opencode plugin: auto-title generation consumes the once-per-session memory injection — main conversation never gets context

[liangrud]

Summary

With the OpenCode plugin (plugin/opencode/agentmemory-capture.ts), the memory-context injection ("once per session", gated by contextInjectedSessions) is consumed by OpenCode's auto-title generation request instead of the real conversation's first turn. The main conversation therefore never receives <agentmemory-instructions> or <agentmemory-context>.

Environment

• agentmemory: current main
• OpenCode plugin: plugin/opencode/agentmemory-capture.ts
• Observed in the opencode request flow itself (see Root cause).

Root cause

When a new session gets its first user message, opencode runs two concurrent LLM requests, both firing experimental.chat.system.transform:

1. Auto-title generation — forked in the session loop at packages/opencode/src/session/prompt.ts:1132-1139; calls llm.stream with sessionID: <the same session> (prompt.ts:233) → LLMRequestPrep.prepare → triggers the hook with { sessionID, model } (src/session/llm/request.ts:69-73).
2. The real user turn — processor.ts:640 → same prepare → same hook, same sessionID.

Both carry the same sessionID, and the plugin's contextInjectedSessions gate is keyed only on sessionID. Whichever request reaches prepare first consumes the once-per-session injection — and the title request (tiny system prompt, small model) virtually always wins the race. The plugin cannot tell them apart.

Evidence

Server log (gen_ai.input.messages, two requests for the same first message):

• title request system prompt: You are a title generator... + <agentmemory-instructions> (+ context) ← injected here
• build / main request system prompt: only the base instructions — no <agentmemory-instructions> / <agentmemory-context>

Suggested fix (upstream, opencode side)

Add the agent name to the experimental.chat.system.transform hook input at packages/opencode/src/session/llm/request.ts:70 (and the plugin type in packages/plugin/src/index.ts), e.g. { sessionID, model, agent: input.agent.name }. Plugins can then reliably skip native/internal agents (input.agent === "title", and future internal agents) instead of string-matching the title prompt. This is additive and backward-compatible for existing plugin consumers.

Workaround (plugin side, brittle)

Plugins can skip the title request by matching the title-generator prompt in output.system[0] (e.g. /^You are (?:a|the) title generator/i), but that silently regresses if opencode ever rewrites the title prompt.

Related

• Feature Request: Native Hooks / Auto-Context Injection for OpenCode #431 — the feature this originated from
• Token input cache rate dropped ~10pp due to per-turn system prompt mutation #720 — per-turn system prompt mutation hurting token caching (same hook)