# Public reply-review schema compatibility

Status: repaired and verified; broader built-in-model/startup/lifecycle work is
tracked in `WO-built-in-free-model-startup-20261012.md`.

The live 1.0.850 bot delivered both private and public `pong`. Public generation
succeeded on Muse, but its separate independent review used a larger schema
that the private single-pass path did not require. The previous live acceptance
covered attention and simple replies, leaving this phase untested.

Real probes against the active installed bundle reproduced two sequential
HTTP 400 errors: unsupported `uniqueItems`, then a forbidden top-level `anyOf`.
The fallback allowed only one schema-repair retry and therefore moved to Step
and Mimo. See the safe [before receipt](evidence/public-review-before-20261012.json).

The repair recognizes the exact observed second diagnostic, lowers only the
root `anyOf` of a complete closed response object, preserves nested alternatives
and tool schemas, and permits one bounded retry per newly learned supported
limitation. The original full host schema, identity/evidence checks and semantic
review remain authoritative; structured output stays buffered until valid.

The [live after receipt](evidence/public-review-after-20261012.json) records
Muse HTTP 400 -> 400 -> 200 on the same model, with valid public-review output at
the production 1,300-token limit. No inference or metadata response was mocked;
synthetic context only, no Telegram sends or tools. The harness now includes
the production review schema, policy and parser; the full Telegram-method
multi-model harness is tracked under the broader work order.

187 focused orchestrator regressions and 21 new public-review boundary tests
passed. Those invariant tests use explicit fixture networking and are separate
from the live acceptance above. The workspace build passed.
