lode

Lode

Decision replay for agentic coding: capture the why, then compound it into reports, reviews, and roadmaps.

See Lode in Action · Documentation · 中文

New: See real Lode outputs — Daily Notes, Weekly Reports, Decision Replay, and Monthly Reviews from 14 months of daily use.

Why

AI makes software exploration cheap. It also makes continuity expensive.

In one coding session, you can compare multiple designs, reject plausible paths, discover hidden constraints, update prompts, change schemas, and make decisions that only make sense because of everything you tried along the way. Git records the final diff. Issue trackers record planned work. Neither keeps the reasoning that made the work possible.

When the session ends, that context usually disappears.

Lode is a local-first, dependency-light habit toolbox for agentic coding's persistent memory. Its primary value path is decision replay: capture decisions, risks, abandoned paths, open questions, and durable artifacts, then replay the reasoning later with cited evidence. Weekly outlines, monthly reviews, and roadmaps are compounding views over the same raw record.

Decision Replay Loop

Lode is packaged around one primary value path:

install -> demo -> capture one session -> query one decision

That loop proves the product: a coding agent can ask why a choice was made and get cited local evidence instead of a vague recap. recall, roadmap, daily, weekly, and monthly outputs come later as compounding views over the same raw record.

Skills

Each skill maps to a habit you already have. In plugin form, activate it with a namespaced command.

Tier 1 — Start here (immediate value, no accumulated data needed)

Skill	When	What it does
/lode:cold-start-interview	First run	Creates ~/.lode/config.yaml with vault path, project identity, language, and report preferences
/lode:capture	Wrap-up or checkpoint	Quietly captures decision/build/repair depth, stage progress, and durable artifacts when relevant
/lode:recall	Session start, after history exists	Recalls recent decisions, risks, open questions, abandoned alternatives, relevant docs, and possible stale intent artifacts

Start with the capture+recall loop for 1-2 weeks. Once raw entries accumulate, Tier 2 skills become richer.

Tier 2 — Week 1+ (better after a few sessions of raw entries)

Skill	When	What it does
/lode:daily	Daily, on demand	Updates Obsidian daily notes from raw entries and git history
/lode:weekly	Weekly, on demand	Builds a PPT preparation reference from raw entries, with a conditional hard-stuff section when evidence exists
/lode:query	Targeted follow-up	Answers "why did we choose this?" with cited decision replay evidence

Tier 3 — Month+ (compounding views, need accumulated history)

Skill	When	What it does
/lode:monthly	Monthly, on demand	Generates a monthly review and candidate rules from repeated evidence
/lode:roadmap	Phase reviews, after multiple decisions	Generates a narrative decision roadmap, including accumulating risks and recurring open questions

Skills are independent. Lode is not a strict pipeline — each skill works on its own, but they share one local storage convention so downstream reports can reuse earlier context.

Recommended Habit Loop

Day 1: Run /lode:cold-start-interview once to configure the vault.

Every session: End with /lode:capture ("收工"), checkpoint long work with /lode:capture checkpoint, and start with /lode:recall ("开工"). These form the core habit loop.

After 1-2 weeks: Use /lode:daily and /lode:weekly for periodic reports, /lode:query for targeted decision follow-up.

After 1+ months: Use /lode:monthly for monthly reviews and /lode:roadmap for decision evolution retrospectives.

Decision Replay

Lode turns "why did we choose this?" into a first-class workflow. Raw entries remain the source of truth, while {vault}/raw/decisions/{slug}.json gives coding agents a compact, cited evidence pack for targeted decision questions. See examples/decision-replay-proof.md for a real Lode-on-Lode dogfood run, including a negative query that correctly returns no answer when the decision history is missing.

First useful path

1. Install Lode.
2. Run the deterministic fixture:

node examples/decision-replay-demo.mjs