opencode2pi 0.0.0-stage → 0.3.2

package/src/health.ts

@@ -0,0 +1,332 @@
+/**
+ * Per-model health probing (docs/PLAN.md §T5).
+ *
+ * The probe speaks to Zen directly rather than through the host engine, for
+ * two reasons. It has to: the whole point of `doctor` is to test the *shape
+ * gate* this extension hardcodes, and routing through the engine would make a
+ * broken gate indistinguishable from a broken model. And it keeps the check
+ * honest about cost — one aborted stream beats a full agent turn.
+ *
+ * Two verdicts, deliberately kept apart:
+ *
+ *   dead  — upstream will never serve this id again (401 / 404 / ModelError /
+ *           geo-block). Sticky: a later 503 does not resurrect it.
+ *   flaky — unavailable *right now* (429 / 5xx / timeout / a rejected
+ *           request). Never promoted to dead on transient evidence, which is
+ *           how a model that is merely having a bad afternoon would get
+ *           wrongly struck off the roster.
+ *
+ * Health is annotation only. It never removes a model from the roster.
+ */
+
+import { mkdir, readFile, writeFile } from 'node:fs/promises'
+import { tmpdir } from 'node:os'
+import { join } from 'node:path'
+
+import type { ProviderModelConfig } from '@oh-my-pi/pi-coding-agent'
+
+import { classifyUpstreamFailure, type UpstreamFailure } from './errors.ts'
+import { sendGateProbe } from './gate.ts'
+
+/**
+ * Probe pacing.
+ *
+ * The anonymous lane meters per egress IP, so a sweep of 13 requests fired at
+ * once spends quota the next sweep needs. Concurrency 2 with a short pause keeps
+ * a sweep from starving its own follow-up.
+ *
+ * Measured 2026-10-06: 429 turned out to be metered *per model*, not per IP —
+ * `big-pickle` answered 200 in the same minute that two other ids answered 429,
+ * serially and 1.5 s apart. So pacing reduces self-inflicted pressure but does
+ * not cure it, which is why a rate-limited verdict gets its own mark instead of
+ * being reported as an unstable model.
+ */
+const PROBE_CONCURRENCY = 2
+const PROBE_GAP_MS = 500
+
+function delay(ms: number): Promise<void> {
+	const { promise, resolve } = Promise.withResolvers<void>()
+	setTimeout(resolve, ms)
+	return promise
+}
+/**
+ * How long a verdict may be shown before `status` stops treating it as current.
+ *
+ * Six hours, not a day: the free pool churns on the order of hours, and a stale
+ * verdict is flagged (`（已过期）`) rather than re-run — a full sweep costs real
+ * inference, so it is the user's call, which is what `probe` is for.
+ */
+export const HEALTH_TTL_MS = 6 * 60 * 60 * 1000
+
+export type ModelHealth = 'ok' | 'flaky' | 'dead' | 'unknown' | 'limited'
+
+export const HEALTH_MARK: Readonly<Record<ModelHealth, string>> = {
+	ok: '✅',
+	flaky: '⚠️',
+	dead: '❌',
+	unknown: '❓',
+	limited: '🚧',
+}
+
+export interface HealthRecord {
+	readonly health: ModelHealth
+	readonly kind: UpstreamFailure | 'OK'
+	readonly detail: string
+	readonly checkedAt: number
+	/** Consecutive `MODEL_GONE` verdicts; two are required before ❌. */
+	readonly terminalFailures: number
+	/** Consecutive transient failures; hints at an outage, never at death. */
+	readonly transientFailures: number
+}
+
+/** Consecutive `ModelError`s required before a model is called dead. */
+export const TERMINAL_FAILURES_TO_CONDEMN = 2
+
+/**
+ * Map a classified failure onto a provisional health verdict.
+ *
+ * Shape and auth rejections are deliberately *not* per-model verdicts: both are
+ * lane-wide, so blaming a model for them would paint ❌ across the entire roster
+ * at the exact moment the gate breaks. Those surface through `doctor`.
+ *
+ * `MODEL_GONE` starts as ⚠️ rather than ❌, and `saveHealth` only promotes it
+ * after {@link TERMINAL_FAILURES_TO_CONDEMN} consecutive hits. That is not
+ * caution for its own sake — measured 2026-10-05, `nemotron-3.ultra-free`
+ * answered `401 ModelError` on one probe and `200` on the next, a minute
+ * apart. Promoting on the first terminal verdict would mark a working model
+ * dead, which is the exact mistake this whole feature exists to avoid.
