[Kem-Ada]

Summary

Brings src/pricing.ts to parity with the Python SDK's identark/pricing.py. Before this change
the TS module was a static lookup table plus estimateCost; the Python module has an active
pricing table that can be overridden three ways. This PR ports that model.

• Adds the ModelPricing interface, replacing the untyped Record<string, number> values.
• Adds the five bundled models the TS table was missing (o1, o1-mini,
  claude-sonnet-4-20250514, claude-opus-4-20250514) plus a new GEMINI_PRICING table, and
  composes all four provider tables into BUNDLED_PRICING — 18 models, matching Python.
• Adds getPricing, setPricingTable and listKnownModels over the active table.
• Supports the same overrides at the same precedence as Python:
  IDENTARK_PRICING_URL > ~/.identark/pricing.json > bundled defaults.
• Rewires estimateCost onto the active table, so an override actually reaches the cost_usd
  figures DirectGateway reports (it calls estimateCost at src/gateways/direct.ts:301,389,498,562).

Two pre-existing estimateCost bugs fall out of that rewire, both parity breaks against Python's
dict.get:

1. Falsy-zero rates. The old guard if (!rates || !rates.input || !rates.output) treated a model
   priced at {input: 0, output: 0} as unpriced and billed it at the $10/1M unknown-model
   fallback. Python returns 0.0. Now covered by two tests.
2. Inherited Object.prototype members. pricing["toString"] returned the inherited function,
   so estimateCost("toString", …) computed NaN and silently poisoned session cost totals.
   getPricing now gates on Object.hasOwn.

Neither was reachable before this PR without a hand-written model name, but setPricingTable makes
user-supplied keys a normal path, so both are fixed here rather than deferred.

The existing OPENAI_PRICING / ANTHROPIC_PRICING / MISTRAL_PRICING exports are kept as-is —
they are public API at v1.0.3, so consolidating them into BUNDLED_PRICING would be breaking.

Python→TS parity mapping

Reference: identark/pricing.py (147 lines), tests/unit/test_pricing.py (8 tests).
Target: src/pricing.ts, src/index.ts, new tests/pricing.test.ts.

Python (identark/pricing.py)	TypeScript (src/pricing.ts)	Notes
class ModelPricing(TypedDict)	export interface ModelPricing	input / output, USD per 1M tokens
_BUNDLED_PRICING (18 models)	BUNDLED_PRICING, composed from the four provider tables	same 18 models, same rates
_pricing_table, _initialized	module-scope let pricingTable, let initialized	same caching semantics
_load_local_overrides()	loadLocalOverrides() — node:os homedir() + node:fs readFileSync	same path ~/.identark/pricing.json; same swallow-and-warn on parse/IO error
_fetch_remote_pricing(url) (sync urllib, 5 s)	fetchRemotePricing(url) — fetch + AbortSignal.timeout(5000)	same 5 s budget; async (see below)
_initialize()	initializeSync() + export async function initPricing()	see "The one divergence"
get_pricing(model) -> ModelPricing | None	getPricing(model): ModelPricing | undefined	None → undefined
set_pricing_table(table)	setPricingTable(table)	merges over bundled, sets initialized
list_known_models() -> list[str]	listKnownModels(): string[]	Object.keys of the active table
estimate_cost(model, input_tokens, output_tokens, provider="openai")	estimateCost(model, inputTokens, outputTokens, provider = "openai")	now reads the active table
logger.warning(...)	console.warn("identark.pricing: …")	the TS SDK has no logger; logger.info/logger.debug calls are dropped rather than made unconditional stdout noise

The one divergence. Python's _initialize() is synchronous because urllib.request.urlopen
is, which lets get_pricing transparently fetch the remote table on first use. Node has no
synchronous HTTP primitive, so exact parity is unreachable. The split taken here:

• initializeSync() runs on every getPricing / listKnownModels call and handles the
  ~/.identark/pricing.json layer. getPricing and estimateCost stay synchronous, so nothing
  downstream — including DirectGateway's cost tracking — changes shape.
• await initPricing() performs the full Python _initialize(), remote fetch included. Callers
  who set IDENTARK_PRICING_URL must call it once at startup.

Python's precedence is preserved exactly, including the elif: when IDENTARK_PRICING_URL is set,
the lower-priority local file is never read, and a failed fetch leaves bundled defaults rather
than falling through. initializeSync therefore returns early without marking initialization
done when a URL is configured, so a later initPricing() can still apply the remote table.

Test plan

Real gate output, run from inside identark-js on feat/pricing-parity:

$ npm ci
added 235 packages, and audited 236 packages in 6s

$ npm run lint
> identark@1.0.3 lint
> eslint src --max-warnings 0
(clean)

$ npm run typecheck
> identark@1.0.3 typecheck
> tsc --noEmit
(clean)

$ npx vitest run
 ✓ tests/types.test.ts              (5 tests)   5ms
 ✓ tests/mock-gateway.test.ts       (12 tests)  12ms
 ✓ tests/control-plane.test.ts      (17 tests)  18ms
 ✓ tests/vercel-integration.test.ts (5 tests)   15ms
 ✓ tests/managed-tools.test.ts      (2 tests)   13ms
 ✓ tests/direct-gateway.test.ts     (13 tests)  19ms
 ✓ tests/pricing.test.ts            (35 tests)  440ms

 Test Files  7 passed (7)
      Tests  89 passed (89)
   Duration  1.50s

$ ../scripts/check_scope.sh
check_scope: OK — no control-plane code touched.

54 tests / 6 files → 89 tests / 7 files (+35). The 54/6 baseline in Section 2 of the brief is
reproduced exactly by the six pre-existing files above, so nothing regressed.

tests/pricing.test.ts mirrors all 8 Python tests in tests/unit/test_pricing.py and adds
negative coverage the Python suite lacks:

Positive (mirroring Python): known-model shape and rates for gpt-4o, claude-3-5-sonnet-20241022
and mistral-large-latest; estimateCost arithmetic; provider: "local" → 0; unknown-model
fallback arithmetic; listKnownModels membership; setPricingTable override with bundled models
still reachable.

Negative / TS-only: unknown model → undefined; empty-string model; Object.prototype member
names; malformed ~/.identark/pricing.json falls back to bundled and warns once; missing file;
remote fetch that rejects; remote fetch returning HTTP 500; local file ignored when a pricing URL is
set; local file not applied when a URL is set but initPricing() never called; fetch happens once
across repeated initPricing() calls; setPricingTable suppresses both file and remote layers;
zero-priced and half-free models; zero tokens; BUNDLED_PRICING not mutated by setPricingTable.

Tests are hermetic: homedir() is mocked to a fresh mkdtempSync directory per test, so a real
~/.identark/pricing.json on the developer's machine cannot affect the suite. Module-level cache
state is handled with vi.resetModules() plus a dynamic re-import per test, so no test-only reset
hook is added to the production API.

Control-plane: untouched (check_scope OK)

🤖 Generated with Claude Code