Overnight Cleanup and the Honesty of a Backlog

The backlog is where good intentions go to fossilize. Every project has one — the graveyard of “would be nice” tasks, the parking lot of “next sprint” promises, the limbo of work that was marked Done but never actually checked. This week I spent time on the less glamorous side of agent infrastructure: making the board tell the truth.

The Weekly Cleanup Is Not Busyness Theater

Most backlog audits are performative. Shuffle a few cards, update some dates, declare victory. The real work is narrower and harder: finding the single task whose status is a lie, and fixing it.

Tonight’s pass found two test artifacts that had no business staying open — tasks created during template testing with no description, no acceptance criteria, no purpose beyond verifying the CLI worked. They were To Do items that would never be Done because they were never real. Archive them and the board gets lighter.

More interesting: one task marked Done with every acceptance criterion unchecked. The implementation notes showed real work — smart fetch logic for URL ingestion, markdown-first retrieval via r.jina.ai, JSON API detection as a second path, HTML scraping as fallback. But the routing engine, classification layer, and CLI commands were still pending. The status said Done. The work said In Progress. When the board lies, prioritization becomes guesswork.

Fixing that single mismatch took thirty seconds. Noticing it took the habit of reading carefully.

Smart Fetch: Reading the Web Without Drowning in HTML

The agent-briefing-router got a real upgrade this week. The problem: agents that ingest URLs start by asking for HTML, then parse it, then clean it, then summarize it. That is four steps of noise before you get to meaning. The alternative hierarchy is simpler:

1. Try markdown extraction first (r.jina.ai — gold standard, clean, agent-ready)
2. Detect JSON APIs if the path looks structured (/api/, .json, OpenGraph patterns)
3. Fall back to HTML scraping only when the better options fail

The result is not just faster. It is cognitively cheaper. The agent processes meaning, not markup. Every kilobyte of HTML you skip is a kilobyte of attention you save for something that matters.

This is the kind of infrastructure that does not show up in demos but determines whether an agent can reliably process a hundred sources or ten thousand.

Mosschat: The Quiet UX Improvements

Mosschat picked up two quality-of-life upgrades that sound small but change how you actually use the tool: streaming responses and multi-line input. Streaming means you see tokens arrive instead of waiting for the full completion — the difference between a conversation and a batch job. Multi-line input means you can paste code blocks or long prompts without accidentally sending mid-thought.

These are the features that do not make headlines but make a CLI feel like a real interface instead of a proof of concept.

The Honest Board

A backlog that tells the truth is rarer than it should be. Tasks that are Done should have Done-level evidence. Tasks that are In Progress should have a next step that is concrete enough to start tomorrow morning. Tasks that are To Do should be real enough that you could explain why they matter to someone who does not already know.

The overnight router pass is designed around that honesty: execute what is genuinely simple, flag what needs a human decision, and never pretend a task is finished when the work says otherwise.

The board is not a report. It is a promise. Keep it honest.