+ *
+ * `REGION_BLOCKED` is the exception: a geo block is deterministic for a given
+ * egress, so one hit is conclusive.
+ *
+ * `RATE_LIMIT` is not a verdict on the model at all. The anonymous lane meters
+ * quota separately per model — measured 2026-10-06, `big-pickle` answered 200
+ * while two other ids answered 429 in the same minute, probed serially — so a
+ * 429 says "this id is out of quota right now", which is a fact about the lane
+ * and not about whether the model works. It gets its own mark so that a quota
+ * pause never masquerades as an unstable model.
+ */
+export function healthFor(kind: UpstreamFailure | 'OK'): ModelHealth {
+	switch (kind) {
+		case 'OK':
+			return 'ok'
+		case 'MODEL_GONE':
+			return 'flaky'
+		// A geo block is deterministic for a given egress: the same request from
+		// the same network keeps saying no.
+		case 'REGION_BLOCKED':
+			return 'dead'
+		case 'RATE_LIMIT':
+			return 'limited'
+		// A request the upstream rejected is the model refusing this particular
+		// round, which is evidence about the request, not a state of the model:
+		// measured on this lane an id answers 400 for one request and works for
+		// the next. An unrecognised failure carries no more
+		// evidence than a 5xx does — the classifier simply did not recognise the
+		// text — so giving it a verdict of its own would invent a fifth state
+		// that only ever showed up in `status`.
+	case 'REQUEST_REJECTED':
+	case 'UNKNOWN':
+	case 'UPSTREAM':
+		return 'flaky'
+		// The shape gate and the credential are lane-wide, so blaming a model
+		// for them would paint ❓ across the whole roster at the exact moment
+		// the gate breaks. They surface through `doctor` instead.
+	case 'SHAPE_REJECTED':
+	case 'AUTH':
+		return 'unknown'
+	}
+}
+
+export interface ProbeResult extends HealthRecord {
+	readonly modelId: string
+	readonly latencyMs: number
+}
+
+async function probeModel(modelId: string): Promise<ProbeResult> {
+	const probe = await sendGateProbe(modelId)
+	const base = { modelId, checkedAt: Date.now(), terminalFailures: 0, transientFailures: 0, latencyMs: probe.latencyMs }
+
+	if (probe.ok) {
+		return { ...base, health: 'ok', kind: 'OK', detail: `HTTP ${probe.status}` }
+	}
+	if (probe.transportError !== undefined) {
+		// A timeout or transport failure says nothing about the model itself:
+		// the 30 s budget can expire on a slow lane, or the request never left.
+		return { ...base, health: 'limited', kind: 'UPSTREAM', detail: `transport: ${probe.transportError}` }
+}
+
+	const kind = classifyUpstreamFailure(probe.status, probe.body).kind
+	return {
+		...base,
+		health: healthFor(kind),
+		kind,
+		detail: `HTTP ${probe.status} ${probe.body.slice(0, 200)}`,
+	}
+}
+
+/** Probe every model with a bounded, paced worker pool, reporting progress as it goes. */
+export async function probeAll(
+	models: readonly ProviderModelConfig[],
+	onProgress?: (done: number, total: number, result: ProbeResult) => void,
+): Promise<ProbeResult[]> {
+	const results: ProbeResult[] = []
+	let cursor = 0
+
+	async function worker(): Promise<void> {
+		while (cursor < models.length) {
+			const model = models[cursor++]
+			if (!model) return
+			const result = await probeModel(model.id)
+			results.push(result)
+			onProgress?.(results.length, models.length, result)
+			await delay(PROBE_GAP_MS)
+		}
+	}
+
+	await Promise.all(Array.from({ length: Math.min(PROBE_CONCURRENCY, models.length) }, worker))
+	results.sort((a, b) => a.modelId.localeCompare(b.modelId))
+	return results
+}
+
+export type HealthStore = Readonly<Record<string, HealthRecord>>
+
+function storeFile(): string {
+	return join(tmpdir(), 'opencode2pi', 'health.json')
+}
+
+export async function loadHealth(): Promise<HealthStore> {
+	try {
+		const parsed: unknown = JSON.parse(await readFile(storeFile(), 'utf8'))
+		if (typeof parsed !== 'object' || parsed === null) return {}
+		const out: Record<string, HealthRecord> = {}
+		for (const [id, value] of Object.entries(parsed)) {
+			if (typeof value !== 'object' || value === null) continue
+			if (!('health' in value) || !('checkedAt' in value)) continue
+			const record = value as Partial<HealthRecord>
+			if (typeof record.checkedAt !== 'number') continue
+			if (
+				record.health !== 'ok' &&
+				record.health !== 'flaky' &&
+				record.health !== 'dead' &&
+				record.health !== 'unknown' &&
+				record.health !== 'limited'
+			)
+				continue
+			out[id] = {
+				health: record.health,
+				kind: record.kind ?? 'UNKNOWN',
+				detail: typeof record.detail === 'string' ? record.detail : '',
+				checkedAt: record.checkedAt,
+				terminalFailures: typeof record.terminalFailures === 'number' ? record.terminalFailures : 0,
+				transientFailures: typeof record.transientFailures === 'number' ? record.transientFailures : 0,
+			}
+		}
+		return out
+	} catch {
+		return {}
+	}
+}
+
+/**
+ * Merge fresh probe results into the store.
