Files
PaperClipAI/tests/runner-e2e/provider-usage.ts
T
DottaandPaperclip b6e690044a feat(evals): allow reviewed bounded provisional accounting (#17365)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Product E2E campaigns include model, setup, storage and
observability costs.
> - Some billing windows close after the tests and reports finish.
> - Requiring settled receipts for retained storage can block a good
campaign for weeks.
> - A reservation alone does not prove that a pending charge has an
upper bound.
> - This change permits provisional completion only with reviewed
evidence for the exact cost configuration.
> - Reports preserve unknown amounts and holds until later
reconciliation.

## Linked Issues or Issue Description

Refs: #17360, #17354

**What existing behavior does this improve?**

The Product E2E scheduled accounting gate and its retained cost reports.

**Current behavior**

Every pending receipt fails finalization. A completed test can wait
through the retained-storage billing window. The journal has dollar
reservations but no explicit qualification of their bounds.

**Proposed behavior**

Permit provisional completion when all pending charges have reviewed
full-lifecycle bounds, no requests remain in flight, no receipts
conflict, and known costs plus held amounts fit the campaign limit. Keep
all existing holds and receipt-finality rules. Missing evidence still
fails.

**Reason and benefit**

An operator can distinguish delayed but bounded billing from unbounded
or missing costs. The campaign can finish without claiming that its
invoice is settled.

**Breaking changes**

The journal accepts optional bound metadata and an invalidation record.
New reports add an optional bounded-pending count. Current readers
retain old journals and reports without granting them bounds. Older code
must not read new bound-bearing journals. The existing daily/weekly
enable gates and budget amounts stay unchanged.

**Additional context**

This implements the operator-approved provisional-cost policy for the
existing eval harness. It does not change product budgets or the
agent-eval roadmap. A targeted GitHub search found no duplicate open PR.

## What Changed

- Accounting owns a narrow evidence contract: a reviewed audit hash,
exact configuration hash and full-lifecycle scope.
- Gateway reservations retain that evidence before dispatch. Changed
limits or rates fail before dispatch. Unexpected routing, model identity
or excessive usage invalidates a bound durably.
- Operating allowances can carry separately reviewed bounds for each
cost component, including its full storage lifetime. No existing
allowance becomes qualified by default.
- Reports and Braintrust experiment metadata distinguish provisional
accounting, complete receipts, blocked accounting and unavailable
evidence. Public output omits private audit hashes.
- Scheduled finalization allows qualified provisional costs and
explicitly rejects spending overruns, including overruns with complete
receipts.
- Add regressions for legacy data, restarts, late reconciliation,
conflicts, malformed evidence, routing uncertainty and exceeded limits.
Document the evidence and migration boundaries.

## Verification

- Full offline eval-support suite: 4,062 Vitest tests passed, two
existing tests skipped; 137 Node tests passed.
- `pnpm test:e2e:runner:typecheck`: passed.
- `git diff --check`: passed.
- Read the retained hosted qualification journal with the new policy.
All six legacy pending charges remain unqualified. Known cost
`$0.421245200`, held `$19.327680000`, and committed `$19.748925200`
remain unchanged. The journal's SHA-256 is unchanged. The original
failed workflow remains failed.
- No paid models, network provider calls, receipt imports or reservation
releases were used for this verification.
- Rebased onto master `4d90634f74`; 142 targeted tests and typecheck
passed.
- Review reproduced a lost OpenAI input-total limit check when cache
details were missing (1 failed, 47 passed). Fix
`5fef9ba7c149951a51eae468614a4b89f0b3d6db` preserves unpriced provider
usage bounds. Then 178 focused tests, full offline support and typecheck
passed.
- Current-head hosted CI passed:
https://github.com/paperclipai/paperclip/actions/runs/40020746585
(attempt 2). Typecheck, build, all general/Runner/serialized tests and
browser gates are green.
- CI attempt 1 hit a PostgreSQL deadlock in unchanged chat-test cleanup;
its skipped flag restoration caused a second assertion failure. One
failed-jobs-only retry passed. Original failure evidence is preserved in
attempt 1. No paid eval campaign was rerun.
- Greptile reviewed exact head
`5fef9ba7c149951a51eae468614a4b89f0b3d6db` at 5/5; its one finding is
fixed and resolved. The optional Paperclip review is paused and supplied
no assessment.

