Add Gemini incident cost ledger diagnostic

Context

#12743 needs final Cloud Billing / service-metrics authority before it can be closed, but the local side of the reconciliation was still reconstructed manually from orchestrator logs and private-content-safe Memory Core length samples.

This ticket captures the narrow implementation leaf: add a reproducible local diagnostic that parses miniSummary backfill logs and computes public-pricing sanity bands from aggregate length statistics, without printing raw Memory Core prompt/response content.

Live latest-open sweep: checked latest 20 open issues at 2026-06-08T11:53:22Z; no equivalent diagnostic-script leaf found. Closest current issue is #12743, which owns the broader billing/token reconciliation and must remain open until Cloud Billing authority lands.

Semantic duplicate sweep: used non-synthesis query_documents(type='ticket') for Gemini incident cost ledger diagnostic script local orchestrator logs because ask_knowledge_base synthesis is part of the active provider-routing/cost-safety incident surface. Results surfaced older/non-equivalent incident and telemetry tickets, not a narrow local ledger diagnostic leaf.

Exact sweep: rg -n "Gemini incident cost ledger|cost ledger diagnostic|miniSummary backfill.*ledger|billing.*token.*ledger|orchestrator.*cost ledger" resources/content/issues resources/content/archive/issues resources/content/discussions returned no matches.

The Problem

The team already has a strong timing/source lead for the 2026-06-08 Gemini spike, but hand-counted log/math evidence is fragile:

• It is easy to mix up one checkout's local log volume with all billable request volume.
• It is easy to leak private memory text if future agents sample prompt/response content directly instead of aggregate lengths.
• It is easy to overstate public-pricing estimates as billing truth, when Cloud Billing SKU/model/token fields remain authoritative.

A small diagnostic script gives future agents a repeatable local ledger and keeps the final #12743 closeout focused on the missing provider-console fields.

The Architectural Reality

• ai/daemons/orchestrator/scheduling/memorySummaryBackfill.mjs emits pending-memory-minisummary:<n> reasons that appear in orchestrator logs.
• ai/services/memory-core/MemoryService.mjs builds miniSummary prompts from stored prompt/response content, so cost estimates need aggregate input-size statistics.
• The correct repo home for one-shot incident diagnostics is ai/scripts/diagnostics/, matching existing scripts such as check-retired-primitives.mjs and review-cost-meter.mjs.
• Unit coverage belongs under test/playwright/unit/ai/scripts/diagnostics/, matching existing diagnostics-script specs.

The Fix

Add a local diagnostics command:

• ai/scripts/diagnostics/gemini-incident-cost-ledger.mjs
• npm run ai:gemini-incident-cost-ledger -- --log <orchestrator.log> ...
• A focused Playwright unit spec using synthetic logs only.

The script should:

• Accept one or more orchestrator log paths.
• Filter by optional ISO start/end window.
• Count miniSummary backfill starts, completions, pending deferrals, active-task deferrals, and pending-start sums.
• Accept aggregate private-content-safe input character statistics such as mean/p99/max.
• Compute low/base/high sanity estimates from configurable chars/token, input price, output price, and output-token assumptions.
• Optionally compare estimate bands with an operator-supplied billing cost.
• Emit both human-readable output and JSON.

Contract Ledger Matrix

Decision Record impact

none. This is incident-forensics tooling, aligned with the local diagnostics script pattern. It does not change provider routing, pricing policy, or AiConfig defaults.

Acceptance Criteria

• A diagnostics script under ai/scripts/diagnostics/ parses miniSummary backfill log events from one or more orchestrator logs.
• The script supports optional ISO window bounds and JSON output.
• The script computes sanity cost bands from aggregate input-length statistics without reading or printing private Memory Core prompt/response content.
• package.json exposes the script via an ai:* command.
• A focused Playwright unit spec verifies log parsing, window filtering, and cost-band math from synthetic log content.
• PR evidence includes a real incident-window run and states that Cloud Billing SKU/model/token data remains the closeout authority for #12743.

Out of Scope

• Closing #12743.
• Pulling Cloud Billing data automatically.
• Sampling raw Memory Core prompt/response content.
• Changing provider defaults or miniSummary backfill scheduling.

Avoided Traps

• Do not use the diagnostic's public-pricing estimate as final billing proof.
• Do not copy private memory text into the issue, PR, or command output.
• Do not make this a daemon or background worker; it is an operator/agent-run diagnostic.

Related

Parent forensic ticket: #12743 Parent cost-safety epic: #12740 Related: #12742, #12746, #12748

Origin Session ID: e8f07ef9-ef7e-4815-8ff4-7abe13720621

Handoff Retrieval Hint: Gemini incident cost ledger miniSummary backfill log parser token math