Harmony-server models (gpt-5.x / openai-codex) reject any Responses
request whose input reproduces reserved control-token spellings such as
`<|channel|>analysis`, failing before generation with invalid_prompt /
"Request blocked". The live agent-loop builders serialized user and
tool-result text verbatim into input_text / function_call_output, so a
tool result carrying the marker (e.g. a broad `omp://` grep) poisoned
the persisted session — every later turn replayed it and failed.
Escape reserved harmony tokens on the transport copy of untrusted
user/tool text in both the shared Responses builder and the codex
builder, gated on `preferredDialect === "harmony"`. The persisted
transcript is untouched and non-harmony models are unaffected. Extends
the existing compaction-payload escaping to the request boundary and
routes both through a shared `escapeHarmonyControlTokens` helper.
Fixes#6913
(cherry picked from commit dfea56588e7aa1c35971c4684221f4a86142e04e)
