opencode adapter appends cumulative turn cost once per step, over-counting multi-step turns

[EvolveAegis]

Bug Description

The opencode adapter's message.updated handler (src/adapters/opencode/plugin.ts:544-579) fires on every step-finish within a turn. The event's .cost is the CUMULATIVE turn cost (the comment at plugin.ts:558-564 states "cost is cumulative for the turn"). Each firing calls db.insertEvent(sessionId, usageEvent), which INSERTs a new agent_usage row (src/session/db.ts:1185). The dedup check (checkDuplicate, keyed on type + data_hash over the last 5 events, db.ts:1175-1178) does not hit because the data string includes the per-step token snapshot, which changes each step.

Result: a 3-step turn writes 3 agent_usage rows, each carrying the cumulative cost so far (e.g. $0.02, $0.04, $0.05). The events are forwarded to the platform with cost_usd in the top-level envelope (src/session/extract.ts:1898, 1941), where the natural additive reading sums them to $0.11 instead of the true $0.05.

The code comment claims "re-emitting is idempotent at the cost column", but there is no cost-column update: each firing appends a row. This is the same cumulative-cost class that the task-usage path explicitly guards against by stripping cost_usd and tagging usage_scope="task_cumulative" (src/session/extract.ts:56-59); the opencode adapter reintroduces it.

Steps To Reproduce

1. Configure context-mode with the opencode adapter.
2. Run a turn that triggers multiple steps (e.g. a tool-calling turn with 3 model calls).
3. Observe: agent_usage events in session_events : one row per step, each with the cumulative cost_usd.
4. Sum cost_usd across the rows for the turn (the natural additive reading, and what the platform receives): the total exceeds the turn's actual cost.

Expected Behavior

The turn's cost should be emitted exactly once (or as deltas), so additive aggregation yields the true cost.

Actual Behavior

Multi-step turns emit one cumulative-cost row per step; additive aggregation over-counts by the sum of the intermediate cumulative values.

Environment

• context-mode main (e27c143)
• opencode adapter
• macOS 15

Suggested fix

Emit the usage event once per turn (e.g. only on the final step-finish / message completion), or upsert/replace the previous agent_usage row for the same assistant message, or store per-step deltas instead of cumulative values on multi-fire rows. Update the misleading "idempotent at the cost column" comment either way.

[EvolveAegis]

Dynamic reproduction and citation corrections:

Dynamic PoC (vitest, 1/1 passed) against parseOpencodeUsage + buildAgentUsageEvent at e27c143: two message.updated events for a 2-step turn (step1 cost 0.02, step2 cumulative cost 0.05) produce two events with distinct data_hash (a4522d71 vs 1d411341), so the dedup window misses and db.insertEvent appends two rows. Additive aggregation sums 0.02 + 0.05 = 0.07, over-counting the true turn cost of 0.05 (40%).

Citation corrections (from adversarial draft review):

• extract.ts:56-59 is wrong: the task-usage guard (strip cost_usd + tag usage_scope="task_cumulative") is at extract.ts:1539-1570, not 56-59.
• db.ts:1175-1178 dedup: the actual check is at 1175-1176; DEDUP_WINDOW = 5 is at db.ts:642.
• extract.ts:1898/1941 is the usage-event builder (sets event.cost_usd); the forward-to-platform path is a later MCP path, and insertEvent itself is DB-only. The local additive over-count stands regardless of the platform forwarding.

Suggested fix stands: emit once per turn, or upsert per assistant message, or store deltas; update the misleading "idempotent at the cost column" comment.