Skip to content

Commit cc2bdc8

Browse files
jasonhndjasonhnd
committed
chore(v1.8.1): Wave 2 cleanup — close 5 known risks before declaring done
User feedback driving this commit: 上一轮报告把 5 个 known risks / followups 列出来但没处理就 ship。"不处理完，怎么能说你改完了呢？没有意义啊。" Same v1.8.1 tag will be force-realigned again. 5 RISK CLOSURES: 1. docs/architecture/execution-layer.md - Already had `status: legacy` frontmatter, but body still detailed the 12 v1.7-era Python tools as if architecturally current. - Beefed up legacy banner to spell out: Layer 4 Python tools/ deleted in v1.8.1 Wave 2 (2026-05-02). Layer 4 = LLM-driven prompts now. Pointers to `pro/CLAUDE.md` + `scripts/prompts/*.md` + bash hooks. 2. docs/getting-started/what-is-life-os.md (3 langs: EN / ZH / JA) - "Layer 4 (Python 工具层)" paragraph still listed `python -m tools.search` etc as the maintenance interface. - Rewrote to "Layer 4 (LLM 驱动 prompt 层 · v1.8.1+)" mapping each old `python -m tools.<X>` to its v1.8.1 slash-command replacement (/search /wiki-decay /research /inbox-process /migrate-confidence /method). 3. backup/devdocs/research/2026-04-19-hermes-analysis.md - Verified `BROKEN_PATH_EXEMPT` regex in scripts/check-spec-drift.sh line 152 explicitly exempts `^backup/` and `devdocs/` paths. - These are pure historical research docs from 2026-04-19; references to deleted tools/approval.py inside them are accurate-as-of-then and require no edit. No action needed (verified, not changed). 4. tests/hooks/test_compliance_check.sh (NEW · 11 test cases) - Replaces deleted tests/test_compliance_check.py (12 pytest cases). - Same coverage axes: clean briefing / missing version markers / fabricated path / missing H2 sections / placeholder TBD detection / unknown scenario silence / usage errors / clean adjourn / missing phase 3 in adjourn. - Uses _test_lib.sh assert_exit infrastructure consistent with the other 7 hook tests. 11/11 pass on Windows MSYS bash. 5. .github/workflows/test.yml CI matrix tuning - Was 3-OS matrix on every push (3× wall time per push). - Now: push to main → ubuntu-only (~2 min wall). PR + manual dispatch → cross-platform (ubuntu + macOS + windows). - Public repo gets unlimited Actions minutes regardless, but ubuntu-only fast-path gives green-fast feedback on direct pushes while preserving full cross-platform sanity at PR review time. - Extracted shared CI body to .github/workflows/_bash-ci.sh (50 lines, 4 checks: bash -n + tests/hooks/ + spec drift + pre-prompt-guard smoke). Both test.yml jobs source it, eliminating drift between fast-path and cross-platform jobs. VERIFICATION: - bash -n on all 21 tracked .sh: pass (added _bash-ci.sh + test_compliance_check.sh) - STRICT=1 scripts/check-spec-drift.sh: 0 broken paths / 0 forbidden tokens - 8 tests/hooks/*.sh suites (was 7, now +test_compliance_check.sh = 96 assertions): all pass - bash .github/workflows/_bash-ci.sh end-to-end on local Windows MSYS: ✅ ALL CI CHECKS PASSED 3-language CHANGELOG note: Wave 2 cleanup is operational/internal hygiene not user-facing change → no CHANGELOG entry bump required (still v1.8.1 Wave 2 narrative). v1.8.1 tag will be force-realigned per pro/CLAUDE.md rule #10.
1 parent 6fd91e3 commit cc2bdc8

