[RELEASE] Repo AI-readiness on the Harness tab (carries #5214)

CHANGELOG.md

@@ -10,6 +10,18 @@
 - **Stuck agents are still spotted.** When one of these tools asks permission to run something, or reports how it went, that now feeds the same checks that notice an agent looping or failing the same thing over and over. A permission you refused is not counted as something the agent did.
 - **What this does not do, said plainly.** It works with the tools that already speak this language, which today are Claude Code and Codex. Everything else still needs the usual install. And what arrives this way is not encrypted by the tool that sends it, so if that matters to you, run ClawMetry inside your own network, where none of it leaves.
 
+### Release: before you blame the agent, look at what you handed it (carries #5214) (2026-08-25)
+- **Who this reaches:** everyone, free on every plan. It is on the Harness tab and needs no key, no config and no network.
+- **Your repo now gets a grade, and it sits next to how often agents got stuck in it.** An agent that flails in a project is often not a bad agent. It is an agent that was handed a folder with nothing telling it what the project is, how to run the tests, or what counts as finished. ClawMetry already knows which folders your agents worked in, and it already records when one of them got stuck. Those two facts now sit on the same screen.
+- **Seven things, and each one names the file we opened to check it.** Is there an instruction file. Is there a way to find the tests. A way to build. A lint or format gate. A CI config. Any skills or commands packaged for the repo. Every result says where it came from, so "Test command discoverable" reads "read from Makefile" rather than asking you to trust it.
+- **We do not run your build to find out whether it works.** The work order asked for it and we are not doing it. Running a build changes your machine and usually reaches the network, and this is a thing that watches, not a thing that acts. So we tell you whether an agent could FIND the build command, and say plainly that we never ran it. There is a test that fails if anyone ever adds code here that shells out.
+- **A thing we cannot read scores nothing, in either direction.** Whether your agent actually loaded your instruction file is a real question and nobody can answer it today: none of the tools report which files they read at the start, and us reading the file only proves that we read it. So it is marked unknown, worth zero, drawn outside the bar with the words "not counted". It is neither a pass you did not earn nor a mark against you.
+- **An inherited default is not a tick.** Every Rust project can run `cargo test`, which tells you nothing about whether anyone in that project meant for it to be run. That gets half credit and a note, not a pass.
+- **A folder can be perfectly clear to one tool and invisible to another.** Claude Code reads `CLAUDE.md`, Cursor reads `.cursor/rules`, eleven of them read `AGENTS.md`. The card shows which tools would find their footing in this repo and which would walk in blind, and the grade follows whichever tool you have selected.
+- **The list of files comes from the same place the Memory and Skills browsers read**, so a tool added next month is scored without anyone updating a second list. A short exclusion covers the files the agent itself writes: a folder that aider has merely been run in is not a folder anyone documented, and counting its chat log as instructions would be a pass nobody earned.
+- **Nothing is invented when there is no history.** A repo no agent has touched says so, rather than reporting that nobody got stuck in it. Those are different facts and a zero would read as the second.
+- **Verified:** 61 tests, and each honesty rule was checked by breaking the code and watching its test go red. Run against the real folders on a real machine, and both states looked at in a browser rather than assumed.
+
 
 ### Release: your agent is watched for more than loops, and the list is sorted by what it costs you (carries #5168) (2026-08-25)
 - **Who this reaches:** everyone. The detection that used to notice a stuck agent now also notices one behaving unlike itself, and the list you look at is ordered by money rather than by whichever check spoke last.