c9f69d9d6d
范围（不含 design-story-foundation、docs/、humanization/README.md 等进行中改动）:
1. 新增 harness/ 控制平面
- skill_harness.py 静态审计：32 个运行时 Skill 的 frontmatter/manifest/文档污染，当前 0 问题
- run_selected.py 选择性执行器：manifest 与磁盘一一对账、依赖阻断、
空跑与 skip-only 失败关闭、AST 测试形状门
- manifests/skills.json：32 个 Skill 的合同责任方与协作领域登记
- manifests/test-inventory.json：81 个测试资产登记
- specs/skill-testing.md 与 README.md：测试分层、证据边界与 harness 职责
2. 实现测试从 .claude/skills/*/scripts/ 迁至 tests/skills/<skill>/
- 71 个测试文件迁移并修复项目根与临时目录运行导入
- 数据库触发器测试宽泛异常收窄为 psycopg.errors.RaiseException
- 抽取离线大测试拆出真实 PG smoke（默认阻断，不计入离线通过）
- 抽取 presence 去重边界拆出独立测试：493 + 78 = 571 项检查不变
3. 运行时文档清理
- 13 个 SKILL.md 移除自测/离线验证段落、测试命令与测试文件事实源表述，
只保留运行时合同；业务运行合同、额度、授权与离线模式均保留
4. SoT 同步
- AGENTS.md：新增 Skill 领域索引（7 个合同责任方分组，覆盖 32 个运行时 Skill）
- 领域 07：测试入口改由 harness/manifests/ 登记，SKILL.md 不承载测试命令
- humanization 覆盖矩阵：活动测试路径同步迁移
验证证据: harness 自测 15 项 + runner 自测 13 项通过；静态审计 32 Skill / 0 问题；
73 个非数据库测试通过；8 个集成条目中 6 个 PostgreSQL 项被依赖门明确阻断；
py_compile 与 git diff --check 通过。未连接 PostgreSQL、网络、真实模型或额度。
已知边界: 真正 skill_behavior_eval 仍为 0，尚未验证任何 Skill 自然语言行为；
evaluate-frozen-replay 的 raw 存储边界冲突留待单独治理。
144 lines
6.6 KiB
YAML
144 lines
6.6 KiB
YAML
schema_version: humanization-research-skill-coverage-v1
|
||
source: ../../../.agents/knowledge/ai-writing-humanization-open-source-research.md
|
||
scope: agent-example自治验证期；研究出处不等于效果证明
|
||
projects:
|
||
- id: humanizer
|
||
mapped_capabilities: [voice_first_baseline, minimal_unique_patch]
|
||
- id: stop-slop
|
||
mapped_capabilities: [five_layer_detection]
|
||
- id: Humanizer-zh
|
||
mapped_capabilities: [preservation_and_modality]
|
||
- id: inkos
|
||
mapped_capabilities: [minimal_unique_patch, bounded_revision_and_original_wins, cross_model_pairwise]
|
||
- id: oh-story-claudecode
|
||
mapped_capabilities: [five_layer_detection, bounded_revision_and_original_wins]
|
||
- id: no-ai-slop
|
||
mapped_capabilities: [detect_edit_separation, pre_generation_guidance]
|
||
- id: PaperSpine
|
||
mapped_capabilities: [detect_edit_separation, bounded_revision_and_original_wins]
|
||
- id: im-not-ai
|
||
mapped_capabilities: [five_layer_detection, holdout_effect_evaluation]
|
||
- id: avoid-ai-writing
|
||
mapped_capabilities: [carrier_scope_and_mask, preservation_and_modality]
|
||
- id: human-writing
|
||
mapped_capabilities: [detect_edit_separation]
|
||
- id: talk-normal
|
||
mapped_capabilities: [pre_generation_guidance, holdout_effect_evaluation]
|
||
- id: AIWriteX
|
||
mapped_capabilities: [minimal_unique_patch]
|
||
- id: humanize-text
|
||
mapped_capabilities: [five_layer_detection, holdout_effect_evaluation]
|
||
- id: shuorenhua
|
||
mapped_capabilities: [carrier_scope_and_mask, sf_snf_boundary_regression]
|
||
- id: academic-humanizer
|
||
mapped_capabilities: [preservation_and_modality]
|
||
- id: speak-human-tw
|
||
mapped_capabilities: [sf_snf_boundary_regression, holdout_effect_evaluation]
|
||
- id: AI_paper
|
||
mapped_capabilities: [holdout_effect_evaluation]
|
||
- id: Openwrite
|
||
mapped_capabilities: [minimal_unique_patch]
|
||
- id: unslop
|
||
mapped_capabilities: [case_to_rule_lifecycle]
|
||
- id: neuro-book
|
||
mapped_capabilities: [five_layer_detection, case_to_rule_lifecycle, holdout_effect_evaluation]
|
||
capabilities:
|
||
- id: detect_edit_separation
|
||
evidence_sources: [no-ai-slop, oh-story-claudecode, academic-humanizer]
|
||
owner_skill: diagnose-ai-flavor
|
||
implementation: humanization/src/deai/diagnose.py
|
||
test: humanization/tests/test_contracts.py
|
||
status: implemented
|
||
- id: five_layer_detection
|
||
evidence_sources: [neuro-book, oh-story-claudecode, avoid-ai-writing]
|
||
owner_skill: diagnose-ai-flavor
|
||
implementation: humanization/src/deai/diagnose.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: implemented
|
||
