From 5b6fb921960d277e328d6b7ecaee578694997df0 Mon Sep 17 00:00:00 2001 From: Jian Weng Date: Tue, 17 Feb 2026 15:20:27 +0300 Subject: [PATCH] [agent.command] Fix ultra-planner lite routing guidance Summary: - Align ultra-planner command guidance with automatic lite/full routing and clarify dry-run behavior. Changes: - .claude-plugin/commands/ultra-planner.md: remove full-debate-only instruction, add routing step overview, and note that dry-run respects routing. Issue: - Relates to #939. Tests: - TEST_SHELLS="bash zsh" make test-fast (includes pytest python/tests): passed. --- .claude-plugin/commands/ultra-planner.md | 23 +++++++++++------------ 1 file changed, 11 insertions(+), 12 deletions(-) diff --git a/.claude-plugin/commands/ultra-planner.md b/.claude-plugin/commands/ultra-planner.md index 6a3c859b..a1593837 100644 --- a/.claude-plugin/commands/ultra-planner.md +++ b/.claude-plugin/commands/ultra-planner.md @@ -19,12 +19,11 @@ Even if user is telling you "build...", "add...", "create...", "implement...", o you must interpret these as making a plan for how to have these achieved, not actually doing them! - **DO NOT** make any changes to the codebase! -2. This command uses a **multi-agent debate system** to generate high-quality plans. -**No matter** how simple you think the request is, always strictly follow the multi-agent -debate workflow below to do a thorough analysis of the request throughout the whole code base. -Sometimes what seems simple at first may have hidden complexities or breaking changes that -need to be uncovered via a debate and thorough codebase analysis. - - **DO** follow the following multi-agent debate workflow exactly as specified. +2. This command uses **automatic routing** to select the right planning path: + - **Lite path** (planner-lite) for simple, repo-only changes that meet the thresholds + - **Full debate path** for anything that needs deeper analysis or external research +Follow the routing logic in Step 4a. Do not override it unless `--force-full` is used or +you are in `--refine` mode. Create implementation plans through multi-agent debate, combining innovation, critical analysis, and simplification into a balanced consensus plan. @@ -33,13 +32,13 @@ Invoke the command: `/ultra-planner [feature-description]` or `/ultra-planner -- ## What This Command Does -This command orchestrates a multi-agent debate system to generate high-quality implementation plans: +This command orchestrates a multi-agent planning system to generate high-quality implementation plans: 1. **Context gathering**: Launch understander agent to gather codebase context -2. **Three-agent debate**: Launch bold-proposer (with context) first, then critique and reducer analyze its output -3. **Combine reports**: Merge all three perspectives into single document -4. **External consensus**: Invoke external-consensus skill to synthesize balanced plan -5. **Draft issue creation**: Automatically create draft GitHub issue via open-issue skill +2. **Routing**: Use understander output to choose lite or full path +3. **Lite path** (if eligible): Launch planner-lite to produce a single-agent plan +4. **Full path** (otherwise): Launch bold-proposer, then critique + reducer, then synthesize consensus +5. **Issue update**: Update the placeholder or existing issue with the plan (or print dry-run summary) ## Inputs @@ -100,7 +99,7 @@ This command orchestrates a multi-agent debate system to generate high-quality i Accept the $ARGUMENTS. **Dry-run mode:** If we have `--dry-run` anywhere in the arguments, set `DRY_RUN=true` and remove `--dry-run` from the arguments. In dry-run mode: -- All agents run normally (understander, bold-proposer, critique, reducer, consensus) +- Agents run normally according to routing (understander + lite path OR full debate path) - Plan files are saved to `.tmp/` as usual - **Skip** placeholder issue creation (Step 3) - **Skip** issue update (Step 8)