hermes - 💡(How to fix) Fix Internal bookkeeping fields leak to strict providers, causing HTTP 400 retry loops

While using opencode-go with Hermes, internal bookkeeping fields (_empty_recovery_synthetic, _thinking_prefill, _empty_terminal_sentinel) are not stripped from messages before API calls. Once they enter conversation history, every subsequent turn fails with a deterministic HTTP 400, triggering a retry loop.

Error

HTTP 400: Error from provider: 2 request validation errors:
Extra inputs are not permitted, field: 'messages[70]._empty_recovery_synthetic', value: True
Extra inputs are not permitted, field: 'messages[71]._empty_recovery_synthetic', value: True

Root cause

Three internal markers are written into message dicts but never stripped at the transport layer:

_thinking_prefill is stripped in some code paths, but _empty_recovery_synthetic and _empty_terminal_sentinel have no stripping logic anywhere. ChatCompletionsTransport.convert_messages() — the transport safety net — also doesn't touch them.

Impact

• Session becomes permanently broken once any of these fields enter messages
• Error classifier retries the deterministic 400, producing "model provider failed after retries"
• Only workaround is starting a new session (history=0)

Related PRs I found

• PR #26054 — targeted the same _empty_recovery_synthetic leak; closed, not merged
• PR #28958 / #29594 — stripped tool_name for strict providers; closed, not merged
• PR #21242 — merged but only addressed persistence, not wire-format stripping

Proposed fix

Strip all three fields in:

1. ChatCompletionsTransport.convert_messages() (transport safety net)
2. conversation_loop.py main loop message builder
3. chat_completion_helpers.py summary/final-response path