Commit
Β·
b9709ff
1
Parent(s):
8abea0c
docs: Deep technical analysis of P2 duplicate report bug
Browse filesComprehensive bug report with:
- Executive summary confirming this is a stack bug (not model)
- Microsoft Agent Framework event type reference
- Event flow trace showing exact duplication path
- Key code paths with line numbers
- Buffer clearing analysis (why comparison isn't possible)
- Edge case analysis
- Verification checklist
Root cause: MagenticFinalResultEvent/WorkflowOutputEvent contains
same content that was already streamed, and no deduplication exists.
docs/bugs/P2_DUPLICATE_REPORT_CONTENT.md
CHANGED
|
@@ -4,6 +4,18 @@
|
|
| 4 |
**Status**: OPEN
|
| 5 |
**Severity**: P2 (UX - Duplicate content confuses users)
|
| 6 |
**Component**: `src/orchestrators/advanced.py` + `src/app.py`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
---
|
| 9 |
|
|
@@ -215,7 +227,116 @@ if has_substantial_streaming:
|
|
| 215 |
|
| 216 |
---
|
| 217 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
## Related
|
| 219 |
|
| 220 |
- **Not related to model quality** - This is a stack bug, not model limitation
|
| 221 |
- P1 Free Tier fix (PR fix/P1-free-tier) enabled streaming, exposing this bug
|
|
|
|
|
|
| 4 |
**Status**: OPEN
|
| 5 |
**Severity**: P2 (UX - Duplicate content confuses users)
|
| 6 |
**Component**: `src/orchestrators/advanced.py` + `src/app.py`
|
| 7 |
+
**Affects**: Both Free Tier (HuggingFace) AND Paid Tier (OpenAI)
|
| 8 |
+
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
## Executive Summary
|
| 12 |
+
|
| 13 |
+
This is a **confirmed stack bug**, NOT a model limitation. The duplicate report appears because:
|
| 14 |
+
|
| 15 |
+
1. Streaming events yield the full report content character-by-character
|
| 16 |
+
2. Final events (`MagenticFinalResultEvent`/`WorkflowOutputEvent`) contain the SAME content
|
| 17 |
+
3. No deduplication exists between streamed content and final event content
|
| 18 |
+
4. Both are appended to the output
|
| 19 |
|
| 20 |
---
|
| 21 |
|
|
|
|
| 227 |
|
| 228 |
---
|
| 229 |
|
| 230 |
+
## Deep Technical Analysis
|
| 231 |
+
|
| 232 |
+
### Microsoft Agent Framework Event Types
|
| 233 |
+
|
| 234 |
+
The framework emits these event types (all inherit from `WorkflowEvent`):
|
| 235 |
+
|
| 236 |
+
| Event Type | Purpose | Key Attributes |
|
| 237 |
+
|------------|---------|----------------|
|
| 238 |
+
| `MagenticAgentDeltaEvent` | Streaming tokens | `text`, `agent_id` |
|
| 239 |
+
| `MagenticAgentMessageEvent` | Agent turn complete | `message` (ChatMessage), `agent_id` |
|
| 240 |
+
| `MagenticFinalResultEvent` | Workflow final result | `message` (ChatMessage) |
|
| 241 |
+
| `MagenticOrchestratorMessageEvent` | Manager bookkeeping | `message`, `kind`, `orchestrator_id` |
|
| 242 |
+
| `WorkflowOutputEvent` | Workflow output | `data`, `source_executor_id` |
|
| 243 |
+
|
| 244 |
+
### Event Flow Trace
|
| 245 |
+
|
| 246 |
+
```
|
| 247 |
+
PHASE 1: Agent Streaming (Reporter)
|
| 248 |
+
βββββββββββββββββββββββββββββββββββββ
|
| 249 |
+
MagenticAgentDeltaEvent(text="##", agent_id="reporter") β yields streaming event
|
| 250 |
+
MagenticAgentDeltaEvent(text=" Summary", agent_id="reporter") β yields streaming event
|
| 251 |
+
MagenticAgentDeltaEvent(text="\n", agent_id="reporter") β yields streaming event
|
| 252 |
+
... (hundreds more delta events)
|
| 253 |
+
MagenticAgentDeltaEvent(text=".", agent_id="reporter") β yields streaming event
|
| 254 |
+
|
| 255 |
+
β Result: Full report content in streaming_buffer (app.py) and current_message_buffer (orchestrator)
|
| 256 |
+
|
| 257 |
+
PHASE 2: Agent Completion
|
| 258 |
+
βββββββββββββββββββββββββββββββββββββ
|
| 259 |
+
MagenticAgentMessageEvent(message=ChatMessage(...), agent_id="reporter")
|
| 260 |
+
β _handle_completion_event() yields: "reporter: [first 200 chars]..."