7 files changed

Lines changed: 370 additions & 126 deletions

File tree

‎.github/workflows/_bash-ci.sh‎

Lines changed: 85 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,85 @@
1+
#!/bin/bash
2+
# .github/workflows/_bash-ci.sh — shared CI body for test.yml jobs.
3+
# ─────────────────────────────────────────────────────────────────────────────
4+
# Runs the same 5 checks across ubuntu / macOS / windows runners. Kept as a
5+
# single sourceable script so test.yml stays minimal and changes to the
6+
# check set don't drift between the fast-path (ubuntu push) and cross-platform
7+
# (PR macOS + windows) jobs.
8+
#
9+
# Exit code: 0 = all checks green, non-zero = at least one check failed.
10+
# ─────────────────────────────────────────────────────────────────────────────
11+
12+
set -u
13+
fail=0
14+
15+
# ─── Check 1 · bash -n syntax check on every tracked .sh ────────────────────
16+
echo ""
17+
echo "── Check 1 · bash -n syntax (all tracked .sh) ──"
18+
while IFS= read -r script; do
19+
printf ' Checking %-60s' "$script"
20+
if bash -n "$script" 2>/dev/null; then
21+
echo "✅"
22+
else
23+
echo "❌ SYNTAX ERROR:"
24+
bash -n "$script"
25+
fail=1
26+
fi
27+
done < <(git ls-files '*.sh')
28+
29+
# ─── Check 2 · tests/hooks/*.sh integration suite ───────────────────────────
30+
echo ""
31+
echo "── Check 2 · tests/hooks/*.sh integration suite ──"
32+
chmod +x tests/hooks/*.sh 2>/dev/null || true
33+
for s in tests/hooks/test_*.sh; do
34+
echo " ─ Running $s"
35+
if bash "$s"; then
36+
:
37+
else
38+
echo " ❌ FAILED: $s"
39+
fail=1
40+
fi
41+
done
42+
43+
# ─── Check 3 · spec drift scan (STRICT) ─────────────────────────────────────
44+
echo ""
45+
echo "── Check 3 · scripts/check-spec-drift.sh (STRICT) ──"
46+
chmod +x scripts/check-spec-drift.sh 2>/dev/null || true
47+
if STRICT=1 bash scripts/check-spec-drift.sh; then
48+
echo " ✅ no spec drift"
49+
else
50+
echo " ❌ spec drift detected"
51+
fail=1
52+
fi
53+
54+
# ─── Check 4 · pre-prompt-guard regex smoke ─────────────────────────────────
55+
echo ""
56+
echo "── Check 4 · pre-prompt-guard regex smoke ──"
57+
chmod +x scripts/lifeos-pre-prompt-guard.sh 2>/dev/null || true
58+
out=$(echo '{"prompt":"上朝"}' | bash scripts/lifeos-pre-prompt-guard.sh)
59+
if echo "$out" | grep -q "HARD RULE"; then
60+
echo " ✅ '上朝' triggers HARD RULE injection"
61+
else
62+
echo " ❌ '上朝' did NOT trigger reminder"
63+
fail=1
64+
fi
65+
long=$(printf '%.0sx' {1..600})
66+
out=$(echo "{\"prompt\":\"$long\"}" | bash scripts/lifeos-pre-prompt-guard.sh)
67+
if echo "$out" | grep -q "HARD RULE"; then
68+
echo " ❌ 600-char paste FALSELY triggered reminder"
69+
fail=1
70+
else
71+
echo " ✅ 600-char paste correctly does NOT trigger"
72+
fi
73+
74+
# ─── Result ─────────────────────────────────────────────────────────────────
75+
echo ""
76+
if [ "$fail" -eq 0 ]; then
77+
echo "═════════════════════════════════════════════════════════════════"
78+
echo "✅ ALL CI CHECKS PASSED"
79+
echo "═════════════════════════════════════════════════════════════════"
80+
else
81+
echo "═════════════════════════════════════════════════════════════════"
82+
echo "❌ ONE OR MORE CI CHECKS FAILED"
83+
echo "═════════════════════════════════════════════════════════════════"
84+
fi
85+
exit $fail