+ *
+ * A single `MODEL_GONE` is not enough to condemn: it is promoted to ❌ only
+ * after {@link TERMINAL_FAILURES_TO_CONDEMN} consecutive terminal verdicts, and
+ * any success clears both counters so a model with a bad afternoon recovers on
+ * its own. A model already marked dead keeps the verdict across a *transient*
+ * re-probe, but a success always clears it.
+ */
+export function mergeHealth(previous: HealthStore, results: readonly ProbeResult[]): HealthStore {
+	const merged: Record<string, HealthRecord> = { ...previous }
+
+	for (const result of results) {
+		// Read the accumulating map, not the pre-call snapshot: two terminal
+		// verdicts for the same id within one batch must count as consecutive.
+		const before = merged[result.modelId]
+		const terminalFailures = result.kind === 'MODEL_GONE' ? (before?.terminalFailures ?? 0) + 1 : 0
+		const transientFailures = result.health === 'ok' ? 0 : (before?.transientFailures ?? 0) + 1
+
+		let health = result.health
+		if (result.kind === 'MODEL_GONE' && terminalFailures >= TERMINAL_FAILURES_TO_CONDEMN) {
+			health = 'dead'
+		} else if (before?.health === 'dead' && result.health !== 'ok') {
+			health = 'dead'
+		}
+
+		merged[result.modelId] = {
+			health,
+			kind: result.kind,
+			detail: result.detail,
+			checkedAt: result.checkedAt,
+			terminalFailures,
+			transientFailures,
+		}
+	}
+	return merged
+}
+
+/**
+ * Drop every model the current roster no longer contains.
+ *
+ * The free lane follows upstream: models get added and withdrawn, so a
+ * merge-only store accumulates the whole history of ids — `no-such-model-xyz`
+ * and other ids that were never real — and grows without bound. A probe covers
+ * exactly the current roster (`probeAll` walks it end to end), so the ids in
+ * `results` *are* the roster, and anything else is a model upstream retired.
+ * Keeping those records would answer no question anyone can still ask.
+ */
+export function pruneHealth(store: HealthStore, keep: ReadonlySet<string>): HealthStore {
+	const pruned: Record<string, HealthRecord> = {}
+	for (const [id, record] of Object.entries(store)) {
+		if (keep.has(id)) pruned[id] = record
+	}
+	return pruned
+}
+
+/**
+ * Merge fresh probe results into the store and write it back.
+ *
+ * `keep` is the roster the caller just looked at, not the ids this batch
+ * covered: `status` probes only the models it has no record of, and pruning to
+ * the batch would throw away every verdict it just refused to re-run.
+ */
+export async function saveHealth(results: readonly ProbeResult[], keep: ReadonlySet<string>): Promise<HealthStore> {
+	const merged = pruneHealth(mergeHealth(await loadHealth(), results), keep)
+
+	try {
+		await mkdir(join(tmpdir(), 'opencode2pi'), { recursive: true })
+		await writeFile(storeFile(), JSON.stringify(merged), 'utf8')
+	} catch {
+		// An unwritable cache only costs a re-probe.
