🪲 BUG-#62: Persist already-shown text when an SSE chunk is malformed (closes #62, #64)

[FernandoCelmer]

Description

• pycodeloop/providers/generic.py — _stream(): wraps json.loads(payload) in try/except json.JSONDecodeError, breaking out of the loop and returning whatever text/tool-calls were already accumulated (stop_reason="malformed_stream") instead of letting the exception discard everything the caller already streamed to the user. Also tracks whether the loop ended via an explicit terminal marker ([DONE] or a finish_reason chunk) — if the iterator just runs out without one (connection dropped mid-response), stop_reason is now "connection_lost" instead of silently defaulting to "stop", which used to make a truncated response indistinguishable from a clean completion.
• tests/providers/test_generic.py — Adds test_streaming_keeps_already_shown_text_on_malformed_chunk and test_streaming_flags_a_connection_dropped_mid_response. Verified both fail without the fix (the first with an unhandled JSONDecodeError, the second with stop_reason == "stop" instead of "connection_lost") and pass with it.

Motivation and Context

Two related integrity bugs in the same streaming path:

• Malformed SSE chunk drops already-streamed text from history #62 — A truncated/corrupt data: line mid-stream raised unhandled. By that point on_delta(...) had already fired for every prior chunk — those deltas already reached the client and rendered on screen — but the exception propagated up, the turn failed, and session.add_assistant(...) was never called, so the text the user already saw never entered the persisted session.
• Silently-dropped stream connection recorded as a clean stop #64 — If the connection dropped without an explicit finish_reason/[DONE] marker, the for loop just ended and stop_reason silently defaulted to "stop", making a truncated response look like the model finished on its own — future turns build on a "complete" message that's actually cut off mid-thought.

Closes #62. Closes #64.

Types of changes

• Bug fix (change that fixes an issue)
• New feature
• Documentation

Checklist

• I have performed a self-review of my own code
• I have added tests that prove my fix is effective
• I have updated the CHANGELOG
• I have updated the documentation accordingly

[FernandoCelmer]

🔍 Code Review

Code issues found: 2

#	Severity	Comment
1	[Blocking]	stop_reason is overwritten when a malformed chunk follows a clean finish_reason
2	[Suggestion]	Missing test for the finish_reason → malformed sequence