Stop dialog replies from repeating their own openings

[iamwavecut]

Problem

Since the dialog moved to NInfer Gemma (2026-10-11), replies collapse into one template in every chat: an opener that praises the person («Имя, ну ты и мастер…! 😂»), a paraphrase of their message («это ты так намекаешь…?»), a recap and a promise with an emoji tail. In 48 h of production replies, 38% of Gemma replies open with such a reaction and 9.1% open like one of the bot's last three replies in the same chat (Qwen on the same contexts: 2.6%).

Causes:

• the Gemma model prior;
• NInfer ignores the dry_* and repeat_penalty fields the bot sends, and its frequency/presence penalties count only the current reply, so nothing penalises a template repeated across replies;
• the bot's own replies in the history act as examples.

Changes

• prompts/chat/_shared_core.prompt: a reply has no intro or outro; no praise or evaluation of the person; no reading their message aloud; plain speech without literary padding; length that fits the message; variety also covers openings and endings.
• openplotva-dialog::reply_shape + aifarm.rs: a final reply that opens like one of the bot's last three replies (repeated_opening), or announces a drawing or generation that no call of the turn started (unexecuted_action), is rejected once per turn and re-sampled with a note. A re-sample is delivered as it is.
• openplotva-telegram::telegram_html_from_stray_markdown, called from prepare_dialog_chat_response: Markdown left in a dialog reply (**bold**, *italic* at word edges, `code`, fences, headings, links, * bullets, *** rules) becomes Telegram HTML. Text inside <code>/<pre> and tag markup stays as it is; censored words like бл**ь stay literal.
• DEFAULT_DIALOG_MODEL_OUTPUT_RETRIES 1 → 5.

Measurement

Replay of 128 captured production dialog requests through the production Gemma (temperature 0.2; two baseline runs matched almost exactly):

metric	baseline	this PR
reaction/praise intro	38.6%	33.3%
praise of the person	7.9%	5.2%
quote echo of the message	7.9%	3.1%
Markdown written by the model	14.9%	8.3%
reply length p50 (chars)	246	154
native tool calls where production called a tool	13/18	10/18
transcript envelope copies (unwrapped downstream)	3/128	8/128

On 2000 production replies with Markdown, the converter leaves visible Markdown in 5.

An earlier prompt draft with literal HTML tag examples, a new final_check and a new identity made Gemma wrap replies in tags, invent links, copy the transcript envelope and drop native tool calls; an ablation found these blocks, and they stay as they were.

Deploy note

Production .env.production had DIALOG_MODEL_OUTPUT_RETRIES=0; it is set to 5 on the server. With 0, a rejected reply would go to the Qwen/Luna fallbacks instead of a Gemma re-sample.

Checks

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test -p openplotva-dialog, -p openplotva-llm, -p openplotva-prompts, -p openplotva-config, -p openplotva-telegram, -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 11 changed file(s).

Generated by 🚫 dangerJS against 924f7bf

[iamwavecut]

Danger prompt warning: the change is intentional. Partial names (shared_core), role markers, element tags and template data shape are unchanged; element texts in _shared_core.prompt changed, one element (human_speech) was added, and resample_note.prompt gained two verdict lines. The system prefix cache resets once per chat after deploy. The prompt was measured on 128 replayed production requests (see the PR description).