fix(reporter): read opencode sessions from opencode.db

[nicko170]

Problem

tokenmaxer backfill opencode returns zero rows on any machine running a current opencode, and reports it as success:

$ tokenmaxer backfill opencode
tokenmaxer: backfilled 0 opencode row(s)
tokenmaxer: backfill complete — 0 row(s) total

opencode >= 1.x stores messages in a SQLite database at <data-dir>/opencode.db. The reporter only knows the older layout — one JSON file per message under <data-dir>/storage/message/<sessionID>/. On opencode 1.18.15 that directory does not exist at all:

$ ls ~/.local/share/opencode
auth.json  log  opencode.db  opencode.db-shm  opencode.db-wal  repos  snapshot  tool-output

So readdirSync(root) throws, collectOpencodeRows hits its catch { continue }, and the source silently contributes nothing. opencode-sessionstart / opencode-sessionend are no-ops for the same reason, so ongoing reporting is affected too, not just backfill.

Fix

Read opencode.db when it is present, and keep the legacy tree as a fallback so machines that have not upgraded — or that still hold pre-upgrade sessions alongside the database — keep reporting. Sessions found in the database win over a same-id legacy directory.

The message.data column holds the same message object the JSON files did (role, modelID, tokens.{input,output,reasoning,cache.{read,write}}, time.created), so parseOpencodeMessages is reused unchanged and both layouts share one parser.

Notes:

• node:sqlite is already used for Cursor's state.vscdb, so no new dependency.
• The database is opened readOnly, and an unreadable file or unexpected schema returns null to fall through to the legacy path rather than throwing.
• sinceMs is applied as a WHERE time_created >= ? predicate, matching the mtime filter the file path uses.

Verification

pnpm check passes — lint, format, typecheck, 374 tests across 42 files.

New src/__tests__/reporter-opencode-db.test.ts covers per-session/per-model summing, started_at being the earliest message, sinceMs filtering, the legacy layout still working, and the empty case. The tests seed a real SQLite database and pin OPENCODE_DATA_DIR/XDG_DATA_HOME/HOME at a temp dir so they do not read the developer's own opencode store.

End-to-end against a real 25-session opencode.db, the reporter's totals match the aggregates opencode maintains independently in its own session table:

	reporter	opencode session table
sessions	25	25
input	255,090,316	255,090,316
output	962,656	962,656

Possibly worth a follow-up

A source whose data directory is entirely absent is indistinguishable from one with no recent activity — both print backfilled 0 row(s) and exit 0. That is what made this quiet for so long. Happy to add a warning when no candidate root exists for a requested source, if you'd want it.