feat: add main harness core

main_harness/.gitignore

@@ -0,0 +1,23 @@
+.DS_Store
+.env
+.env.*
+!.env.example
+.venv/
+__pycache__/
+.pytest_cache/
+*.py[cod]
+*.pem
+*.key
+*.p12
+*.pfx
+tmp/
+Data/
+backup/
+run_records/
+harness_ablation_runs/
+result/
+results/
+outputs/
+dist/
+build/
+*.egg-info/

main_harness/README.md

@@ -0,0 +1,71 @@
+# main_harness
+
+PawBench 的 Harness attribution 与 Feature ablation 模块。此目录只包含源码和测试，不包含 task data、API key 或运行记录。
+
+## 工作流
+
+```text
+PawBench result + trajectory
+  -> Reasoning backend 输出 H/M/Ex code 与 evidence
+  -> bridge 校验 H-code 并映射 F-code
+  -> 生成一个 accepted H-to-F Feature development request
+  -> Qwen3.8-Max through Claude Code 按 skill 实现并写 admission receipt
+  -> 独立校验 receipt 后才允许打开一个 Feature
+  -> paired ablation 与 holdout result 回传 PawBench
+```
+
+- `M-code`：模型可观察行为，不映射 Feature。
+- `Ex-code`：task、scorer 或外部服务问题，不自动映射 Feature。
+- `H-code`：Harness 机制问题，可根据 evidence 选择 0-2 个 F-code 做 ablation。
+
+## 主要代码
+
+| 路径 | 作用 |
+| --- | --- |
+| `scripts/feature_taxonomy.py` | H/M/Ex 定义、15 个 Feature 与 H-to-F mapping。 |
+| `scripts/pawbench_output_adapter.py` | 将 PawBench output 转成 attribution input。 |
+| `scripts/bridge_attribution_to_harness_core.py` | 校验 H-code，生成 Feature switch 建议。 |
+| `candidates/agentscope/` | 默认 AgentScope Harness 及受治理的 Qwen3.8-Max Feature 开发桥接。 |
+| `tests/` | taxonomy、adapter、bridge、report 与 security 测试。 |
+
+## PawBench 接入
+
+Reasoning backend 需为每个 task 提供：
+
+```json
+{
+  "task_id": "example-task",
+  "codes": ["H2"],
+  "evidence": "direct trajectory evidence"
+}
+```
+
+Bridge 输出 `recommended_feature_ids` 与 `recommended_switch_keys`。在 Feature
+未启用时，先用 `agentscope-develop-feature prepare` 固化一个 accepted H-to-F
+request，再由 Qwen3.8-Max through Claude Code 在本地 skill 约束下实现。只有
+`FEATURE_ADMISSION_RECEIPT.json` 独立通过全部 gate，调用方才可以打开一个
+Feature 并复跑。所有归因必须基于 trajectory evidence，不能由 score 直接推断。
+
+## 安全边界
+
+- API key 只从环境变量读取，不写入源码。
+- provider key 与 base URL 按同一 namespace 绑定，避免 key 发往错误 endpoint。
+- `Data/`、`tmp/`、cache、run records 和 ablation outputs 不进入代码提交。
+- API-backed run 可能把 prompt、model output 和 response ID 写入本地记录；共享前必须再次脱敏。
+
+## 测试
+
+在本目录执行：
+
+```bash
+python -m pytest -q
+python scripts/run_feature_contracts.py --candidate agentscope --pretty
+python -m pytest candidates/agentscope/tests/test_feature_development.py -q
+python candidates/agentscope/scripts/v2_ablation_matrix.py
+```
+
+API-backed run 由 `DASHSCOPE_API_KEY`、`OPENAI_API_KEY`，或成对的 `LLM_API_KEY` + `LLM_BASE_URL` 配置。Qwen3.8-Max 的 Claude Code route 使用 `DASHSCOPE_API_KEY` 加 `DASHSCOPE_ANTHROPIC_BASE_URL`，或从 `DASHSCOPE_BASE_URL` 推导；不要把 key 写入参数、URL、receipt 或文件。
+
+## TODO
+
+当前 H-code 与 Feature mapping 是 hard-coded 的统一 taxonomy。下一步按 PawBench task category 分别设计 Feature，并用真实 trajectory 做 task-specific ablation 验证。

