feat(desktop): configurable group chat round cap + model-aware limits + token budget guard

[beplee]

Summary

• Add group-chat-config.ts: config.yaml-driven group_chat.* settings + free-model detection
• Make GROUP_CHAT_MAX_ROUNDS config-aware: free models (:free suffix or free_models[] allowlist) get unlimited rounds up to HARD_CAP=20, paid models default to max_rounds: 3
• Add token-budget guard: cumulative room-log tokens exceeding group_chat.token_budget (default 80k chars) force exitKind='capped' independent of round count — the real safety net for pathological cascades
• Add getGroupChatConfig(), isFreeModel(), getEffectiveRoundCap(), getCurrentModelName() helpers
• Add config.yaml schema in cli-config.yaml.example under group_chat:

Why

Hardcoded 3-round cap is too tight for free-model rooms (where API cost is zero) and too loose for paid-model rooms (where each round burns real money). A single config block drives everything automatically.

Config

group_chat:
  max_rounds: 3              # paid models
  max_rounds_free: null      # null = unlimited (HARD_CAP=20 backstop)
  token_budget: 80000        # hard ceiling per drive (chars, ~20k tokens)
  free_models:               # quality-tier allowlist
    - upstage/solar-pro4
    - meituan/longcat-2.0
    - tencent/hy3
  hard_cap: 20               # absolute ceiling even for free models

Safety

• Token budget is the real backstop — fires regardless of model type or round count
• HARD_CAP=20 prevents unbounded free-model cascades (20 rounds × 6 bots = 120 messages max)
• Paid models always capped at max_rounds (default 3)
• exitKind='capped' logged to activity feed for every termination path

Test plan

• Paid model → 3-round cap fires
• Free model (allowlist) → 20-round cap fires
• Token budget exhaustion → mid-drive cap fires
• Model switch mid-drive → cap re-evaluates
• 6-bot free room → halts cleanly at token budget
• Empty free_models + :free suffix → allowlist is sole gate (no bypass)

Refs: #96726, #96842 (closed — superseeded by this port to new module structure)

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Overall: Feature that makes GROUP_CHAT_MAX_ROUNDS=3 config-driven and model-aware: free models ( :free or free_models[] allowlist) get hard_cap=20 up to unlimited, paid models stay max_rounds=3, plus a hard token-budget guard (80k chars ≈20k tokens) that exits capped independent of rounds. Addresses #96726 follow-up and #96842.

What it does

• apps/desktop/src/plugins/hermes-bots/group-chat-config.ts:59 new module: GroupChatConfig (max_rounds, max_rounds_free=null, hard_cap=20, token_budget=80000, free_models:[]), getGroupChatConfig() via host.request('config.get',{key:group_chat}) with defaults, isFreeModel(name,cfg) (/:free$/i or substring allowlist), getEffectiveRoundCap (null/Infinity→hard_cap, else min(cap,hard_cap)), getCurrentModelName() via model.options, estimateTokens≈len/4.
• apps/desktop/src/plugins/hermes-bots/group-rounds.ts:502 resolves effectiveCap at drive start and token-budgets each round: roomLogAtRoundStart → driveTokens = sum(estimateTokens(e.text)), if >= tokenBudget → exitKind=capped + recordGroupActivity capped and return; loop now for (round < effectiveCap).
• cli-config.yaml.example:1554 documents group_chat: block.

Non-blocking notes

• Config read is async once per drive — paid-model cap evaluates before first round; mid-drive model switch only re-evaluated next drive (not per-round re-read) — acceptable as drives short; getCurrentModelName remains available for future per-round re-eval.
• estimateTokens≈chars/4 is rough lower-bound; token_budget default 80k chars labeled chars but treated as token budget — ensure docs clarify it's char-based budget (≈20k tokens).
• free_models[] substring match is case-insensitive includes — tencent/hy3 will match tencent/hy3-preview etc.; allowlist semantics vs exact ID — document substring behavior so operators don't expect exact match.
• Test plan checkboxes are manual; consider adding vitest for isFreeModel/getEffectiveRoundCap/budget guard.

Non-blocking — please use your judgment.