fix(processor): Design step 'model produced invalid content' — retry + max_tokens (ADO #43771)

[Shreyas-Microsoft]

Purpose

Surgical fix for ADO #43771 — [Demo] [Container-Migration] - Design step fails with "The model produced invalid content".

The fix is 2 production files, 11 net lines.

Root cause

The Design step's ResultGenerator agent produces a complex nested JSON schema (Design_ExtendedBooleanResult). The previous max_tokens=12_000 cap was truncating that output mid-stream, which causes Azure OpenAI's own response-validation to reject it with the message "The model produced invalid content". The orchestrator's retry decorator wasn't recognising that error as transient, so the Design step failed outright.

Change

File	Change
src/processor/src/libs/agent_framework/azure_openai_response_retry.py	_looks_like_rate_limit() now also returns True for messages containing "model produced invalid content" / "invalid content". This causes AzureOpenAIResponseClientWithRetry._inner_get_response() (and its streaming variant) to retry on this transient AOAI response-validation failure instead of bubbling it up. Up to 8 retries with exponential backoff + jitter (configurable via AOAI_429_* env vars).
src/processor/src/libs/base/orchestrator_base.py	ResultGenerator agent's max_tokens bumped from 12_000 → 25_000 to give the nested-JSON serialiser headroom and prevent the truncation that triggers the error in the first place.

Verification

• Live REPL against the new predicate:
  • _looks_like_rate_limit("The model produced invalid content") → True ✅
  • _looks_like_rate_limit("Invalid content was returned") → True ✅
  • _looks_like_rate_limit("429 Too Many Requests") → True (unchanged) ✅
  • _looks_like_rate_limit("Some random error") → False (no over-retry) ✅
• ResultGenerator flows through orchestrator_base.py:189 for the Design step (src/processor/src/steps/design/orchestration/design_orchestrator.py:233), so the max_tokens bump applies exactly where the failure occurs.
• Backend unit tests: 585 passed, 93.28% coverage (gate 82%).
• Processor unit tests: 812 passed, 87.43% coverage (gate 82%).

Commit

• edf076b — fix(processor): retry 'model produced invalid content' and bump ResultGenerator max_tokens (Bug #43771)

Does this introduce a breaking change?

• Yes
• No (the retry predicate gains an extra transient-error case but never loses one; the max_tokens bump is a per-agent ceiling increase only)

Golden Path Validation

• Backend tests: 585 passed, 93.28% coverage (gate 82%)
• Processor tests: 812 passed, 87.43% coverage (gate 82%)

Deployment Validation

• Not applicable — code-only change to the processor. No infrastructure or deployment changes.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
TOTAL	3097	208	93%

report-only-changed-files is enabled. No files were changed during this commit :)

Tests	Skipped	Failures	Errors	Time
588	0 💤	0 ❌	0 🔥	22.990s ⏱️

[github-actions]

Processor Coverage Report •

File	Stmts	Miss	Cover	Missing
src/processor/src/libs/agent_framework
azure_openai_response_retry.py	363	132	63%	87, 197, 205, 227, 253, 263–265, 389, 396–399, 401–403, 409–411, 428, 449–450, 455–458, 461–464, 466–469, 471, 481–483, 485–488, 490–493, 495–499, 501, 509–511, 513, 529–531, 536, 540–543, 547, 550, 558–559, 565–566, 571, 573, 594, 600–602, 606–607, 618, 620–623, 627, 630, 638–639, 645–647, 649–652, 654–665, 668, 675–676, 701–702, 709–710, 712, 716, 719–720, 724–725, 727–729, 737, 739–741, 743–745, 747–748, 758
src/processor/src/libs/base
orchestrator_base.py	165	49	70%	62, 68, 71–74, 80–81, 83–84, 127, 138, 143, 148–149, 153, 161, 163, 165, 172–173, 180, 182, 189, 193, 202, 206, 211, 213–214, 216, 315–317, 320, 326, 333–335, 362–364, 367, 374, 381–383, 434–435
TOTAL	5727	720	87%

Tests	Skipped	Failures	Errors	Time
812	0 💤	0 ❌	0 🔥	19.944s ⏱️