+	}
+	return merged
+}
+
+export function isStale(record: HealthRecord | undefined): boolean {
+	return record === undefined || Date.now() - record.checkedAt > HEALTH_TTL_MS
+}
+
+/** Canonical order, so the summary reads the same way on every run. */
+const VERDICT_ORDER: readonly ModelHealth[] = ['ok', 'flaky', 'limited', 'unknown', 'dead']
+
+/**
+ * Summarize the roster's verdicts as they are *displayed*.
+ *
+ * The store, never the raw probe results, is the source of truth here: a second
+ * consecutive `ModelError` promotes ⚠️ to ❌ in {@link mergeHealth} only, so
+ * counting raw results makes the summary contradict the lines right below it. A
+ * model with no record is not a verdict and is not counted as one — its roster
+ * line says `未探测`, which is a statement about our coverage, not about the
+ * model.
+ */
+export function summarizeHealth(roster: readonly { readonly id: string }[], store: HealthStore): string {
+	const counts: Record<ModelHealth, number> = { ok: 0, flaky: 0, dead: 0, unknown: 0, limited: 0 }
+	for (const model of roster) {
+		const health = store[model.id]?.health
+		if (health) counts[health]++
+	}
+	return VERDICT_ORDER.filter((health) => counts[health] > 0)
+		.map((health) => `${HEALTH_MARK[health]} ${counts[health]}`)
+		.join('  ')
+}

package/src/host.d.ts

@@ -0,0 +1,15 @@
+/**
+ * omp resolves `@earendil-works/*` at runtime to its own built-in pi-ai copy
+ * through a legacy-compat shim (docs/FINDINGS.md §3.4), so the extension must
+ * import through that specifier or it would register its custom API into a
+ * second, unrelated registry instance.
+ *
+ * The npm package of the same name is an older build that predates
+ * `registerCustomApi` (docs/FINDINGS.md §3.3), so it is deliberately NOT a
+ * dependency: installing it would shadow this declaration and type the
+ * extension against an API surface the host does not have. Instead, point the
+ * specifier at the host's own type package, which ships with omp itself.
+ */
+declare module '@earendil-works/pi-ai' {
+	export * from '@oh-my-pi/pi-ai'
+}

package/src/index.ts

@@ -0,0 +1,254 @@
+/**
+ * opencode2pi — OpenCode Zen's anonymous free lane as an omp/pi provider.
+ *
+ * The built-in `opencode-zen` provider requires a paid `OPENCODE_API_KEY`.
+ * This extension needs none: the upstream gate is a request-shape check, not
+ * a credential check. Three conditions, all confirmed live on 2026-10-05:
+ *
+ *   1. `Authorization: Bearer public`  (anonymous credential)
+ *   2. canonical session id  `^ses_[0-9a-f]{12}[0-9A-Za-z]{14}$`
+ *   3. an agent-shaped body: `tools` containing `bash` and `read`
+ *
+ * Non-streaming requests are rejected with FreeTierError. Host streaming is
+ * inherent to the openai-completions engine, so condition 3 is the only body
+ * change this extension has to make.
+ *
+ * Integration seam (verified against omp 18.6.1 — see docs/FINDINGS.md):
+ *   - `pi.registerProvider` config accepts `onPayload` / `prepareRequest`
+ *     WITHOUT ERROR but NEVER CALLS THEM. Do not use them.
+ *   - The working seam is `registerCustomApi(apiId, streamSimple, sourceId)`,
+ *     which receives `(model, context, options)` — a structured context, not a
+ *     serialized body. Shape the context, then delegate to the host engine
+ *     with `model.api` swapped back to `openai-completions` to avoid recursion.
+ *   - `toolChoice` lives on the *options* object, not on `Context`; setting it
+ *     on the context is a silent no-op.