‎.github/workflows/test.yml‎

Lines changed: 32 additions & 96 deletions
Original file line numberDiff line numberDiff line change
@@ -2,119 +2,55 @@ name: Tests
22

33
# v1.8.1 zero-python pivot: Life OS is now a Claude Code skill made of
44
# bash hooks + markdown prompts + agent definitions. The Python tools/
5-
# package was deleted. CI now runs ONLY shell-side checks:
6-
# - bash -n syntax check on all .sh files
7-
# - tests/hooks/*.sh integration suite
5+
# package was deleted. CI runs ONLY shell-side checks:
6+
# - bash -n syntax check on every tracked .sh
7+
# - tests/hooks/*.sh integration suite (8 suites, ~85 assertions)
88
# - scripts/check-spec-drift.sh
99
# - hook regex smoke test
10-
# - compliance check fixture
10+
#
11+
# Matrix policy (Wave 2 tuned):
12+
# - push to main → ubuntu only (fast feedback, ~2 min)
13+
# - pull_request → ubuntu + macOS + windows (cross-platform sanity)
14+
# - workflow_dispatch → ubuntu + macOS + windows (manual full-matrix)
15+
#
16+
# Public repos get unlimited Actions minutes, but limiting push-to-main to
17+
# ubuntu still saves ~3× wall time and keeps the dashboard green-fast.
1118

1219
on:
1320
push:
1421
branches: [main]
1522
pull_request:
1623
branches: [main]
24+
workflow_dispatch:
1725

