fix: move project evaluation to claude sonnet 5.5 after bedrock retired sonnet 4 (lfx-cdp-insights-ops#820)

[ulemons]

Summary

Since 2026-10-16 04:00 UTC every project evaluation fails with Bedrock's "This model version has reached the end of its life" for us.anthropic.claude-sonnet-4-20250514-v1:0. The onboarding bot still parses requests (Haiku 4.5) but every new non-LF project ends in "⚠️ something went wrong on our side", and the nightly discovery run produced 15 error rows and nothing else. The same model is behind the member-enrichment workflows (processMemberSources is already failing in cdp-alerts).

Step 1 of 2 (this PR): move only LlmQueryType.PROJECT_EVALUATION to Claude Sonnet 5.5 as a quick production check of the new model. Step 2: once an evaluation succeeds on it, a follow-up PR moves the other seven query types and the two hard-coded ids in the merge-suggestions workflows, as asked by the CDP team.

Why Sonnet 5.5 and not 4.5: on Bedrock Sonnet 4.5 is already "Legacy" with EOL on 2027-04-08; Sonnet 5.5 launched 2026-09-28 with EOL no sooner than 2027-09-28, supports the Invoke API with anthropic_version: bedrock-2023-05-31 from us-east-1 (what LlmService uses), and costs $2/$10 per MTok instead of $3/$15.

Changes

• enums/llm.ts: CLAUDE_SONNET_5_5 = 'us.anthropic.claude-sonnet-5-5' (US geo inference id from the AWS model card).
• llm.ts: region and pricing for it; PROJECT_EVALUATION points to it. Every other query type stays on Sonnet 4 for now.

Before deploying

Bedrock lists Sonnet 5.5 access as "See Access" rather than "Open": make sure the production AWS account has the model enabled (Bedrock → Model access), e.g. aws bedrock get-inference-profile --inference-profile-identifier us.anthropic.claude-sonnet-5-5 --region us-east-1. If it is not available, the fallback is Sonnet 5 (us.anthropic.claude-sonnet-5, open access).

Test plan

• vitest types, pnpm lint --deny-warnings, pnpm tsc-check
• Deploy backend (hosts the evaluation endpoint)
• Press "Try again" on the charmbracelet/wish thread: the evaluation runs and the row leaves error
• llmPromptHistory shows a project_evaluation row with model us.anthropic.claude-sonnet-5-5

Refs linuxfoundation/lfx-cdp-insights-ops#820

🤖 Generated with Claude Code

[cursor]

PR Summary

Medium Risk
Changes production LLM model and response parsing for project evaluation; mis-parsed multi-block responses or missing Bedrock model access would break evaluations until fixed.

Overview
Project evaluation is moved off retired Claude Sonnet 4 to Claude Sonnet 5.5 on Bedrock, with region/pricing wired for the new model id. Only LlmQueryType.PROJECT_EVALUATION changes in this PR; other query types stay on Sonnet 4 for a follow-up.

That query’s Bedrock args swap temperature: 0 for thinking: { type: 'between_tools' }, matching Sonnet 5.5’s extended reasoning mode.

LlmService no longer assumes the answer is body.content[0].text; it uses firstTextBlock so the parsed answer is the first type: 'text' block when the response includes non-text blocks (e.g. thinking) ahead of the final text.

^{Reviewed by Cursor Bugbot for commit 37c51af. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Copilot]

🟡 Changes recommended

The pricing configuration understates costs for the selected US geographic inference profile.

1 open finding

• Correct US geographic profile pricing for Claude Sonnet 4.5 · New

What changed in this PR

Moves project evaluation to Claude Sonnet 4.5 to restore evaluations after Sonnet 4 retirement, leaving other query types unchanged.

Changes:

• Adds the Sonnet 4.5 model identifier.
• Registers its region and pricing and selects it for project evaluation.

File	Description
services/​libs/​types/​src/​llm.ts	Adds model mappings and switches project evaluation.
services/​libs/​types/​src/​enums/​llm.ts	Defines the Sonnet 4.5 US inference profile.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 2804145. Configure here.}

[Copilot]

🟡 Changes recommended

Request parameters and response parsing remain incompatible with the new model and can prevent successful evaluations.

2 open findings

• Extract text blocks by type from Sonnet responses · New
• Remove unsupported temperature override for Sonnet 5.5 · New

1 resolved since last review

• Correct US geographic profile pricing for Claude Sonnet 4.5

🧠 Review effort: Balanced

[Copilot]

🔵 Needs a closer look

Production model access and a successful end-to-end evaluation remain unverified.

2 open findings

• Remove unsupported temperature override for Sonnet 5.5
• Extract text blocks by type from Sonnet responses

🧠 Review effort: Balanced

[Copilot]

🔵 Needs a closer look

Production model access and a successful end-to-end evaluation still require confirmation.

0 open findings

2 resolved since last review

• Remove unsupported temperature override for Sonnet 5.5
• Extract text blocks by type from Sonnet responses

🧠 Review effort: Balanced