## Risks

- Evidence hashes bind an audit to a configuration; they do not prove
the audit is true. Operators must review real provider limits, all price
tiers, full resource lifetimes and enforceable infrastructure limits
before configuring bounds.
- Provisional completion is not invoice settlement. Later billing can
reveal overruns, which must be reconciled and inspected. The ledger
cannot stop external billing.
- Existing pilot journals, configured allowances and price cards have no
new bound approval. This PR alone does not qualify or enable recurring
paid work.
- The gateway cannot qualify a canceled Braintrust request without route
confirmation. It preserves that unknown hold.
- No new dependency, secret, repository variable, workflow, schedule or
cloud resource changes.

## Model Used

OpenAI Codex, GPT-6 family. The exact deployment ID and context-window
size are not exposed in this session. Used reasoning, repository tools
and code execution. No paid evaluation model calls were made for this
change.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [x] All Paperclip CI gates are green
- [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-10-23 13:57:57 -07:00

174 lines
10 KiB
TypeScript

import { usdToUnits } from "../../packages/shared/src/money.js";
import type { EvalCostObservation } from "./cost-ledger.js";
type Tokens = NonNullable<EvalCostObservation["tokens"]>;
type Row = Record<string, unknown>;
const row = (value: unknown): Row => value && typeof value === "object" && !Array.isArray(value) ? value as Row : {};
const count = (value: unknown): number | undefined => typeof value === "number" && Number.isSafeInteger(value) && value >= 0 ? value : undefined;
export interface ProviderUsageReceipt {
provider: "openai" | "anthropic";
tokens: Tokens;
complete: boolean;
tokenBounds?: { input: number; output: number; valid: boolean };
providerResponseId?: string;
model?: string;
cacheWrite5m?: number;
cacheWrite1h?: number;
cacheWrite30m?: number;
cacheWrite12h?: number;
cacheWriteAttribution?: { source: "request_policy"; version: "openai-prompt-cache-2026-10-20"; ttl: "30m" | "12h" };
}
function openAiRequestCachePolicy(payload: unknown): { model: string; ttl: "30m" | "12h" } | undefined {
const body = row(payload);
if (body.model !== "gpt-5.6-sol" || body.previous_response_id !== undefined || body.conversation !== undefined || body.prompt !== undefined
|| body.context_management !== undefined || !(typeof body.input === "string" || Array.isArray(body.input))) return;
const options = row(body.prompt_cache_options);
if (body.prompt_cache_options !== undefined && body.prompt_cache_options !== options) return;
if (Object.keys(options).some(key => !["mode", "ttl"].includes(key)) || options.mode !== undefined && options.mode !== "implicit") return;
const ttl = options.ttl === undefined ? "30m" : options.ttl;
if (ttl !== "30m" && ttl !== "12h") return;
const hasUnqualifiedInput = (value: unknown): boolean => {
if (Array.isArray(value)) return value.some(hasUnqualifiedInput);
if (!value || typeof value !== "object") return false;
const item = row(value);
return "prompt_cache_breakpoint" in item || item.type === "compaction" || item.type === "item_reference"
|| Object.values(item).some(hasUnqualifiedInput);
};
if (hasUnqualifiedInput(body.input)) return;
return { model: body.model, ttl };
}
function tokenBounds(provider: "openai" | "anthropic", usage: Row) {
const input = row(usage.input_tokens_details), output = row(usage.output_tokens_details);
const writes = provider === "openai" ? row(input.cache_write_tokens_details) : row(usage.cache_creation);
const rawWrite = provider === "openai" ? input.cache_write_tokens : usage.cache_creation_input_tokens;