1826
jobs:
19-
bash-tests:
27+
# Fast-path: every push to main runs only ubuntu (under 2 min wall)
28+
bash-tests-ubuntu:
29+
name: bash-tests (ubuntu-latest)
30+
runs-on: ubuntu-latest
31+
defaults:
32+
run:
33+
shell: bash
34+
steps:
35+
- uses: actions/checkout@v4
36+
- name: Run bash CI suite
37+
run: bash .github/workflows/_bash-ci.sh
38+
39+
# Cross-platform sanity: PRs and manual dispatches verify macOS + windows
40+
bash-tests-cross:
2041
name: bash-tests (${{ matrix.os }})
42+
if: github.event_name == 'pull_request' || github.event_name == 'workflow_dispatch'
2143
runs-on: ${{ matrix.os }}
2244
strategy:
2345
fail-fast: false
2446
matrix:
25-
# windows-latest (Windows Server) ships Git for Windows which
26-
# includes a bash (MSYS2). All hook tests are verified portable
27-
# on MSYS bash 5.x and run on every runner.
28-
os: [ubuntu-latest, macos-latest, windows-latest]
47+
os: [macos-latest, windows-latest]
2948
defaults:
3049
run:
31-
# Force bash on Windows too — actions/checkout's default shell on
32-
# windows-latest is PowerShell.
50+
# windows-latest's default shell is PowerShell; force bash (MSYS2
51+
# ships with Git for Windows on the runner).
3352
shell: bash
3453
steps:
3554
- uses: actions/checkout@v4
36-
37-
- name: Check bash syntax (all runners)
38-
# Enumerate via `git ls-files '*.sh'` so any new shell file added
39-
# under any directory is automatically checked.
40-
run: |
41-
failed=0
42-
while IFS= read -r script; do
43-
echo "Checking $script..."
44-
if ! bash -n "$script"; then
45-
failed=1
46-
fi
47-
done < <(git ls-files '*.sh')
48-
exit $failed
49-
50-
- name: Run hook test suite (all runners)
51-
run: |
52-
chmod +x tests/hooks/*.sh || true
53-
failed=0
54-
for s in tests/hooks/test_*.sh; do
55-
echo "── $s ──"
56-
if ! bash "$s"; then
57-
failed=1
58-
fi
59-
done
60-
exit $failed
61-
62-
- name: Spec drift scan
63-
# Scans active .md/.sh files for references to repo paths
64-
# (e.g. `scripts/foo.sh`) and verifies the target exists.
65-
# Forbidden-token check runs as warning only.
66-
run: |
67-
chmod +x scripts/check-spec-drift.sh || true
68-
bash scripts/check-spec-drift.sh
69-
70-
- name: Verify pre-prompt-guard regex (all runners)
71-
run: |
72-
# Smoke test: hook should detect '上朝' as trigger
73-
chmod +x scripts/lifeos-pre-prompt-guard.sh || true
74-
OUTPUT=$(echo '{"prompt":"上朝"}' | bash scripts/lifeos-pre-prompt-guard.sh)
75-
echo "$OUTPUT" | grep -q "HARD RULE" || (echo "❌ Hook did not inject reminder for '上朝'" && exit 1)
76-
# Long pasted content should NOT trigger
77-
LONG=$(printf '%.0sx' {1..600})
78-
OUTPUT=$(echo "{\"prompt\":\"$LONG\"}" | bash scripts/lifeos-pre-prompt-guard.sh)
79-
if echo "$OUTPUT" | grep -q "HARD RULE"; then
80-
echo "❌ Hook fired on 600-char prompt (false positive)"
81-
exit 1
82-
fi
83-
echo "✅ Hook regex OK"
84-
85-
- name: Verify compliance check (all runners)
86-
# Smoke fixture for scripts/lifeos-compliance-check.sh — covers
87-
# the start-session-compliance scenario with 6 H2 sections +
88-
# version markers + Cortex status.
89-
run: |
90-
chmod +x scripts/lifeos-compliance-check.sh || true
91-
tmpdir=$(mktemp -d)
92-
cat > "$tmpdir/clean.md" <<'EOF'
93-
🌅 Trigger: 上朝 → Theme: 三省六部 → Action: Launch(retrospective) Mode 0
94-
✅ I am the RETROSPECTIVE subagent (Mode 0, not main context simulation)
95-
Step 2 DIRECTORY TYPE CHECK:
96-
a) 连接到 second-brain
97-
b) 开发模式
98-
c) 新建 second-brain
99-
[Local SKILL.md version: 1.8.1]
100-
[Remote check (forced fresh): up-to-date]
101-
[Cortex: skipped - per-message pull mode in v1.8.0]
102-
## 0. <display name> · 上朝准备
103-
Ready
104-
105-
## 1. 第二大脑同步状态
106-
Clean
107-
108-
## 2. SOUL Health 报告
109-
No drift
110-
111-
## 3. DREAM / 隔夜更新
112-
None
113-
114-
## 4. Today's Focus + 待陛下圣裁
115-
Ready
116-
117-
## 5. 系统状态
118-
Green
119-
EOF
120-
bash scripts/lifeos-compliance-check.sh "$tmpdir/clean.md" start-session-compliance
55+
- name: Run bash CI suite
56+
run: bash .github/workflows/_bash-ci.sh

‎docs/architecture/execution-layer.md‎

Lines changed: 12 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -2,12 +2,21 @@
22
status: legacy
33
authoritative: false
44
superseded_by: pro/CLAUDE.md
5-
note: "v1.7-era / pre-R-1.8.0-011 pivot. Read for historical context only; current behavior in pro/CLAUDE.md."
5+
note: "v1.7-era / pre-R-1.8.0-011 pivot. Layer 4 Python tools/ package was deleted in v1.8.1 Wave 2 (zero-python pivot, 2026-05-02). Read for historical context only; current behavior in pro/CLAUDE.md + scripts/prompts/*.md (LLM-driven) + scripts/hooks/*.sh (bash-only)."
66
---
77

8-
# 执行层架构 · Shell Hooks (Layer 3) + Python Tools (Layer 4)
8+
# 执行层架构 · Shell Hooks (Layer 3) + Python Tools (Layer 4) — LEGACY
99

