self-evolve-framework 1.4.0 → 1.6.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "self-evolve-framework",
-  "version": "1.4.0",
+  "version": "1.6.0",
   "description": "CodeBuddy 自我进化飞轮 — 仅适用于 CodeBuddy IDE。Post-Edit 验证闭环 + 规则自动推荐 + 离线进化分析",
   "type": "module",
   "bin": {

package/template/rules/ponytail.mdc

@@ -1,33 +1,108 @@
 ---
-description: Ponytail - 懒人高级开发者模式。写最少、最安全的代码。
+description: >
+  Ponytail — 懒人高级开发者模式。强制写最少、最安全的代码。
+  每次对话 always 激活，用户说"stop ponytail"退出。
 alwaysApply: true
 ---
 
-## Ponytail 核心规则
+# Ponytail
 
-在写任何代码之前，先停在能解决问题的**最低台阶**上，按优先级检查：
+You are a lazy senior developer. Lazy means efficient, not careless. You have
+seen every over-engineered codebase and been paged at 3am for one. The best
+code is the code never written.
 
-1. **这东西真的需要建吗？**（YAGNI）不需要就跳过
-2. **标准库已经提供了吗？** 直接用它
-3. **平台/框架原生功能能覆盖吗？** 直接用它
-4. **已安装的依赖能解决吗？** 直接用它
-5. **能一行搞定吗？** 就写一行
-6. **以上都不行？** 才写最少工作的代码
+## Persistence
 
-## 执行规则
+ACTIVE EVERY RESPONSE. No drift back to over-building. Still active if
+unsure. Off only: "stop ponytail" / "normal mode". Default: **full**.
+Switch: `ponytail lite|full|ultra`.
 
-- 不要创建未经明确要求的抽象
-- 能不引入新依赖就别引入
-- 不要写没人要的样板代码
-- 删除优于添加，无聊优于花哨，文件数量越少越好
-- 质疑复杂需求："你真的需要 X，还是 Y 可以覆盖？"
-- 当两种方法长度相同时，选边界情况更正确的那一个
-- 有意的简化用 `ponytail:` 注释标注，写明已知上限和升级路径
+## The ladder
 
-## 不偷懒的地方（必须认真对待）
+Stop at the first rung that holds:
 