note: regex/handler/density机械执行；semantic仍需外部模型或人工
|
||
- id: carrier_scope_and_mask
|
||
evidence_sources: [shuorenhua, avoid-ai-writing, neuro-book]
|
||
owner_skill: diagnose-ai-flavor
|
||
implementation: humanization/src/deai/carriers.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: implemented
|
||
- id: voice_first_baseline
|
||
evidence_sources: [humanizer, shuorenhua, inkos]
|
||
owner_skill: establish-voice-baseline
|
||
implementation: .claude/skills/establish-voice-baseline/scripts/establish_voice_baseline.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: implemented
|
||
note: 角色策略仍需planner/作者补充和确认
|
||
- id: pre_generation_guidance
|
||
evidence_sources: [no-ai-slop, neuro-book, humanizer]
|
||
owner_skill: prevent-ai-flavor
|
||
implementation: .claude/skills/prevent-ai-flavor/scripts/prevent_ai_flavor.py
|
||
test: tests/skills/assemble-context/test_assemble_writer_context.py
|
||
status: implemented
|
||
- id: sf_snf_boundary_regression
|
||
evidence_sources: [speak-human-tw, shuorenhua, neuro-book]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: humanization/samples
|
||
test: humanization/src/deai/evaluation.py
|
||
status: implemented
|
||
- id: minimal_unique_patch
|
||
evidence_sources: [inkos, Openwrite, shuorenhua]
|
||
owner_skill: revise-ai-flavor
|
||
implementation: humanization/src/deai/patch.py
|
||
test: humanization/tests/test_contracts.py
|
||
status: implemented
|
||
- id: preservation_and_modality
|
||
evidence_sources: [humanizer, academic-humanizer, shuorenhua]
|
||
owner_skill: revise-ai-flavor
|
||
implementation: humanization/src/deai/gates.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: implemented
|
||
note: POV/因果/伏笔仍需事实快照和语义detector
|
||
- id: bounded_revision_and_original_wins
|
||
evidence_sources: [inkos, oh-story-claudecode, neuro-book]
|
||
owner_skill: revise-ai-flavor
|
||
implementation: humanization/src/deai/pipeline.py
|
||
test: tests/skills/revise-ai-flavor/test_revise_ai_flavor.py
|
||
status: implemented
|
||
note: 当前以最多3轮合同、复扫和pairwise no_gain实现；跨轮最佳快照编排仍由上层负责
|
||
- id: source_revalidation
|
||
evidence_sources: [shuorenhua, neuro-book]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: .claude/skills/capture-ai-flavor-cases/scripts/capture_cases.py
|
||
test: tests/skills/capture-ai-flavor-cases/test_capture_cases.py
|
||
status: implemented
|
||
- id: case_to_rule_lifecycle
|
||
evidence_sources: [unslop, shuorenhua, neuro-book]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: .claude/skills/capture-ai-flavor-cases/scripts/mine_ai_flavor.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: implemented
|
||
note: 状态写回数据库；规则文件激活需显式输出和人工审批
|
||
- id: cross_model_pairwise
|
||
evidence_sources: [inkos, neuro-book, oh-story-claudecode]
|
||
owner_skill: revise-ai-flavor
|
||
implementation: humanization/src/deai/pairwise.py
|
||
test: humanization/tests/test_contracts.py
|
||
status: implemented
|
||
- id: holdout_effect_evaluation
|
||
evidence_sources: [neuro-book, speak-human-tw, im-not-ai]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: humanization/src/deai/evaluation.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: partial
|
||
note: 合同回放和holdout计数门已建；真实多作者多题材holdout尚未形成
|
||
- id: model_drift_deprecation
|
||
evidence_sources: [inkos, im-not-ai, neuro-book]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: humanization/src/deai/evaluation.py
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: partial
|
||
note: 人工审批的deprecate输出已建；自动跨模型巡检尚未接入生产调度
|
||
- id: independent_reader_blind_eval
|
||
evidence_sources: [neuro-book, inkos, humanizer]
|
||
owner_skill: capture-ai-flavor-cases
|
||
implementation: humanization/eval
|
||
test: humanization/tests/test_humanization_v2.py
|
||
status: pending
|
||
note: pairwise合同存在，但独立读者规模化盲评未完成
|