OpenCode¶

Conversation database¶

opencode.db is a relational session → message → part schema (Drizzle). A conversation turn is reconstructed by joining a part row up to its message (for the role) and session (for the title and working directory). User text parts participate in prompt scope under --exhaustive; assistant and reasoning parts require --scope conversations or --scope all.

session table — one row per session:

message table — id, session_id (FK), and a data JSON column. Assistant messages carry a top-level modelID/providerID/path.cwd; user messages instead nest the selected model under model.modelID, so agentgrep reads modelID and falls back to model.modelID:

{"role": "assistant", "modelID": "...", "providerID": "...",
 "time": {"created": 1779999665000}, "path": {"cwd": "..."}}

part table — id, message_id (FK), session_id, and a data JSON column holding one content part. The searchable text lives here:

A part’s kind is derived from the joined message role (user → prompt, otherwise history). Tool, file, snapshot, patch, and step-marker parts are metadata and stay outside default search. Message timestamps are unix-milliseconds and are normalized to ISO-8601.

The same opencode.db file also carries OpenCode’s v2 event-sourced tables — session_input, session_message, event/event_sequence, and todo. The event table is now populated on stable installs and mirrors the part transcript text, so agentgrep leaves it unsearched to avoid duplicate hits; the canonical transcript stays in session/message/part. The secret-bearing account, account_state, control_account, and credential tables are present but never enumerated — the adapter reads only text-bearing part rows.

The legacy pre-migration layout (one JSON file per session, message, and part under storage/) is documented but no longer searched — current installs migrate it into opencode.db on startup.

Prompt history¶

opencode.prompt_history is OpenCode’s recall log, one JSON object per line:

{"input": "Review [Pasted ~43 lines]", "mode": "normal",
 "parts": [{"type": "text", "text": "...",
            "source": {"text": {"start": 7, "end": 25,
                                "value": "[Pasted ~43 lines]"}}}]}

input is the prompt as typed. A pasted block shows there as a placeholder, and its parts entry holds the pasted text with the placeholder’s span; agentgrep splices the paste back in, so pasted text is searchable. mode is kept as metadata.

The log has no timestamp and no session id, and agentgrep invents neither. Its records sort after dated ones, a date filter excludes them, and has:timestamp is false for them. Plain search reads this log alone, so a prompt appears once. Under --exhaustive or --scope all, opencode.db is read too, and a prompt present in both appears once per store, the same way Codex’s prompt history stands beside its session transcripts and Claude Code’s history.jsonl beside its project transcripts. The log is not a copy of the database: prompts recalled there can be absent from opencode.db entirely.

1.19.4¶

Observed 2026-10-25: no store changes since 1.18.30 (2026-09-11).

1.18.30¶

Observed 2026-09-11: no store changes since 1.18.15 (2026-08-08).