const rawRead = provider === "openai" ? input.cached_tokens : usage.cache_read_input_tokens;
const ttls = provider === "openai" ? [writes.ttl_30m_tokens, writes.ttl_12h_tokens]
: [writes.ephemeral_5m_input_tokens, writes.ephemeral_1h_input_tokens];
const inputs = [usage.input_tokens, rawRead, rawWrite, ...ttls];
const outputs = [usage.output_tokens, ...(provider === "openai" ? [output.reasoning_tokens] : [])];
const value = (raw: unknown) => count(raw) ?? 0;
const cached = value(rawRead) + Math.max(value(rawWrite), value(ttls[0]) + value(ttls[1]));
const inputBound = provider === "openai" ? Math.max(value(usage.input_tokens), cached) : value(usage.input_tokens) + cached;
const outputBound = Math.max(...outputs.map(value));
return { input: inputBound, output: outputBound,
valid: [...inputs, ...outputs].every(raw => raw === undefined || count(raw) !== undefined)
&& Number.isSafeInteger(inputBound) && Number.isSafeInteger(outputBound) };
}
export function normalizeProviderUsage(provider: "openai" | "anthropic", payload: unknown): ProviderUsageReceipt {
const body = row(payload), usage = row(body.usage);
const output = count(usage.output_tokens);
let tokens: Tokens;
let cacheWrite5m: number | undefined, cacheWrite1h: number | undefined;
let cacheWrite30m: number | undefined, cacheWrite12h: number | undefined;
if (provider === "openai") {
const totalInput = count(usage.input_tokens);
const details = row(usage.input_tokens_details);
const cachedRead = count(details.cached_tokens);
let cachedWrite = count(details.cache_write_tokens);
const writes = row(details.cache_write_tokens_details);
cacheWrite30m = count(writes.ttl_30m_tokens);
cacheWrite12h = count(writes.ttl_12h_tokens);
if (cachedWrite !== undefined && cacheWrite30m !== undefined && cacheWrite12h !== undefined && cacheWrite30m + cacheWrite12h !== cachedWrite) cachedWrite = undefined;
const reasoning = count(row(usage.output_tokens_details).reasoning_tokens);
tokens = { input: totalInput !== undefined && cachedRead !== undefined && cachedWrite !== undefined && totalInput >= cachedRead + cachedWrite ? totalInput - cachedRead - cachedWrite : undefined,
cachedRead, cachedWrite, output, reasoning };
} else {
const cache = row(usage.cache_creation);
cacheWrite5m = count(cache.ephemeral_5m_input_tokens);
cacheWrite1h = count(cache.ephemeral_1h_input_tokens);
tokens = { input: count(usage.input_tokens), cachedRead: count(usage.cache_read_input_tokens), cachedWrite: count(usage.cache_creation_input_tokens), output };
if (tokens.cachedWrite !== undefined && cacheWrite5m !== undefined && cacheWrite1h !== undefined && cacheWrite5m + cacheWrite1h !== tokens.cachedWrite) tokens.cachedWrite = undefined;
}
const complete = tokens.input !== undefined && tokens.cachedRead !== undefined && tokens.cachedWrite !== undefined && output !== undefined
&& !(tokens.reasoning !== undefined && tokens.reasoning > output);
return { provider, tokenBounds: tokenBounds(provider, usage), tokens: Object.fromEntries(Object.entries(tokens).filter(([, value]) => value !== undefined)), complete,
...(typeof body.id === "string" ? { providerResponseId: body.id } : {}),
...(typeof body.model === "string" ? { model: body.model } : {}),
...(cacheWrite5m !== undefined ? { cacheWrite5m } : {}), ...(cacheWrite1h !== undefined ? { cacheWrite1h } : {}),
...(cacheWrite30m !== undefined ? { cacheWrite30m } : {}), ...(cacheWrite12h !== undefined ? { cacheWrite12h } : {}) };
}
export class ProviderUsageAccumulator {
private body: Row = {};
private usage: Row = {};
private final = false;
private readonly cachePolicy: ReturnType<typeof openAiRequestCachePolicy>;
constructor(readonly provider: "openai" | "anthropic", request?: unknown) {
