feat: add live data pull for opengradient token by kylexqian · Pull Request #55 · OpenGradient/tee-gateway

kylexqian · 2026-04-16T09:32:18Z

Add live data pull for opengradient token. Currently set to ETH as our token is not yet listed
Added more CI/CD tests surrounding pulling data from coingecko
This also includes a TTL of the coin check time as 2 minutes

…ation Replaces the hardcoded Decimal("1") mock with a real CoinGecko ETH/USD price fetch (ETH used as OPG proxy until OPG is listed). Caches the price for 2 minutes so at most one network call is made per TTL window. Falls back to last-known-good price on refresh failure, or a hard-coded $2000 floor if no price has ever been fetched. Config constants (TTL, CoinGecko ID, fallback price) are centralised in config.py so switching to real OPG requires only a one-line change. Adds test_util.py with 18 tests covering the fetch, cache, fallback, and full dynamic_session_cost_calculator pipeline. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test_pricing.py was relying on the old Decimal("1") mock price; patch get_token_a_price_usd to $1 via setUpModule/tearDownModule so the hardcoded expected wei values remain valid as pure USD pricing unit tests. Add tests/test_integration.py (pytest -m integration) with 4 live tests: - _fetch_opg_price_usd returns a positive Decimal from real CoinGecko - ETH price falls within a sanity range ($100–$100 000) - Second call within TTL returns cached value (no second network call) - Full dynamic_session_cost_calculator pipeline with live price Add a separate integration-tests CI job in test.yml and register the integration marker in pyproject.toml. Also add pythonpath = ["."] to pytest config so tests/ can import tee_gateway when run standalone. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add OPG_PRICE_SANITY_MIN_USD / OPG_PRICE_SANITY_MAX_USD to config.py ($0.000001–$1 000 000) so the integration test stays in sync when switching from the ETH proxy to real OPG without touching test code. Widen lower bound from $100 to $0.000001 to accommodate a potentially low OPG token price. Rename test to test_price_is_within_sanity_bounds. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

All mentions of ETH/ethereum in comments, log messages, error strings, and test helpers now refer to OPG_PRICE_COINGECKO_ID so that switching to the real OPG token requires only a one-line change in config.py. - util.py: docstring, error message, and log line use the coin ID var - test_util.py: mock response body and URL assertion key off the config - test_integration.py: comments and print output use the coin ID var - config.py: comments no longer mention ETH specifically Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

That method no longer exists in the current og-x402 library — the patch was a no-op and mypy flagged it as attr-defined on PaymentMiddleware. In the current library the session_cost_calculator is called inside _accumulate_session_cost(), which is wrapped by a broad try/except in StreamingSessionResponse.close(). Replace the dead code with a comment explaining the current behaviour and where to patch if stricter error propagation is needed in future. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add OPG_PRICE_FETCH_RETRIES = 3 to config.py. _fetch_opg_price_usd now loops up to that many attempts, logging a warning on each failure and re-raising the last exception if all attempts are exhausted. Add two tests: succeeds-on-last-attempt and raises-after-all-retries, both keyed off the config constant so they stay correct if the default changes. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add validate_pricing_preflight(model) to util.py. It checks two things before any LLM call is made: 1. model is in the registry (get_model_config raises on unknown models) 2. token price is positive (guards against a degenerate price feed state) Both controllers call it immediately after parsing the request and return 400 Bad Request if it raises, so an unrecognised model or broken price feed is rejected before inference rather than silently producing free inference after the response has already been sent. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- config.py: use "opengradient" as OPG_PRICE_COINGECKO_ID; drop OPG_PRICE_HARD_FALLBACK_USD and OPG_PRICE_SANITY_MIN_USD - util.py: raise clear ValueError when CoinGecko returns no price for the coin (token listed but no trading price yet); remove stale-cache and hard-fallback branches in get_token_a_price_usd — failures now propagate so inference is blocked with a 400 rather than silently using a made-up price - test_util.py: replace hard-fallback test with a "no usd key" test; update stale-cache test to assert the error propagates - test_integration.py: restructure around OPG having no price yet — verify the slug is recognised, verify the correct error is raised, skip price-dependent tests until OPG begins trading Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- _fetch_opg_price_usd: don't retry ValueError — no-price and non-positive-price are deterministic CoinGecko responses that won't change on retry; only network/transport exceptions get retried now - validate_pricing_preflight: remove dead price <= 0 guard (impossible since get_token_a_price_usd() now raises instead of returning a fallback) and fix stale docstring that still mentioned the removed hard fallback - test_raises_when_coin_has_no_price: assert call_count == 1 to lock in the no-retry behaviour Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

