<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <id>https://stackassay.com/stacks/releases</id>
  <title>Stack Assay: Stack release notes</title>
  <subtitle>What changed in every version of every Stack.</subtitle>
  <link rel="self" href="https://stackassay.com/stacks/releases.xml"/>
  <link rel="alternate" type="text/html" href="https://stackassay.com/stacks/"/>
  <author><name>Stack Assay</name></author>
  <updated>2026-11-21T00:00:00Z</updated>
  <entry>
    <id>https://stackassay.com/stacks/software-factories/releases/0.2.1</id>
    <title>Software factories Stack 0.2.1</title>
    <link rel="alternate" type="text/html" href="https://stackassay.com/stacks/#software-factories"/>
    <published>2026-11-21T00:00:00Z</published>
    <updated>2026-11-21T00:00:00Z</updated>
    <content type="html">&lt;ul&gt;&lt;li&gt;Same commands, flags, config file and defaults, now built on the shared Stack kit that every Stack Assay Stack uses (the sandbox, egress gateway, meter, hidden checks, `validate`, task loader and licence check), so fixes reach every Stack through one release path.&lt;/li&gt;&lt;li&gt;No defaults changed.&lt;/li&gt;&lt;li&gt;Copying the agent's work back never writes a file the agent was not shown: one already on your machine but not in its copy (a gitignored file), or one git ignores. The whole write-back is refused instead.&lt;/li&gt;&lt;li&gt;The AI review and check output in the PR body sit in a code block the text cannot close, so a review cannot write its own Markdown or fake a gate result.&lt;/li&gt;&lt;li&gt;The egress gateway forwards only the OpenRouter model and catalogue paths the agent uses; key, credit and cost-lookup endpoints are refused, so an agent cannot read your account through the key it is borrowing.&lt;/li&gt;&lt;li&gt;The agent no longer sees common credential files anywhere in the repo (`.env` files at any depth, `.npmrc`, `.netrc`, `.pypirc`, `*.pem`, `*.key`, SSH private keys, `.aws/`, `.ssh/`), and a change to one of them, or to `.factory-runs/`, is refused and blocks the run instead of reaching the PR.&lt;/li&gt;&lt;li&gt;A task's `task.json` may also give `"prompt"` (as well as `"ticket"` or `"tickets"`) and `"tests"`, its own test command, which `factory validate` uses instead of the config's.&lt;/li&gt;&lt;li&gt;Docker networks and containers the factory creates are labelled `factory=1` (was `stack-sff=1`); their names and the 10.232.0.0/24 subnet are unchanged. Stage ids in run records start with `stage-`.&lt;/li&gt;&lt;li&gt;`factory doctor` says when a newer version is on the registry, using your licence key (it sends the key only to the Stack Assay registry); offline it skips the check.&lt;/li&gt;&lt;li&gt;The README has an Updates section: upgrading, Dependabot and Renovate with the Stack Assay registry, and renewals.&lt;/li&gt;&lt;li&gt;`factory help` (and `--help`, `-h`) prints the usage and exits 0; an unknown command says so and exits 1.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Upgrade: &lt;code&gt;npm i @stackassay/stack-software-factories@latest&lt;/code&gt;&lt;/p&gt;</content>
  </entry>
  <entry>
    <id>https://stackassay.com/stacks/agent-harnesses/releases/0.1.1</id>
    <title>Agent harnesses Stack 0.1.1</title>
    <link rel="alternate" type="text/html" href="https://stackassay.com/stacks/#agent-harnesses"/>
    <published>2026-11-20T00:00:00Z</published>
    <updated>2026-11-20T00:00:00Z</updated>
    <content type="html">&lt;ul&gt;&lt;li&gt;Installs with `npm i` from the Stack Assay registry: 0.1.0 could not, because the registry dropped the note that the kit is bundled inside the package and npm went looking for it.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Upgrade: &lt;code&gt;npm i @stackassay/stack-agent-harnesses@latest&lt;/code&gt;&lt;/p&gt;</content>
