Jason 2026-11-19: Claude models through an API key may run on any harness; OAuth Claude models only
on claude-code. The name check is gone. The catalogue reads every Claude OAuth account's model names,
whatever the account's state, and holds a group model with one of those names to claude-code, since
the proxy routes by name across every account. route() has no account data, so it holds upstream
claude only; admission and Triage hold the rest through the catalogue.
cliproxy holds a model to claude-code when its upstream is claude, when an OAuth Claude account
offers the same name, or when the name contains "claude"; every other proxy model runs on
claude-code and pi. The catalogue, the launch (sessionModel and the models.json list) and the
provider route apply the one rule and refuse with model-harness-incompatible, so a stored
Claude-on-pi route is refused, never rewritten. The other route's pi provider is mosaic-proxy-other.
Triage: the route asks the provider for the transport on the stored harness and holds the
harness's env() to one contract (the key, unchanged, under exactly one of ANTHROPIC_AUTH_TOKEN or
ANTHROPIC_API_KEY; the URL only as ANTHROPIC_BASE_URL), with no list of harness names. The host
gives the module the real proxy and transport, checks each file it writes under the session's HOME
(regular, one link, no key) and binds it read-only after the run directory, and refuses a route env
with no key name or two. PATH starts with node_modules/.bin, as a bus session's does.
Tests: claude-code runs a z.ai and a codex proxy model against a stand-in (real claude binary);
pi answers Triage through a stand-in under bwrap; every new fence has a test.
A cliproxy organization preset offers a model from the owner's catalogue and says in plain words
why it can't be used: no proxy, no riding bus answered, or no catalogue model is available. The
organization service passes an optional model to a preset's overlay; other presets ignore it.
Ride all, an instance-owner action on the Model proxy page, adds cliproxy to each organization
through that organization's own plugins.set. It skips and names an organization whose change log
shows cliproxy removed, whose bus is down, or whose bus has no proxy client. Proxy admin status
names it as the next step. A fixture organization is listed as using stand-in models; one more
confirmed click moves a stopped one to the owner's models through the organization service's
existing whole-document update at the revision read.
A fenced proxy then refuses each riding organization's launches with proxy-uncertain, a direct
anthropic route included.
A bus started before the instance had a proxy checks once at the next accounts, page, catalogue
or fence request, and loads the client through the start path's checks, loader and lock. A client
got that way, or by the bus's own init, is checked again at each such request and dropped when
proxy.json goes away, no longer loads, moves port, the key stops loading, the lock refuses, or a
dev fixture the host would refuse is in use. While there is no client, the bus names the cause in
one plain line on not-configured and on the unavailable catalogue, and the Console shows it.
The Triage host got route: null. It now gets a route read fresh at each
admit from the route store, the key file, each active organization's
model catalogue (held to checkRoutes) and the harness module. The spawn
env adds only ANTHROPIC_BASE_URL (the instance proxy) and
ANTHROPIC_AUTH_TOKEN (the Triage key); the host sets every MOSAIC_ name.
The key file reader moves to packages/cli/src/triage-key-file.mjs so the
Console reads the key with the key verbs' checks without loading the
proxy management client; triage-key.mjs re-exports it. Each refusal (no
route, no key, unsafe key, proxy down, model not offered, no catalogue,
harness incompatible) launches nothing and answers 409 with one plain
line. Every admit runs the host's fences on the answer, a session
running or not.
The cap fails closed: two turns in a row with unknown usage end the
session with a host line; the next message starts a new run with a
fresh session directory, and its usage counts again.
#1873, lane triage-wiring
GET and POST /api/triage/route, owner only behind the proxy-admin gate, with one
Origin and Sec-Fetch-Site same-origin on a write and Fetch Metadata on a read.
GET returns the stored route or says it is not set, plus the merged catalogue
choices (model keys, availability, compatible installed harnesses) with no
provider, upstream, module, key or path. POST takes model, harness, cap, time
zone and revision, checks the route with checkRoutes against the merged
catalogue (offered, available, compatible, installed), requires a known time
zone with no default, and writes through the store's writeRoute, so the lock,
the revision compare and the receipt are the store's. A stale revision answers
409 with the stored revision. updatedBy is the signed-in owner.
The Settings > Triage page shows only to the owner /api/config marks proxyAdmin.
Model and harness selects (the harness list is the chosen model's), a cap in
tokens, reset 00:00 read only, a time zone, Save. Not set before a first save;
a refusal is one line and keeps the fields; a stale save offers Reload route; a
second Save while one runs does nothing; one line says saving does not start
Triage. cli.mjs keeps route: null for the Triage host.
Each turn's usage goes to <dataRoot>/triage/usage.jsonl, one line per
usage.end, 0600 and owned by the running user; a link, wrong mode or
owner, or a non-file refuses. The host holds the stored cap through
daily-quota mayLaunch and report before each start and each queued
message, lets a crossing turn finish, drains the queue with a host
line, sums the file again after a restart (a torn line counts as an
unknown turn), and GET /api/triage adds usageToday, cap and resetsAt.
With no route stored the fields say "not set". cli.mjs keeps route
null. (#1873, lane triage-cap)
The file owner check has its own test through a file-owner seam, for
route.json on read and write and for route-receipts.jsonl on write. The
directory, temp file and a new receipts file get their modes on the open
descriptor, so a narrow umask cannot leave a route the next read refuses;
an existing receipts file is never repaired. A failed temp write refuses
and removes the temp file. A failure after the rename refuses as
triage-route-unrecorded and names the new revision. The input zone check
uses the stored cap's rule, so +05:00 refuses as an unknown zone. A
changed cap keeps its since, as an organization's token limits do, so
lowering the cap never drops tokens already counted (#1873, lane
triage-route-store).
<dataRoot>/triage/route.json holds Triage's model route and daily cap as names only:
a catalogue model key, a harness name and a daily-quota cap leaf, with no env value
and no key. The directory is 0700 and the file 0600, owned by the running user; a
link, a hard link, a non-file, a wrong mode or owner, an unknown field or an unknown
time zone refuses. A write is compare-and-set on revision under an exclusive lock,
atomic (temp file, fsync, rename, directory fsync), and appends one receipt line.
The reset is 00:00 in a zone the owner gives, with no default. The file location,
field names and reset rule each live in one place.
mosaic triage route show|set are human verbs and refuse an agent run before any read.
Nothing is wired into the Triage host: the Console keeps route: null. Held for
Jason's ruling on the design's four questions (#1873, lane triage-route-store).
The workflows read listed the triage template, which only the client hid.
The read handler now drops every template the organization vocabulary marks
system-only, and the Console's read supplier in cli.mjs drops them too. The
client filter stays. (#1873, lane triage-followups)