Context Ledger

Approach

One index.html with inline CSS and vanilla JS, no framework or bundler, because the whole thing is a form, a list, and a small rule-evaluation function — a build step would add nothing but risk. Ledger entries (decisions and rules) are plain JS objects persisted to localStorage under one key; on load the page reads that key and falls back to a bundled seed array on first visit or after a 'reset to seed' click, so the demonstration data and the user's own edits share one code path. Each rule carries an optional machine-checkable enforcement: {type: 'forbidden_pattern'|'required_pattern', pattern, flags, scope} where scope is a simple glob (e.g. 'src/**') converted to a regex to match file paths inside a unified-diff-style text blob. The pre-commit gate is a pure function: given a commit's diff text and the current in-memory rule list, it extracts changed file paths, applies each enforced rule's regex against the relevant file's hunk, and returns a pass/fail verdict per rule plus an overall blocked/clean result — so adding, editing, or deleting a rule changes future gate runs immediately, with no hardcoded outcomes tied to specific commit IDs. A handful of preset sample commits (some clean, some violating one or more rules, one boundary case) ship alongside a free-text textarea so a visitor can also paste their own diff-shaped text and run the same real evaluation against it.

The source post

https://x.com/DanKornas/status/2113770372797440121

Scoring

Pick

Krusch Context MCP is a local SQLite MCP server that gives coding agents memory of project decisions and rules across sessions, with pre-commit checks. The MCP server itself isn't a web artifact, so this reshapes the idea around its actual mechanism rather than its transport: a page where you record decisions/rules for a fake 'project', they persist in the browser's own storage across a reload, and a pre-commit action runs those rules for real and blocks on a violation. Nothing here is mocked or scripted to look live — the persistence and the rule check are both real client-side computation, so it clears the honesty bar that sank driftwatch's live-answer fakery. It needs no API key, no account, and no build-time network call, so self-containedness is a clean 5. It beats the other two reshape-able candidates (the Adyen registry and the agent-patterns tutorial) on surprise, since 'memory that outlives the session and a gate that actually enforces something' reads as a sharper, less-seen demo than a directory browser or a patterns explainer. Everything else on the list was rejected outright rather than reshaped.

source: https://x.com/DanKornas/status/2113770372797440121

Review

mean 4.50/5

No change: the gate passed on attempt 1, the artifact's core claim checked out against the actual source rather than NOTES.md's description of it, and nothing in NOTES.md or the last 7 nights (all shipped, no repeated build-stage failures) points to a repeatable fix worth codifying in lessons or routing.

Cost

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

Gate

Stages

Notes

Omitted for lack of

None. All planned functionality runs locally in the browser without a service.