main_harness/candidates/agentscope/.gitignore

@@ -0,0 +1,12 @@
+.venv/
+.env
+.env.*
+!.env.example
+__pycache__/
+.pytest_cache/
+*.pyc
+tmp/
+dist/
+build/
+*.egg-info/
+.DS_Store

main_harness/candidates/agentscope/FEATURE_CHANGE_STANDARD.md

@@ -0,0 +1,62 @@
+# AgentScope Feature Change Standard
+
+This is the normative procedure for adding, strengthening, disabling, or
+removing an AgentScope Feature. A Feature is admitted only when its OFF path,
+runtime seams, evidence, and release gates are all explicit.
+
+## Boundary
+
+A Feature is a harness-owned, independently switchable intervention around one
+fixed AgentScope model and tool loop. It may change workspace policy, tool
+contract, runtime policy, observability, or context and memory at its declared
+seam. It must not change model identity or sampling, benchmark task or split,
+Harbor authority, external verifier, accepted attribution evidence, or an
+unrelated Feature.
+
+## Required contract
+
+Before editing, bind the request to:
+
+- stable Feature ID, English and Chinese names, and one H1-H5 owner;
+- causal mechanism, spatial dependencies, writes, conflicts, and OFF behavior;
+- apply, trace, cleanup, cancellation, retry, and resume phases;
+- a Feature-independent observable and Feature-owned trace event;
+- safety floors that remain when the Feature is OFF;
+- compatibility for config, state, trace, artifacts, CLI, and package;
+- a frozen paired experiment and holdout gate;
+- retirement/rollback behavior.
+
+Active display names must not use `/` to join concepts. Historical aliases may
+remain only as explicit compatibility provenance.
+
+## Admission workflow
+
+1. The attribution bridge accepts an optimization-arm H-to-F decision.
+2. `agentscope-develop-feature prepare` persists a request and a compiled
+   repository-skill receipt.
+3. `agentscope-develop-feature run --execute` invokes only Qwen3.8-Max through
+   Claude Code with strict empty MCP configuration and the local skill.
+4. The coding agent implements and tests OFF before ON, then writes the exact
+   `FEATURE_ADMISSION_RECEIPT.json`.
+5. `agentscope-develop-feature verify` independently validates the request and
+   receipt. Only then may the caller activate the nominated Feature.
+
+## Mandatory gates
+
+The receipt must contain exactly these passing gates:
+
+`boundary`, `off_equivalence`, `spatial_contract`, `temporal_contract`,
+`safety`, `causal_pair`, `holdout`, `compatibility`, and `benchmark_matrix`.
+
+For every claimed benchmark, retain a one-Feature paired row for all 15
+Features. A blocked or mechanism-not-observed row is not a causal pass, and an
+absent row blocks a support claim. Published runs and frozen inputs are never
+rewritten to make an admission pass.
+
+## Removal symmetry
+
+Removal uses the same Feature identity, seams, OFF/absence behavior, frozen
+comparison, trace, safety, holdout, compatibility, package, and matrix gates.
+Deleting a call site alone is not removal: registrations, consumers, state,
+configuration, tests, documentation, package exports, and old artifact readers
+must remain valid.

