Added
- The OpenAI-compatible tool-loop helper can explicitly load a custom preset
fromproviders.jsonwith its own endpoint, credential and model allowlist.
Custom presets do not yet register orchestration seats or cost-report rows.
Fixed
-
Debate counts the native Sonnet advisor only on an allowed Claude Code route
with a completed round artifact. Codex requires two external advisor results
instead of counting an unverified host subagent as Sonnet. -
Claude Code discovers native workflow skills while Codex loads its separate
adapted skill tree. All existing skill names, explicit invocation controls,
aliases and starter-pack entries remain available. -
Define stops after failed decision, context or completion writes, preserving
earlier state and requiring repair before presenting the result as saved. -
Research prompts request full IDs (
PLAT-1931,probe-4) to avoid false numeric-claim failures.
Merged changes: #1356,
#1369,
#1371,
#1370, and
#1375.
The debate repair fixes #1373.
Custom orchestration seats remain tracked in
#1299, and broader runtime
and installed-host acceptance work remains in
#1304.
Full changelog: v11.25.2...v11.26.0