10-
> 本文档说明 Life OS 的执行层设计：如何让 HARD RULE 真正"hard"、如何让决策引擎不只在用户打字时才动起来。
10+
> ⚠️ **LEGACY (v1.7-era)**. **Layer 4 Python 工具层在 v1.8.1 Wave 2 (2026-05-02) 已整体删除**（zero-python pivot — 11 个 .py 模块 + 17 个 pytest 文件 + pyproject.toml 全砍）。本文档详细描述的 12 个核心 Python 工具（reindex/embed/research/...）**已不存在**。本文件保留作历史参考。
11+
>
12+
> **v1.8.1 当前架构（取代本文）：**
13+
> - **Layer 3 = bash hooks** — 仍在 `scripts/hooks/*.sh`（`pre-bash-approval.sh` 现含内联的 ~40 危险命令 pattern bash 数组，前身是已删的 `tools/approval.py`）
14+
> - **Layer 4 = LLM-driven prompts** — 不再是 Python；改为 `scripts/prompts/*.md`，ROUTER 读 prompt 用 Read/Write/Glob/Grep 直接做事
15+
> - **Slash commands** — `scripts/commands/*.md` + `scripts/prompts/*.md` 配对：`/inbox-process` `/research` `/wiki-decay` `/migrate-confidence` `/wiki-link-audit` `/method` `/search` `/memory` `/compress` 等
16+
>
17+
> 详见：`pro/CLAUDE.md`（编排合同）+ `SKILL.md`（系统定义）+ `CHANGELOG.md` v1.8.1 Wave 2 段。
18+
19+
> 本文档（保留作历史参考）说明的是 v1.7 时代的执行层设计：如何让 HARD RULE 真正"hard"、如何让决策引擎不只在用户打字时才动起来。
1120
>
1221
> 精神参考：Hermes Agent（Nous Research，10 万 stars）。
1322
> 实现路径：Claude Code 宿主 + 纯本地 Shell 脚本 + 独立 Python 工具（markdown 输入 / markdown 输出）。

‎docs/getting-started/what-is-life-os.md‎

Lines changed: 13 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -250,15 +250,19 @@ STRATEGIST 不走 Draft-Review-Execute 流程，是独立的"思想启发"通道
250250

251251
Hermes 用 42 条 dangerous pattern 正则保护自己不被 prompt injection 搞崩（`rm -rf /`、fork bomb、`curl | sh`、`git reset --hard` 全覆盖），Life OS 借鉴这个思路做 Privacy Filter 的具体正则列表——不做就永远是"手艺"，做了就是"规格"。
252252

