See Lode in Action
14MONTHS OF DAILY USE
8SKILLS
0MOCKS
Every example below is real Lode output, lightly sanitized. No demos, no synthetic data.
01 / Daily Note
Every day, automatically structured.
What to notice
Four-level hierarchy: project → module → change type → specific task. Generated from raw entries + git history. No manual writing.
markdown
### 2026.05.05
- [Content Pipeline]
（API 层）
【结构变更】
- [x] （routes → errors/helpers）集中 API 错误码至 AppError 枚举，抽取共享路由辅助函数，
统一三模式（Direct/Callback/Interactive）错误响应结构（+240/-231 行）
- 修复 2 个 P0 bug：exception handler 访问 task_id 崩溃、interactive fireCallback 参数错位
- 引入模型继承，消除 contracts 中的重复定义
【能力升级】
- [x] （API → Direct）Direct 模式新增 callback_url 支持，Pipeline 完成后通过
BackgroundTasks 触发回调通知（+17/-4 行）
- [x] （API → Testing）新增 mock callback server 工具，支持本地验证完整 webhook 流程（+58 行）
### 2026.05.06
- [Content Pipeline]
（Orchestrator → Core → Schemas → API）
【结构变更】
- [x] pipeline_runner 大规模拆分，按职责分离为 4 个子模块，新增 RegistrySnapshot 类型约束，
TTL 缓存替代 lru_cache，健康检查增加 LLM provider 可达性探测（+1348/-1040 行）
- pipeline_runner 从 1829→802 行，单文件膨胀问题解决
- 原子 JSON 写入防止 registry 数据损坏
（Schemas）
【问题定位】
- [x] LLM 输出模型 extra 策略从 forbid 放宽至 ignore，防止 LLM 输出字段错位导致 stage 崩溃
（+10/-10 行）
- 修复 extract_characters 因 variant_label 层级错误而 crash 的生产问题
### 2026.05.07
- [Content Pipeline]
（Props → Characters → Registry → Output）
【能力升级】
- [x] 道具提升为一等实体，新增独立 ItemProfile 模型（item_id/scene_prompt/holders/state_variants），
在 extract_characters 中联合提取，registry v2→v3 迁移并兼容旧数据（+359/-21 行）
- 联合提取方案避免新增 pipeline stage，单次 LLM 调用零额外开销02 / Weekly Report
PPT-ready weekly outline, zero manual writing.
What to notice
Slide-structured reports with problem → solution narrative. Each row in the risk table traces back to an original raw entry.
markdown
# 2026-W21 Content Pipeline Weekly Report
> Signal coverage: High — 11 raw entries, 3 git-only gaps (minor config)
## Slide 2: Overview
- **English Script Pipeline Compatibility:** Fixed character name localization, binding
validation failures, on-screen text translation, and rolling mode inconsistencies triggered
by English-language scripts across the full pipeline
- **Content Fidelity & Efficiency:** Enhanced text/numeric retention rules based on
cross-script analysis, optimized batch prompt token waste, fixed over-aggressive dialogue rules
- **Location Name Three-Layer Defense:** Built a cascade from prompt constraints to
code-level enforcement across parse → analyze_scenes
## Slide 3: Problem — English Script Pipeline Breaks
| Problem | Break point | Impact |
| Character names localized | extract_characters prompt requires Chinese | Agent K → 特工 K, binding set breaks |
| Canonical name truncation | LLM returns "Guardians" not "Guardians (Squad)" | Strict set containment check fails |
| On-screen text translated | "原文" (original) is ambiguous | English panel text → Chinese |
| Rolling mode inconsistency | Parse LLM translates names unstably | Same character gets different names across episodes |
| Dialogue prefix won't strip | canonical ≠ raw + rf-string bug | 22% of dialogue entries keep full prefix |
## Slide 4: Technical Solution
Phase 1 (May 18) — Character name & binding chain fix:
parse (characters: Agent K, keep original)
│
▼
extract_characters (prompt: don't translate English names)
│ name="Agent K" (not "特工 K")
▼
generate_content binding
│ LLM returns: "Guardians"
│ CharacterNameMatcher: Guardians → Guardians (Squad) ✓
▼
content: on-screen text verbatim from source_text
"SYSTEM ALERT" → "SYSTEM ALERT" (no translation)
Phase 2 (May 21) — rolling mode & dialogue prefix:
parse.py removed:
_batch_translate_character_names (~100 lines)
→ character name determinism: same input = same name = forever
local_flow_builder.py:
_strip_dialogue_prefix(speaker, speaker_raw)
→ dual canonical + raw matching pattern
→ "Dr. Chen Wei" + "Chen Wei (os):" both match03 / Decision Replay
"Why did we choose this?" — with cited evidence.
