You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This seems to be an efficient way to run Tailscale, but something is so convoluted that I cannot even get it to run on v2.
Initializing server
Version 2.3.4
{"level":"info","file":"/config/tsdproxy.yaml","time":"2026-11-12T18:36:56Z","message":"loading configuration"}
error: yaml: unmarshal errors:
line 20: field dashboard not found in type config.config
line 26: field proxyAccessLog not found in type config.LogConfig
Nothing in issues, so let's try v3!
I go make an OAuth credential. Policy/ACL is not a scope.
Thanks for the detailed report — there are two separate things here:
1. The v2 config errors. Your tsdproxy.yaml has two field mistakes (probably v1-era or hand-edited):
dashboard: (line 20) is not a config field in v2 or v3 — the dashboard is served on the main HTTP port, there is nothing to configure.
proxyAccessLog (line 26) is a top-level field, not a child of log:. It should read:
log:
level: infojson: falseproxyAccessLog: true
tsdproxy errors out on unknown fields on purpose (to catch typos). A proper v1→v2 upgrade guide is tracked in #486.
2. "Policy/ACL is not a scope" (v3). You hit a Tailscale rename. In the OAuth clients page, the policy scopes are now listed as Policy File (policy_file:read / policy_file) instead of "Policy/ACL" — select those. Our v3 docs still say "Policy/ACL", which is a docs bug on our side.
There's also a possible code issue behind it: tsdproxy's API client requests the literal scope IDs policy:read / policy:write, and the official Tailscale scope reference documents policy_file:read / policy_file (new) and acl / acl:read (legacy) — neither policy:read nor policy:write appears there. Confirming and fixing that needs testing with a real OAuth client, so I've flagged it to @almeidapaulopt rather than guessing at scope IDs. If you create the client with the Policy File scopes and tsdproxy still fails on the ACL step, please paste the exact error here — it will tell us whether the token exchange rejects the requested scope names.
(Unrelated to what you hit, but if you're on v3: #520 tracks a separate auth-key generation quirk.)
You're right to push back — I went and checked the source instead of trusting memory, and our docs are wrong, not you.
Verified against the v2.3.4 config struct (internal/config/config.go): there is nodashboard server-config field in v2 (nor v3 — the yaml:"dashboard" tag that exists belongs to tsdproxy.template presets, a different thing). The dashboard's only knob is the top-leveladminAllowLocalhost flag, and proxyAccessLog is top-level too, not a child of log:.
Our own Getting Started example contains both mistakes:
dashboard:
adminAllowLocalhost: true # ← invalid: no such sectionlog:
level: infojson: falseproxyAccessLog: true # ← invalid: belongs at top level
tsdproxy is strict on purpose (unknown fields = startup error, to catch typos), so your config did exactly the right thing: it refused to boot. Working v2.3.4 form:
(If you don't need those knobs, drop both lines — defaults are sane: proxyAccessLog: true, and adminAllowLocalhost auto-enables when running in Docker.)
The docs fix is on us — the same wrong block sits in the current and the v3 getting-started pages. The OAuth scope point from my earlier reply stands: the docs' "Policy/ACL" name is outdated on Tailscale's side too, and the code-level scope check (policy:read/policy:write) still needs confirming against a real OAuth client.
Labeling this documentation. Sorry for the runaround in the first reply.
Docs fix is up: #561 — corrects the invalid dashboard: section and the proxyAccessLog placement in the getting-started pages (current + v3) and in the v3 from-v2 upgrade guide, verified against the v2.3.4 strict loader and v3's unmarshalNormalized.
Follow-up found while verifying: on main/v3, keys that exist elsewhere in the schema but are misplaced are currently silently dropped instead of rejected (v2.3.4 rejected them) — filed as #562 with a fix in #563.
Activity
almeidapaulopt commentedon Nov 12, 2026
Thanks for the detailed report — there are two separate things here:
1. The v2 config errors. Your
tsdproxy.yamlhas two field mistakes (probably v1-era or hand-edited):dashboard:(line 20) is not a config field in v2 or v3 — the dashboard is served on the main HTTP port, there is nothing to configure.proxyAccessLog(line 26) is a top-level field, not a child oflog:. It should read:tsdproxy errors out on unknown fields on purpose (to catch typos). A proper v1→v2 upgrade guide is tracked in #486.
2. "Policy/ACL is not a scope" (v3). You hit a Tailscale rename. In the OAuth clients page, the policy scopes are now listed as Policy File (
policy_file:read/policy_file) instead of "Policy/ACL" — select those. Our v3 docs still say "Policy/ACL", which is a docs bug on our side.There's also a possible code issue behind it: tsdproxy's API client requests the literal scope IDs
policy:read/policy:write, and the official Tailscale scope reference documentspolicy_file:read/policy_file(new) andacl/acl:read(legacy) — neitherpolicy:readnorpolicy:writeappears there. Confirming and fixing that needs testing with a real OAuth client, so I've flagged it to @almeidapaulopt rather than guessing at scope IDs. If you create the client with the Policy File scopes and tsdproxy still fails on the ACL step, please paste the exact error here — it will tell us whether the token exchange rejects the requested scope names.(Unrelated to what you hit, but if you're on v3: #520 tracks a separate auth-key generation quirk.)
reavpC6K commentedon Nov 12, 2026
I want to think this is AI slop because the dashboard bit is literally there in Getting Started.
almeidapaulopt commentedon Nov 13, 2026
You're right to push back — I went and checked the source instead of trusting memory, and our docs are wrong, not you.
Verified against the v2.3.4 config struct (
internal/config/config.go): there is nodashboardserver-config field in v2 (nor v3 — theyaml:"dashboard"tag that exists belongs totsdproxy.templatepresets, a different thing). The dashboard's only knob is the top-leveladminAllowLocalhostflag, andproxyAccessLogis top-level too, not a child oflog:.Our own Getting Started example contains both mistakes:
tsdproxy is strict on purpose (unknown fields = startup error, to catch typos), so your config did exactly the right thing: it refused to boot. Working v2.3.4 form:
(If you don't need those knobs, drop both lines — defaults are sane:
proxyAccessLog: true, andadminAllowLocalhostauto-enables when running in Docker.)The docs fix is on us — the same wrong block sits in the current and the v3 getting-started pages. The OAuth scope point from my earlier reply stands: the docs' "Policy/ACL" name is outdated on Tailscale's side too, and the code-level scope check (
policy:read/policy:write) still needs confirming against a real OAuth client.Labeling this
documentation. Sorry for the runaround in the first reply.almeidapaulopt commentedon Nov 13, 2026
Docs fix is up: #561 — corrects the invalid
dashboard:section and theproxyAccessLogplacement in the getting-started pages (current + v3) and in the v3 from-v2 upgrade guide, verified against the v2.3.4 strict loader and v3'sunmarshalNormalized.Follow-up found while verifying: on main/v3, keys that exist elsewhere in the schema but are misplaced are currently silently dropped instead of rejected (v2.3.4 rejected them) — filed as #562 with a fix in #563.