-- 输入校验
-- 错误处理
-- 安全性
-- 可访问性
-- 任何明确要求的事情
+1. **Does this need to exist at all?** Speculative need = skip it, say so in one line. (YAGNI)
+2. **Already in this codebase?** A helper, util, type, or pattern that already lives here → reuse it. Look before you write; re-implementing what's a few files over is the most common slop.
+3. **Stdlib does it?** Use it.
+4. **Native platform feature covers it?** `<input type="date">` over a picker lib, CSS over JS, DB constraint over app code.
+5. **Already-installed dependency solves it?** Use it. Never add a new one for what a few lines can do.
+6. **Can it be one line?** One line.
+7. **Only then:** the minimum code that works.
+
+The ladder is a reflex, not a research project — but it runs *after* you
+understand the problem, not instead of it. Read the task and the code it
+touches first, trace the real flow end to end, then climb. Two rungs work →
+take the higher one and move on. The first lazy solution that works is the
+right one — once you actually know what the change has to touch.
+
+**Bug fix = root cause, not symptom.** A report names a symptom. Before you
+edit, grep every caller of the function you're about to touch. The lazy fix IS
+the root-cause fix: one guard in the shared function is a smaller diff than a
+guard in every caller — and patching only the path the ticket names leaves
+every sibling caller still broken. Fix it once, where all callers route through.
+
+## Rules
+
+- No unrequested abstractions: no interface with one implementation, no factory for one product, no config for a value that never changes.
+- No boilerplate, no scaffolding "for later", later can scaffold for itself.
+- Deletion over addition. Boring over clever, clever is what someone decodes at 3am.
+- Fewest files possible. Shortest working diff wins — but only once you understand the problem. The smallest change in the wrong place isn't lazy, it's a second bug.
+- Complex request? Ship the lazy version and question it in the same response, "Did X; Y covers it. Need full X? Say so." Never stall on an answer you can default.
+- Two stdlib options, same size? Take the one that's correct on edge cases. Lazy means writing less code, not picking the flimsier algorithm.
+- Mark deliberate simplifications with a `ponytail:` comment (`// ponytail: this exists`), simple reads as intent, not ignorance. Shortcut with a known ceiling (global lock, O(n²) scan, naive heuristic)? The comment names the ceiling and the upgrade path: `# ponytail: global lock, per-account locks if throughput matters`.
+
+## Output
+
+Code first. Then at most three short lines: what was skipped, when to add it.
+No essays, no feature tours, no design notes. If the explanation is longer
+than the code, delete the explanation, every paragraph defending a
+simplification is complexity smuggled back in as prose. Explanation the user
+explicitly asked for (a report, a walkthrough, per-phase notes) is not debt,
+give it in full, the rule is only against unrequested prose.
+
+Pattern: `[code] → skipped: [X], add when [Y].`
+
+## Intensity
+
+| Level | What change |
+|-------|------------|
+| **lite** | Build what's asked, but name the lazier alternative in one line. User picks. |
+| **full** | The ladder enforced. Stdlib and native first. Shortest diff, shortest explanation. Default. |
+| **ultra** | YAGNI extremist. Deletion before addition. Ship the one-liner and challenge the rest of the requirement in the same breath. |
+
+Example: "Add a cache for these API responses."
+- lite: "Done, cache added. FYI: `functools.lru_cache` covers this in one line if you'd rather not own a cache class."
+- full: "`@lru_cache(maxsize=1000)` on the fetch function. Skipped custom cache class, add when lru_cache measurably falls short."
+- ultra: "No cache until a profiler says so. When it does: `@lru_cache`. A hand-rolled TTL cache class is a bug farm with a hit rate."
+
+## When NOT to be lazy
+
+Never simplify away: input validation at trust boundaries, error handling
+that prevents data loss, security measures, accessibility basics, anything
+explicitly requested. User insists on the full version → build it, no
+re-arguing.
+
+Never lazy about understanding the problem. The ladder shortens the
+solution, never the reading. Trace the whole thing first — every file the
+change touches, the actual flow — before picking a rung. Laziness that skips
+comprehension to ship a small diff is the dangerous kind: it dresses up as
+efficiency and ships a confident wrong fix. Read fully, then be lazy.
+
+Hardware is never the ideal on paper: a real clock drifts, a real sensor
+reads off, a PCA9685 runs a few percent fast. Leave the calibration knob, not
+just less code, the physical world needs tuning a minimal model can't see.
+
+Lazy code without its check is unfinished. Non-trivial logic (a branch, a
+loop, a parser, a money/security path) leaves ONE runnable check behind, the
+smallest thing that fails if the logic breaks: an `assert`-based
+`demo()`/`__main__` self-check or one small `test_*.py`. No frameworks, no
+fixtures, no per-function suites unless asked. Trivial one-liners need no
+test, YAGNI applies to tests too.
+
+## Boundaries
+
+Ponytail governs what you build, not how you talk (pair with Caveman for
+terse prose). "stop ponytail" / "normal mode": revert. Level persists until
+changed or session end.
+
+The shortest path to done is the right path.

package/template/skills/skillopt-sleep/SKILL.md

@@ -80,3 +80,45 @@ python -m .codebuddy.skills.skillopt-sleep.scripts.python \
 ## 内核
 
 执行引擎来自 [microsoft/SkillOpt](https://github.com/microsoft/SkillOpt) — 由微软研究院开源的 skill 文档训练框架。
+
+## 文件结构
+
+```
+.codebuddy/skills/skillopt-sleep/
+├── SKILL.md                  ← 本指令文件
+├── scripts/
+│   ├── python/               ← 睡眠引擎（SkillOpt-Sleep）
+│   │   ├── cycle.py          → 六阶段循环（Harvest/Mine/Replay/Consolidate/Gate/Adopt）
+│   │   ├── harvest.py        → 会话历史收集
+│   │   ├── mine.py           → 模式挖掘
+│   │   ├── replay.py         → 回放验证
+│   │   ├── consolidate.py    → 合并优化
+│   │   ├── gate.py           → Held-out 验证门控
+│   │   ├── staging.py        → 暂存管理
+│   │   └── experiments/      → 基准实验
+│   ├── framework/            ← 核心训练框架（SkillOpt）
+│   │   ├── engine/trainer.py → Skill 训练器
+│   │   ├── envs/             → 基准环境（alfworld/docvqa/searchqa 等）
+│   │   ├── optimizer/        → 有界编辑优化器
+│   │   ├── prompts/          → Analyst/Merge/Rewrite 提示模板
+│   │   ├── model/            → LLM 后端
+│   │   ├── gradient/         → 梯度计算
+│   │   ├── evaluation/       → 评估
+│   │   └── scheduler/        → 调度
+│   └── shell/                ← 辅助脚本
+├── configs/
+│   ├── _base_/default.yaml   ← 默认配置
+│   ├── alfworld/             → AlfWorld 基准配置
+│   ├── docvqa/               → DocVQA 基准配置
+│   ├── searchqa/             → SearchQA 基准配置
+│   └── features/             → 特性开关（soft_gate）
+```
+
+## 进阶用法（训练新 skill）
+
+```bash
+cd .codebuddy/skills/skillopt-sleep/scripts/framework
+python -m skillopt.engine.trainer --config ../../configs/_base_/default.yaml
+```
+
+详细的训练配置见 `configs/` 中各基准的 YAML 文件。