What to notice
Not a summary — an evidence pack. Reasoning, rejected alternatives, open questions, and impact. Every captured decision becomes queryable.
json
{
"id": "lode:2026-W19:002",
"confidence": "explicit",
"week": "2026-W19",
"timestamp": "2026-05-05T11:30:00+08:00",
"decision": "Added zero-config mode to session-recap with decision-recording schema enhancement across all skills",
"why": "Adoption barrier caused by requiring configuration before first value delivery. Schema lacked decision context needed for project self-awareness.",
"chosen": "Zero-config for session-recap only → single skill is the natural first-touch hook, other skills gain data once vault is configured.",
"rejected": [
{
"option": "Terminal-only output (no formatting)",
"reason": "would not demonstrate the full richness of Lode entries"
},
{
"option": "Zero-config for all 5 skills",
"reason": "weekly/monthly skills need accumulated data to be useful"
}
],
"impact": "First-use experience transforms from configure-first to value-first. Raw entries now carry exploration signals that enable roadmap visibility across weekly and monthly reviews.",
"open_questions": [
"Will the zero-config Markdown output consistently follow the template across different AI agents?",
"Should other skills eventually gain lightweight zero-config modes?"
],
"evidence_refs": ["abc1234"]
}04 / Monthly Review
One month, one page — with signal extraction.
What to notice
20 work days, 126 completed tasks — compounded from daily notes into statistics, work distribution, and key change tracking. Evidence mode: strict.
markdown
# 2026-04 Work Summary
---
> **20** work days | **126** completed tasks | **0** in progress
---
## Overview
Fast-paced month, advancing two directions: content generation and content automation.
First half focused on architecture simplification and data model unification, clearing
structural barriers for cross-episode execution. Second half entered full content
automation: three leaps from monolithic pipeline to cross-episode rolling execution, from
batch mode to per-character concurrency, from single-thread to three-mode API service.
## Work Distribution
| Category | Count | Share |
| Capability | 17 | 30% |
| Structural | 15 | 26% |
| Bug fixes | 14 | 25% |
| Configuration | 5 | 9% |
| Documentation | 2 | 4% |
| Other | 4 | 7% |
## Key Changes (net >300 lines)
| Change | Net lines | Note |
| Pipeline v2.7 cross-episode | +2792/-430 | Largest change this month |
| Content binding system upgrade | +1716/-128 | Layered binding + grouping |
| Per-character extraction | +1376/-35 | Fixed batch 12/21 miss |
| Timeline scene-first refactor | +654/-3185 | Net -2531 lines (simplified) |
| Scene analysis fact constraints | +634/-253 | Strict evidence mode |
| Rolling pipeline 3-layer refactor | +610/-243 | 3-layer architecture |
> Based on Daily Note.md 2026-04 raw records and skeleton.json statistics.
> Evidence mode: strict. Each summary traceable to original daily entry.How It Works
Six skills, one habit loop.
| Skill | When | What it does |
|---|---|---|
/lode:capture | Every session wrap-up ("收工", "done") | Classifies archetype, captures decision/repair depth, indexes artifacts |
/lode:recall | Session start ("开工") | Recalls decisions, risks, open questions, relevant docs |
/lode:query | Targeted follow-up | "Why did we choose this?" with cited evidence |
/lode:daily | Daily, on demand | Structured daily notes from raw entries + git history |
/lode:weekly | Weekly, on demand | PPT-ready weekly outlines with technical narrative |
/lode:monthly | Monthly, on demand | Monthly reviews with statistics and candidate rules |
All examples above are real Lode outputs from daily use, 2025-03 to 2026-05. Internal URLs and proprietary details have been removed.