fix: propagate reasoning effort through completion runtime #189

Merged

lost-rob0t merged 5 commits from fix/185-reasoning-effort into main

2026-08-25 05:20:31 +00:00

lost-rob0t commented

2026-08-25 05:17:04 +00:00

(Migrated from github.com)

Closes #185.

What changed

carries reasoning_effort/1 into direct model requests
makes the root planner inherit the selected effort by default
adds planner_reasoning_effort/1 as an explicit trusted override
enforces host-selected reasoning across nested rlm, parallel, and retry model steps so model-produced options cannot silently downgrade it
preserves legacy request shape when no reasoning option is configured
exposes --reasoning-effort and --planner-reasoning-effort in the CLI
documents the closed effort enum

Verification

focused rlm_completion tests
focused CLI tests
runtime load checks
git diff --check
dedicated issue-185 patch gate passed before cleanup

This unblocks expert-KB routing such as openai/gpt-5.6-luna with reasoning_effort(max) from being an enforced runtime selection rather than routing metadata.

Closes #185. ## What changed - carries `reasoning_effort/1` into direct model requests - makes the root planner inherit the selected effort by default - adds `planner_reasoning_effort/1` as an explicit trusted override - enforces host-selected reasoning across nested `rlm`, `parallel`, and `retry` model steps so model-produced options cannot silently downgrade it - preserves legacy request shape when no reasoning option is configured - exposes `--reasoning-effort` and `--planner-reasoning-effort` in the CLI - documents the closed effort enum ## Verification - focused `rlm_completion` tests - focused CLI tests - runtime load checks - `git diff --check` - dedicated issue-185 patch gate passed before cleanup This unblocks expert-KB routing such as `openai/gpt-5.6-luna` with `reasoning_effort(max)` from being an enforced runtime selection rather than routing metadata.