this.cachePolicy = provider === "openai" ? openAiRequestCachePolicy(request) : undefined;
}
observe(event: unknown) {
const value = row(event);
if (this.provider === "openai") {
const response = row(value.response);
if (["response.completed", "response.incomplete", "response.failed"].includes(String(value.type))) {
this.body = response;
this.usage = row(response.usage);
this.final = true;
} else if (value.object === "response") {
this.body = value;
this.usage = row(value.usage);
this.final = ["completed", "incomplete", "failed"].includes(String(value.status));
}
} else {
if (value.type === "message_start") {
this.body = row(value.message);
this.usage = { ...this.usage, ...row(this.body.usage) };
} else if (value.type === "message_delta") this.usage = { ...this.usage, ...row(value.usage) };
else if (value.type === "message_stop") this.final = true;
else if (value.type === "message") { this.body = value; this.usage = row(value.usage); this.final = true; }
}
}
receipt() {
const receipt = normalizeProviderUsage(this.provider, { ...this.body, usage: this.usage });
const details = row(this.usage.input_tokens_details);
if (this.cachePolicy && receipt.model === this.cachePolicy.model && receipt.complete && receipt.tokens.cachedWrite! > 0
&& !("cache_write_tokens_details" in details)) {
receipt.cacheWrite30m = this.cachePolicy.ttl === "30m" ? receipt.tokens.cachedWrite : 0;
receipt.cacheWrite12h = this.cachePolicy.ttl === "12h" ? receipt.tokens.cachedWrite : 0;
receipt.cacheWriteAttribution = { source: "request_policy", version: "openai-prompt-cache-2026-10-20", ttl: this.cachePolicy.ttl };
}
return { ...receipt, final: this.final };
}
}
export interface EvalTokenRates {
inputUsdPerMillion: string;
cachedReadUsdPerMillion: string;
outputUsdPerMillion: string;
cacheWrite5mUsdPerMillion?: string;
cacheWrite1hUsdPerMillion?: string;
cacheWrite30mUsdPerMillion?: string;
cacheWrite12hUsdPerMillion?: string;
}
export interface EvalModelPrice extends EvalTokenRates {
provider: "openai" | "anthropic";
model: string;
version: string;
sourceUrl: string;
longContext?: { aboveInputTokens: number; rates: EvalTokenRates };
}
export function estimateProviderCost(receipt: ProviderUsageReceipt, price: EvalModelPrice): string | null {
if (!receipt.complete || receipt.model !== price.model || receipt.provider !== price.provider) return null;
const { tokens } = receipt;
const totalInput = tokens.input! + tokens.cachedRead! + tokens.cachedWrite!;
if (!Number.isSafeInteger(totalInput)) return null;
const rates = price.longContext && totalInput > price.longContext.aboveInputTokens ? price.longContext.rates : price;
const priced = (count: number, rate: string) => {
if (!/^(?:0|[1-9][0-9]*)(?:\.[0-9]{1,9})?$/.test(rate)) throw new Error("Invalid nonnegative model price");
return BigInt(count) * usdToUnits(rate);
};
let numerator = priced(tokens.input!, rates.inputUsdPerMillion) + priced(tokens.cachedRead!, rates.cachedReadUsdPerMillion) + priced(tokens.output!, rates.outputUsdPerMillion);
if (tokens.cachedWrite! > 0) {
const counts = receipt.provider === "openai" ? [receipt.cacheWrite30m, receipt.cacheWrite12h] : [receipt.cacheWrite5m, receipt.cacheWrite1h];
const writeRates = receipt.provider === "openai" ? [rates.cacheWrite30mUsdPerMillion, rates.cacheWrite12hUsdPerMillion] : [rates.cacheWrite5mUsdPerMillion, rates.cacheWrite1hUsdPerMillion];
if (counts[0] === undefined || counts[1] === undefined || counts[0] + counts[1] !== tokens.cachedWrite || writeRates[0] === undefined || writeRates[1] === undefined) return null;
numerator += priced(counts[0], writeRates[0]) + priced(counts[1], writeRates[1]);
}
const units = (numerator + 999_999n) / 1_000_000n;
return `${units / 1_000_000_000n}.${String(units % 1_000_000_000n).padStart(9, "0")}`;
}