opencode

Overview ¶

Package opencode implements the event-level adapter for the opencode CLI.

opencode stores per-message usage both in a SQLite database and as JSON files. Under each data directory we read BOTH:

• SQLite "opencode.db" (or the first "opencode-<token>.db") — table `message(id, session_id, data)` where `data` is the message JSON.
• JSON files under "storage/message/**/*.json" (the same shape).

Both carry the same per-message `data` payload:

{id, sessionID, providerID, modelID, time:{created:<ms>},
 tokens:{input, output, reasoning, cache:{read, write}, total},
 cost, path:{cwd, root}}

Token mapping (opencode reports cache read/write as separate buckets, like Anthropic): Input=tokens.input, Output=tokens.output, CacheCreation=tokens.cache.write, CacheRead=tokens.cache.read, Reasoning=tokens.reasoning, and Total is reconciled against tokens.total via tokenutil.ApplyTotalFallback.

Reasoning is normalized to ADDITIVE output. Current writers subtract it from output; older writers retained it inside output. A positive total that exactly accounts for input/output/cache identifies that overlap (see buildEvent).

The persisted dedup key is "opencode|<message id>", so the SQLite row and the JSON file for the same message collapse to one stored event (DB is discovered first, so it wins on INSERT OR IGNORE).

CRITICAL: strictly read-only. JSON files are opened O_RDONLY; the database is opened with a read-only DSN (mode=ro plus query_only(1)) — never immutable=1, because opencode writes this database live and keeps a large WAL an immutable reader cannot see (see collectDB). Nothing under the agent's directories is created, locked, or modified.