docs: record reasoning effort verification and deployment

This commit is contained in:
2026-10-20 23:53:13 +01:00
parent 724de27561
commit 8899dd150f
+26
View File
@@ -0,0 +1,26 @@
# Per-stage reasoning effort — 2026-10-20
Model stage configs accept `reasoning_effort`: default/omitted, none (alias off),
minimal, low, medium, high, xhigh, max. Definitions retain the authored value;
the runtime normalizes defaults and the off alias without rewriting versions.
OpenCode V2 passes a native model variant on session creation, for both agent
and text-only stages. OpenCode V1 rejects explicit effort. Pi passes `--thinking`
and verifies the selected model's reasoning capability and effective level before
admitting the task, rejecting silent clamping. Capability is model/version-specific;
current Pi supports max, while older/unsupported models may reject or clamp it.
Omission changes neither transport defaults nor Pi's existing RPC sequence.
Full Go race tests, build/vet, web/installer regressions, pipeline persistence and
model-override tests passed. Native transport tests use isolated fixtures, not
paid provider inference. The managed authoring skill was refreshed with the
current-source installer.
Runtime code `724de275612798c255873a5bbda31183ca360d9e` was deployed as a continuation
of the authorized server update. Active-work checks excluded only this verified
conversation; a SQLite/binary rollback backup was retained. The running server
reports that clean build and validates `reasoning_effort=high`. No pipeline was
published, edited or executed, and no live model generation was used for this
verification. ntfy, OpenCode, and the web service were not restarted.
The upstream push remains blocked by macOS Keychain credential access (`-25308`).