Commit Graph
6 Commits
Author SHA1 Message Date
helix-nine 1a4bd6f251 chore: exclude the vendored upstream from prettier 2026-08-16 22:41:41 -06:00
helix-nine ee18089d82 docs: rewrite README as a technical reference and refresh the package docs 2026-08-16 22:41:41 -06:00
helix-nine a77cba492c document the unauthenticated endpoints, drop TODO.md for issues
Nothing on the proxy is authenticated, so whoever can reach the port can
spend the user's PPQ.AI balance — true of every version, and not
something either doc said. Enabling PPQ_DATA_DIR adds a second
unauthenticated route, POST /setup/api-key, which replaces the stored
key (upstream PayPerQ/ppq-private-mode-proxy#23).

README states both and whose each is; instructions.md tells the user
plainly to expose the interface only to people they trust, which is the
part they can act on.

TODO.md is deleted rather than emptied. Its live-inference item is
closed — PayPerQ confirmed billed end-to-end against glm-5-2 and
llama3-3-70b — and the other two are now issues #2 (backup/restore,
unverified: it needs the server password, and the 0.4 file split gives
it two new ways to fail) and #3 (bump when the auth patch lands, and
retract the warnings it makes false). The scaffold's worklist has been
worked; a file left behind to say so is not worth reading.
2026-08-12 10:06:57 -06:00
helix-nine 8397c7b1eb adopt 0.4's persistent key store and browser setup form
The earlier bump took 0.4.1 for its model catalog and status page but
left PPQ_DATA_DIR unset, which disabled the key form the release exists
to add. Enable it, and let the proxy own the key outright.

PPQ_API_KEY is no longer passed. The volume's `proxy/` subpath mounts at
/data as PPQ_DATA_DIR, so upstream's key store is the single source of
truth: the Status Page and the Configure PPQ API Key action write the
same config.json, and there is no StartOS-side copy to revert what the
browser saved. A chown oneshot hands /data to the image's `node` user so
the page can write there.

Verbose Logging moves to store.json outside the mount, because upstream
rewrites config.json wholesale when a key is saved and would drop it.

main.ts watches config.json — upstream loads it once at startup, so the
watch is what makes the action take effect: the write restarts the
daemon, which reloads the file. A key saved on the page applies
immediately and converges through the same restart.

The task drops from critical to important. A critical task suspends
every other control, so it would have left the user unable to start the
service and reach the page's own form — and with PPQ_DATA_DIR set the
proxy no longer exits keyless; it attests, serves the page, and answers
401 until a key exists. main.ts clears the task when a key appears,
since saving on the page never touches the action that would clear it.

The action's key field is now validated against upstream's own
^sk-[A-Za-z0-9]{16,64}$ shape and no longer refuses a blank submission,
a key having become optional.
2026-08-12 09:58:35 -06:00
helix-nine 42b2841945 bump upstream to 0.4.1, add the status page UI, fix CI branch triggers
Community-registry audit ahead of listing.

Upstream 0.4.1 renames the model catalog (private/kimi-k2-6 →
private/kimi-k3, drops qwen3-vl-30b) and serves a status page at GET /,
so the package was documenting a stale model list and hiding a whole
surface. Export that page as a `ui` interface alongside the existing
`api` on one host, and unmask the API interface — its URL carries no
credentials, so masking hid the address for nothing.

PPQ_DATA_DIR stays unset deliberately. With it set, upstream persists a
browser-entered key to its own config.json while the StartOS-supplied
PPQ_API_KEY still wins at every restart, silently reverting the user.
StartOS owns the key via the action; the status page's form renders
disabled.

CI never ran: build.yml's PR target and tagAndRelease.yml's push
trigger both named `master` on a repo whose only branch is `main`.
Point them at `main` and reset all three workflows to the package
template, dropping the guard job retired in start-technologies@ed0dca2.

packageRepo now names the Start9-Community fork, which is where CI
runs, where the org's registry variables live, and where releases are
cut.

Also: drop three .catch handlers on file-model reads, which swallowed a
genuine read failure into "no key configured" — readOnce already
returns null for a missing file; pass null rather than an empty Mounts
for the unmounted subcontainer; inline two single-use interface consts;
and rewrite README/instructions against the writing guides, fixing a
YAML block that claimed a /data mount the package deliberately does not
make.
2026-08-12 09:58:35 -06:00
helix-nine 5146a2670f ci: stop handing PR builds the signing key
A PR build only compiles and packs, and build.yml falls back to
`start-cli init-key` when the secret is absent, so passing DEV_KEY only put the
real signing key on a runner that executes branch-authored code.
2026-08-11 15:54:21 -06:00