[3intermute]

Problem

OpenCode v2 replaced the SQLite schema: sessions moved from session/message to session_v2/session_message (same opencode.db). The collector's query fails with no such table: session; run_query() returns None and the failure is swallowed, so on OpenCode v2 abtop silently shows zero sessions — no error in the TUI. v0.5.5 (latest release) and current main are both affected.

This bit the wider ecosystem the same way (ccusage #1644, tokscale #1155).

Fix (one file, src/collector/opencode.rs)

• Query session_v2 joined to project; read token totals from the row (tokens_input/tokens_output/tokens_cache_read/tokens_cache_write) instead of json_extract aggregation over the message table. One sqlite3 fork per refresh instead of two, and no scan of message blobs.
• session_v2.model is a JSON blob ({"id":"glm-5.3","providerID":"zai-key2",...}), not a plain string — extract $.id/$.providerID so the TUI keeps rendering provider/model; both still go through sanitize_db_field.
• Filter parent_id IS NULL: in a long-lived DB 84% of session_v2 rows (4599/5474 here) are subagent child sessions that never map to a user-launched process — without the filter they flood the top-20 and break PID claiming.
• Turn count from session_message rows with type='assistant', served by the covering index (session_id, type, seq).

PID↔cwd matching, status derivation, and sanitization are unchanged.

Verification

On a live 2.7GB WAL DB (OpenCode 2.2.5, actively written):

• The exact query runs in ~27ms (EXPLAIN QUERY PLAN confirms the covering index; outer scan is ~880 parent rows).
• abtop --json before: "sessions": []. After: all 7 running opencode TUI processes matched by cwd, correct zai-key2/glm-5.3 model strings, tokens/cache/turn counts populated.
• cargo test (222 passed) and cargo clippy -- -D warnings clean.

Note: context-% for OpenCode still over-reports for long sessions because it divides cumulative lifetime tokens by the window — pre-existing behavior, unchanged here.