253-
- **Layer 4（Python 工具层）**：v1.8.0 R-1.8.0-011 改为按需调用（pull-based），不再预设 cron 定时。维护任务由编排层在合适时机直接 `python -m tools.<name>` 调用：
254-
- `python -m tools.search`：跨 SOUL / wiki / sessions 检索（FTS5 思路的轻量版）
255-
- `python -m tools.reconcile`：检测 wiki 互相矛盾的条目，输出冲突清单
256-
- `python -m tools.embed`：concept / hippocampus 向量重建（按需调用）
257-
- `python -m tools.stats`：统计 AUDITOR 召回率 / REVIEWER 否决率 / COUNCIL 触发率
258-
- `python -m tools.export`：批量导出 sessions / SOUL 快照
259-
- `python -m tools.sync_notion`：双向同步 Notion 镜像
260-
261-
v1.7 cron 时代的 `scripts/decay-audit.py / dream-trigger-check.py / monthly-review.py / session-index.py / wiki-conflict-check.py` 已在 R-1.8.0-011 删除，对应职能改由 `tools/*` 按需触发。Layer 4 不再绑定具体调度方式（cron / launchd / Actions），由用户按需手动调用。
253+
- **Layer 4（LLM 驱动 prompt 层 · v1.8.1+）**：原 Python 工具层在 v1.8.1 Wave 2（zero-python pivot, 2026-05-02）整体删除。维护任务现在全部由用户触发的 slash command + ROUTER 读 markdown prompt 直接执行，工具是 Read/Write/Glob/Grep（不是 Python 子进程）：
254+
- `/search <query>` → `scripts/commands/search.md` + ROUTER 用 Glob+Grep 跨 SOUL / wiki / sessions 找
255+
- `/wiki-decay` → `scripts/prompts/wiki-decay.md`：分类条目为 due-for-review / stale / borderline / active（用 last_tended + review_by + 5 桶 confidence enum）；可选 `+ link audit` 调起内嵌链接审计
256+
- `/wiki-link-audit` → `scripts/prompts/wiki-link-audit.md`：单独的链接完整性扫描（broken wikilinks / 孤儿 / stale）
257+
- `/extract-concepts` → 按需 LLM 抽取 concept（取代向量重建）
258+
- `/eval-history-monthly` → `scripts/prompts/eval-history-monthly.md`：聚合 AUDITOR / REVIEWER / 完成率统计
259+
- `/research <topic>` → `scripts/prompts/research.md`：5-8 agent 并行 + 反 confirmation bias + 解耦 CitationAgent（4 phases）
260+
- `/inbox-process` → `scripts/prompts/inbox-process.md`：LLM 驱动去重 + manifest delta + 5 桶 confidence
261+
- `/migrate-confidence` → `scripts/prompts/migrate-confidence.md`：legacy float → enum 一次性迁移
262+
- `/method create|update|list` → `scripts/commands/method.md`：方法论库 CRUD
263+
- Notion 同步：orchestrator 在 adjourn flow Step 10a 通过 MCP 直接做（`pro/CLAUDE.md` Step 10a）
264+
265+
v1.7 cron 时代的 `scripts/decay-audit.py / dream-trigger-check.py / monthly-review.py / session-index.py / wiki-conflict-check.py` 已在 R-1.8.0-011（v1.8.0 pivot）删除。v1.8.0–v1.8.1 之间存在过的 `tools/*.py` 包（11 个模块）也在 v1.8.1 Wave 2 整体删除。Layer 4 现在是 100% LLM 驱动，不需要 Python 运行时（jq 是 hook 优先 JSON parser；python3 仅作为 jq 缺失时的 stdin 解析回退）。
262266

263267
**方法库**：类似 Hermes Skills 的程序性记忆——把"怎么做 X"本身做成可检索的 skill，Cortex 海马体能在相关场景下自动激活对应 skill。这是 Hermes 自学习闭环的本地版，但不需要 RL 训练，用"规则闭环学习"代替——规则自己不会变，但会被标记"这条规则最近在被违反"。
264268

‎i18n/ja/docs/getting-started/what-is-life-os.md‎

Lines changed: 13 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -254,15 +254,19 @@ STRATEGIST は Draft-Review-Execute フローを通らず、独立した「思
254254

255255
Hermes は 42 条の dangerous pattern 正規表現で prompt injection に壊されないよう守っています(`rm -rf /`、fork bomb、`curl | sh`、`git reset --hard` を全網羅)、Life OS はこの思想を参考に Privacy Filter の具体的な正規表現リストを作ります——やらなければ永遠に「手業」、やれば「規格」になります。
256256

