fix(query): recover from OpenAI-compat length truncations; forward non-interactive to headless

[ericleepi314]

Two model-agnostic fixes extracted from the DeepSeek-flash harness work (feat/deepseek-flash-harness) — the parts that benefit every provider, with all DeepSeek-gated and eval-specific changes left behind.

1. Recover from OpenAI-compat finish_reason="length" truncations

Every stop_reason consumer in the query loop keys on the internal (Anthropic) word "max_tokens" — the withheld-content escalation lane (→ 64K retry), the "resume, break into smaller pieces" recovery nudge, and _is_withheld_max_output_tokens. None of them matched the OpenAI wire's "length", so a truncated response from any OpenAI-compatible provider (openai, openrouter, zai, deepseek, …) fell through as a normal end-of-turn: no escalation, no resume nudge, and a headless run ended with a silently-clipped answer.

_call_model_sync now maps finish_reason == "length" → stop_reason = "max_tokens" (Anthropic already emits "max_tokens", so this only normalizes the other wire).

The _api_error = "max_output_tokens" tag is now applied only to tool-free truncations. A truncation that still carries a surviving tool call must be yielded — headless persistence (on_message) is the only route into the saved session — not withheld while its tools execute anyway. This also closes the same latent leak on the Anthropic wire, where max_tokens + tool_use has always been possible.

2. Forward non-interactive mode to the headless prompt

build_effective_system_prompt (the headless/TUI cutover builder) never forwarded options.is_non_interactive_session into build_full_system_prompt_blocks, so the headless (-p) model never received the # Non-Interactive Mode section and behaved as if a human would reply. Now forwarded; interactive callers (TUI bridge, agent-server) leave the flag unset and are unchanged. Model-agnostic.

Tests

• tests/test_openai_length_stop_reason.py — the 64K escalation fires on finish_reason="length"; a "stop" turn is untouched; a truncation carrying a surviving tool call is yielded (not withheld).
• tests/test_headless_non_interactive_section.py — section present when the flag is set, absent otherwise, null-safe when tool_context has no .options.

Verification

• Focused + wider sweep: 2345 passed, 9 skipped, 0 failures.
• Existing max_tokens escalation tests stay green (they use finish_reason="max_tokens" + tool_uses=None, so the tool-free guard still fires).
• Only "length" is remapped — stop / tool_calls / content_filter pass through unchanged.

🤖 Generated with Claude Code