Summary

A session's persisted cumulative_token_usage (in ~/.acpx/sessions/<id>.json) — and therefore getStatus().usage.cumulative — is always {}, even though the agent reports a full per-turn token breakdown every turn. This affects every ACP agent I checked: claude-agent-acp, codex-acp, and cursor-agent. cumulative_cost is captured fine; only the token breakdown is lost.

Root cause

The conversation reducer captures token usage only from usage_update._meta.usage:

• src/session/conversation-model.ts → usageToTokenUsage() reads update._meta?.usage (falling back to top-level input_tokens/…).

But none of the agents put the breakdown there. They emit usage_update as { used, size } (claude adds cost on the final one), and report the actual token breakdown on the session/prompt response under result.usage:

• claude-agent-acp: result.usage = { inputTokens, outputTokens, cachedReadTokens, cachedWriteTokens, totalTokens }
• codex-acp: result.usage = { inputTokens, outputTokens, cachedReadTokens, thoughtTokens, totalTokens } (plus _meta.quota detail)
• cursor-agent: response schema is { stopReason, usage: { inputTokens, outputTokens, cachedReadTokens, cachedWriteTokens, … } }

The data is already available to the runtime — src/cli/output/output.ts consumes it for the printed usage line:

const usageLine = this.formatUsageLine(asRecord(result.usage));

…but result.usage is never folded into the conversation's cumulative_token_usage / request_token_usage. So the persisted record (and getStatus()) shows nothing, while the CLI's own usage line shows the real numbers.

Reproduction

1. Run any prompt with claude-agent-acp, codex-acp, or cursor-agent.
2. Inspect ~/.acpx/sessions/<id>.json → cumulative_token_usage is {} (note cumulative_cost is populated for claude).
3. The session's <id>.stream.ndjson confirms every usage_update is { used, size } (no _meta.usage), while the session/prompt response carries the breakdown under result.usage.

Suggested fix

Fold the prompt response's result.usage into cumulative_token_usage / request_token_usage (mapping cachedReadTokens → cache_read_input_tokens, cachedWriteTokens → cache_creation_input_tokens, thoughtTokens → thought_tokens) — the same result.usage the output line already reads. The existing usage_update._meta.usage path can stay for agents that populate it.