agsearch indexes Claude Code and Codex CLI today. Other agents that keep transcripts on disk can be added the same way: a parser plus an entry in sources, with no change to search or ranking.
Gemini CLI
Sessions live at:
~/.gemini/tmp/<project>/chats/session-<timestamp>-<id>.json
One JSON object per session:
sessionId, projectHash, startTime, lastUpdated, kind, messages[]
messages[]: id, timestamp, type, content
That maps onto the existing 8-field row schema almost directly.
Read the format from source, not from samples. Gemini CLI is open source, so the schema is defined rather than inferred:
packages/core/src/services/chatRecordingTypes.tspackages/core/src/services/chatRecordingService.ts
There are a couple of small third-party history viewers, but they are reverse-engineered and a weaker reference than the types above.
Worth having a handful of real sessions across a few projects to test against before merging.
opencode
sst/opencode persists sessions through a storage/ module (packages/opencode/src/storage/storage.ts, schema.ts). The on-disk layout is not confirmed yet, but it is open source, so the same approach applies.
Gemini CLI first: its format is already confirmed.
Out of scope here
Pi — no transcripts on disk. ~/.pi holds agent/auth.json and nothing else, so there is nothing to index. Worth revisiting if that changes.
Cursor — state.vscdb is a large SQLite database (blob-encoded JSON, undocumented schema that moves between releases). Readable with stdlib sqlite3, but it is a second storage model rather than another parser, so it deserves its own issue and its own decision.
Contributions welcome. If you use one of these and want it supported, say so here.
agsearch indexes Claude Code and Codex CLI today. Other agents that keep transcripts on disk can be added the same way: a parser plus an entry in
sources, with no change to search or ranking.Gemini CLI
Sessions live at:
One JSON object per session:
That maps onto the existing 8-field row schema almost directly.
Read the format from source, not from samples. Gemini CLI is open source, so the schema is defined rather than inferred:
packages/core/src/services/chatRecordingTypes.tspackages/core/src/services/chatRecordingService.tsThere are a couple of small third-party history viewers, but they are reverse-engineered and a weaker reference than the types above.
Worth having a handful of real sessions across a few projects to test against before merging.
opencode
sst/opencode persists sessions through a
storage/module (packages/opencode/src/storage/storage.ts,schema.ts). The on-disk layout is not confirmed yet, but it is open source, so the same approach applies.Gemini CLI first: its format is already confirmed.
Out of scope here
Pi — no transcripts on disk.
~/.piholdsagent/auth.jsonand nothing else, so there is nothing to index. Worth revisiting if that changes.Cursor —
state.vscdbis a large SQLite database (blob-encoded JSON, undocumented schema that moves between releases). Readable with stdlibsqlite3, but it is a second storage model rather than another parser, so it deserves its own issue and its own decision.Contributions welcome. If you use one of these and want it supported, say so here.