1. Telegram xdist mock pollution (37 tests): Add tests/gateway/conftest.py
with a shared _ensure_telegram_mock() that runs at collection time.
Under pytest-xdist, test_telegram_caption_merge.py (bare top-level
import, no mock) would trigger the ImportError fallback in
gateway/platforms/telegram.py, caching ChatType=None and Update=Any
for the entire worker — cascading into 37 downstream failures.
2. VIRTUAL_ENV env var leak (4 tests): TestDetectVenvDir tests monkeypatched
sys.prefix but didn't clear VIRTUAL_ENV. After commit 50c35dca added a
VIRTUAL_ENV check to _detect_venv_dir(), CI's real venv leaked through.
3. Copilot base_url missing (1 test): _resolve_runtime_from_pool_entry()
set api_mode for copilot but didn't add the base_url fallback — unlike
openrouter, anthropic, and codex which all have one. Production bug.
4. Stale vision model assertion (1 test): _PROVIDER_VISION_MODELS added
zai -> glm-5v-turbo but the test still expected the main model glm-5.1.
5. Reasoning item id intentionally stripped (1 test): Production code at
run_agent.py:3738 deliberately excludes 'id' from reasoning items
(store=False causes API 404). Test was asserting the old behavior.
6. context_length warning not reaching custom_providers (1 test): The test
didn't pass base_url to AIAgent, so self.base_url was empty and the
custom_providers URL comparison at line 1302 never matched.
7. Matrix room ID URL-encoding (1 test): Production code now URL-encodes
room IDs (!room:example.com -> %21room%3Aexample.com) but the test
assertion wasn't updated.
8. Google Workspace calendar tests (2 tests): Tests assert on +agenda CLI
args that don't exist in the production calendar_list() function. They
only 'passed' before because _gws_binary() returned None, the Python
SDK fallback ran, googleapiclient import failed, SystemExit was raised,
and post-exit assertions were never reached. Skip when gws not installed.
Remaining 4 failures (test_run_progress_topics.py) are pre-existing flaky
tests that fail inconsistently under xdist — confirmed on clean main.
Previously, non-integer context_length values (e.g. '256K') in
config.yaml were silently ignored, causing the agent to fall back
to 128K auto-detection with no user feedback. This was confusing
for users with custom LiteLLM endpoints expecting larger context.
Now prints a clear stderr warning and logs at WARNING level when
model.context_length or custom_providers[].models.<model>.context_length
cannot be parsed as an integer, telling users to use plain integers
(e.g. 256000 instead of '256K').
Reported by community user ChFarhan via Discord.
