fix(run_agent): surface traceback frame for empty-message errors in retry display

[AndreKurait]

What does this PR do?

Fixes the retry loop's 📝 Error: display showing a blank line when an exception has an empty string representation. Walks the traceback and surfaces the last frame instead.

Motivation — real-world encounter

While running the agent against bedrock/global.anthropic.claude-opus-4-7, the retry loop logged:

❌ Error on attempt 1/3 after 0.63s (elapsed: 0.63s): RuntimeError: Unexpected event order, got error before "message_start"
📝 Error:

The 📝 Error: line was empty. That RuntimeError comes from anthropic/lib/streaming/_messages.py:454 when Bedrock returns an error event as the first stream event — the SDK's accumulator rejects it without forwarding the payload. There are several other paths with the same symptom:

• Bare raise AssertionError() (no message).
• Bare raise with no argument.
• Third-party SDK guards that raise empty exceptions.
• Any code doing raise RuntimeError() for flow control.

In all cases, today the user sees 📝 Error: followed by nothing — forcing them to attach a debugger to learn where it came from.

Fix

In AIAgent._summarize_api_error(), when str(error).strip() is empty, format the last traceback frame:

if not raw.strip():
    import traceback as _tb
    tb = getattr(error, "__traceback__", None)
    if tb is not None:
        frames = _tb.extract_tb(tb)
        if frames:
            last = frames[-1]
            return (
                f"{type(error).__name__} at {last.filename}:{last.lineno} "
                f"in {last.name}() — {last.line or '(no source)'}"
            )
    return f"{type(error).__name__} (no message, no traceback)"

Now the display becomes:

📝 Error: RuntimeError at /path/venv/.../anthropic/lib/streaming/_messages.py:454 in _process_event() — raise RuntimeError(f'Unexpected event order...')

…which is immediately actionable.

Which type of PR is this?

• Bug fix (non-breaking change which fixes an issue)
• Documentation update (none)

How has this been tested?

New test file tests/agent/test_summarize_api_error.py:

pytest tests/agent/test_summarize_api_error.py -v
# 4 passed in 4.88s

Covers:

1. Bare AssertionError() — shows frame location + source line.
2. Empty-message error with no traceback — falls back to "<Type> (no message, no traceback)".
3. Whitespace-only message — treated as empty, frame shown.
4. Non-empty message — fall-through unchanged (regression guard).

Risk

• Zero impact on non-empty errors: not raw.strip() short-circuits immediately.
• traceback is stdlib, imported lazily inside the guarded branch.
• getattr(error, "__traceback__", None) handles exceptions constructed but never raised.
• No behaviour change for any existing caller.

Related

Sibling PRs addressing Bedrock/Anthropic runtime robustness:

• fix(bedrock): evict cached boto3 client on stale-connection errors #14710 — evict stale cached boto3 clients between calls
• fix(bedrock): resolve context length via static table before custom-endpoint probe #14721 — resolve Bedrock model context length correctly
• feat(config): expose "max" as a valid reasoning effort level #14717 — add "max" reasoning effort level
• fix(switch_model): guard _fallback_chain access on partially-constructed agents #14731 — guard _fallback_chain on partially-constructed agents

A follow-up PR will intercept the specific "Unexpected event order" RuntimeError at the Bedrock adapter layer and re-raise it with the discarded payload — this PR is the universal safety net that also catches the many other empty-error paths in the codebase.

Checklist

• My changes follow the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation (N/A — internal)
• My changes generate no new warnings
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally with my changes
• Any dependent changes have been merged and published

[teknium1]

Thanks for addressing an actionable diagnostics gap. The underlying blank-summary behavior is still present on current main: run_agent.py:2192-2195 returns an empty string for a bare empty-message exception, and agent/conversation_loop.py:2954-2971 renders that directly after 📝 Error:.

Problems

• The proposed early fallback would now bypass current main's provider-payload recovery at run_agent.py:2171-2190. That path handles an empty SDK message by extracting and redacting response.text (added in 093f567f0), so traceback formatting must be the final fallback rather than the first branch after raw = str(error).

Suggested changes

• Place the traceback/type fallback immediately before the final raw-string fallback at run_agent.py:2192.
• Merge the regression coverage into tests/run_agent/test_summarize_api_error.py and add an empty-message error with useful response.text, asserting the response payload still wins.

This is an automated hermes-sweeper review.