[eli-r-ph]

💡 Motivation and Context

Capture corrects clock skew from each request's PostHog-Request-Timestamp, so every physical send needs a fresh stamp.
The capture session's HTTP adapter retried a failed connection up to twice inside one SDK attempt, reusing the stamped headers.
After slow connection failures, capture saw a stale timestamp and shifted every event in that request later by the time spent connecting.

A raising before_send hook was logged with its exception message, which can hold the data the hook was removing.
The migration guide also listed three of the four legacy option properties.

The event encoder copied every event to replace NaN and Infinity with null before encoding, which made each encode 1.5-1.7x slower even when the event had no such value.

Changes

• Capture uses its own session whose adapter never retries, so each connection try is an SDK attempt with its own attempt number and timestamp. It is rebuilt after fork and when socket options change, like the other sessions.
  • The adapter's read retries never applied to capture: urllib3 does not retry a read error on a POST. Only connection retries change.
  • Other requests (remote config, flag definitions, metrics, tracing) keep the existing session.
• A raising before_send logs one error line with the event name and exception type, and the traceback at debug level, in Client and AsyncPosthog.
• A before_send that returns something other than a dict or None logs its type and drops the event.
• The event encoder encodes strictly first and copies the event only when that fails on a non-finite number. Output is unchanged.
• The migration guide adds $product_tour_id, the one-connection-try-per-attempt change, and the new hook log.
• The migration guide adds a reverse proxy checklist: forward both /i/v1 paths, pass Authorization and the PostHog-* headers, and accept 20 MB bodies.

💚 How did you test it?

• New test: a refused connection is tried exactly once per SDK attempt through the real session stack (it failed with 9 tries for 3 attempts before this change).
• Hook tests assert the exact error line, no traceback on it, and one debug traceback, for both clients.
• Session tests cover no adapter retries, socket options and the fork rebuild for the capture session.
• Encoder tests check the exact output with allow_nan on and off, and that events without NaN or Infinity skip the copy (fails before this change).
• ruff format --check, ruff check, mypy with the baseline, and the full test suite pass.

📝 Checklist

• I reviewed the submitted code.
• I added tests to verify the changes.
• I updated the docs if needed.
• No breaking change or entry added to the changelog.

If releasing new changes

• Ran sampo add to generate a changeset file

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

Written with Claude Code (Claude Opus 5.5, claude-opus-5-5) under the direction of the assignee.
The capture session reuses the flags session's no-retry builder rather than changing the general session, so non-capture requests keep their behavior.

🤖 Generated with Claude Code