Cost X-ray

Produce a ranked, evidence-backed report on where this agent's model spend goes and what to change. ACP sits in the request path, so it sees composition (cache blend, context breakdown, per-tool output bytes, loop share) — not just totals. Your job: pull the data, run the decompositions below, and deliver a report the user can act on before scaling to production.

Data sources (in order of preference)

1. ACP MCP tools if connected: acp_optimize (composition), acp_trace (one run's steps), acp_recommendations (policy findings), acp_cost (spend by model).
2. REST fallback with the key in ~/.acp/credentials (or ~/.acp/proxy-key):
   • GET https://api.agenticcontrolplane.com/api/v1/runs?window=7d — run rollups
   • GET https://api.agenticcontrolplane.com/api/v1/runs/{runKey} — one run with ordered steps
   • GET https://api.agenticcontrolplane.com/api/v1/introspect/cost-insights?window=7d — ranked recommendations Auth: Authorization: Bearer <key>.

Analysis steps

1. Headline: total cost, run count, window. If byoAuth, label costs "@ API rates" (subscription traffic priced at what it would cost).
2. Cache economics — decompose before concluding. A low headline hit rate has at least four distinct causes with different fixes. Compute from step data (model, tMs, promptTokens, cachedTokens):
   • per-model hit rate — Anthropic caches are per-model; if one model is cold, switching is the cause → route or batch by model;
   • hit rate by gap since the previous same-model call (<1min / 1–5min / 5–60min) — decay over gaps = TTL expiry → schedule work closer together or accept it;
   • full-price share immediately after a model switch — high = switching cost, low = rules it out;
   • uniform ~50% across models and gaps = structural: likely cache writes of new tool output being counted as misses (heavy tool-result appending), or genuine prefix instability. Say which is more consistent with the context composition, and say what you cannot distinguish from the data.
3. Context composition: system vs history vs tool-results share. If tool results dominate, name the top tools by re-read bytes (toolResultBytesByTool) and the worst offender's MB.
4. Loop share: report the number. Only recommend a subagent split for autonomous agents (background/api tier) — for interactive agents the growing loop IS the work.
5. Failure waste: only from runOutcome === "failed" runs. Runs without a terminal outcome NEVER get a waste claim.
6. Recommendations: merge /cost-insights items, ranked by impactUsd.

Report rules (honesty is the product)

• Window-observed dollars only unless the fix is a deterministic config change (model routing, cap) — only levers project to monthly.
• State the basis on every claim ("based on N runs / M steps over D days"). Step arrays cap at 250 — say so when a run is truncated.
• Name capture gaps instead of showing zeros. Empty contextBreakdown/topTools = the proxy isn't in the model path → tell the user how to add it (point the model client's base URL at the ACP proxy). Empty costByModel on older runs = pre-capture data.
• Never claim intent you can't observe — an errored step a run recovered from is not waste.

Report shape

# Cost X-ray — <agent> — last <window>
<headline: $X across N runs (@ API rates?)>

## What your money buys        ← composition + top tools
## Cache economics             ← decomposed, with the cause you ruled IN and OUT
## What to change before shipping   ← ranked, each with $ at stake + concrete action
## What we couldn't see        ← capture gaps + how to close them

Deliver as markdown. Offer to save it to a file if the user wants to share it.