fix(query): send reasoning effort on the OpenAI-compatible wire

[ericleepi314]

The bug

--effort was a silent no-op for every OpenAI-compatible provider on the headless -p path — the path the terminal-bench harness drives.

_call_model_sync emitted the level only inside its is_anthropic branch, as output_config.effort. The OpenAI-compatible wire takes a top-level reasoning_effort body field, and nothing emitted it. Captured against a local capture server:

--effort max --provider openrouter
  → body: {messages, model, stream, stream_options, tools}

No effort field of any kind. No error, no log line — the job config would faithfully record effort=max while the request carried nothing.

This invalidates the effort setting on past non-Anthropic evals. eval/harbor/jobs/tb21-deepseek-max is configured effort=max; DeepSeek is OpenAI-compatible; it ran at DeepSeek's default. Any comparison against a new run at a real effort level needs that baseline re-run.

What's in here

The fix. Adds the missing branch, and drops the xhigh clamp on this wire — _model_supports_xhigh_effort is an allowlist of Anthropic model-name substrings, so it matched nothing here and silently downgraded every xhigh to high. xhigh is a first-class OpenAI level (their guide lists none|minimal|low|medium|high|xhigh|max and recommends xhigh for long agentic runs).

Deletes _EffortProvider. Fixing the above collided with it: the wrapper injected the same field with setdefault while routing returned thinking_effort=None, so query.py filled the key from settings.effort first and the wrapper no-op'd. An explicit /effort max went out as medium — inverting the documented precedence. _turn_effort_routing now just forwards the level for both families. One level, one injection site.

Registers openai/gpt-5.6-luna — the OpenRouter id, which is what reaches the provider when Harbor runs the model (it splits --model on the first slash only, so openrouter/openai/gpt-5.6-luna → --provider openrouter --model openai/gpt-5.6-luna). MODEL_CONFIGS is keyed by bare name, so that id matched nothing and fell back to 200K against a 1M-class window. The bare gpt-5.6-* rows from #773 are left alone, including their ordering-based namespace handling. The resolver still does not strip a <vendor>/ prefix ("decision #1"); get_model_config's docstring now records why, with both failure modes named.

Pricing gains Luna's rates including the >272K long-context tier, since a 1M-window model on a benchmark will cross it and cost is a number the eval reports.

⚠️ Behavior change worth knowing

compact.py calls context.provider.chat_async() directly, bypassing the wire boundary, and the deleted wrapper's chat_async override was what carried effort there. Compaction summaries on OpenAI-compatible providers now run at the provider default rather than the session's /effort. Safe direction, and it matches the Anthropic side, which never got output_config on that path either.

Verification

• Wire, before and after the rebase: max → max, xhigh → xhigh, absent when unset; Anthropic still gets output_config and never reasoning_effort.
• Effort is honored, not just accepted — reasoning tokens rise 148 → 154 → 266 → 516 across low/medium/high/max on openai/gpt-5.6-luna. A 200 OK would not have shown this.
• Two live tasks at effort=max, independently verified: a git stage+commit, and writing run_length.py + pytest tests (5 passed on a clean re-run). Cost accounting exact: 74,179 in + 2,801 out → $0.0090985.
• Full suite green. The 27 MCP-auth failures in a fresh worktree venv are pre-existing — reproduced identically on pristine main with this branch's changes stashed.
• Both fixes and every new guard mutation-tested: disabling each makes the intended tests fail, and restoring makes them pass.

Reviewed by the critic subagent across three rounds. It caught the precedence inversion above, and a differential test of mine that was a tautology (it passed with the code under test deleted) — both fixed; the replacement asserts hardcoded expectations and guards itself with assertNotIn(model_id, MODEL_CONFIGS).

Running the model

See eval/harbor/RUN_LUNA_TB21.md. Containers install from PyPI or git+, so benchmarking a non-Anthropic model at a specific effort needs --ak source= pinned at this commit or later.

🤖 Generated with Claude Code