main_harness/candidates/agentscope/README.md

@@ -0,0 +1,87 @@
+# PawBench AgentScope Harness
+
+默认的 Harness backend。AgentScope 提供 agent、model adapter、tool 与 ReAct
+loop；PawBench Core 提供 Feature switch、workspace isolation、trace
+normalization 与 verification。`agentscope-develop-feature` 是受治理的开发
+桥接：它不会自行打开 Feature，也不会改写 attribution 结论。
+
+## Feature mapping
+
+| H-code | Feature |
+| --- | --- |
+| H1 Environment and Workspace | F1.1 Workspace Binding, F1.2 Readiness and Reset, F1.3 Isolation and Permissions |
+| H2 Tool Contract | F2.1 Action Contract, F2.2 Tool Availability, F2.3 Result and Error Feedback |
+| H3 Runtime and Loop | F3.1 Completion and Termination, F3.2 Budget and Guards, F3.3 Recovery and Resume |
+| H4 Observability and Acceptance | F4.1 Diagnostic Trace, F4.2 State and Artifact Deltas, F4.3 Verification |
+| H5 Context and Memory | F5.1 Context Assembly, F5.2 Persistent Memory, F5.3 Compaction |
+
+`Ex-3` 是外部 provider 或 hosted service 故障，不自动映射 F-code。只有
+Harness 对故障的处理也有直接 evidence 时，才额外归因 H/F。
+
+## 受治理的 Feature 开发
+
+默认 coding agent 是 **Qwen3.8-Max through Claude Code**。它只接收一个已
+接受的 optimization-arm H-to-F request，并使用仓库内的
+`implement-attributed-feature` skill。执行顺序固定为：
+
+```text
+attribution bridge
+  -> accepted H-to-F request
+  -> prepare skill receipt
+  -> Qwen3.8-Max through Claude Code
+  -> independent admission receipt validation
+  -> caller activates exactly one admitted Feature
+  -> paired benchmark and holdout evaluation
+```
+
+`run --execute` 才允许 Claude Code 改写本地 workspace。子进程只获得
+DashScope gateway 所需的凭据；GitHub、Alibaba 与其他主机 token 不会继承，
+且不会写入 receipt。需要配置：
+
+```bash
+export DASHSCOPE_API_KEY='...'
+export DASHSCOPE_ANTHROPIC_BASE_URL='https://.../apps/anthropic'
+# 或使用 DASHSCOPE_BASE_URL=https://.../compatible-mode/v1 自动推导
+```
+
+从 `main_harness/` 运行。`prepare`、`run` 与 `verify` 总是通过
+`--workspace-root` 读取 PawBench checkout 内的 canonical manifest 和 local
+skill；evidence 必须来自 optimization arm，任何含 `holdout` 的路径都会被拒绝。
+已安装的 command 可以在 checkout 外显示 `--help`，但不能脱离 checkout 执行开发。
+
+```bash
+python -m pip install -e candidates/agentscope
+
+agentscope-develop-feature prepare \
+  --workspace-root .. \
+  --output-dir main_harness/candidates/agentscope/tmp/feature_development/round_01 \
+  --h-code H2 --feature-id F2.2 \
+  --selection-reason 'accepted three-peer attribution evidence' \
+  --evidence main_harness/candidates/agentscope/tmp/round_01/ATTRIBUTION_SUMMARY.json
+
+agentscope-develop-feature run --workspace-root .. --execute \
+  --request main_harness/candidates/agentscope/tmp/feature_development/round_01/FEATURE_DEVELOPMENT_REQUEST.json
+
+agentscope-develop-feature verify --workspace-root .. \
+  --request main_harness/candidates/agentscope/tmp/feature_development/round_01/FEATURE_DEVELOPMENT_REQUEST.json
+```
+
+`FEATURE_CHANGE_STANDARD.md`、`feature_manifest.json` 与
+`feature_value_contract.json` 是实现和验收的共同约束。receipt 必须使
+boundary、OFF equivalence、spatial、temporal、safety、causal pair、holdout、
+compatibility 和 benchmark matrix 九个 gate 全部通过。
+
+## 本地测试
+
+在 `main_harness/` 执行：
+
+```bash
+python -m pytest candidates/agentscope/tests -q
+python candidates/agentscope/scripts/local_demo.py
+python candidates/agentscope/scripts/feature_impact_smoke.py --iterations 10
+python candidates/agentscope/scripts/v2_ablation_matrix.py
+```
+
+API-backed ablation 脚本保持原有 DashScope-compatible endpoint 配置。每次
+run 创建新的 workspace 与 trace；`--resume` 仅用于明确支持续跑的脚本。
+运行产物、request、receipt 和 token 都不得提交。