)

Claude feature	Reason
PreToolUse (file enrichment)	OpenCode tool.execute.before doesn't have stdout→context pipe; also disabled by default (#143)
SubagentStart/Stop	OpenCode Task tool lifecycle not exposable via plugin events
Notification / TaskCompleted	Claude-specific events with no OpenCode equivalent
Claude Memory Bridge (MEMORY.md sync)	OpenCode uses AGENTS.md; could be implemented as a skill
/forget and /session-history commands	Lower priority; can add later

PreCompact implementation detail

OpenCode's experimental.session.compacting hook has output.context: string[] — this maps directly to Claude's PreCompact stdout injection. The plugin calls GET /agentmemory/context and pushes the result into output.context.

Plugin registration

{ "plugin": ["./plugins/agentmemory-capture.ts"] }

Environment

• agentmemory: v0.9.4 (Docker, iii-engine 0.11.6)
• OpenCode + @opencode-ai/plugin v1.1.25
• Full stack: Ollama embeddings + OpenRouter LLM + HTTP→stdio MCP bridge

[rohitg00]

Thanks @cl0ckt0wer — comprehensive writeup, the per-hook mapping table is exactly the kind of compatibility audit we need for every new agent surface.

This is being implemented in:

• feat: OpenCode plugin with 22 auto-capture hooks (closes #236, #244) #237 — full OpenCode plugin (22 hooks, two slash commands, plugin README with the same gap analysis you wrote here)
• fix(plugin): add session instruction injection and consolidation pipeline (closes #233) #239 — three follow-up gap fixes (system-prompt instruction injection via experimental.chat.system.transform, consolidation pipeline on session.deleted, the clco→oco port sweep)

I've asked over there to consolidate the two PRs into one (they touch the same files), but either way this issue gets fully addressed.

Re your "Gaps (not yet implemented)" table — those are tracked separately:

• PreToolUse / file enrichment: tool.execute.before lacking a stdout→context pipe is a real upstream limitation. We'll keep the hook in the plugin but it'll be a no-op until OpenCode ships an equivalent. Filed as a follow-up consideration once feat: OpenCode plugin with 22 auto-capture hooks (closes #236, #244) #237 / fix(plugin): add session instruction injection and consolidation pipeline (closes #233) #239 land.
• SubagentStart/Stop: same — upstream Task-tool lifecycle isn't exposable. Skip until OpenCode adds events.
• MEMORY.md (AGENTS.md) bridge: agreed, this is a skill, not a hook. Reasonable for the next OpenCode pass.
• /forget and /session-history: low-priority slash commands; bundle into the next plugin update.

Holding open until #237 + #239 land. Once they do, will close as fixed-by-#237/#239 with full credit to your audit work here.