moai/.planning/phases/05-multi-model-discussions/05-03-PLAN.md

phase	plan	type
05-multi-model-discussions	03	execute

Implement /discuss mode with sequential rounds, context building, and /next, /stop commands (M5 milestone).

Purpose: Enable structured multi-round discussions where each model sees prior responses. Output: Working /discuss, /next, /stop commands with full conversation context passed to each model.

<execution_context> ~/.claude/get-shit-done/workflows/execute-phase.md ~/.claude/get-shit-done/templates/summary.md </execution_context>

@.planning/PROJECT.md @.planning/ROADMAP.md @.planning/STATE.md

Prior plan context:

@.planning/phases/05-multi-model-discussions/05-02-SUMMARY.md

Key files:

@src/moai/core/orchestrator.py @src/moai/core/services/discussion.py @src/moai/bot/handlers/discussion.py @SPEC.md (system prompts and discussion flow)

Tech stack available: python-telegram-bot, openai (async), sqlalchemy Established patterns: query_models_parallel, Discussion/Round/Message persistence, typing indicator Constraining decisions:

• 05-02: Orchestrator pattern established
• 04-02: Typing indicator for AI calls

Task 1: Add context building and sequential round execution to orchestrator src/moai/core/orchestrator.py Add to orchestrator.py: - build_context(discussion: Discussion) -> list[dict] - Converts all rounds/messages to OpenAI message format - Returns list of {"role": "assistant"/"user", "content": "**Model:** response"} - Models see their own responses as assistant, others' as user (simplified: all prior as user context) - Include original question as first user message

• async run_discussion_round(discussion: Discussion, models: list[str], project_name: str) -> dict[str, str]
  • Builds context from all prior rounds
  • Calls each model SEQUENTIALLY (not parallel) so each sees previous in same round
  • Returns dict mapping model → response
  • Creates Round(type=SEQUENTIAL) and Messages via discussion service

Sequential means: Claude responds, then GPT sees Claude's response AND responds, then Gemini sees both. Use asyncio loop, not gather, to ensure sequential execution within round. Import orchestrator, verify build_context and run_discussion_round exist orchestrator.py has build_context and run_discussion_round with sequential model calls

Task 2: Implement /discuss command with round limit src/moai/bot/handlers/discussion.py, src/moai/bot/handlers/__init__.py Add to discussion.py: - discuss_command(update, context) handler for "/discuss [rounds]" - Requires selected project with models - Requires active discussion (from /open) or starts new one with inline question - Parses optional rounds argument (default: 3 from project settings or hardcoded) - Stores round_limit and current_round in context.user_data["discussion_state"] - Runs first round via run_discussion_round - Displays round results with "**Round N:**" header - Shows "Round 1/N complete. Use /next or /stop"

Register /discuss handler. Store discussion_id in user_data for /next and /stop to reference. After /open, run /discuss 3, verify round 1 executes with sequential responses /discuss starts sequential discussion, stores state for continuation, displays formatted output

Task 3: Implement /next and /stop commands src/moai/bot/handlers/discussion.py, src/moai/bot/handlers/__init__.py Add to discussion.py: - next_command(update, context) handler for "/next" - Reads discussion_state from user_data - Error if no active discussion or round_limit reached - Increments round counter, runs run_discussion_round - If final round: auto-complete discussion, show "Discussion complete (N rounds)" - Otherwise: show "Round N/M complete. Use /next or /stop"

• stop_command(update, context) handler for "/stop"
  • Reads discussion_state from user_data
  • Completes discussion early via complete_discussion service
  • Clears discussion_state from user_data
  • Shows "Discussion stopped at round N. Use /consensus to summarize."

Register both handlers in init.py. Run full flow: /open → /discuss 2 → /next → verify round 2 runs → /stop or let it complete /next advances rounds with context, /stop ends early, both clear state appropriately

Before declaring plan complete: - [ ] Full flow works: /open → /discuss 3 → /next → /next → auto-completes - [ ] /stop works mid-discussion - [ ] Each round shows sequential responses (Claude first, then GPT seeing Claude, etc.) - [ ] Round counter displays correctly (Round 1/3, Round 2/3, etc.) - [ ] Discussion marked COMPLETED when finished - [ ] Error messages for: no discussion, round limit reached

<success_criteria>

• /discuss starts sequential multi-round discussion
• /next advances with full context passed to models
• /stop ends discussion early
• Models see all prior responses in context
• M5 milestone (Discuss mode sequential) complete </success_criteria>

After completion, create `.planning/phases/05-multi-model-discussions/05-03-SUMMARY.md`