fix(agent): trigger fallback when content filter terminates stream mid-delivery (#32421)

[Tranquil-Flow]

What does this PR do?

When a provider's output safety filter (e.g. MiniMax new_sensitive, Azure content_filter, Bedrock guardrail_intervened) kills a streaming response mid-delivery, the partial stream stub is now tagged with _content_filter_terminated=True. The conversation loop detects this tag and immediately activates fallback_providers instead of retrying against the same provider (which would hit the same filter).

Root cause: interruptible_streaming_api_call already handled stream errors by building a partial SimpleNamespace stub, but it didn't distinguish content-filter errors from transient network errors. The loop then retried against the same provider, which produced the same filter rejection.

Fix:

1. agent/chat_completion_helpers.py: Extract is_content_filter_stream_error() helper that matches known provider content-filter error keywords (new_sensitive, content_filter, guardrail_intervened, safety, recitation, refusal) and error code 1027.
2. agent/conversation_loop.py: Check _content_filter_terminated flag on partial stream stubs before continuation retry; if set, trigger try_activate_fallback().

Related Issue

Fixes #32421

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/chat_completion_helpers.py: Added _CONTENT_FILTER_STREAM_ERROR_PATTERNS and is_content_filter_stream_error() helper
• agent/conversation_loop.py: Added content-filter tag check before continuation retry, with fallback activation
• tests/agent/test_32421_content_filter_fallback.py: 22 tests covering keyword detection (16 parametrized positive + negative cases), stub tagging, loop tag reading, fallback activation, no-fallback warning, and false-positive guards

How to Test

1. pytest tests/agent/test_32421_content_filter_fallback.py -v
2. All 22 tests should pass
3. Fail-without-fix: revert agent/chat_completion_helpers.py and agent/conversation_loop.py to upstream/main → tests fail with ImportError: cannot import name 'is_content_filter_stream_error'

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 15.6

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A
• I've updated tool descriptions/schemas if I changed tool behavior — or N/A

[teknium1]

Fixed on main via PR #54064 (merge commit 4f16950), which salvages this approach along with #37851 and #33845.

Your stub-tag approach (_content_filter_terminated + a loop check that routes to fallback) is exactly what landed — credited in the merged PR. The one change: instead of a second _CONTENT_FILTER_STREAM_ERROR_PATTERNS list, the merged version runs the swallowed error through the existing error_classifier (content_policy_blocked) at the swallow point, so there's a single source of truth for "what counts as a content filter."

Verified live end-to-end: before the fix _try_activate_fallback was called 0 times on a MiniMax new_sensitive stall; after, fallback fires on the first stub and the fallback provider completes the turn. Thanks for the fix!