[iamwavecut]

Problem

After #257 (deployed 2026-10-13 22:43Z) the repeated-opening guard re-sampled about one turn in five, and at temperature 0.2 the second sample usually came back with the same opener («Хасн, ну ты и философ!» → «Хасн, ну ты и философ!»). With DIALOG_MODEL_OUTPUT_RETRIES=5 other validators burned up to six Gemma calls per turn before the fallback, and some turns ran out of budget. The text that reaches the chat is also what the next turn imitates, so a delivered template keeps the chat infected.

Changes

• openplotva-dialog::reply_shape::review_reply cuts, when at least three words remain:
  • a first sentence (up to two) that only reacts or rates the person: laughter and interjections alone, «Имя, ну ты и…», «да ты…!», «ты просто…!», «Ой, всё, …!», also after a name or «Слушай,» set off by a comma;
  • a sentence that quotes a run of the current message back with a question or a gloss («это ты…», «намекаешь», «звучит как»), and a «Типа / То есть / Значит…» sentence right after it;
  • a closing promise of a generation that no call of the turn started.
• It re-samples once only when the opening still repeats one of the bot's last eight replies, or when the reply is only a promise.
• Custom personas are not reviewed.
• aifarm.rs delivers the reviewed text, so the history stores it too.
• _shared_core.prompt no_greetings: do not start a reply with the person's name.
• DEFAULT_DIALOG_MODEL_OUTPUT_RETRIES back to 1; production env was set to 1 on 2026-10-13 23:3xZ.

Measurement

• Review over 4000 production replies (30 h, custom personas skipped): 84% kept, 16% trimmed, 0.7% re-sampled (was ~20% re-sampled after Stop dialog replies from repeating their own openings #257). A read of 55 random trimmed replies found no broken meaning; JSON replies and «бл**ь»-style text stay as they are.
• no_greetings replay on 122 production requests through production Gemma: replies starting with a vocative name 25% → 20%; tool calls on the 12 tool turns 10 → 10 (native 7 → 6); reply length unchanged.

Checks

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p openplotva-dialog, -p openplotva-llm, -p openplotva-config, -p openplotva-prompts, -p openplotva-app --lib dialog_jobs

[github-actions]

	Warnings
⚠️	Prompt files are cache-sensitive contracts; confirm prompt shape and role markers were preserved intentionally.

	Messages
📖	Danger checked 4 changed file(s).

Generated by 🚫 dangerJS against 1d1776f

[iamwavecut]

Danger prompt warning: intentional. Only the text of the no_greetings element in _shared_core.prompt changed; tags, partials, role markers and template data are unchanged. Replay on 122 production requests is in the PR description.