+ */
+
+import { createHash } from 'node:crypto'
+
+import { createAssistantMessageEventStream, registerCustomApi, streamSimple } from '@earendil-works/pi-ai'
+import type { AssistantMessage, Context, Model, SimpleStreamOptions } from '@earendil-works/pi-ai'
+import type { ExtensionAPI, ProviderModelConfig } from '@oh-my-pi/pi-coding-agent'
+
+import { discoverFreeModels } from './discovery.ts'
+import { handleCommand } from './commands.ts'
+import { classifyUpstreamFailure, formatFailure } from './errors.ts'
+import { hidden, loadFilters, onHiddenChange, publishFilters, publishHealth } from './filters.ts'
+import { loadHealth } from './health.ts'
+import { PROCESS_SESSION, sessionForRequest, sessionHeaders } from './session.ts'
+import { SEED_MODELS, UNAVAILABLE, VERIFIED_FREE } from './seed.ts'
+
+const PROVIDER = 'opencode-zen-free'
+const API_ID = 'openai-completions-zen-free'
+const ZEN_BASE = 'https://opencode.ai/zen/v1'
+const ANONYMOUS_KEY = 'public'
+const SOURCE_ID = 'opencode2pi'
+
+const PROJECT_ID = 'prj_' + createHash('sha256').update('omp:default-project').digest('hex').slice(0, 24)
+
+const DISGUISE_HEADERS: Record<string, string> = {
+	'user-agent': 'opencode/1.18.31 (win32 x64; node24.19.0)',
+	'x-opencode-client': 'cli',
+	'x-opencode-project': PROJECT_ID,
+}
+
+/** The upstream gate demands a `bash` and a `read` function in `tools`. */
+const GATE_TOOL_NAMES = ['bash', 'read'] as const
+
+const GATE_TOOLS = GATE_TOOL_NAMES.map((name) => ({
+	name,
+	description: 'Reserved for the host runtime; do not call it.',
+	parameters: { type: 'object' as const, properties: {} },
+}))
+
+/**
+ * Ensure the gate tools are present.
+ *
+ * The host ships real `bash` and `read` tools, so in practice this is a no-op
+ * and the stubs never reach the model. It only bites when a caller runs with no
+ * tools at all (a bare completion), where the gate would otherwise fail — so in
+ * that case the stubs arrive with `toolChoice: 'none'` to stop the model from
+ * calling a stub the host cannot service.
+ */
+function withGateTools(
+	context: Context,
+	options: SimpleStreamOptions,
+): { context: Context; options: SimpleStreamOptions } {
+	const existing = context.tools ?? []
+	const names = new Set(existing.map((tool) => tool.name))
+	const missing = GATE_TOOLS.filter((tool) => !names.has(tool.name))
+	if (missing.length === 0) return { context, options }
+
+	if (existing.length > 0) return { context: { ...context, tools: [...existing, ...missing] }, options }
+
+	return {
+		context: { ...context, tools: [...missing] },
+		options: { ...options, toolChoice: 'none' },
+	}
+}
+
+/**
+ * Replace a raw provider error with one that names the actual failure mode.
+ *
+ * `errorClassificationMessage` is preserved (and seeded from the original text
+ * when absent) because that is the field the host's recovery logic reads; only
+ * the human-facing `errorMessage` is rewritten. Without that split the
+ * classification would be display-only but would also erase the host's own
+ * retry decisions.
+ */
+function explainFailure(modelId: string, message: AssistantMessage): AssistantMessage {
+	const original = message.errorMessage ?? ''
+	const classification = classifyUpstreamFailure(message.errorStatus, original)
+	return {
+		...message,
+		errorMessage: formatFailure(modelId, classification),
+		errorClassificationMessage: message.errorClassificationMessage ?? original,
+	}
+}
+
+let log: ExtensionAPI['logger'] | undefined
+let warnedMissingSession = false
+
+/**
+ * Custom API handler: shapes the context, pins the conversation's session id,
+ * classifies failures, then delegates the actual streaming to the host engine.
+ */
+registerCustomApi(API_ID, (model: Model, context: Context, options?: SimpleStreamOptions) => {
+	const incoming = options ?? {}
+	const gated = withGateTools(context, incoming)
+
+	const session = sessionForRequest(options)
+	if (!session.derived && !warnedMissingSession) {
+		warnedMissingSession = true
+		log?.warn(
+			`[${PROVIDER}] host supplied no session identity; using process-level session ${session.id}`,
+		)
+	}
+
+	// The session headers ride on the model we hand the engine rather than on
+	// a wrapped `options.fetch`: `fetch` is optional, and a missing wrapper
+	// would silently drop the session header and fail the upstream gate with a
+	// bare 403. Merging onto the model's own headers also preserves whatever
+	// the host already resolved there.
+	const builtin: Model = {
+		...model,
+		api: 'openai-completions' as Model['api'],
+		headers: { ...model.headers, ...sessionHeaders(session.id) },
+	}
+	const forwarded = gated.options
+
+	const upstream = streamSimple(builtin, gated.context, forwarded)
+
+	// Failures arrive as a stream *event*, not a rejected promise, so
+	// classification means forwarding the stream and rewriting that event.
