* Studio: add --secure Cloudflare-only mode and revamp API usage examples
--secure / --not-secure on `unsloth studio` and `unsloth studio run`:
- --secure binds 127.0.0.1, requires the Cloudflare tunnel, and advertises only
the Cloudflare link. cloudflared reaches the server over localhost, so the raw
port is never exposed on a public interface.
- If the tunnel cannot start, fail closed with a clear message instead of
silently leaving a raw 0.0.0.0 link.
- Default stays not-secure (no behavior change); coexists with the existing
--cloudflare/--no-cloudflare flag. Host defaults are unchanged.
- /api/health (authed) now reports the live tunnel URL.
API usage examples (Profile > API):
- Example tabs for curl, Python, curl + tools, Python + tools, plus an OS row
(Linux/macOS/WSL vs Windows) auto-detected from the platform.
- Windows curl passes the JSON body via a file so PowerShell does not strip the
quotes when calling curl.exe.
- Python + tools forwards enable_tools/enabled_tools through extra_body and
guards chunk.choices, since tool-lifecycle events carry no choices.
- Shows the loaded model name and the real API key while it is still revealed.
- A Cloudflare Tunnel toggle (default on) shows the public tunnel URL and uses
it as the base_url in the examples when a tunnel is running.
Tests cover the tunnel start gate and the --secure flag on both commands.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate --secure tools on public exposure and harden API examples
In secure mode the server binds loopback but is reachable via the public
Cloudflare tunnel, so resolve the tool policy against the public exposure
(0.0.0.0) rather than the loopback bind. This keeps server-side tools off by
default and prompts before enabling them, instead of inheriting the loopback
default of on. The startup tool notice now names the public surface.
Also reject --secure with --no-cloudflare directly in run_server and the
run.py argparse (not only the CLI), JSON-encode interpolated model names so
Windows paths and quotes cannot produce invalid JSON or broken snippets, and
force-refresh /api/health on the API panel so a tunnel that starts after the
first health read still surfaces its URL.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: API examples show direct host when tunnel toggle is off; move Copy onto code
The Cloudflare Tunnel toggle had no visible effect when Studio was opened
through the tunnel: the off state fell back to window.location.origin, which
equals the tunnel URL in that case. /api/health now reports the direct
host:port (server_url), and the API panel uses it for the off state so it shows
the real non-tunnel base. Also move the Copy button out of the tab row and onto
the code block.
* Studio: highlight API examples, add advanced tabs, fix tunnel toggle row
Syntax-highlight the curl/PowerShell/Python snippets with the app's shared
shiki plugin (bash/powershell/python). Add 'curl + advanced' and
'Python + advanced' tabs that set temperature/top_p/top_k/min_p/
repetition_penalty/max_tokens, enable thinking, and turn on all tools.
The Cloudflare Tunnel row no longer shifts the code block: the tunnel URL is
always rendered (dimmed when off) so toggling keeps the row height constant.
Key the highlighted block on its content so it remounts when only the base URL
changes (the renderer's block memo otherwise kept a stale URL).
* Studio: rename API tunnel toggle to Secure HTTPS, hint --secure when exposed
Rename the API examples toggle from Cloudflare Tunnel to Secure HTTPS. When the
server was not launched with --secure, show an info tooltip noting the raw
0.0.0.0 port is still globally reachable and pointing at --secure. /api/health
now reports whether --secure was used so the hint is hidden in secure mode.
* Studio: force tools off for plain network/secure launches
The plain 'unsloth studio --secure' (and '-H 0.0.0.0') launcher re-execs run.py
and never installed a tool policy, so the process default (honor per-request
enable_tools) let any API-key holder run Python/terminal tools over the public
endpoint. Force the policy off at the run.py entrypoint when network-reachable
(0.0.0.0 or --secure); 'unsloth studio run' still installs its own resolved
policy and does not go through this path.
* Studio: apply default tool policy in run_server, not the run.py entrypoint
The plain launcher runs from the studio venv and calls run_server directly, so
it never hit the run.py __main__ guard. Move the network/secure default-off tool
policy into run_server so every launch path (plain, --secure, direct run.py)
gets it; the run subcommand still overrides it with its resolved policy.
* Studio: clarify --secure help text on the network exposure tradeoff
Spell out in --help (both unsloth studio and unsloth studio run, plus the
run.py argparse) that --not-secure also serves the raw 0.0.0.0 port reachable
from anywhere on the network, matching the API panel's Secure HTTPS hint.
* Studio: cache API-key PBKDF2 derivation to cut per-request /v1 auth overhead
validate_api_key re-ran the 100k-round PBKDF2 on every authenticated
request, adding ~15ms to each /v1 call made with an sk-unsloth- key.
Benchmarked against the bare llama-server it proxies to, API-key requests
carried ~22ms of fixed overhead vs ~7ms for the JWT path; the gap was
entirely this redundant key derivation (Pydantic validation measured
0.005ms, so it is not a factor).
The raw-key to hash mapping is a pure deterministic function of the fixed
server salt, so memoize it per process, keyed by a salted HMAC of the key
(never the key or a recoverable digest). The cached value equals what is
already stored at rest. Revocation and expiry remain enforced by the
SQLite read on every call, so a cache hit only skips the KDF, never the
active or expiry checks. Only keys that exist in the DB are cached, so
unknown-key spam cannot grow it.
After the change the API-key /v1 overhead drops to ~8ms, at parity with
JWT, while the at-rest PBKDF2 hashing is unchanged.
Adds test_api_key_expiry.py covering API-key and JWT expiry enforcement
and the new cache: it skips the KDF on repeat and still rejects revoked
or expired keys.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten comments across the secure-tunnel and API-key changes
Condense multi-line comments and docstrings to one or two lines, drop the
ones that restate obvious code, and remove an orphaned test section header.
Comment-only: verified with comment_tools.py check (9/9 code unchanged), the
auth/secure-tunnel/CLI test suites, and a clean frontend typecheck and build.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
40c8ad78b9