test_tool_forwarding.py: the chat controller now runs validate_pricing_preflight before every request, which hits CoinGecko for real. These tests mock the model and TEE keys but not the price feed, so they were getting 400s or 429s. Added the same module-level get_token_a_price_usd patch used in test_pricing.py. test_integration.py: _opg_has_price() makes one CoinGecko call at collection time; test_fetch_raises_clear_error_when_no_price then made a second immediate call, triggering a 429. The test now handles HTTPError 429 by skipping rather than failing — rate-limiting is transient and not the thing being tested. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

ani/token-update removed BASE_TESTNET_OPG_ADDRESS and BASE_TESTNET_NETWORK; update hardcoded asset address in test_util.py and the import in test_integration.py to use BASE_MAINNET_OPG_ADDRESS. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Copilot

Pull request overview

This PR adds a live CoinGecko-backed USD price feed for the OPG payment token (with caching/TTL) and wires a pricing preflight into request handlers so pricing failures are detected before any LLM call. It also updates payment/network constants to Base mainnet, bumps og-x402, and expands CI test coverage with offline unit tests plus an opt-in live integration test job.

Changes:

Implement live OPG/USD fetch via CoinGecko with retry + 2-minute TTL cache, and add validate_pricing_preflight() to gate requests early.
Switch gateway payment configuration from Base testnet to Base mainnet (network + token + recipient), and add the ERC20 approval gas sponsoring extension to x402 routes.
Add offline unit tests for price fetching/caching + dynamic pricing, add a live-network integration test file, and split CI into unit vs integration jobs.

Reviewed changes

Copilot reviewed 12 out of 13 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
`tee_gateway/util.py`	Adds CoinGecko fetch + cached token pricing and a pricing preflight helper.
`tee_gateway/config.py`	Introduces CoinGecko feed configuration (TTL/retries/coin id/sanity max).
`tee_gateway/controllers/chat_controller.py`	Calls pricing preflight before handling chat requests.
`tee_gateway/controllers/completions_controller.py`	Calls pricing preflight before handling completion requests.
`tee_gateway/definitions.py`	Moves network/token constants to Base mainnet and updates recipient/token address constants.
`tee_gateway/__main__.py`	Updates x402 registration/routes for mainnet and adds gas sponsoring extension.
`tee_gateway/test/test_util.py`	New offline unit tests for fetch/caching and dynamic cost math with mocked network.
`tests/test_integration.py`	New live-network integration tests targeting CoinGecko behavior.
`tests/test_pricing.py`	Pins token price to keep pricing unit tests deterministic and updates token address constant.
`tee_gateway/test/test_tool_forwarding.py`	Pins token price to avoid live network calls during tool-forwarding tests.
`.github/workflows/test.yml`	Runs new unit test file and adds a separate integration test job.
`pyproject.toml`	Bumps `og-x402` constraint and adds pytest integration marker config.
`uv.lock`	Locks updated dependency graph including `og-x402` and `nest-asyncio`.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-19T10:39:24Z

+import json  # noqa: E402
+import urllib.request  # noqa: E402
+
+from tee_gateway.config import (  # noqa: E402
+    OPG_PRICE_CACHE_TTL_SECONDS,
+    OPG_PRICE_COINGECKO_ID,
+    OPG_PRICE_FETCH_RETRIES,
+)
 from tee_gateway.definitions import (  # noqa: E402
    ASSET_DECIMALS_BY_ADDRESS,
 )
 from tee_gateway.model_registry import get_model_config  # noqa: E402


These imports are introduced mid-file and suppressed with multiple # noqa: E402. This makes the module harder to read/maintain and can mask real import-order issues. Consider moving the new CoinGecko/config imports up with the rest of the imports at the top of the file (or isolating the price-feed code into its own module) so E402 suppression isn’t needed.

definitely fix this

Copilot · 2026-04-19T10:39:24Z

+
+
+class TestGetTokenAPriceUSD(unittest.TestCase):
+    """get_token_a_price_usd must respect the TTL and fallback gracefully."""


