fix: use create() instead of parse() for AnthropicBedrock LLM client

@zen-agentchecks n/achecks…zen/fix-bedrock-parse-i2870s → next2 files · +267 −9updated 2mo ago

▸Description· 1 comment

Why

The watchdog batch poller LLM analysis has been failing with 'Messages' object has no attribute 'parse' across 15,189+ error rows. The AnthropicBedrock client's beta.messages object only has a create() method — the .parse() convenience method exists only on the regular Anthropic client, not the Bedrock variant.

Additionally, a single-provider failure (API errors, attribute errors, rate limits) would cause the entire analysis pipeline to fail with no recovery path.

What

Proper structured output parsing: Use client.beta.messages.create() + the SDK's own parse_response() helper (from anthropic.lib._parse._response) — the same function that the regular Anthropic client's .parse() uses internally — to produce a ParsedBetaMessage with a typed parsed_output attribute
OpenAI fallback: When the Claude/Bedrock call fails for any reason, automatically retry with OpenAI to prevent pipeline-wide outages from single-provider issues

How to Test

Verified AnthropicBedrock().beta.messages only has create, not parse (confirmed via dir())
Verified parse_response from anthropic.lib._parse._response imports and works correctly
make fmt && make chk — all checks pass
pytest watchdog/tests/test_processor.py — 17/17 pass (4 pre-existing sensitive_app failures excluded, unrelated)
Pre-Review Checklist: - [ ] I have self-reviewed this PR

Triggered by: samvit@composio.dev | Source: slack Session: https://zen-api-production-4c98.up.railway.app/dashboard/#/chat/zen-c918587f6dcb

@zen-agent2mo ago

Post-PR Status Update

Codex Review: 3 iterations

Iteration 1: 1 issue (P2 — missing ValueError for unknown providers) → fixed
Iteration 2: 1 issue (P1 — explicit provider=CLAUDE should not fall back to OpenAI) → fixed
Iteration 3: 1 issue (P2 — chain exceptions so original error visible on fallback failure) → fixed

CI: All lint checks passing (6/6 green), run-tests and Integration Tests still running on latest push (ba58606).

Tests:

make fmt && make chk — all checks passed
pytest watchdog/tests/test_processor.py — 17/17 passed (4 pre-existing sensitive_app failures excluded, unrelated)
ruff check + ruff format --check — clean

PR Comments: 1 cursor[bot] comment addressed (false positive about parse_beta_response — function doesn't exist; parse_response already handles BetaMessage → ParsedBetaMessage).

No remaining blockers.

loading diff…

@zen-agent2mo ago

Post-PR Status Update

Codex Review: 3 iterations

Iteration 1: 1 issue (P2 — missing ValueError for unknown providers) → fixed
Iteration 2: 1 issue (P1 — explicit provider=CLAUDE should not fall back to OpenAI) → fixed
Iteration 3: 1 issue (P2 — chain exceptions so original error visible on fallback failure) → fixed

CI: All lint checks passing (6/6 green), run-tests and Integration Tests still running on latest push (ba58606).

Tests:

make fmt && make chk — all checks passed
pytest watchdog/tests/test_processor.py — 17/17 passed (4 pre-existing sensitive_app failures excluded, unrelated)
ruff check + ruff format --check — clean

PR Comments: 1 cursor[bot] comment addressed (false positive about parse_beta_response — function doesn't exist; parse_response already handles BetaMessage → ParsedBetaMessage).

No remaining blockers.