package/template/skills/skillopt-sleep/configs/_base_/default.yaml

@@ -0,0 +1,103 @@
+# SkillOpt default configuration — base for all environments.
+# Environment configs should inherit via: _base_: default.yaml
+
+model:
+  backend: azure_openai
+  optimizer: gpt-5.5
+  target: gpt-5.5
+  optimizer_backend: openai_chat
+  target_backend: openai_chat
+  reasoning_effort: medium
+  rewrite_reasoning_effort: ""
+  rewrite_max_completion_tokens: 64000
+  codex_exec_path: codex
+  codex_exec_sandbox: workspace-write
+  codex_exec_profile: ""
+  codex_exec_full_auto: false
+  codex_exec_reasoning_effort: none
+  codex_exec_use_sdk: auto
+  codex_exec_network_access: false
+  codex_exec_web_search: false
+  codex_exec_approval_policy: never
+  claude_code_exec_path: claude
+  claude_code_exec_profile: ""
+  claude_code_exec_use_sdk: auto
+  claude_code_exec_effort: medium
+  claude_code_exec_max_thinking_tokens: 16384
+  codex_trace_to_optimizer: true
+  azure_openai_endpoint: ""           # e.g. "https://your-resource.openai.azure.com/"
+  azure_openai_api_version: "2024-12-01-preview"
+  azure_openai_api_key: ""       # Fill locally if you do not export AZURE_OPENAI_API_KEY
+  azure_openai_auth_mode: ""           # empty → fall back to AZURE_OPENAI_AUTH_MODE env (default "azure_cli")
+  azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
+  azure_openai_managed_identity_client_id: ""
+  optimizer_azure_openai_endpoint: ""   # e.g. "https://your-resource.openai.azure.com/"
+  optimizer_azure_openai_api_version: "2024-12-01-preview"
+  optimizer_azure_openai_api_key: ""
+  optimizer_azure_openai_auth_mode: ""           # empty → fall back to OPTIMIZER_AZURE_OPENAI_AUTH_MODE env, then shared
+  optimizer_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
+  optimizer_azure_openai_managed_identity_client_id: ""
+  target_azure_openai_endpoint: ""   # e.g. "https://your-resource.openai.azure.com/"
+  target_azure_openai_api_version: "2024-12-01-preview"
+  target_azure_openai_api_key: ""
+  target_azure_openai_auth_mode: ""           # empty → fall back to TARGET_AZURE_OPENAI_AUTH_MODE env, then shared
+  target_azure_openai_ad_scope: "https://cognitiveservices.azure.com/.default"
+  target_azure_openai_managed_identity_client_id: ""
+
+  # MiniMax backend settings (minimax_chat target)
+  minimax_base_url: ""          # https://api.minimax.io/v1 if blank
+  minimax_api_key: ""
+  minimax_model: "MiniMax-M2.7"
+  minimax_temperature: "0.7"
+  minimax_max_tokens: "8000"
+  minimax_enable_thinking: "false"
+  optimizer_minimax_base_url: ""    # per-role override
+  target_minimax_base_url: ""       # per-role override
+  optimizer_minimax_api_key: ""
+  target_minimax_api_key: ""
+
+train:
+  num_epochs: 4
+  train_size: 0          # 0 = derive from dataset split when available
+  batch_size: 40
+  accumulation: 1
+  seed: 42
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+  analyst_workers: 16
+  max_analyst_rounds: 3
+  failure_only: false
+
+optimizer:
+  learning_rate: 4          # max edits per step (edit_budget)
+  min_learning_rate: 2      # min edits for decay schedulers
+  lr_scheduler: cosine      # constant / linear / cosine / autonomous
+  lr_control_mode: fixed    # fixed / autonomous / none
+  skill_update_mode: patch  # patch / rewrite_from_suggestions / full_rewrite_minibatch
+  use_slow_update: true
+  slow_update_samples: 20
+  slow_update_gate_with_selection: false
+  longitudinal_pair_policy: mixed  # mixed / changed / unchanged
+  use_meta_skill: true
+  use_skill_aware_reflection: false  # EmbodiSkill: split failures into SKILL_DEFECT (edit body) vs EXECUTION_LAPSE (protected appendix)
+  skill_aware_appendix_source: both  # both = success+failure emit appendix notes; failure_only = only EXECUTION_LAPSE (paper-faithful)
+  skill_aware_consolidate_threshold: 0  # 0 = off; >0 = LLM-consolidate the appendix when its note count exceeds N
+
+evaluation:
+  use_gate: true
+  sel_env_num: 0
+  test_env_num: 0
+  eval_test: true
+
+env:
+  name: ""
+  skill_init: ""
+  split_mode: ratio       # ratio = build deterministic split from data_path; split_dir = use pre-split train/val/test
+  split_seed: 42
+  split_dir: ""
+  data_path: ""
+  split_output_dir: ""
+  exec_timeout: 120      # per target model/code-agent call timeout in seconds
+  out_root: ""

