03 — Observability baseline

Type: AFK Label: ready-for-agent

What to build

Structured logs, metrics, and tracing for every chat turn — landed before the tool-loop work so we can debug it. Logs the shape of every turn, never the content (no plaintext user messages, no tool-result payloads).

• Structured CloudWatch logs: one log line per turn with conversationId, userId, tokensIn, tokensOut, latencyMs, bedrockRequestId, and message-shape counts (number of user / assistant / tool-use / tool-result blocks). No plaintext.
• CloudWatch metrics: turns-per-user, tokens-in / tokens-out, end-to-end turn latency, Bedrock error count. Per-user dimensions for cost attribution.
• AWS X-Ray tracing on the chat Lambda, instrumented through to the Bedrock client call.

Acceptance criteria

• Every turn produces exactly one structured log line in CloudWatch with the fields above.
• CloudWatch metrics exist and update in real-time as turns are processed.
• An X-Ray trace for a single turn shows the Lambda invocation and the Bedrock subsegment.
• Smoke test: ten conversations from two different users produce distinguishable entries in logs and per-user metric dimensions.
• No log line contains user message text, system prompt text, or model output text.

Blocked by

• 02-bedrock-text-conversation-streaming