Proxy → native server flag: parity confirmed, proxy retired (0.75)

What was run

First run on the native mechanism: llama.cpp's --reasoning-budget 4000 set in the llama-swap config (server restarted 2026-07-02), replacing the external proxy. Same two signal tasks × 10 attempts (20 trials), thinking ON, port 8020 (direct — from here on, direct runs are budget-capped server-side; the trial JSON can't tell them apart from old uncapped runs, see RUNS.md).

Results

Mean 0.75: regex-log 8/10, openssl-selfsigned-cert 7/10. Verified from the transcripts: - Largest single <think> block ≈ 3,141 tokens (median ~146) vs ~20,000 in the uncapped runs — the cap engages, and the server nudges </think> closed before 4000 rather than hard-splicing at it like the proxy did. - 0 × stopReason: length across all 278 assistant messages — the runaway-turn failure mode is gone at this task scale.

Interpretation

Clean swap: same bounding, statistically identical score (0.75 vs proxy's 0.77), one less process. The server flag is the mechanism going forward. Remaining failures are agent-level (many-turn meandering) or model baseline — not reasoning-budget issues.

Next steps

Stop tuning on 2 tasks — run the full 89-task suite to get breadth, find the failure modes the smoke set can't show, and collect data for the reasoning-budget sweep.