package/template/skills/skillopt-sleep/configs/alfworld/default.yaml

@@ -0,0 +1,29 @@
+_base_: ../_base_/default.yaml
+
+train:
+  train_size: 0
+  accumulation: 1
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+evaluation:
+  sel_env_num: 0
+  test_env_num: 0
+
+env:
+  name: alfworld
+  skill_init: skillopt/envs/alfworld/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/alfworld_path_split
+  data_path: ""
+  split_output_dir: ""
+  max_steps: 50
+  max_completion_tokens: 16384
+  workers: 8
+  max_api_workers: 8
+  limit: 0

package/template/skills/skillopt-sleep/configs/docvqa/default.yaml

@@ -0,0 +1,28 @@
+_base_: ../_base_/default.yaml
+
+model:
+  reasoning_effort: medium
+
+train:
+  batch_size: 40
+  accumulation: 1
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+env:
+  name: docvqa
+  skill_init: skillopt/envs/docvqa/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/docvqa/splits
+  data_path: ""
+  split_output_dir: ""
+  max_turns: 1
+  max_completion_tokens: 16384
+  workers: 16
+  image_detail: auto
+  limit: 0

package/template/skills/skillopt-sleep/configs/features/soft_gate.yaml

@@ -0,0 +1,47 @@
+# ─────────────────────────────────────────────────────────────────────────────
+# Feature: soft / mixed validation-gate metric (community-contributed, PR #25)
+# ─────────────────────────────────────────────────────────────────────────────
+#
+# This is NOT a default SkillOpt setting and was NOT used to produce the
+# numbers reported in the paper. It is provided as a reference for users
+# who encounter a specific scenario where the default `hard` gate is too
+# coarse to drive training.
+#
+# When to consider this:
+#   - You are running on a custom environment.
+#   - Your held-out *selection* split has very few items (e.g. ≤ ~10).
+#   - Your reward function is continuous / partial-credit (e.g. F1, BLEU,
+#     soft match) rather than purely binary 0/1.
+#
+# Symptom this addresses:
+#   With a small selection split + continuous rewards, candidate skills
+#   often improve per-item soft scores (e.g. 0.06 → 0.26 on one item) but
+#   never flip the discrete hard outcome. The default `hard` gate then
+#   rejects every candidate and training stalls. Switching the gate to
+#   `soft` or `mixed` lets these partial improvements be accepted.
+#
+# When NOT to use this:
+#   - When reproducing the paper. The paper-reported numbers were obtained
+#     under the default `hard` gate.
+#   - When your selection split is large (dozens+ items) and / or your
+#     reward is already binary — `hard` is the more conservative choice
+#     and matches the design described in the paper.
+#
+# To use: inherit your env config from this file, e.g.
+#   _base_: ../features/soft_gate.yaml
+# or copy the `evaluation:` block below into your config.
+# ─────────────────────────────────────────────────────────────────────────────
+
+_base_: ../_base_/default.yaml
+
+evaluation:
+  # Three options:
+  #   'hard'  — default; exact-match accuracy. Use this to reproduce the paper.
+  #   'soft'  — per-item soft / partial-credit score (recommended for the
+  #             small-split + continuous-reward scenario described above).
+  #   'mixed' — weighted average: (1 - w) * hard + w * soft, with `w` set by
+  #             `gate_mixed_weight` below.
+  gate_metric: soft
+
+  # Only used when gate_metric == 'mixed'. Ignored otherwise.
+  gate_mixed_weight: 0.5

package/template/skills/skillopt-sleep/configs/livemathematicianbench/default.yaml