The docstring says get_token_a_price_usd "must respect the TTL and fallback gracefully", but the implementation and other tests explicitly expect refresh failures to raise immediately (no fallback). Update this docstring to match the actual behavior to avoid confusion when maintaining the cache semantics.

Suggested change

"""get_token_a_price_usd must respect the TTL and fallback gracefully."""

"""get_token_a_price_usd must respect the TTL and raise on refresh failures."""

Copilot · 2026-04-19T10:39:24Z

+# CoinGecko coin ID for the OPG token.
+# https://www.coingecko.com/en/coins/opengradient
+OPG_PRICE_COINGECKO_ID: str = "opengradient"
+


PR description mentions the live-price pull is "currently set to ETH" because OPG isn't listed yet, but the code is configured to query CoinGecko ID "opengradient". Either update the PR description (or switch config to ETH if that’s still the intended temporary behavior) so implementation and documentation match.

Copilot · 2026-04-19T10:39:25Z

+    try:
+        validate_pricing_preflight(chat_request.model)
+    except ValueError as exc:
+        return {"error": "Bad Request", "message": str(exc)}, 400


validate_pricing_preflight() can raise non-ValueError exceptions (e.g. urllib.error.URLError / OSError) when CoinGecko is unreachable. This try/except only catches ValueError, so transient network failures will bubble up as an unhandled 500 rather than a controlled client error. Consider catching broader exceptions from validate_pricing_preflight and returning an appropriate status (e.g. 503) while still using 400 for true client input errors (unknown model, no price data).

Suggested change

return {"error": "Bad Request", "message": str(exc)}, 400

return {"error": "Bad Request", "message": str(exc)}, 400

except Exception as exc:

logger.exception(

"Pricing preflight validation failed due to an upstream/service error"

)

return {

"error": "Service Unavailable",

"message": "Pricing validation is temporarily unavailable. Please try again later.",

}, 503

Copilot · 2026-04-19T10:39:25Z

+    try:
+        validate_pricing_preflight(body.model)
+    except ValueError as exc:
+        return {"error": "Bad Request", "message": str(exc)}, 400


validate_pricing_preflight() may raise non-ValueError exceptions (e.g. URLError/OSError) when the CoinGecko fetch fails. Since only ValueError is caught here (and this call is outside the main try/except below), network failures can become unhandled 500s. Catch broader exceptions from validate_pricing_preflight (and map them to something like 503) while reserving 400 for invalid request inputs.

Suggested change

return {"error": "Bad Request", "message": str(exc)}, 400

return {"error": "Bad Request", "message": str(exc)}, 400

except Exception as exc:

logger.error(f"Pricing preflight error: {str(exc)}", exc_info=True)

return {

"error": "Service Unavailable",

"message": "Pricing validation unavailable",

}, 503

Copilot · 2026-04-19T10:39:25Z

    "POST /v1/completions": RouteConfig(
        accepts=[
            PaymentOption(
                scheme="upto",
                pay_to=EVM_PAYMENT_ADDRESS,
                price=AssetAmount(
-                    amount=COMPLETIONS_OPG_SESSION_MAX_SPEND,
-                    asset=BASE_OPG_ADDRESS,
+                    amount=CHAT_COMPLETIONS_OPG_SESSION_MAX_SPEND,
+                    asset=BASE_MAINNET_OPG_ADDRESS,
                    extra={


The /v1/completions RouteConfig is using CHAT_COMPLETIONS_OPG_SESSION_MAX_SPEND for its static session cap. definitions.py has a separate COMPLETIONS_OPG_SESSION_MAX_SPEND constant; using the chat cap here is likely accidental and will be wrong if the two caps diverge. Import and use COMPLETIONS_OPG_SESSION_MAX_SPEND for the completions route.

