What changed in Mr Onion, for someone using it. Each entry links the
decision record behind it; docs/decisions/README.md
indexes them all.
Versioning (TFP-257, decided 2026-10-27): semantic versioning, below 1.0 while the report format can still change. The minor version goes up when a phase completes; the patch version for fixes between phases. 1.0 comes when the report and the CI command are declared stable, after Phase 47.
navigator.sendBeacon, a keepalive fetch and
fetchLater() in the page, in every frame. A site’s beacons are still
noted, not a failure, and any other such write still fails the test as a
guard breach (TFP-284). That holds when a page saved its own reference to
one of them at load, as an analytics script can: they are wrapped before
the page’s scripts run, and the wrappers change nothing until the guard
arms (#35, TFP-287). A
request built from another frame’s Request is stopped too, and a page
can no longer switch the guard off. A fetchLater() a page queues while
it loads is still sent: known limitation 12.Mr Onion can run as one step of another project’s CI. It crawls the
site, generates the suite and runs it, then ends with an exit code the
pipeline acts on
(Phase 47;
the guide is docs/CI_INTEGRATION.md).
emil3663/mr-onion/action@v0.4.0. Give it the URL
your job reaches, such as a server the job starts on localhost. It
runs Mr Onion’s image as the job’s own user, attaches the HTML report and
JUnit results, writes a job summary, and fails the step on failing tests
(TFP-248).ghcr.io/emil3663/mr-onion:0.4.0: Node
24.14.1, Playwright 1.59.1 and its Chromium 147, on Ubuntu 24.04. The
same site gives the same suite on any machine. It runs as any user
(--user), so its files in a mounted folder are yours, not root’s.
Every release runs the built image and checks what it writes before
publishing it (TFP-247, TFP-248).mr-onion run does what the UI does, through the same code, with
no UI and no server. Exit 0 means no failures, 1 test failures, 2
couldn’t run (never a pass), and 3 invalid configuration. It writes
report.json, report.html, junit.xml, the suite that ran, and
evidence for failing tests. --server starts your app first and stops
it afterwards, whatever the result. The command was in 0.3.0’s code but
not in its notes (TFP-245).--baseline latest, only failures an earlier run
didn’t have fail the build. Known and fixed tests are listed. A run is
taken as latest only if it exited 0, or was a first run
(--baseline latest with no earlier run), failures included. One that
failed on a new failure (even with --report-only), failed without
--baseline, or couldn’t run never is. A run with no recorded verdict
(saved by the app, or before 0.3.0) is taken too. This was also in
0.3.0’s code (TFP-246).The app, its reports, the command, the settings, the data folder, the GitHub repo and the Claude Code plugin all take the new name (Phase 48). Ticket and test IDs, saved runs, baselines and site suites are unchanged.
mr-onion (mr-onion run, npm run mr-onion). The
default output folder is ./mr-onion-out. testforge and
npm run testforge still work until 1.0: they print one line saying the
name has changed, then run the same command with the same exit code.MRONION_* in place of TESTFORGE_*, with the
rest of each name unchanged (MRONION_DATA_DIR, MRONION_OLLAMA_MODEL,
…). The old names still work until 1.0, with one notice per setting
naming both. If both are set, the new one wins, unless the new one is
blank. Every setting Mr Onion sets for a run reaches its tests under both
names, so a suite saved before the rename runs as it did, interactive tests
included.%LOCALAPPDATA%\TestForge
becomes %LOCALAPPDATA%\MrOnion (macOS: ~/Library/Application Support/MrOnion;
Linux: ~/.local/share/mr-onion). Your site suites, candidates and
gap-writer settings move with it, and one notice says where they went.
The runs the app lists are in the repo’s runs/ folder, which the move
doesn’t touch. A folder you set
yourself (--data-dir, MRONION_DATA_DIR or TESTFORGE_DATA_DIR) is
never moved. If the move can’t happen (a file in the folder held open),
the old folder is used and the next start tries again.MRONION_INTERACTIVE), and the test data and probe routes they send to
a site use the new name.mr-onion-qwen7b, built from
ollama/Modelfile.mr-onion-qwen7b. A model you built as
testforge-qwen7b is used when the new tag is missing, and the
generation record says so; ollama cp testforge-qwen7b mr-onion-qwen7b
gives it the new name with no rebuild.Reinstall the Claude Code plugin. It is now mr-onion-skills in
mr-onion-skills-marketplace. Once your checkout has the renamed
marketplace, testforge-skills@testforge-skills-marketplace no longer
resolves (“not found in marketplace”). Updating the marketplace isn’t
enough, as it keeps its old registered name; remove it and add it again,
from your checkout’s folder:
claude plugin uninstall testforge-skills@testforge-skills-marketplace
claude plugin marketplace remove testforge-skills-marketplace
claude plugin marketplace add <your checkout>
claude plugin install mr-onion-skills@mr-onion-skills-marketplace
(In Claude Code, the same with /plugin.) The skills are the same twelve,
and the repo’s .claude/settings.json enables the new name.
github.com/emil3663/mr-onion. Old links and clones
redirect.TESTFORGE_DATA_DIR to the moved folder (%LOCALAPPDATA%\MrOnion and
so on) so 0.2.x finds your site suites and candidates.HTTP_PROXY, HTTPS_PROXY, NO_PROXY
and NODE_EXTRA_CA_CERTS when they are set. Mr Onion doesn’t yet set
Playwright’s own proxy option, so this hasn’t been shown to carry a run
through a proxy (TFP-209).artifacts-<id> folders left by interrupted runs:
older than an hour, of no active run, and needed by no saved run. A folder
whose saved run can’t be read is kept
(TFP-242).error-context.md
twice, one link of which answered HTTP 400. Runs saved before this keep the
broken link; the working one still serves the file
(TFP-243).The first tagged version.
/api/health, each run’s run.json and the report’s last line now give
the TestForge version and commit that ran it. Runs from before say the
version wasn’t recorded (TFP-257).next, among them Proxy bypasses and remote code execution on
Windows-hosted servers. The app’s own
dependencies have no known advisory, and the nightly fails if a high or
critical one appears (TFP-266)..rsc or segment-prefetch spelling of an API path can’t reach the API
from another site or host: each is pinned in the regression suite
(TFP-266)." or \ no longer fails its
check on TestForge’s own error instead of the site’s (TFP-267).AbortError when their context
closes while requests wait. The cap is proven at the server: never more
than 5 arrivals in a second.phase-39-run-concurrency-decision-record.md.Not tagged (owner decision, 2026-10-28): Phase 39 completes before the
version stamp lands, so the first tag is v0.2.0. This entry stays as the
baseline it describes.
Everything built from Phase 16 to Phase 42, except Phase 39 (above). Before this, Phases 1–15 (April 2026) moved TestForge from a FastAPI backend to Next.js (history).
127.0.0.1, and every /api/* request is refused
unless its Host is local, its Origin is TestForge’s own, and a write is
JSON. Before this, any device on the network, and possibly a web page open
in the same browser, could make TestForge run code (TFP-261). See
SECURITY.md.testforge-skills Claude Code plugin: 12 testing and engineering
skills, usable without the app (skills/README.md).