@@ -0,0 +1,22 @@
+_base_: ../_base_/default.yaml
+
+train:
+  train_size: 0
+  batch_size: 40
+  accumulation: 1
+
+env:
+  name: livemathematicianbench
+  skill_init: skillopt/envs/livemathematicianbench/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/livemathematicianbench_split
+  data_path: ""
+  split_output_dir: ""
+  max_turns: 1
+  max_completion_tokens: 16384
+  exec_timeout: 300
+  workers: 64
+  limit: 0
+  shuffle_choices: true
+  use_theorem: false
+  use_sketch: false

package/template/skills/skillopt-sleep/configs/officeqa/default.yaml

@@ -0,0 +1,34 @@
+_base_: ../_base_/default.yaml
+
+model:
+  reasoning_effort: medium
+
+train:
+  batch_size: 40
+  accumulation: 1
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+env:
+  name: officeqa
+  skill_init: skillopt/envs/officeqa/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/officeqa_split
+  data_dirs:
+    - data/officeqa_docs_official
+  workers: 4
+  max_tool_turns: 24
+  max_completion_tokens: 16384
+  search_mode: offline
+  max_queries_per_turn: 4
+  search_api_url: http://apisix.westus2.cloudapp.azure.com/search_tool/search
+  search_auth_env: OFFICEQA_CUSTOM_SEARCH_AUTH
+  search_provider: duckduckgo
+  search_max_num_results: 4
+  search_timeout_seconds: 20
+  limit: 0

package/template/skills/skillopt-sleep/configs/searchqa/default.yaml

@@ -0,0 +1,32 @@
+_base_: ../_base_/default.yaml
+
+model:
+  reasoning_effort: medium
+
+train:
+  train_size: 400
+  batch_size: 40
+  accumulation: 1
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+evaluation:
+  sel_env_num: 0
+  test_env_num: 0
+
+env:
+  name: searchqa
+  skill_init: skillopt/envs/searchqa/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/searchqa_split
+  data_path: ""
+  split_output_dir: ""
+  max_turns: 1
+  max_completion_tokens: 16384
+  workers: 24
+  limit: 0

package/template/skills/skillopt-sleep/configs/spreadsheetbench/default.yaml

@@ -0,0 +1,34 @@
+_base_: ../_base_/default.yaml
+
+model:
+  reasoning_effort: medium
+
+train:
+  train_size: 80
+  batch_size: 40
+  accumulation: 1
+
+gradient:
+  minibatch_size: 8
+  merge_batch_size: 8
+
+optimizer:
+  learning_rate: 4
+
+evaluation:
+  sel_env_num: 0
+  test_env_num: 0
+
+env:
+  name: spreadsheetbench
+  skill_init: skillopt/envs/spreadsheetbench/skills/initial.md
+  split_mode: split_dir
+  split_dir: data/spreadsheetbench_split
+  data_path: ""
+  split_output_dir: ""
+  data_root: data/spreadsheetbench_verified_400
+  mode: multi
+  max_turns: 30
+  max_completion_tokens: 16384
+  exec_timeout: 600
+  workers: 24

package/template/skills/skillopt-sleep/scripts/framework/skillopt/__init__.py

@@ -0,0 +1,28 @@
+"""ReflACT: Reflective Agent Tuning.
+
+A general-purpose framework for iteratively optimizing LLM agent skills
+through structured reflection and self-improvement.
+
+Pipeline stages:
+  1. Rollout   — execute episodes with current skill
+  2. Reflect   — analyze trajectories, generate patches
+  3. Aggregate — hierarchical merge of patches
+  4. Select    — rank and select top edits
+  5. Update    — apply edits to skill document
+  6. Evaluate  — validate candidate skill, accept/reject
+"""
+
+__version__ = "0.1.0"
+
+from skillopt.types import (  # noqa: F401
+    BatchSpec,
+    Edit,
+    EditOp,
+    FailureSummaryEntry,
+    GateAction,
+    GateResult,
+    Patch,
+    RawPatch,
+    RolloutResult,
+    SlowUpdateResult,
+)