| 1 | package agent |
| 2 | |
| 3 | import "testing" |
| 4 | |
| 5 | // TestDeliveryClassificationMatrix protects the boundary between advisory |
| 6 | // troubleshooting, observable read-only work, and mutation delivery. |
| 7 | func TestDeliveryClassificationMatrix(t *testing.T) { |
| 8 | tests := []struct { |
| 9 | name string |
| 10 | input string |
| 11 | wantEvidence bool |
| 12 | wantMutation bool |
| 13 | }{ |
| 14 | {"make sense", "why does this make sense?", false, false}, |
| 15 | {"node prose", "why does the node selection matter?", false, false}, |
| 16 | {"swift prose", "why does Swift concurrency work this way?", false, false}, |
| 17 | {"go prose", "why does this go wrong so often?", false, false}, |
| 18 | {"remote url", "why can't I open https://github.com?", false, false}, |
| 19 | {"remote analysis", "can you analyze why Outlook won't sync?", false, false}, |
| 20 | {"how to install", "how do I install the plugin?", false, false}, |
| 21 | {"chinese how to install", "如何安装插件", false, false}, |
| 22 | {"chinese how to modify", "怎么修改 WPS 设置", false, false}, |
| 23 | {"python prose", "why is Python popular?", false, false}, |
| 24 | {"docker prose", "why is Docker popular?", false, false}, |
| 25 | {"pytest prose", "why is pytest popular?", false, false}, |
| 26 | {"backtick python prose", "why is `Python` popular?", false, false}, |
| 27 | {"backtick code identifier", "what does `context.Context` mean?", false, false}, |
| 28 | {"double dash prose", "why is this -- strangely -- happening?", false, false}, |
| 29 | {"advisory upgrade", "为什么升级插件后失败", false, false}, |
| 30 | {"advisory adjustment", "为什么调整设置后没有效果", false, false}, |
| 31 | {"advisory change application", "为什么申请修改失败", false, false}, |
| 32 | {"negated change", "please don't make the requested changes", false, false}, |
| 33 | {"negated chinese", "请勿修改代码", false, false}, |
| 34 | {"negated inspection", "Explain why it fails. Do not inspect or change files.", false, false}, |
| 35 | {"negated chinese inspection", "解释失败原因，不要检查或修改文件", false, false}, |
| 36 | {"review only", "review only and do not fix anything", true, false}, |
| 37 | {"markdown anchor", "why does README.md render incorrectly?", true, false}, |
| 38 | {"git command", "why does git diff --check fail?", true, false}, |
| 39 | {"custom command", "why does mytool --strict fail?", true, false}, |
| 40 | {"repository anchor", "can you analyze this repository and explain why it fails?", true, false}, |
| 41 | {"observable audit", "audit the current configuration", true, false}, |
| 42 | {"make changes", "make the necessary changes", true, true}, |
| 43 | {"thanks then update", "thanks for fixing that, now update the tests", true, true}, |
| 44 | {"chinese thanks then update", "谢谢你，请继续修改配置", true, true}, |
| 45 | {"thanks then review", "thanks for the help; now review this pull request", true, false}, |
| 46 | {"modify", "please modify the config", true, true}, |
| 47 | {"push", "push the branch", true, true}, |
| 48 | {"commit", "commit the fix", true, true}, |
| 49 | {"move", "move the file", true, true}, |
| 50 | {"bump", "bump the dependency", true, true}, |
| 51 | {"advice then fix", "why does it fail and fix it", true, true}, |
| 52 | {"polite fix before advice", "could you please fix why it fails", true, true}, |
| 53 | {"chinese polite fix before advice", "请你帮我修复为什么会失败", true, true}, |
| 54 | {"chinese advice then fix", "为什么失败然后修复它", true, true}, |
| 55 | {"negated install then patch", "I cannot install dependencies but patch the parser", true, true}, |
| 56 | {"shared negation", "I cannot install and update dependencies", false, false}, |
| 57 | {"reset negation", "I cannot install dependencies and please update config", true, true}, |
| 58 | {"chinese reset", "无法安装依赖请在现有配置中修改", true, true}, |
| 59 | {"negated request", "不想请团队修改代码", false, false}, |
| 60 | {"negated application", "禁止申请修改配置", false, false}, |
| 61 | {"deferred conversation token", "Remember ORBIT-42 and answer on the next turn.", false, false}, |
| 62 | {"chinese deferred conversation token", "记住 ORBIT-42，下一轮回答", false, false}, |
| 63 | {"deferred token with durable negation", "Remember ORBIT-42 for the next turn. Do not save it permanently.", false, false}, |
| 64 | {"chinese deferred token with durable negation", "记住 ORBIT-42，下一轮回答。不要写入文件或长期记忆。", false, false}, |
| 65 | {"long conversational context", "Please keep this code in mind because I will ask you about it in my next message", false, false}, |
| 66 | {"durable memory", "Remember ORBIT-42 permanently across sessions", true, true}, |
| 67 | {"chinese durable memory", "请永久记住 ORBIT-42，跨会话也要保留", true, true}, |
| 68 | {"durable memory advice", "How do I save a preference permanently across sessions?", false, false}, |
| 69 | } |
| 70 | for _, tt := range tests { |
| 71 | t.Run(tt.name, func(t *testing.T) { |
| 72 | if got := deliveryTaskNeedsEvidence(tt.input); got != tt.wantEvidence { |
| 73 | t.Errorf("deliveryTaskNeedsEvidence(%q) = %v, want %v", tt.input, got, tt.wantEvidence) |
| 74 | } |
| 75 | if got := deliveryTaskNeedsMutation(tt.input); got != tt.wantMutation { |
| 76 | t.Errorf("deliveryTaskNeedsMutation(%q) = %v, want %v", tt.input, got, tt.wantMutation) |
| 77 | } |
| 78 | }) |
| 79 | } |
| 80 | } |
| 81 |