FAQ

Are the costs Pounce shows my actual bill?

No. Treat every cost in Pounce as representative, not actual. Use your provider’s billing page as the source of truth.

Pounce only ever shows a dollar figure that the coding agent itself reported — it never prices tokens on its own. That still leaves gaps that make the number approximate:

• It covers a single thread, not your account. Other threads, other tools, and other machines aren’t in it.
• It can cover part of a thread. A leading ~ means the figure includes only the turns Pounce drove (see below).
• Agents report cost at different moments and granularities, and some don’t report it at all.
• Subscription plans, credits, discounts, and taxes are applied by your provider after the fact and are invisible here.

Why does my Claude thread show tokens but no cost?

Because Claude Code doesn’t write cost to disk.

It records token counts on every assistant message in its transcript, but the dollar figure exists only on the envelope that closes a live turn — and that envelope is never saved. Read the transcript afterwards and the dollars are simply gone.

So Pounce captures it as turns happen: when the Bridge runs a turn, it banks the agent-reported cost in a local ledger at ~/.pounce/usage.jsonl. That means:

• Turns you ran from Pounce have a real, agent-reported cost.
• Turns you ran in a terminal contribute tokens but no cost — Pounce never saw the envelope.
• A thread that’s a mix of both shows ~ to mark the total as partial.
• A thread with no Pounce-driven turns shows tokens and no cost at all.

Nothing is estimated to fill those gaps. Earlier versions of Pounce multiplied tokens by a built-in price table; that table silently drifted from real pricing and presented a guess as a fact, so it was removed.

What does each agent actually report?

Agent	Tokens	Cost	Also reports
OpenCode	Yes	Yes — full history, in USD	—
Claude Code	Yes — full history	Only for turns Pounce drove	Context window
Codex	Yes — full history	Never	Plan type, rate-limit usage
Cursor	No	No	—

Codex bills against a ChatGPT plan rather than per request, so there is no dollar amount for it to report. Instead of inventing one, Pounce shows what Codex does state — how much of your rate-limit window you’ve consumed, e.g. 120.7M · 26% of 5h.

What are cache tokens, and why is my total so large?

Coding agents re-send your conversation on every turn and rely on prompt caching, so cached reads dominate long threads — tens of millions of tokens is normal and does not mean you were billed at the full input rate. Cached reads are typically far cheaper than fresh input. Pounce includes them in the total because they’re real tokens the agent processed.

Can I export a session?

Yes. The Bridge serves any thread as ATIF (Agent Trajectory Interchange Format), a cross-framework JSON format for agent trajectories:

curl -H "Authorization: Bearer $BRIDGE_TOKEN" \
  "http://localhost:8099/v1/trajectory?agent=claude&thread=<thread-id>&download=1" \
  -o session.atif.json

The export contains the same messages, reasoning, tool calls, and tool results you see in the app, grouped into trajectory steps, plus a final_metrics block. cost_usd appears there only when it’s a real agent-reported figure, so a consumer summing many trajectories won’t accumulate guesses.

This is useful for attaching a session to a bug report, feeding it to an evaluation harness, or handing it to tooling that already understands ATIF.