|
| 261 |
+
β Clears current_message_buffer
|
| 262 |
+
β app.py flushes streaming_buffer to response_parts with "π‘ **STREAMING**:" prefix
|
| 263 |
+
|
| 264 |
+
PHASE 3: Workflow Termination (THE BUG)
|
| 265 |
+
βββββββββββββββββββββββββββββββββββββ
|
| 266 |
+
MagenticFinalResultEvent(message=ChatMessage(...)) β Contains SAME full report!
|
| 267 |
+
OR
|
| 268 |
+
WorkflowOutputEvent(data=ChatMessage(...)) β Contains SAME full report!
|
| 269 |
+
|
| 270 |
+
β _process_event() extracts text with _extract_text()
|
| 271 |
+
β Returns AgentEvent(type="complete", message=FULL_REPORT)
|
| 272 |
+
β app.py appends FULL_REPORT to response_parts (NO prefix)
|
| 273 |
+
|
| 274 |
+
RESULT: Report appears twice:
|
| 275 |
+
1. "π‘ **STREAMING**: [full report]"
|
| 276 |
+
2. "[full report again]"
|
| 277 |
+
```
|
| 278 |
+
|
| 279 |
+
### Key Code Paths
|
| 280 |
+
|
| 281 |
+
**`advanced.py` lines 299-345 (main loop):**
|
| 282 |
+
```python
|
| 283 |
+
# Buffer is cleared HERE (line 337) after MagenticAgentMessageEvent
|
| 284 |
+
current_message_buffer = ""
|
| 285 |
+
|
| 286 |
+
# But MagenticFinalResultEvent comes AFTER and _process_event has no buffer context!
|
| 287 |
+
agent_event = self._process_event(event, iteration) # line 341
|
| 288 |
+
if agent_event:
|
| 289 |
+
yield agent_event # line 345 - yields duplicate!
|
| 290 |
+
```
|
| 291 |
+
|
| 292 |
+
**`advanced.py` lines 532-539 (_process_event):**
|
| 293 |
+
```python
|
| 294 |
+
elif isinstance(event, MagenticFinalResultEvent):
|
| 295 |
+
text = self._extract_text(event.message) # Extracts FULL content
|
| 296 |
+
return AgentEvent(type="complete", message=text) # Returns FULL content
|
| 297 |
+
```
|
| 298 |
+
|
| 299 |
+
**`app.py` lines 229-232 (UI handling):**
|
| 300 |
+
```python
|
| 301 |
+
if event.type == "complete":
|
| 302 |
+
response_parts.append(event.message) # Appends to existing streamed content!
|
| 303 |
+
yield "\n\n".join(response_parts)
|
| 304 |
+
```
|
| 305 |
+
|
| 306 |
+
### Why Buffer Clearing Doesn't Help
|
| 307 |
+
|
| 308 |
+
The `current_message_buffer` is cleared (line 337) BEFORE the final events arrive. So even if we wanted to compare, we've already lost the reference:
|
| 309 |
+
|
| 310 |
+
```python
|
| 311 |
+
# Line 327-338: Handle MagenticAgentMessageEvent
|
| 312 |
+
iteration += 1
|
| 313 |
+
comp_event, prog_event = self._handle_completion_event(...)
|
| 314 |
+
yield comp_event
|
| 315 |
+
yield prog_event
|
| 316 |
+
current_message_buffer = "" # CLEARED!
|
| 317 |
+
continue
|
| 318 |
+
|
| 319 |
+
# Line 341-345: Handle final events (buffer is empty now!)
|
| 320 |
+
agent_event = self._process_event(event, iteration) # No buffer context
|
| 321 |
+
```
|
| 322 |
+
|
| 323 |
+
### Potential Edge Cases
|
| 324 |
+
|
| 325 |
+
1. **Tool-only turns**: If agent makes tool calls without text, buffer is empty β fallback text used
|
| 326 |
+
2. **Multiple agents streaming**: Buffer clears on agent switch (line 311-313) β OK
|
| 327 |
+
3. **Timeout**: Uses `_handle_timeout()` which invokes ReportAgent directly β Different path
|
| 328 |
+
4. **No final event**: Falls back to "Research completed..." message (line 354-363) β OK
|
| 329 |
+
|
| 330 |
+
### Verification Needed
|
| 331 |
+
|
| 332 |
+
- [ ] Confirm `MagenticFinalResultEvent` vs `WorkflowOutputEvent` - which is emitted?
|
| 333 |
+
- [ ] Confirm bug occurs on both Free and Paid tiers
|
| 334 |
+
- [ ] Measure content length match between streaming and final event
|
| 335 |
+
|
| 336 |
+
---
|
| 337 |
+
|
| 338 |
## Related
|
| 339 |
|
| 340 |
- **Not related to model quality** - This is a stack bug, not model limitation
|
| 341 |
- P1 Free Tier fix (PR fix/P1-free-tier) enabled streaming, exposing this bug
|
| 342 |
+
- SPEC-17 Accumulator Pattern addressed repr bug but created this side effect
|