+	const relayed = createAssistantMessageEventStream()
+	void (async () => {
+		try {
+			for await (const event of upstream) {
+				relayed.push(
+					event.type === 'error'
+						? { ...event, error: explainFailure(model.id, event.error) }
+						: event,
+				)
+			}
+			relayed.end(await upstream.result())
+		} catch (error) {
+			relayed.fail(error)
+		}
+	})()
+	return relayed
+}, SOURCE_ID)
+
+/**
+ * The catalog we publish. Starts as the seed — the only thing that can exist
+ * before discovery — and is replaced by the discovered list once it lands.
+ */
+let catalog: readonly ProviderModelConfig[] = SEED_MODELS
+
+/**
+ * Write the provider registry from the catalog minus what the user hid.
+ *
+ * This is the *only* place the model list is written, and it writes the whole
+ * list every time: re-registering a provider name replaces its entire model
+ * slice (host contract, `src/config/model-registry.ts`), which is what makes a
+ * filter choice reach `/model` and `--model`.
+ *
+ * Discovery is ours to run, not the host's. Registering `fetchDynamicModels`
+ * looked equivalent and is not: the host caches a dynamic result for 24 h
+ * under a fingerprint of its *own* empty static list, so a warm cache never
+ * calls the fetcher again, and a re-registration inside a session never re-runs
+ * discovery — the models merged at startup stay in the registry no matter what
+ * we filter. Both were measured: a hidden model came straight back after
+ * `/opencode2pi filter`, and one that had never been in the seed survived.
+ */
+function applyProvider(pi: ExtensionAPI): void {
+	const excluded = hidden()
+	const models = catalog.filter((model) => !excluded.has(model.id))
+	if (models.length === 0) {
+		// The host reads a provider with no models as a URL override and keeps the
+		// previous list, so a filter that hides everything would silently do
+		// nothing. Say so instead.
+		log?.warn(`[${PROVIDER}] 过滤规则隐藏了全部模型，模型列表不会缩小；用 /opencode2pi filter 放宽`)
+	}
+
+	pi.registerProvider(PROVIDER, {
+		baseUrl: ZEN_BASE,
+		api: API_ID,
+		apiKey: ANONYMOUS_KEY,
+		authHeader: true,
+		headers: DISGUISE_HEADERS,
+		models: [...models],
+	})
+}
+
+/** Re-discover the free lane and republish. A failure keeps what we have. */
+async function refreshCatalog(pi: ExtensionAPI): Promise<void> {
+	const { models, diagnostics } = await discoverFreeModels()
+	for (const line of diagnostics) log?.info(`[${PROVIDER}] ${line}`)
+	if (models.length === 0) return
+	catalog = models
+	applyProvider(pi)
+}
+
+export default async function (pi: ExtensionAPI): Promise<void> {
+	log = pi.logger
+
+	// The stored choices are part of what the roster is, so they are adopted
+	// before the first registration rather than applied on the next probe.
+	publishFilters(await loadFilters())
+	publishHealth(await loadHealth())
+	onHiddenChange(() => applyProvider(pi))
+	applyProvider(pi)
+
+	// The host awaits extension factories and only then drains the registration
+	// queue (`extensibility/extensions/loader.ts`), so waiting here is what lets
+	// `--model provider/id` resolve an id that discovery found and the seed does
+	// not carry. The seed is already registered, so a discovery failure costs
+	// nothing but the wait — and that wait is the one the host imposed on itself
+	// back when it owned discovery.
+	await refreshCatalog(pi)
+
+	pi.registerCommand('opencode2pi', {
+		description:
+			'opencode2pi 诊断：doctor 检查形状闸门，status 显示上次探测结果，probe 重新探测全部模型，filter 设置隐藏规则',
+		handler: handleCommand,
+	})
+
+	pi.logger.info(
+		`[${PROVIDER}] registered · process session ${PROCESS_SESSION} · seed ${SEED_MODELS.length} models · ` +
+			`verified-free ${Object.keys(VERIFIED_FREE).length} · known-rejected ${Object.keys(UNAVAILABLE).length} · ` +
+			`hidden ${hidden().size}`,
+	)
+}
+
+export {
+	explainFailure,
+	withGateTools,
+	PROVIDER,
+	API_ID,
+	SOURCE_ID,
+	PROCESS_SESSION,
+	SEED_MODELS,
+	UNAVAILABLE,
+	VERIFIED_FREE,
+}