Assistant response lost on mid-stream timeout

[FernandoCelmer]

Problem

During long interactions (model with slow reasoning), the assistant's response can be lost entirely from the history — the user sends a message, waits, and only gets a generic error with no trace of the response.

Root cause

pycodeloop/providers/generic.py: a single timeout (default 60s, line ~194/336) covers the socket read for every chunk of the SSE stream. If the model takes longer than 60s "thinking" without emitting any bytes, the read in _stream() (lines ~403-410) raises TimeoutError mid-stream.

When that happens:

1. The accumulated text (text) and partial tool_calls (pending) are local variables inside _stream() — they're discarded, the exception propagates without returning anything.
2. agent.py (line ~44) treats TimeoutError as retryable and resends the request from scratch via _complete(), hitting the same 60s ceiling again, up to _MAX_RETRIES = 3 (line ~27).
3. After exhausting retries, the exception propagates with no try/except through Agent.run() / CodeLoop.run().
4. pycodeloop/cli/chat.py (line ~432) catches the exception generically, logs "✗ Error: ..." and stops — session.add_assistant() is never called.

Result: the user's message stays in the history with no corresponding response — it looks like the reply "disappeared".

Secondary finding

Session._repair_dangling_tool_calls() (pycodeloop/core/session.py, lines ~59-92) already handles orphaned tool_calls from a lost connection (inserts a synthetic "Cancelled — connection was lost before this tool ran." message), but there's no equivalent handling for when the entire text response never makes it into the session.

Tertiary finding

Agent._run_tool_calls() (line ~279): when a parallel tool call times out, executor.shutdown(wait=False) leaves the thread running in the background, but its real result is never used — another silent work-loss spot.

Proposed fix

1. In _stream(), catch timeout/network errors mid-read and return a partial ProviderResponse with stop_reason="connection_lost" instead of letting the raw exception propagate — reusing the pattern already in place for saw_terminal_marker=False, but only when something was actually accumulated (otherwise keep raising so the existing retry logic in agent.py still applies to a pure connection failure).
2. Increase the default timeout (60s is short for models with long reasoning).

[FernandoCelmer]

Fixed in #82.