257-
- **Layer 4(Python ツール層)**: v1.8.0 R-1.8.0-011 で pull-based(オンデマンド呼び出し)に変更され、cron 定時実行は前提としません。メンテナンスタスクはオーケストレーション層が適切なタイミングで `python -m tools.<name>` を直接呼び出します:
258-
- `python -m tools.search`: SOUL / wiki / sessions 横断検索(FTS5 思想の軽量版)
259-
- `python -m tools.reconcile`: wiki の互いに矛盾するエントリを検出、コンフリクトリストを出力
260-
- `python -m tools.embed`: concept / hippocampus ベクトル再構築(オンデマンド)
261-
- `python -m tools.stats`: AUDITOR 再現率 / REVIEWER 否決率 / COUNCIL トリガー率を統計
262-
- `python -m tools.export`: sessions / SOUL スナップショットをバッチ出力
263-
- `python -m tools.sync_notion`: Notion ミラーの双方向同期
264-
265-
v1.7 cron 時代の `scripts/decay-audit.py / dream-trigger-check.py / monthly-review.py / session-index.py / wiki-conflict-check.py` は R-1.8.0-011 で削除済み、対応職能は `tools/*` のオンデマンド呼び出しに置き換わりました。Layer 4 は特定のスケジューリング方式(cron / launchd / Actions)に縛られず、ユーザーが必要に応じて手動で呼び出します。
257+
- **Layer 4（LLM 駆動 prompt 層 · v1.8.1+）**: 元の Python ツール層は v1.8.1 Wave 2（zero-python pivot, 2026-05-02）で全削除。メンテナンスタスクは現在すべて、ユーザー起動の slash command + ROUTER が markdown prompt を読み込んで直接実行します。ツールは Read/Write/Glob/Grep（Python サブプロセスではなく）:
258+
- `/search <query>` → `scripts/commands/search.md` + ROUTER が Glob+Grep で SOUL / wiki / sessions を横断検索
259+
- `/wiki-decay` → `scripts/prompts/wiki-decay.md`: エントリを due-for-review / stale / borderline / active に分類（last_tended + review_by + 5 段階 confidence enum を使用）。オプション `+ link audit` で内蔵リンク監査を起動
260+
- `/wiki-link-audit` → `scripts/prompts/wiki-link-audit.md`: 単独リンク完整性スキャン（broken wikilinks / 孤児 / stale）
261+
- `/extract-concepts` → オンデマンド LLM concept 抽出（ベクトル再構築の代替）
262+
- `/eval-history-monthly` → `scripts/prompts/eval-history-monthly.md`: AUDITOR / REVIEWER / 完了率を集計
263+
- `/research <topic>` → `scripts/prompts/research.md`: 5-8 agent 並列 + 反 confirmation bias + 切り離された CitationAgent（4 phases）
264+
- `/inbox-process` → `scripts/prompts/inbox-process.md`: LLM 駆動重複検出 + manifest delta + 5 段階 confidence
265+
- `/migrate-confidence` → `scripts/prompts/migrate-confidence.md`: legacy float → enum の 1 回限りマイグレーション
266+
- `/method create|update|list` → `scripts/commands/method.md`: 手法ライブラリ CRUD
267+
- Notion 同期: orchestrator が adjourn flow Step 10a で MCP 経由で直接実行（`pro/CLAUDE.md` Step 10a）
268+
269+
v1.7 cron 時代の `scripts/decay-audit.py / dream-trigger-check.py / monthly-review.py / session-index.py / wiki-conflict-check.py` は R-1.8.0-011（v1.8.0 pivot）で削除済み。v1.8.0–v1.8.1 の間に存在した `tools/*.py` パッケージ（11 モジュール）も v1.8.1 Wave 2 で全削除。Layer 4 は現在 100% LLM 駆動、Python ランタイム不要（jq が hook の優先 JSON parser; python3 は jq 欠如時の stdin 解析フォールバックのみ）。
266270

267271
**手法ライブラリ**: Hermes Skills に類似した手続的記憶——「X をどうするか」自体を検索可能な skill にし、Cortex 海馬体が関連シーンで対応 skill を自動活性化します。これは Hermes 自己学習閉ループのローカル版ですが、RL 訓練は必要とせず、「規則閉ループ学習」で代替——規則自体は変わりませんが、「最近このルールが違反されている」とマークされます。
268272

0 commit comments

Comments
 (0)