  </entry>
  <entry>
    <id>https://stackassay.com/stacks/agent-harnesses/releases/0.1.0</id>
    <title>Agent harnesses Stack 0.1.0</title>
    <link rel="alternate" type="text/html" href="https://stackassay.com/stacks/#agent-harnesses"/>
    <published>2026-11-20T00:00:00Z</published>
    <updated>2026-11-20T00:00:00Z</updated>
    <content type="html">&lt;ul&gt;&lt;li&gt;Ships under the Stack Assay Commercial Licence (`LICENSE`) and trusts the live licence signing key, so a purchased licence reads as licensed.&lt;/li&gt;&lt;li&gt;First version: `harness run`, `assay`, `probe`, `validate`, `init`, `doctor`, `image` and `licence`.&lt;/li&gt;&lt;li&gt;Claude Code 2.1.326 by default, Ori Code 0.17.0 as the open-source comparison, mini-swe-agent 2.4.6 and Hermes v0.23.2 available; all pinned, all run only in the locked-down sandbox behind the key-holding egress gateway.&lt;/li&gt;&lt;li&gt;The normalised run record with exact cost from OpenRouter's records, cost per solved task, median time and cache hit rate; run-budget guardrails (dry run, `--max-usd` cap, pilot, watchdog, `.stack/ledger.jsonl`).&lt;/li&gt;&lt;li&gt;`harness reprice` fills in cost records that OpenRouter publishes late; until then a run counts at its estimate against the cap and the ledger, and every assay reports the key's usage before and after.&lt;/li&gt;&lt;li&gt;Fourteen worked examples (the Stack Assay harness lab's task set), each with a reference solution and a planted wrong answer that `harness validate` proves the checks catch.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Upgrade: &lt;code&gt;npm i @stackassay/stack-agent-harnesses@latest&lt;/code&gt;&lt;/p&gt;</content>
  </entry>
  <entry>
    <id>https://stackassay.com/stacks/software-factories/releases/0.1.0</id>
    <title>Software factories Stack 0.1.0</title>
    <link rel="alternate" type="text/html" href="https://stackassay.com/stacks/#software-factories"/>
    <published>2026-11-20T00:00:00Z</published>
    <updated>2026-11-20T00:00:00Z</updated>
    <content type="html">&lt;ul&gt;&lt;li&gt;First release, published to the Stack Assay registry.&lt;/li&gt;&lt;li&gt;Ships under the Stack Assay Commercial Licence (`LICENSE`), replacing the prototype notice, and trusts the live licence signing key, so a purchased licence reads as licensed.&lt;/li&gt;&lt;li&gt;One agent per ticket in a locked-down sandbox behind an OpenRouter-only egress gateway (BYOK), with planner and reviewer stages off by default.&lt;/li&gt;&lt;li&gt;A merge gate that never merges on green alone: hidden checks, the agent's own tests and an optional AI reviewer, ending in *ready for review*, *flagged* or *blocked*. Ships as a GitHub Action.&lt;/li&gt;&lt;li&gt;`make assay` re-runs the lab's measurement on your own tasks and writes the lab's run record.&lt;/li&gt;&lt;li&gt;`factory validate`: proves each hidden check fails the starting repo, passes `solution/` and catches every `wrong/&amp;lt;name&amp;gt;/` before you trust it. `examples/house-rules` (ledger and parcels, 18 wrong answers) is a template.&lt;/li&gt;&lt;li&gt;`factory licence [--refresh]`: an offline licence check with 14 days' grace. It never blocks a run; unlicensed runs are labelled.&lt;/li&gt;&lt;li&gt;npm packaging from the Stack Assay registry, with `factory init`, `factory image` and `factory doctor`.&lt;/li&gt;&lt;li&gt;Grading uses a per-run nonce in the environment so a build can't forge a passing check, plus two more fixes from a security review.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Upgrade: &lt;code&gt;npm i @stackassay/stack-software-factories@latest&lt;/code&gt;&lt;/p&gt;</content>
  </entry>
</feed>
