gledach signal-patterns

The gledach shelf · 8 of 12 slots filled

Why these eight · All repositories

The shelf Repository

gledach / the shelf / signal-patterns

signal-patterns · slot 02 · reference

How to build tools that
watch the public web.
Nothing to install.

What to watch, what it costs, how to build it, and the engineering decisions that only become obvious after something has broken.

Read the guides Repository

MIT licensed · no package, no version · 4 guides · 8 patterns

the thesis, in one line ordered by lead time 
infrastructure  →  hiring  →  documentation  →  changelog  →  marketing  →  press
   (weeks)         (months,      (hours)         (hours)      (at launch)   (after)
                    vague)
A company is only as visible as what it exposes publicly, and the useful signals are the ones it exposes before it means to. Most tools start at the press end, which is the last place anything appears and where every competitor is already looking.

What this is

Not a framework. The reasoning is the reusable part.

There is no package to install and nothing to version, because the code around a decision like this usually is not the part worth keeping. The reference implementation is signal, which does all of it against a real market.

Watch what leaks

Certificates, documentation and changelogs beat press releases. Certificate logs run twenty to thirty days ahead, and changelog pages change on 73% of monitors within ninety days.

Hand it to an agent

A dashboard needs someone to open it. What you actually want is an answer to a question you have not thought to ask yet, which means structured output, stable ids, and the ability to say I am not in a position to answer.

Two open standards settle it

MCP is how an agent queries you, Agent Skills is how it learns to use you. Both of the most-used open personal-agent runtimes of 2026 speak both, and both are local first, so a local process with a queryable store fits them with no work.

Four guides

Start wherever the question you have lives.

Each one is a markdown file in the repository. Nothing here renders differently to how it reads on GitHub, which is deliberate.

01

What is watchable

A catalogue of public sources ranked by lead time, cost and precision. Certificate logs run twenty to thirty days ahead, changelog pages change on 73% of monitors within ninety days, news is the last place anything appears. Includes what to refuse.

Read it →

02

What it costs

Model capabilities as of September 2026, and the cost levers ranked by return. The top three are not model choice: batch APIs at a flat half, prefix caching at a tenth on reads, and not buying output you discard.

Read it →

03

Building a watcher

The shape of one, and the decisions that are easy to get wrong: the choice of hash id, where state lives, cadence, and what it does when it fails.

Read it →

04

Being consumable by agents

MCP so an agent can query you, Agent Skills so it knows how to use you, and why permission must not scale with capability.

Read it →

Eight patterns

Each one earned its place by failing first.

The bias throughout: an AI pipeline fails silently and pleasantly. The dashboard fills up, nothing throws, and it looks healthy right until you check. Most of these are machinery for turning a quiet failure into a loud one.

# Pattern The failure it prevents
1 Store chokepoint Schema changes with unbounded blast radius
2 Canonical and mirror Regeneration silently destroying human-written content
3 Never store an uncomputed verdict A dead model writing guesses into permanent storage
4 Read-only agent surface An agent spending your money because it found a button
5 The coverage block An agent reporting nothing happened when collection is broken
6 Zone separation The lethal trifecta, when your input is attacker-reachable
7 Assertions as guardrails Every pattern above quietly eroding
8 Collect, do not write Untestable sources, and shared logic wrong in several ways at once

Where one says the incident, that is a real one, with real numbers, from a single deployment over about two months.

The shortest version

If you read nothing else.

Eight lines. Every one of them is the compressed form of something that went wrong first.

  • Watch what leaks, not what is announced. Certificates, docs and changelogs beat press releases.

  • Cadence is the cost dial, not model choice. A per-subject job on the wrong tier cost one deployment around two hundred dollars a month for no added freshness.

  • Batch and cache before you optimise anything else. A flat half, then reads at a tenth.

  • A failed model writes nothing. A halted run announces itself. A run that writes guesses does not.

  • Return I do not know as data. Empty and broken look identical to an agent.

  • Agents read by default. Every paid or state-changing action is opt in, behind a shared ceiling, and permission must not scale with capability.

  • A skill is a supply chain. 341 malicious skills were found among roughly 13,000 in one 2026 marketplace audit.

  • Every rule above needs an assertion, or it is a comment.

If one of these is wrong, or you have a counter-example where it cost more than it saved, open an issue. A pattern that has never been argued with has not been tested. New sources worth watching are equally welcome, especially ones with a lead time and a measured precision attached.

Elsewhere on the shelf

Same author. Different shapes.

This page is the reasoning. The tools beside it are what that reasoning turned into, and signal is the one it was written alongside.

Slot 01 · live preview

signal

The reference implementation: ten watchers and one inbox on whichever market you name, answerable by an agent over MCP.

Open the page →

Slot 03 · shipped

whenrun

Day-ahead power prices turned into the window your job should run in, then it starts the job there.

Open the page →

Slot 04 · shipped

phonehome

Which domains every device on your network contacts, and when one starts talking somewhere new.

Open the page →

Slot 05 · shipped

recall

Four public registers checked against a list of the things you own, with every match graded.

Open the page →

Slot 06 · shipped

talk

Which API hosts your local coding agents connect to, and whether one of them is leaking sockets.

Open the page →

Slot 07 · shipped

snare

Remembers the hooks a coding agent runs at startup, and the editor tasks that run when a folder opens. Prints only when that set changes. It does not delete a hook.

Open the page →

Slot 08 · shipped

leash

Takes a snapshot before an agent session and another after, then shows what changed outside that repo: files, startup entries, services, packages and ports. It does not delete, kill or block anything.

Open the page →