adambalogh · 2026-04-19T10:40:52Z

            ),
        ],
+        extensions={
+            **declare_erc20_approval_gas_sponsoring_extension(),


is this new?

adambalogh · 2026-04-19T10:46:31Z

+    try:
+        validate_pricing_preflight(body.model)
+    except ValueError as exc:
+        return {"error": "Bad Request", "message": str(exc)}, 400


why not just let the exception propagate?

normally this would be 500 no? since it's outside of the user's control

adambalogh · 2026-04-19T10:52:13Z

+    "last_good": None,
    "updated_at": 0.0,
 }
 _token_price_lock = threading.Lock()


is there really a need for a lock?

adambalogh · 2026-04-19T10:54:56Z

+# StreamingSessionResponse.close().  That close() wraps the call in a broad
+# try/except, so any exception raised by the calculator (e.g. ValueError for
+# an unrecognised model or missing usage data) is logged but otherwise
+# swallowed — the session cost is simply not incremented for that request.


why is this?

adambalogh · 2026-04-19T10:56:14Z

+#   "updated_at" – epoch seconds of last successful fetch (float)
 _token_price_cache: dict[str, Any] = {
-    "value": Decimal("1"),
+    "last_good": None,


please create a namedtuple at least instead of raw dicts

also it wouldn't hurt to create a class instead of global variables for storing these if possible

adambalogh · 2026-04-19T10:57:25Z

+    inference after the response has already been sent.
+    """
+    get_model_config(model)  # raises ValueError for unknown models
+    get_token_a_price_usd()  # raises if price is unavailable


this should run in the background not add potentially 5s delay to requests

adambalogh · 2026-04-19T10:59:02Z

-# response to the client instead of an incorrect silent charge.
+# A previous version of this file monkey-patched _resolve_session_request_cost
+# to let exceptions propagate, but that method no longer exists in the library.
+# If stricter error propagation is needed in future, patch _accumulate_session_cost


isn't this some regression?

adambalogh

please clean the code up

Log at ERROR level when validate_pricing_preflight raises — covers both the CoinGecko no-price case in util.py and the catch sites in the chat and completions controllers, so the root cause is visible in server logs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

kylexqian · 2026-04-21T00:08:37Z

Changed approach in #56

dixitaniket and others added 26 commits April 1, 2026 18:53

init token update

0f9726f

base mainnet changes

dac78ea

addr updates

ce3a491

dep updates

015edb9

Merge remote-tracking branch 'origin/main' into ani/token-update

d81fbc7

updates

6bd4075

token addr update

7d3ec36

Merge branch 'main' into ani/token-update

4e8f3f3

Remove unused USDC var and lint

793cc93

extension change

fa2c545

mainnet changes

baf01e9

Lint

e9f0d27

More lint changes

edda294

Fix lint

bada03e

Fix lint

5207da9

kylexqian force-pushed the claude/admiring-goldberg branch from e686d86 to 2bb2ed6 Compare April 19, 2026 06:26

adambalogh requested a review from Copilot April 19, 2026 10:34

Copilot started reviewing on behalf of adambalogh April 19, 2026 10:35 View session

Copilot AI reviewed Apr 19, 2026

View reviewed changes

adambalogh reviewed Apr 19, 2026

View reviewed changes

adambalogh requested changes Apr 19, 2026

View reviewed changes

kylexqian closed this Apr 21, 2026



		class TestGetTokenAPriceUSD(unittest.TestCase):
		"""get_token_a_price_usd must respect the TTL and fallback gracefully."""

-        return {"error": "Bad Request", "message": str(exc)}, 400
+        return {"error": "Bad Request", "message": str(exc)}, 400
+    except Exception as exc:
+        logger.exception(
+            "Pricing preflight validation failed due to an upstream/service error"
+        )
+        return {
+            "error": "Service Unavailable",
+            "message": "Pricing validation is temporarily unavailable. Please try again later.",
+        }, 503

Conversation

kylexqian commented Apr 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 19, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

adambalogh Apr 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

adambalogh left a comment

Choose a reason for hiding this comment

Uh oh!

kylexqian commented Apr 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

kylexqian commented Apr 16, 2026 •

edited

Loading

adambalogh Apr 19, 2026 •

edited

Loading