From bd41fa416956112d26cdfbd742d68ba7f38b5a4b Mon Sep 17 00:00:00 2001 From: Rui Hu Date: Tue, 16 Jun 2026 17:19:50 +0800 Subject: [PATCH] save before change --- agents/rui-orchestrator.md | 2 +- agents/rui-review-pr.md | 8 ++++++-- agents/rui-self-qa.md | 8 ++++++++ 3 files changed, 15 insertions(+), 3 deletions(-) diff --git a/agents/rui-orchestrator.md b/agents/rui-orchestrator.md index 1b2d5db..ca7ab3e 100644 --- a/agents/rui-orchestrator.md +++ b/agents/rui-orchestrator.md @@ -2,7 +2,7 @@ description: Top-level coordination agent that routes complex or multi-domain tasks to the appropriate specialized subagent, and handles simple requests directly. mode: primary permission: - edit: deny + edit: allow bash: "*": allow skill: diff --git a/agents/rui-review-pr.md b/agents/rui-review-pr.md index 9e132ee..05961d0 100644 --- a/agents/rui-review-pr.md +++ b/agents/rui-review-pr.md @@ -17,8 +17,8 @@ permission: task: "*": deny "rui-review-pr-focus": allow -model: bifrost/ct/kimi-k2.6 -temperature: 1.0 +model: bifrost/ct/qwen3.6-35b-a3b +temperature: 0.6 --- You are a PR review orchestration agent. Your job is to coordinate a thorough, @@ -72,6 +72,10 @@ When dispatching each agent, explicitly instruct it: Also, you MUST forward whatever context the user gives you to the `rui-review-pr-focus` sub-agents. +If the agent returns empty result, it's likely being cut-off in the middle of the session. +In this case, you should resume the agent to continue. If an agent being resume multiple times +still giving empty result, try dispatch a new agent. If still giving empty result, ignore it and moving on. + ## Step 3: Aggregate Findings After all subagents complete: diff --git a/agents/rui-self-qa.md b/agents/rui-self-qa.md index cf5d005..be8ed31 100644 --- a/agents/rui-self-qa.md +++ b/agents/rui-self-qa.md @@ -47,6 +47,10 @@ post it to Forgejo yet at this stage. DO NOT reuse the review session from previous turns. Always start fresh when dispatching review agent. +If the agent returns empty result, it's likely being cut-off in the middle of the session. +In this case, you should resume the agent to continue. If an agent being resume multiple times +still giving empty result, try dispatch a new agent. + ### Step 2: Evaluate Verdict - If the verdict is **Approve**: the loop is complete. Proceed to Final Report. @@ -76,6 +80,10 @@ need it for the consolidated Forgejo post. DO NOT reuse the fix session from previous turns. Always start fresh when dispatching fix agent. +If the agent returns empty result, it's likely being cut-off in the middle of the session. +In this case, you should resume the agent to continue. If an agent being resume multiple times +still giving empty result, try dispatch a new agent. + ### Step 4: Increment and Check Limit Increment the iteration counter.