OpenCode conversation identity

このトピックはまだ日本語に翻訳されていないため、英語の原文を表示しています。 英語の原文を読む

OpenCode and OpenCode Go reuse explicit conversation IDs across turns. The resolver checks x-opencode-session, existing affinity/session headers, native CLI session/thread headers, request metadata, and then top-level session/thread fields. Claude’s JSON-encoded metadata.user_id is inspected for a session ID; an opaque account/user ID alone is never treated as a conversation ID.

Values containing control characters, empty IDs and IDs longer than 256 characters are ignored. JSON metadata parsing is bounded. Before translation, the executor header normalizer preserves this identity only for the OpenCode providers; it does not inject OpenCode headers for other providers.

The existing canonical ses_ encoding and conversation fingerprint fallback remain unchanged. With no explicit ID, the first user message and other existing fingerprint inputs continue to determine continuity. This is not a new authorization boundary or proof of upstream content isolation.

The implementation lives in open-sse/utils/opencodeSessionIdentity.ts, open-sse/utils/opencodeHeaders.ts and open-sse/handlers/chatCore/executorClientHeaders.ts. Regression tests cover native IDs, Claude metadata, precedence, invalid values and the existing fingerprint behavior without paid upstream calls.

---

undefined (a58000c7685f)